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
121 // This slice records the initializing tasks that need to be
122 // done to start up the runtime. It is built by the linker.
123 var runtime_inittasks []*initTask
125 // main_init_done is a signal used by cgocallbackg that initialization
126 // has been completed. It is made before _cgo_notify_runtime_init_done,
127 // so all cgo calls can rely on it existing. When main_init is complete,
128 // it is closed, meaning cgocallbackg can reliably receive from it.
129 var main_init_done chan bool
131 //go:linkname main_main main.main
134 // mainStarted indicates that the main M has started.
137 // runtimeInitTime is the nanotime() at which the runtime started.
138 var runtimeInitTime int64
140 // Value to use for signal mask for newly created M's.
141 var initSigmask sigset
143 // The main goroutine.
147 // Racectx of m0->g0 is used only as the parent of the main goroutine.
148 // It must not be used for anything else.
151 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
152 // Using decimal instead of binary GB and MB because
153 // they look nicer in the stack overflow failure message.
154 if goarch.PtrSize == 8 {
155 maxstacksize = 1000000000
157 maxstacksize = 250000000
160 // An upper limit for max stack size. Used to avoid random crashes
161 // after calling SetMaxStack and trying to allocate a stack that is too big,
162 // since stackalloc works with 32-bit sizes.
163 maxstackceiling = 2 * maxstacksize
165 // Allow newproc to start new Ms.
168 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
170 newm(sysmon, nil, -1)
174 // Lock the main goroutine onto this, the main OS thread,
175 // during initialization. Most programs won't care, but a few
176 // do require certain calls to be made by the main thread.
177 // Those can arrange for main.main to run in the main thread
178 // by calling runtime.LockOSThread during initialization
179 // to preserve the lock.
183 throw("runtime.main not on m0")
186 // Record when the world started.
187 // Must be before doInit for tracing init.
188 runtimeInitTime = nanotime()
189 if runtimeInitTime == 0 {
190 throw("nanotime returning zero")
193 if debug.inittrace != 0 {
194 inittrace.id = getg().goid
195 inittrace.active = true
198 doInit(runtime_inittasks) // Must be before defer.
200 // Defer unlock so that runtime.Goexit during init does the unlock too.
210 main_init_done = make(chan bool)
212 if _cgo_pthread_key_created == nil {
213 throw("_cgo_pthread_key_created missing")
216 if _cgo_thread_start == nil {
217 throw("_cgo_thread_start missing")
219 if GOOS != "windows" {
220 if _cgo_setenv == nil {
221 throw("_cgo_setenv missing")
223 if _cgo_unsetenv == nil {
224 throw("_cgo_unsetenv missing")
227 if _cgo_notify_runtime_init_done == nil {
228 throw("_cgo_notify_runtime_init_done missing")
231 // Set the x_crosscall2_ptr C function pointer variable point to crosscall2.
232 if set_crosscall2 == nil {
233 throw("set_crosscall2 missing")
237 // Start the template thread in case we enter Go from
238 // a C-created thread and need to create a new thread.
239 startTemplateThread()
240 cgocall(_cgo_notify_runtime_init_done, nil)
243 // Run the initializing tasks. Depending on build mode this
244 // list can arrive a few different ways, but it will always
245 // contain the init tasks computed by the linker for all the
246 // packages in the program (excluding those added at runtime
247 // by package plugin).
248 for _, m := range activeModules() {
252 // Disable init tracing after main init done to avoid overhead
253 // of collecting statistics in malloc and newproc
254 inittrace.active = false
256 close(main_init_done)
261 if isarchive || islibrary {
262 // A program compiled with -buildmode=c-archive or c-shared
263 // has a main, but it is not executed.
266 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
269 runExitHooks(0) // run hooks now, since racefini does not return
273 // Make racy client program work: if panicking on
274 // another goroutine at the same time as main returns,
275 // let the other goroutine finish printing the panic trace.
276 // Once it does, it will exit. See issues 3934 and 20018.
277 if runningPanicDefers.Load() != 0 {
278 // Running deferred functions should not take long.
279 for c := 0; c < 1000; c++ {
280 if runningPanicDefers.Load() == 0 {
286 if panicking.Load() != 0 {
287 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
298 // os_beforeExit is called from os.Exit(0).
300 //go:linkname os_beforeExit os.runtime_beforeExit
301 func os_beforeExit(exitCode int) {
302 runExitHooks(exitCode)
303 if exitCode == 0 && raceenabled {
308 // start forcegc helper goroutine
313 func forcegchelper() {
315 lockInit(&forcegc.lock, lockRankForcegc)
318 if forcegc.idle.Load() {
319 throw("forcegc: phase error")
321 forcegc.idle.Store(true)
322 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
323 // this goroutine is explicitly resumed by sysmon
324 if debug.gctrace > 0 {
327 // Time-triggered, fully concurrent.
328 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
332 // Gosched yields the processor, allowing other goroutines to run. It does not
333 // suspend the current goroutine, so execution resumes automatically.
341 // goschedguarded yields the processor like gosched, but also checks
342 // for forbidden states and opts out of the yield in those cases.
345 func goschedguarded() {
346 mcall(goschedguarded_m)
349 // goschedIfBusy yields the processor like gosched, but only does so if
350 // there are no idle Ps or if we're on the only P and there's nothing in
351 // the run queue. In both cases, there is freely available idle time.
354 func goschedIfBusy() {
356 // Call gosched if gp.preempt is set; we may be in a tight loop that
357 // doesn't otherwise yield.
358 if !gp.preempt && sched.npidle.Load() > 0 {
364 // Puts the current goroutine into a waiting state and calls unlockf on the
367 // If unlockf returns false, the goroutine is resumed.
369 // unlockf must not access this G's stack, as it may be moved between
370 // the call to gopark and the call to unlockf.
372 // Note that because unlockf is called after putting the G into a waiting
373 // state, the G may have already been readied by the time unlockf is called
374 // unless there is external synchronization preventing the G from being
375 // readied. If unlockf returns false, it must guarantee that the G cannot be
376 // externally readied.
378 // Reason explains why the goroutine has been parked. It is displayed in stack
379 // traces and heap dumps. Reasons should be unique and descriptive. Do not
380 // re-use reasons, add new ones.
381 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
382 if reason != waitReasonSleep {
383 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
387 status := readgstatus(gp)
388 if status != _Grunning && status != _Gscanrunning {
389 throw("gopark: bad g status")
392 mp.waitunlockf = unlockf
393 gp.waitreason = reason
394 mp.waittraceev = traceEv
395 mp.waittraceskip = traceskip
397 // can't do anything that might move the G between Ms here.
401 // Puts the current goroutine into a waiting state and unlocks the lock.
402 // The goroutine can be made runnable again by calling goready(gp).
403 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
404 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
407 func goready(gp *g, traceskip int) {
409 ready(gp, traceskip, true)
414 func acquireSudog() *sudog {
415 // Delicate dance: the semaphore implementation calls
416 // acquireSudog, acquireSudog calls new(sudog),
417 // new calls malloc, malloc can call the garbage collector,
418 // and the garbage collector calls the semaphore implementation
420 // Break the cycle by doing acquirem/releasem around new(sudog).
421 // The acquirem/releasem increments m.locks during new(sudog),
422 // which keeps the garbage collector from being invoked.
425 if len(pp.sudogcache) == 0 {
426 lock(&sched.sudoglock)
427 // First, try to grab a batch from central cache.
428 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
429 s := sched.sudogcache
430 sched.sudogcache = s.next
432 pp.sudogcache = append(pp.sudogcache, s)
434 unlock(&sched.sudoglock)
435 // If the central cache is empty, allocate a new one.
436 if len(pp.sudogcache) == 0 {
437 pp.sudogcache = append(pp.sudogcache, new(sudog))
440 n := len(pp.sudogcache)
441 s := pp.sudogcache[n-1]
442 pp.sudogcache[n-1] = nil
443 pp.sudogcache = pp.sudogcache[:n-1]
445 throw("acquireSudog: found s.elem != nil in cache")
452 func releaseSudog(s *sudog) {
454 throw("runtime: sudog with non-nil elem")
457 throw("runtime: sudog with non-false isSelect")
460 throw("runtime: sudog with non-nil next")
463 throw("runtime: sudog with non-nil prev")
465 if s.waitlink != nil {
466 throw("runtime: sudog with non-nil waitlink")
469 throw("runtime: sudog with non-nil c")
473 throw("runtime: releaseSudog with non-nil gp.param")
475 mp := acquirem() // avoid rescheduling to another P
477 if len(pp.sudogcache) == cap(pp.sudogcache) {
478 // Transfer half of local cache to the central cache.
479 var first, last *sudog
480 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
481 n := len(pp.sudogcache)
482 p := pp.sudogcache[n-1]
483 pp.sudogcache[n-1] = nil
484 pp.sudogcache = pp.sudogcache[:n-1]
492 lock(&sched.sudoglock)
493 last.next = sched.sudogcache
494 sched.sudogcache = first
495 unlock(&sched.sudoglock)
497 pp.sudogcache = append(pp.sudogcache, s)
501 // called from assembly.
502 func badmcall(fn func(*g)) {
503 throw("runtime: mcall called on m->g0 stack")
506 func badmcall2(fn func(*g)) {
507 throw("runtime: mcall function returned")
510 func badreflectcall() {
511 panic(plainError("arg size to reflect.call more than 1GB"))
515 //go:nowritebarrierrec
516 func badmorestackg0() {
517 writeErrStr("fatal: morestack on g0\n")
521 //go:nowritebarrierrec
522 func badmorestackgsignal() {
523 writeErrStr("fatal: morestack on gsignal\n")
531 func lockedOSThread() bool {
533 return gp.lockedm != 0 && gp.m.lockedg != 0
537 // allgs contains all Gs ever created (including dead Gs), and thus
540 // Access via the slice is protected by allglock or stop-the-world.
541 // Readers that cannot take the lock may (carefully!) use the atomic
546 // allglen and allgptr are atomic variables that contain len(allgs) and
547 // &allgs[0] respectively. Proper ordering depends on totally-ordered
548 // loads and stores. Writes are protected by allglock.
550 // allgptr is updated before allglen. Readers should read allglen
551 // before allgptr to ensure that allglen is always <= len(allgptr). New
552 // Gs appended during the race can be missed. For a consistent view of
553 // all Gs, allglock must be held.
555 // allgptr copies should always be stored as a concrete type or
556 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
557 // even if it points to a stale array.
562 func allgadd(gp *g) {
563 if readgstatus(gp) == _Gidle {
564 throw("allgadd: bad status Gidle")
568 allgs = append(allgs, gp)
569 if &allgs[0] != allgptr {
570 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
572 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
576 // allGsSnapshot returns a snapshot of the slice of all Gs.
578 // The world must be stopped or allglock must be held.
579 func allGsSnapshot() []*g {
580 assertWorldStoppedOrLockHeld(&allglock)
582 // Because the world is stopped or allglock is held, allgadd
583 // cannot happen concurrently with this. allgs grows
584 // monotonically and existing entries never change, so we can
585 // simply return a copy of the slice header. For added safety,
586 // we trim everything past len because that can still change.
587 return allgs[:len(allgs):len(allgs)]
590 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
591 func atomicAllG() (**g, uintptr) {
592 length := atomic.Loaduintptr(&allglen)
593 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
597 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
598 func atomicAllGIndex(ptr **g, i uintptr) *g {
599 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
602 // forEachG calls fn on every G from allgs.
604 // forEachG takes a lock to exclude concurrent addition of new Gs.
605 func forEachG(fn func(gp *g)) {
607 for _, gp := range allgs {
613 // forEachGRace calls fn on every G from allgs.
615 // forEachGRace avoids locking, but does not exclude addition of new Gs during
616 // execution, which may be missed.
617 func forEachGRace(fn func(gp *g)) {
618 ptr, length := atomicAllG()
619 for i := uintptr(0); i < length; i++ {
620 gp := atomicAllGIndex(ptr, i)
627 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
628 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
632 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
633 // value of the GODEBUG environment variable.
634 func cpuinit(env string) {
636 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
637 cpu.DebugOptions = true
641 // Support cpu feature variables are used in code generated by the compiler
642 // to guard execution of instructions that can not be assumed to be always supported.
645 x86HasPOPCNT = cpu.X86.HasPOPCNT
646 x86HasSSE41 = cpu.X86.HasSSE41
647 x86HasFMA = cpu.X86.HasFMA
650 armHasVFPv4 = cpu.ARM.HasVFPv4
653 arm64HasATOMICS = cpu.ARM64.HasATOMICS
657 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
658 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
659 // early before much of the runtime is initialized.
660 func getGodebugEarly() string {
661 const prefix = "GODEBUG="
664 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
665 // Similar to goenv_unix but extracts the environment value for
667 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
669 for argv_index(argv, argc+1+n) != nil {
673 for i := int32(0); i < n; i++ {
674 p := argv_index(argv, argc+1+i)
675 s := unsafe.String(p, findnull(p))
677 if hasPrefix(s, prefix) {
678 env = gostring(p)[len(prefix):]
686 // The bootstrap sequence is:
690 // make & queue new G
691 // call runtime·mstart
693 // The new G calls runtime·main.
695 lockInit(&sched.lock, lockRankSched)
696 lockInit(&sched.sysmonlock, lockRankSysmon)
697 lockInit(&sched.deferlock, lockRankDefer)
698 lockInit(&sched.sudoglock, lockRankSudog)
699 lockInit(&deadlock, lockRankDeadlock)
700 lockInit(&paniclk, lockRankPanic)
701 lockInit(&allglock, lockRankAllg)
702 lockInit(&allpLock, lockRankAllp)
703 lockInit(&reflectOffs.lock, lockRankReflectOffs)
704 lockInit(&finlock, lockRankFin)
705 lockInit(&cpuprof.lock, lockRankCpuprof)
707 // Enforce that this lock is always a leaf lock.
708 // All of this lock's critical sections should be
710 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
712 // raceinit must be the first call to race detector.
713 // In particular, it must be done before mallocinit below calls racemapshadow.
716 gp.racectx, raceprocctx0 = raceinit()
719 sched.maxmcount = 10000
721 // The world starts stopped.
727 godebug := getGodebugEarly()
728 initPageTrace(godebug) // must run after mallocinit but before anything allocates
729 cpuinit(godebug) // must run before alginit
730 alginit() // maps, hash, fastrand must not be used before this call
731 fastrandinit() // must run before mcommoninit
732 mcommoninit(gp.m, -1)
733 modulesinit() // provides activeModules
734 typelinksinit() // uses maps, activeModules
735 itabsinit() // uses activeModules
736 stkobjinit() // must run before GC starts
738 sigsave(&gp.m.sigmask)
739 initSigmask = gp.m.sigmask
746 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
747 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
748 // set to true by the linker, it means that nothing is consuming the profile, it is
749 // safe to set MemProfileRate to 0.
750 if disableMemoryProfiling {
755 sched.lastpoll.Store(nanotime())
757 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
760 if procresize(procs) != nil {
761 throw("unknown runnable goroutine during bootstrap")
765 // World is effectively started now, as P's can run.
768 if buildVersion == "" {
769 // Condition should never trigger. This code just serves
770 // to ensure runtime·buildVersion is kept in the resulting binary.
771 buildVersion = "unknown"
773 if len(modinfo) == 1 {
774 // Condition should never trigger. This code just serves
775 // to ensure runtime·modinfo is kept in the resulting binary.
780 func dumpgstatus(gp *g) {
782 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
783 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
786 // sched.lock must be held.
788 assertLockHeld(&sched.lock)
790 // Exclude extra M's, which are used for cgocallback from threads
793 // The purpose of the SetMaxThreads limit is to avoid accidental fork
794 // bomb from something like millions of goroutines blocking on system
795 // calls, causing the runtime to create millions of threads. By
796 // definition, this isn't a problem for threads created in C, so we
797 // exclude them from the limit. See https://go.dev/issue/60004.
798 count := mcount() - int32(extraMInUse.Load()) - int32(extraMLength.Load())
799 if count > sched.maxmcount {
800 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
801 throw("thread exhaustion")
805 // mReserveID returns the next ID to use for a new m. This new m is immediately
806 // considered 'running' by checkdead.
808 // sched.lock must be held.
809 func mReserveID() int64 {
810 assertLockHeld(&sched.lock)
812 if sched.mnext+1 < sched.mnext {
813 throw("runtime: thread ID overflow")
821 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
822 func mcommoninit(mp *m, id int64) {
825 // g0 stack won't make sense for user (and is not necessary unwindable).
827 callers(1, mp.createstack[:])
838 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
839 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
843 // Same behavior as for 1.17.
844 // TODO: Simplify this.
845 if goarch.BigEndian {
846 mp.fastrand = uint64(lo)<<32 | uint64(hi)
848 mp.fastrand = uint64(hi)<<32 | uint64(lo)
852 if mp.gsignal != nil {
853 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
856 // Add to allm so garbage collector doesn't free g->m
857 // when it is just in a register or thread-local storage.
860 // NumCgoCall() iterates over allm w/o schedlock,
861 // so we need to publish it safely.
862 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
865 // Allocate memory to hold a cgo traceback if the cgo call crashes.
866 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
867 mp.cgoCallers = new(cgoCallers)
871 func (mp *m) becomeSpinning() {
873 sched.nmspinning.Add(1)
874 sched.needspinning.Store(0)
877 func (mp *m) hasCgoOnStack() bool {
878 return mp.ncgo > 0 || mp.isextra
881 var fastrandseed uintptr
883 func fastrandinit() {
884 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
888 // Mark gp ready to run.
889 func ready(gp *g, traceskip int, next bool) {
891 traceGoUnpark(gp, traceskip)
894 status := readgstatus(gp)
897 mp := acquirem() // disable preemption because it can be holding p in a local var
898 if status&^_Gscan != _Gwaiting {
900 throw("bad g->status in ready")
903 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
904 casgstatus(gp, _Gwaiting, _Grunnable)
905 runqput(mp.p.ptr(), gp, next)
910 // freezeStopWait is a large value that freezetheworld sets
911 // sched.stopwait to in order to request that all Gs permanently stop.
912 const freezeStopWait = 0x7fffffff
914 // freezing is set to non-zero if the runtime is trying to freeze the
916 var freezing atomic.Bool
918 // Similar to stopTheWorld but best-effort and can be called several times.
919 // There is no reverse operation, used during crashing.
920 // This function must not lock any mutexes.
921 func freezetheworld() {
923 // stopwait and preemption requests can be lost
924 // due to races with concurrently executing threads,
925 // so try several times
926 for i := 0; i < 5; i++ {
927 // this should tell the scheduler to not start any new goroutines
928 sched.stopwait = freezeStopWait
929 sched.gcwaiting.Store(true)
930 // this should stop running goroutines
932 break // no running goroutines
942 // All reads and writes of g's status go through readgstatus, casgstatus
943 // castogscanstatus, casfrom_Gscanstatus.
946 func readgstatus(gp *g) uint32 {
947 return gp.atomicstatus.Load()
950 // The Gscanstatuses are acting like locks and this releases them.
951 // If it proves to be a performance hit we should be able to make these
952 // simple atomic stores but for now we are going to throw if
953 // we see an inconsistent state.
954 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
957 // Check that transition is valid.
960 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
962 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
968 if newval == oldval&^_Gscan {
969 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
973 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
975 throw("casfrom_Gscanstatus: gp->status is not in scan state")
977 releaseLockRank(lockRankGscan)
980 // This will return false if the gp is not in the expected status and the cas fails.
981 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
982 func castogscanstatus(gp *g, oldval, newval uint32) bool {
988 if newval == oldval|_Gscan {
989 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
991 acquireLockRank(lockRankGscan)
997 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
998 throw("castogscanstatus")
1002 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
1003 // various latencies on every transition instead of sampling them.
1004 var casgstatusAlwaysTrack = false
1006 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
1007 // and casfrom_Gscanstatus instead.
1008 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
1009 // put it in the Gscan state is finished.
1012 func casgstatus(gp *g, oldval, newval uint32) {
1013 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
1014 systemstack(func() {
1015 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
1016 throw("casgstatus: bad incoming values")
1020 acquireLockRank(lockRankGscan)
1021 releaseLockRank(lockRankGscan)
1023 // See https://golang.org/cl/21503 for justification of the yield delay.
1024 const yieldDelay = 5 * 1000
1027 // loop if gp->atomicstatus is in a scan state giving
1028 // GC time to finish and change the state to oldval.
1029 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1030 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1031 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1034 nextYield = nanotime() + yieldDelay
1036 if nanotime() < nextYield {
1037 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1042 nextYield = nanotime() + yieldDelay/2
1046 if oldval == _Grunning {
1047 // Track every gTrackingPeriod time a goroutine transitions out of running.
1048 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1057 // Handle various kinds of tracking.
1060 // - Time spent in runnable.
1061 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1064 // We transitioned out of runnable, so measure how much
1065 // time we spent in this state and add it to
1068 gp.runnableTime += now - gp.trackingStamp
1069 gp.trackingStamp = 0
1071 if !gp.waitreason.isMutexWait() {
1072 // Not blocking on a lock.
1075 // Blocking on a lock, measure it. Note that because we're
1076 // sampling, we have to multiply by our sampling period to get
1077 // a more representative estimate of the absolute value.
1078 // gTrackingPeriod also represents an accurate sampling period
1079 // because we can only enter this state from _Grunning.
1081 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1082 gp.trackingStamp = 0
1086 if !gp.waitreason.isMutexWait() {
1087 // Not blocking on a lock.
1090 // Blocking on a lock. Write down the timestamp.
1092 gp.trackingStamp = now
1094 // We just transitioned into runnable, so record what
1095 // time that happened.
1097 gp.trackingStamp = now
1099 // We're transitioning into running, so turn off
1100 // tracking and record how much time we spent in
1103 sched.timeToRun.record(gp.runnableTime)
1108 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1110 // Use this over casgstatus when possible to ensure that a waitreason is set.
1111 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1112 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1113 gp.waitreason = reason
1114 casgstatus(gp, old, _Gwaiting)
1117 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1118 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1119 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1120 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1121 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1124 func casgcopystack(gp *g) uint32 {
1126 oldstatus := readgstatus(gp) &^ _Gscan
1127 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1128 throw("copystack: bad status, not Gwaiting or Grunnable")
1130 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1136 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1138 // TODO(austin): This is the only status operation that both changes
1139 // the status and locks the _Gscan bit. Rethink this.
1140 func casGToPreemptScan(gp *g, old, new uint32) {
1141 if old != _Grunning || new != _Gscan|_Gpreempted {
1142 throw("bad g transition")
1144 acquireLockRank(lockRankGscan)
1145 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1149 // casGFromPreempted attempts to transition gp from _Gpreempted to
1150 // _Gwaiting. If successful, the caller is responsible for
1151 // re-scheduling gp.
1152 func casGFromPreempted(gp *g, old, new uint32) bool {
1153 if old != _Gpreempted || new != _Gwaiting {
1154 throw("bad g transition")
1156 gp.waitreason = waitReasonPreempted
1157 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1160 // stopTheWorld stops all P's from executing goroutines, interrupting
1161 // all goroutines at GC safe points and records reason as the reason
1162 // for the stop. On return, only the current goroutine's P is running.
1163 // stopTheWorld must not be called from a system stack and the caller
1164 // must not hold worldsema. The caller must call startTheWorld when
1165 // other P's should resume execution.
1167 // stopTheWorld is safe for multiple goroutines to call at the
1168 // same time. Each will execute its own stop, and the stops will
1171 // This is also used by routines that do stack dumps. If the system is
1172 // in panic or being exited, this may not reliably stop all
1174 func stopTheWorld(reason string) {
1175 semacquire(&worldsema)
1177 gp.m.preemptoff = reason
1178 systemstack(func() {
1179 // Mark the goroutine which called stopTheWorld preemptible so its
1180 // stack may be scanned.
1181 // This lets a mark worker scan us while we try to stop the world
1182 // since otherwise we could get in a mutual preemption deadlock.
1183 // We must not modify anything on the G stack because a stack shrink
1184 // may occur. A stack shrink is otherwise OK though because in order
1185 // to return from this function (and to leave the system stack) we
1186 // must have preempted all goroutines, including any attempting
1187 // to scan our stack, in which case, any stack shrinking will
1188 // have already completed by the time we exit.
1189 // Don't provide a wait reason because we're still executing.
1190 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1191 stopTheWorldWithSema()
1192 casgstatus(gp, _Gwaiting, _Grunning)
1196 // startTheWorld undoes the effects of stopTheWorld.
1197 func startTheWorld() {
1198 systemstack(func() { startTheWorldWithSema(false) })
1200 // worldsema must be held over startTheWorldWithSema to ensure
1201 // gomaxprocs cannot change while worldsema is held.
1203 // Release worldsema with direct handoff to the next waiter, but
1204 // acquirem so that semrelease1 doesn't try to yield our time.
1206 // Otherwise if e.g. ReadMemStats is being called in a loop,
1207 // it might stomp on other attempts to stop the world, such as
1208 // for starting or ending GC. The operation this blocks is
1209 // so heavy-weight that we should just try to be as fair as
1212 // We don't want to just allow us to get preempted between now
1213 // and releasing the semaphore because then we keep everyone
1214 // (including, for example, GCs) waiting longer.
1217 semrelease1(&worldsema, true, 0)
1221 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1222 // until the GC is not running. It also blocks a GC from starting
1223 // until startTheWorldGC is called.
1224 func stopTheWorldGC(reason string) {
1226 stopTheWorld(reason)
1229 // startTheWorldGC undoes the effects of stopTheWorldGC.
1230 func startTheWorldGC() {
1235 // Holding worldsema grants an M the right to try to stop the world.
1236 var worldsema uint32 = 1
1238 // Holding gcsema grants the M the right to block a GC, and blocks
1239 // until the current GC is done. In particular, it prevents gomaxprocs
1240 // from changing concurrently.
1242 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1243 // being changed/enabled during a GC, remove this.
1244 var gcsema uint32 = 1
1246 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1247 // The caller is responsible for acquiring worldsema and disabling
1248 // preemption first and then should stopTheWorldWithSema on the system
1251 // semacquire(&worldsema, 0)
1252 // m.preemptoff = "reason"
1253 // systemstack(stopTheWorldWithSema)
1255 // When finished, the caller must either call startTheWorld or undo
1256 // these three operations separately:
1258 // m.preemptoff = ""
1259 // systemstack(startTheWorldWithSema)
1260 // semrelease(&worldsema)
1262 // It is allowed to acquire worldsema once and then execute multiple
1263 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1264 // Other P's are able to execute between successive calls to
1265 // startTheWorldWithSema and stopTheWorldWithSema.
1266 // Holding worldsema causes any other goroutines invoking
1267 // stopTheWorld to block.
1268 func stopTheWorldWithSema() {
1271 // If we hold a lock, then we won't be able to stop another M
1272 // that is blocked trying to acquire the lock.
1274 throw("stopTheWorld: holding locks")
1278 sched.stopwait = gomaxprocs
1279 sched.gcwaiting.Store(true)
1282 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1284 // try to retake all P's in Psyscall status
1285 for _, pp := range allp {
1287 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1299 pp, _ := pidleget(now)
1303 pp.status = _Pgcstop
1306 wait := sched.stopwait > 0
1309 // wait for remaining P's to stop voluntarily
1312 // wait for 100us, then try to re-preempt in case of any races
1313 if notetsleep(&sched.stopnote, 100*1000) {
1314 noteclear(&sched.stopnote)
1323 if sched.stopwait != 0 {
1324 bad = "stopTheWorld: not stopped (stopwait != 0)"
1326 for _, pp := range allp {
1327 if pp.status != _Pgcstop {
1328 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1332 if freezing.Load() {
1333 // Some other thread is panicking. This can cause the
1334 // sanity checks above to fail if the panic happens in
1335 // the signal handler on a stopped thread. Either way,
1336 // we should halt this thread.
1347 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1348 assertWorldStopped()
1350 mp := acquirem() // disable preemption because it can be holding p in a local var
1351 if netpollinited() {
1352 list := netpoll(0) // non-blocking
1362 p1 := procresize(procs)
1363 sched.gcwaiting.Store(false)
1364 if sched.sysmonwait.Load() {
1365 sched.sysmonwait.Store(false)
1366 notewakeup(&sched.sysmonnote)
1379 throw("startTheWorld: inconsistent mp->nextp")
1382 notewakeup(&mp.park)
1384 // Start M to run P. Do not start another M below.
1389 // Capture start-the-world time before doing clean-up tasks.
1390 startTime := nanotime()
1395 // Wakeup an additional proc in case we have excessive runnable goroutines
1396 // in local queues or in the global queue. If we don't, the proc will park itself.
1397 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1405 // usesLibcall indicates whether this runtime performs system calls
1407 func usesLibcall() bool {
1409 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1412 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1417 // mStackIsSystemAllocated indicates whether this runtime starts on a
1418 // system-allocated stack.
1419 func mStackIsSystemAllocated() bool {
1421 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1425 case "386", "amd64", "arm", "arm64":
1432 // mstart is the entry-point for new Ms.
1433 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1436 // mstart0 is the Go entry-point for new Ms.
1437 // This must not split the stack because we may not even have stack
1438 // bounds set up yet.
1440 // May run during STW (because it doesn't have a P yet), so write
1441 // barriers are not allowed.
1444 //go:nowritebarrierrec
1448 osStack := gp.stack.lo == 0
1450 // Initialize stack bounds from system stack.
1451 // Cgo may have left stack size in stack.hi.
1452 // minit may update the stack bounds.
1454 // Note: these bounds may not be very accurate.
1455 // We set hi to &size, but there are things above
1456 // it. The 1024 is supposed to compensate this,
1457 // but is somewhat arbitrary.
1460 size = 8192 * sys.StackGuardMultiplier
1462 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1463 gp.stack.lo = gp.stack.hi - size + 1024
1465 // Initialize stack guard so that we can start calling regular
1467 gp.stackguard0 = gp.stack.lo + stackGuard
1468 // This is the g0, so we can also call go:systemstack
1469 // functions, which check stackguard1.
1470 gp.stackguard1 = gp.stackguard0
1473 // Exit this thread.
1474 if mStackIsSystemAllocated() {
1475 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1476 // the stack, but put it in gp.stack before mstart,
1477 // so the logic above hasn't set osStack yet.
1483 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1484 // so that we can set up g0.sched to return to the call of mstart1 above.
1491 throw("bad runtime·mstart")
1494 // Set up m.g0.sched as a label returning to just
1495 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1496 // We're never coming back to mstart1 after we call schedule,
1497 // so other calls can reuse the current frame.
1498 // And goexit0 does a gogo that needs to return from mstart1
1499 // and let mstart0 exit the thread.
1500 gp.sched.g = guintptr(unsafe.Pointer(gp))
1501 gp.sched.pc = getcallerpc()
1502 gp.sched.sp = getcallersp()
1507 // Install signal handlers; after minit so that minit can
1508 // prepare the thread to be able to handle the signals.
1513 if fn := gp.m.mstartfn; fn != nil {
1518 acquirep(gp.m.nextp.ptr())
1524 // mstartm0 implements part of mstart1 that only runs on the m0.
1526 // Write barriers are allowed here because we know the GC can't be
1527 // running yet, so they'll be no-ops.
1529 //go:yeswritebarrierrec
1531 // Create an extra M for callbacks on threads not created by Go.
1532 // An extra M is also needed on Windows for callbacks created by
1533 // syscall.NewCallback. See issue #6751 for details.
1534 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1541 // mPark causes a thread to park itself, returning once woken.
1546 notesleep(&gp.m.park)
1547 noteclear(&gp.m.park)
1550 // mexit tears down and exits the current thread.
1552 // Don't call this directly to exit the thread, since it must run at
1553 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1554 // unwind the stack to the point that exits the thread.
1556 // It is entered with m.p != nil, so write barriers are allowed. It
1557 // will release the P before exiting.
1559 //go:yeswritebarrierrec
1560 func mexit(osStack bool) {
1564 // This is the main thread. Just wedge it.
1566 // On Linux, exiting the main thread puts the process
1567 // into a non-waitable zombie state. On Plan 9,
1568 // exiting the main thread unblocks wait even though
1569 // other threads are still running. On Solaris we can
1570 // neither exitThread nor return from mstart. Other
1571 // bad things probably happen on other platforms.
1573 // We could try to clean up this M more before wedging
1574 // it, but that complicates signal handling.
1575 handoffp(releasep())
1581 throw("locked m0 woke up")
1587 // Free the gsignal stack.
1588 if mp.gsignal != nil {
1589 stackfree(mp.gsignal.stack)
1590 // On some platforms, when calling into VDSO (e.g. nanotime)
1591 // we store our g on the gsignal stack, if there is one.
1592 // Now the stack is freed, unlink it from the m, so we
1593 // won't write to it when calling VDSO code.
1597 // Remove m from allm.
1599 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1605 throw("m not found in allm")
1607 // Delay reaping m until it's done with the stack.
1609 // Put mp on the free list, though it will not be reaped while freeWait
1610 // is freeMWait. mp is no longer reachable via allm, so even if it is
1611 // on an OS stack, we must keep a reference to mp alive so that the GC
1612 // doesn't free mp while we are still using it.
1614 // Note that the free list must not be linked through alllink because
1615 // some functions walk allm without locking, so may be using alllink.
1616 mp.freeWait.Store(freeMWait)
1617 mp.freelink = sched.freem
1621 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1624 handoffp(releasep())
1625 // After this point we must not have write barriers.
1627 // Invoke the deadlock detector. This must happen after
1628 // handoffp because it may have started a new M to take our
1635 if GOOS == "darwin" || GOOS == "ios" {
1636 // Make sure pendingPreemptSignals is correct when an M exits.
1638 if mp.signalPending.Load() != 0 {
1639 pendingPreemptSignals.Add(-1)
1643 // Destroy all allocated resources. After this is called, we may no
1644 // longer take any locks.
1648 // No more uses of mp, so it is safe to drop the reference.
1649 mp.freeWait.Store(freeMRef)
1651 // Return from mstart and let the system thread
1652 // library free the g0 stack and terminate the thread.
1656 // mstart is the thread's entry point, so there's nothing to
1657 // return to. Exit the thread directly. exitThread will clear
1658 // m.freeWait when it's done with the stack and the m can be
1660 exitThread(&mp.freeWait)
1663 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1664 // If a P is currently executing code, this will bring the P to a GC
1665 // safe point and execute fn on that P. If the P is not executing code
1666 // (it is idle or in a syscall), this will call fn(p) directly while
1667 // preventing the P from exiting its state. This does not ensure that
1668 // fn will run on every CPU executing Go code, but it acts as a global
1669 // memory barrier. GC uses this as a "ragged barrier."
1671 // The caller must hold worldsema.
1674 func forEachP(fn func(*p)) {
1676 pp := getg().m.p.ptr()
1679 if sched.safePointWait != 0 {
1680 throw("forEachP: sched.safePointWait != 0")
1682 sched.safePointWait = gomaxprocs - 1
1683 sched.safePointFn = fn
1685 // Ask all Ps to run the safe point function.
1686 for _, p2 := range allp {
1688 atomic.Store(&p2.runSafePointFn, 1)
1693 // Any P entering _Pidle or _Psyscall from now on will observe
1694 // p.runSafePointFn == 1 and will call runSafePointFn when
1695 // changing its status to _Pidle/_Psyscall.
1697 // Run safe point function for all idle Ps. sched.pidle will
1698 // not change because we hold sched.lock.
1699 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1700 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1702 sched.safePointWait--
1706 wait := sched.safePointWait > 0
1709 // Run fn for the current P.
1712 // Force Ps currently in _Psyscall into _Pidle and hand them
1713 // off to induce safe point function execution.
1714 for _, p2 := range allp {
1716 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1726 // Wait for remaining Ps to run fn.
1729 // Wait for 100us, then try to re-preempt in
1730 // case of any races.
1732 // Requires system stack.
1733 if notetsleep(&sched.safePointNote, 100*1000) {
1734 noteclear(&sched.safePointNote)
1740 if sched.safePointWait != 0 {
1741 throw("forEachP: not done")
1743 for _, p2 := range allp {
1744 if p2.runSafePointFn != 0 {
1745 throw("forEachP: P did not run fn")
1750 sched.safePointFn = nil
1755 // runSafePointFn runs the safe point function, if any, for this P.
1756 // This should be called like
1758 // if getg().m.p.runSafePointFn != 0 {
1762 // runSafePointFn must be checked on any transition in to _Pidle or
1763 // _Psyscall to avoid a race where forEachP sees that the P is running
1764 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1765 // nor the P run the safe-point function.
1766 func runSafePointFn() {
1767 p := getg().m.p.ptr()
1768 // Resolve the race between forEachP running the safe-point
1769 // function on this P's behalf and this P running the
1770 // safe-point function directly.
1771 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1774 sched.safePointFn(p)
1776 sched.safePointWait--
1777 if sched.safePointWait == 0 {
1778 notewakeup(&sched.safePointNote)
1783 // When running with cgo, we call _cgo_thread_start
1784 // to start threads for us so that we can play nicely with
1786 var cgoThreadStart unsafe.Pointer
1788 type cgothreadstart struct {
1794 // Allocate a new m unassociated with any thread.
1795 // Can use p for allocation context if needed.
1796 // fn is recorded as the new m's m.mstartfn.
1797 // id is optional pre-allocated m ID. Omit by passing -1.
1799 // This function is allowed to have write barriers even if the caller
1800 // isn't because it borrows pp.
1802 //go:yeswritebarrierrec
1803 func allocm(pp *p, fn func(), id int64) *m {
1806 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1807 // disable preemption to ensure it is not stolen, which would make the
1808 // caller lose ownership.
1813 acquirep(pp) // temporarily borrow p for mallocs in this function
1816 // Release the free M list. We need to do this somewhere and
1817 // this may free up a stack we can use.
1818 if sched.freem != nil {
1821 for freem := sched.freem; freem != nil; {
1822 wait := freem.freeWait.Load()
1823 if wait == freeMWait {
1824 next := freem.freelink
1825 freem.freelink = newList
1830 // Free the stack if needed. For freeMRef, there is
1831 // nothing to do except drop freem from the sched.freem
1833 if wait == freeMStack {
1834 // stackfree must be on the system stack, but allocm is
1835 // reachable off the system stack transitively from
1837 systemstack(func() {
1838 stackfree(freem.g0.stack)
1841 freem = freem.freelink
1843 sched.freem = newList
1851 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1852 // Windows and Plan 9 will layout sched stack on OS stack.
1853 if iscgo || mStackIsSystemAllocated() {
1856 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1860 if pp == gp.m.p.ptr() {
1865 allocmLock.runlock()
1869 // needm is called when a cgo callback happens on a
1870 // thread without an m (a thread not created by Go).
1871 // In this case, needm is expected to find an m to use
1872 // and return with m, g initialized correctly.
1873 // Since m and g are not set now (likely nil, but see below)
1874 // needm is limited in what routines it can call. In particular
1875 // it can only call nosplit functions (textflag 7) and cannot
1876 // do any scheduling that requires an m.
1878 // In order to avoid needing heavy lifting here, we adopt
1879 // the following strategy: there is a stack of available m's
1880 // that can be stolen. Using compare-and-swap
1881 // to pop from the stack has ABA races, so we simulate
1882 // a lock by doing an exchange (via Casuintptr) to steal the stack
1883 // head and replace the top pointer with MLOCKED (1).
1884 // This serves as a simple spin lock that we can use even
1885 // without an m. The thread that locks the stack in this way
1886 // unlocks the stack by storing a valid stack head pointer.
1888 // In order to make sure that there is always an m structure
1889 // available to be stolen, we maintain the invariant that there
1890 // is always one more than needed. At the beginning of the
1891 // program (if cgo is in use) the list is seeded with a single m.
1892 // If needm finds that it has taken the last m off the list, its job
1893 // is - once it has installed its own m so that it can do things like
1894 // allocate memory - to create a spare m and put it on the list.
1896 // Each of these extra m's also has a g0 and a curg that are
1897 // pressed into service as the scheduling stack and current
1898 // goroutine for the duration of the cgo callback.
1900 // It calls dropm to put the m back on the list,
1901 // 1. when the callback is done with the m in non-pthread platforms,
1902 // 2. or when the C thread exiting on pthread platforms.
1904 // The signal argument indicates whether we're called from a signal
1908 func needm(signal bool) {
1909 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1910 // Can happen if C/C++ code calls Go from a global ctor.
1911 // Can also happen on Windows if a global ctor uses a
1912 // callback created by syscall.NewCallback. See issue #6751
1915 // Can not throw, because scheduler is not initialized yet.
1916 writeErrStr("fatal error: cgo callback before cgo call\n")
1920 // Save and block signals before getting an M.
1921 // The signal handler may call needm itself,
1922 // and we must avoid a deadlock. Also, once g is installed,
1923 // any incoming signals will try to execute,
1924 // but we won't have the sigaltstack settings and other data
1925 // set up appropriately until the end of minit, which will
1926 // unblock the signals. This is the same dance as when
1927 // starting a new m to run Go code via newosproc.
1932 // nilokay=false is safe here because of the invariant above,
1933 // that the extra list always contains or will soon contain
1935 mp, last := getExtraM(false)
1937 // Set needextram when we've just emptied the list,
1938 // so that the eventual call into cgocallbackg will
1939 // allocate a new m for the extra list. We delay the
1940 // allocation until then so that it can be done
1941 // after exitsyscall makes sure it is okay to be
1942 // running at all (that is, there's no garbage collection
1943 // running right now).
1944 mp.needextram = last
1946 // Store the original signal mask for use by minit.
1947 mp.sigmask = sigmask
1949 // Install TLS on some platforms (previously setg
1950 // would do this if necessary).
1953 // Install g (= m->g0) and set the stack bounds
1954 // to match the current stack. If we don't actually know
1955 // how big the stack is, like we don't know how big any
1956 // scheduling stack is, but we assume there's at least 32 kB.
1957 // If we can get a more accurate stack bound from pthread,
1961 gp.stack.hi = getcallersp() + 1024
1962 gp.stack.lo = getcallersp() - 32*1024
1963 if !signal && _cgo_getstackbound != nil {
1964 // Don't adjust if called from the signal handler.
1965 // We are on the signal stack, not the pthread stack.
1966 // (We could get the stack bounds from sigaltstack, but
1967 // we're getting out of the signal handler very soon
1968 // anyway. Not worth it.)
1969 var bounds [2]uintptr
1970 asmcgocall(_cgo_getstackbound, unsafe.Pointer(&bounds))
1971 // getstackbound is an unsupported no-op on Windows.
1973 gp.stack.lo = bounds[0]
1974 gp.stack.hi = bounds[1]
1977 gp.stackguard0 = gp.stack.lo + stackGuard
1979 // Should mark we are already in Go now.
1980 // Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
1981 // which means the extram list may be empty, that will cause a deadlock.
1982 mp.isExtraInC = false
1984 // Initialize this thread to use the m.
1988 // mp.curg is now a real goroutine.
1989 casgstatus(mp.curg, _Gdead, _Gsyscall)
1993 // Acquire an extra m and bind it to the C thread when a pthread key has been created.
1996 func needAndBindM() {
1999 if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
2004 // newextram allocates m's and puts them on the extra list.
2005 // It is called with a working local m, so that it can do things
2006 // like call schedlock and allocate.
2008 c := extraMWaiters.Swap(0)
2010 for i := uint32(0); i < c; i++ {
2013 } else if extraMLength.Load() == 0 {
2014 // Make sure there is at least one extra M.
2019 // oneNewExtraM allocates an m and puts it on the extra list.
2020 func oneNewExtraM() {
2021 // Create extra goroutine locked to extra m.
2022 // The goroutine is the context in which the cgo callback will run.
2023 // The sched.pc will never be returned to, but setting it to
2024 // goexit makes clear to the traceback routines where
2025 // the goroutine stack ends.
2026 mp := allocm(nil, nil, -1)
2028 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
2029 gp.sched.sp = gp.stack.hi
2030 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
2032 gp.sched.g = guintptr(unsafe.Pointer(gp))
2033 gp.syscallpc = gp.sched.pc
2034 gp.syscallsp = gp.sched.sp
2035 gp.stktopsp = gp.sched.sp
2036 // malg returns status as _Gidle. Change to _Gdead before
2037 // adding to allg where GC can see it. We use _Gdead to hide
2038 // this from tracebacks and stack scans since it isn't a
2039 // "real" goroutine until needm grabs it.
2040 casgstatus(gp, _Gidle, _Gdead)
2044 // mark we are in C by default.
2045 mp.isExtraInC = true
2049 gp.goid = sched.goidgen.Add(1)
2051 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
2054 traceOneNewExtraM(gp)
2056 // put on allg for garbage collector
2059 // gp is now on the allg list, but we don't want it to be
2060 // counted by gcount. It would be more "proper" to increment
2061 // sched.ngfree, but that requires locking. Incrementing ngsys
2062 // has the same effect.
2065 // Add m to the extra list.
2069 // dropm puts the current m back onto the extra list.
2071 // 1. On systems without pthreads, like Windows
2072 // dropm is called when a cgo callback has called needm but is now
2073 // done with the callback and returning back into the non-Go thread.
2075 // The main expense here is the call to signalstack to release the
2076 // m's signal stack, and then the call to needm on the next callback
2077 // from this thread. It is tempting to try to save the m for next time,
2078 // which would eliminate both these costs, but there might not be
2079 // a next time: the current thread (which Go does not control) might exit.
2080 // If we saved the m for that thread, there would be an m leak each time
2081 // such a thread exited. Instead, we acquire and release an m on each
2082 // call. These should typically not be scheduling operations, just a few
2083 // atomics, so the cost should be small.
2085 // 2. On systems with pthreads
2086 // dropm is called while a non-Go thread is exiting.
2087 // We allocate a pthread per-thread variable using pthread_key_create,
2088 // to register a thread-exit-time destructor.
2089 // And store the g into a thread-specific value associated with the pthread key,
2090 // when first return back to C.
2091 // So that the destructor would invoke dropm while the non-Go thread is exiting.
2092 // This is much faster since it avoids expensive signal-related syscalls.
2094 // NOTE: this always runs without a P, so, nowritebarrierrec required.
2096 //go:nowritebarrierrec
2098 // Clear m and g, and return m to the extra list.
2099 // After the call to setg we can only call nosplit functions
2100 // with no pointer manipulation.
2103 // Return mp.curg to dead state.
2104 casgstatus(mp.curg, _Gsyscall, _Gdead)
2105 mp.curg.preemptStop = false
2108 // Block signals before unminit.
2109 // Unminit unregisters the signal handling stack (but needs g on some systems).
2110 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2111 // It's important not to try to handle a signal between those two steps.
2112 sigmask := mp.sigmask
2120 msigrestore(sigmask)
2123 // bindm store the g0 of the current m into a thread-specific value.
2125 // We allocate a pthread per-thread variable using pthread_key_create,
2126 // to register a thread-exit-time destructor.
2127 // We are here setting the thread-specific value of the pthread key, to enable the destructor.
2128 // So that the pthread_key_destructor would dropm while the C thread is exiting.
2130 // And the saved g will be used in pthread_key_destructor,
2131 // since the g stored in the TLS by Go might be cleared in some platforms,
2132 // before the destructor invoked, so, we restore g by the stored g, before dropm.
2134 // We store g0 instead of m, to make the assembly code simpler,
2135 // since we need to restore g0 in runtime.cgocallback.
2137 // On systems without pthreads, like Windows, bindm shouldn't be used.
2139 // NOTE: this always runs without a P, so, nowritebarrierrec required.
2142 //go:nowritebarrierrec
2144 if GOOS == "windows" || GOOS == "plan9" {
2145 fatal("bindm in unexpected GOOS")
2149 fatal("the current g is not g0")
2151 if _cgo_bindm != nil {
2152 asmcgocall(_cgo_bindm, unsafe.Pointer(g))
2156 // A helper function for EnsureDropM.
2157 func getm() uintptr {
2158 return uintptr(unsafe.Pointer(getg().m))
2162 // Locking linked list of extra M's, via mp.schedlink. Must be accessed
2163 // only via lockextra/unlockextra.
2165 // Can't be atomic.Pointer[m] because we use an invalid pointer as a
2166 // "locked" sentinel value. M's on this list remain visible to the GC
2167 // because their mp.curg is on allgs.
2168 extraM atomic.Uintptr
2169 // Number of M's in the extraM list.
2170 extraMLength atomic.Uint32
2171 // Number of waiters in lockextra.
2172 extraMWaiters atomic.Uint32
2174 // Number of extra M's in use by threads.
2175 extraMInUse atomic.Uint32
2178 // lockextra locks the extra list and returns the list head.
2179 // The caller must unlock the list by storing a new list head
2180 // to extram. If nilokay is true, then lockextra will
2181 // return a nil list head if that's what it finds. If nilokay is false,
2182 // lockextra will keep waiting until the list head is no longer nil.
2185 func lockextra(nilokay bool) *m {
2190 old := extraM.Load()
2195 if old == 0 && !nilokay {
2197 // Add 1 to the number of threads
2198 // waiting for an M.
2199 // This is cleared by newextram.
2200 extraMWaiters.Add(1)
2206 if extraM.CompareAndSwap(old, locked) {
2208 return (*m)(unsafe.Pointer(old))
2216 func unlockextra(mp *m, delta int32) {
2217 extraMLength.Add(delta)
2218 extraM.Store(uintptr(unsafe.Pointer(mp)))
2221 // Return an M from the extra M list. Returns last == true if the list becomes
2222 // empty because of this call.
2225 func getExtraM(nilokay bool) (mp *m, last bool) {
2226 mp = lockextra(nilokay)
2231 unlockextra(mp.schedlink.ptr(), -1)
2232 return mp, mp.schedlink.ptr() == nil
2235 // Returns an extra M back to the list. mp must be from getExtraM. Newly
2236 // allocated M's should use addExtraM.
2239 func putExtraM(mp *m) {
2244 // Adds a newly allocated M to the extra M list.
2247 func addExtraM(mp *m) {
2248 mnext := lockextra(true)
2249 mp.schedlink.set(mnext)
2254 // allocmLock is locked for read when creating new Ms in allocm and their
2255 // addition to allm. Thus acquiring this lock for write blocks the
2256 // creation of new Ms.
2259 // execLock serializes exec and clone to avoid bugs or unspecified
2260 // behaviour around exec'ing while creating/destroying threads. See
2265 // These errors are reported (via writeErrStr) by some OS-specific
2266 // versions of newosproc and newosproc0.
2268 failthreadcreate = "runtime: failed to create new OS thread\n"
2269 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2272 // newmHandoff contains a list of m structures that need new OS threads.
2273 // This is used by newm in situations where newm itself can't safely
2274 // start an OS thread.
2275 var newmHandoff struct {
2278 // newm points to a list of M structures that need new OS
2279 // threads. The list is linked through m.schedlink.
2282 // waiting indicates that wake needs to be notified when an m
2283 // is put on the list.
2287 // haveTemplateThread indicates that the templateThread has
2288 // been started. This is not protected by lock. Use cas to set
2290 haveTemplateThread uint32
2293 // Create a new m. It will start off with a call to fn, or else the scheduler.
2294 // fn needs to be static and not a heap allocated closure.
2295 // May run with m.p==nil, so write barriers are not allowed.
2297 // id is optional pre-allocated m ID. Omit by passing -1.
2299 //go:nowritebarrierrec
2300 func newm(fn func(), pp *p, id int64) {
2301 // allocm adds a new M to allm, but they do not start until created by
2302 // the OS in newm1 or the template thread.
2304 // doAllThreadsSyscall requires that every M in allm will eventually
2305 // start and be signal-able, even with a STW.
2307 // Disable preemption here until we start the thread to ensure that
2308 // newm is not preempted between allocm and starting the new thread,
2309 // ensuring that anything added to allm is guaranteed to eventually
2313 mp := allocm(pp, fn, id)
2315 mp.sigmask = initSigmask
2316 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2317 // We're on a locked M or a thread that may have been
2318 // started by C. The kernel state of this thread may
2319 // be strange (the user may have locked it for that
2320 // purpose). We don't want to clone that into another
2321 // thread. Instead, ask a known-good thread to create
2322 // the thread for us.
2324 // This is disabled on Plan 9. See golang.org/issue/22227.
2326 // TODO: This may be unnecessary on Windows, which
2327 // doesn't model thread creation off fork.
2328 lock(&newmHandoff.lock)
2329 if newmHandoff.haveTemplateThread == 0 {
2330 throw("on a locked thread with no template thread")
2332 mp.schedlink = newmHandoff.newm
2333 newmHandoff.newm.set(mp)
2334 if newmHandoff.waiting {
2335 newmHandoff.waiting = false
2336 notewakeup(&newmHandoff.wake)
2338 unlock(&newmHandoff.lock)
2339 // The M has not started yet, but the template thread does not
2340 // participate in STW, so it will always process queued Ms and
2341 // it is safe to releasem.
2351 var ts cgothreadstart
2352 if _cgo_thread_start == nil {
2353 throw("_cgo_thread_start missing")
2356 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2357 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2359 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2362 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2364 execLock.rlock() // Prevent process clone.
2365 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2369 execLock.rlock() // Prevent process clone.
2374 // startTemplateThread starts the template thread if it is not already
2377 // The calling thread must itself be in a known-good state.
2378 func startTemplateThread() {
2379 if GOARCH == "wasm" { // no threads on wasm yet
2383 // Disable preemption to guarantee that the template thread will be
2384 // created before a park once haveTemplateThread is set.
2386 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2390 newm(templateThread, nil, -1)
2394 // templateThread is a thread in a known-good state that exists solely
2395 // to start new threads in known-good states when the calling thread
2396 // may not be in a good state.
2398 // Many programs never need this, so templateThread is started lazily
2399 // when we first enter a state that might lead to running on a thread
2400 // in an unknown state.
2402 // templateThread runs on an M without a P, so it must not have write
2405 //go:nowritebarrierrec
2406 func templateThread() {
2413 lock(&newmHandoff.lock)
2414 for newmHandoff.newm != 0 {
2415 newm := newmHandoff.newm.ptr()
2416 newmHandoff.newm = 0
2417 unlock(&newmHandoff.lock)
2419 next := newm.schedlink.ptr()
2424 lock(&newmHandoff.lock)
2426 newmHandoff.waiting = true
2427 noteclear(&newmHandoff.wake)
2428 unlock(&newmHandoff.lock)
2429 notesleep(&newmHandoff.wake)
2433 // Stops execution of the current m until new work is available.
2434 // Returns with acquired P.
2438 if gp.m.locks != 0 {
2439 throw("stopm holding locks")
2442 throw("stopm holding p")
2445 throw("stopm spinning")
2452 acquirep(gp.m.nextp.ptr())
2457 // startm's caller incremented nmspinning. Set the new M's spinning.
2458 getg().m.spinning = true
2461 // Schedules some M to run the p (creates an M if necessary).
2462 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2463 // May run with m.p==nil, so write barriers are not allowed.
2464 // If spinning is set, the caller has incremented nmspinning and must provide a
2465 // P. startm will set m.spinning in the newly started M.
2467 // Callers passing a non-nil P must call from a non-preemptible context. See
2468 // comment on acquirem below.
2470 // Argument lockheld indicates whether the caller already acquired the
2471 // scheduler lock. Callers holding the lock when making the call must pass
2472 // true. The lock might be temporarily dropped, but will be reacquired before
2475 // Must not have write barriers because this may be called without a P.
2477 //go:nowritebarrierrec
2478 func startm(pp *p, spinning, lockheld bool) {
2479 // Disable preemption.
2481 // Every owned P must have an owner that will eventually stop it in the
2482 // event of a GC stop request. startm takes transient ownership of a P
2483 // (either from argument or pidleget below) and transfers ownership to
2484 // a started M, which will be responsible for performing the stop.
2486 // Preemption must be disabled during this transient ownership,
2487 // otherwise the P this is running on may enter GC stop while still
2488 // holding the transient P, leaving that P in limbo and deadlocking the
2491 // Callers passing a non-nil P must already be in non-preemptible
2492 // context, otherwise such preemption could occur on function entry to
2493 // startm. Callers passing a nil P may be preemptible, so we must
2494 // disable preemption before acquiring a P from pidleget below.
2501 // TODO(prattmic): All remaining calls to this function
2502 // with _p_ == nil could be cleaned up to find a P
2503 // before calling startm.
2504 throw("startm: P required for spinning=true")
2517 // No M is available, we must drop sched.lock and call newm.
2518 // However, we already own a P to assign to the M.
2520 // Once sched.lock is released, another G (e.g., in a syscall),
2521 // could find no idle P while checkdead finds a runnable G but
2522 // no running M's because this new M hasn't started yet, thus
2523 // throwing in an apparent deadlock.
2524 // This apparent deadlock is possible when startm is called
2525 // from sysmon, which doesn't count as a running M.
2527 // Avoid this situation by pre-allocating the ID for the new M,
2528 // thus marking it as 'running' before we drop sched.lock. This
2529 // new M will eventually run the scheduler to execute any
2536 // The caller incremented nmspinning, so set m.spinning in the new M.
2544 // Ownership transfer of pp committed by start in newm.
2545 // Preemption is now safe.
2553 throw("startm: m is spinning")
2556 throw("startm: m has p")
2558 if spinning && !runqempty(pp) {
2559 throw("startm: p has runnable gs")
2561 // The caller incremented nmspinning, so set m.spinning in the new M.
2562 nmp.spinning = spinning
2564 notewakeup(&nmp.park)
2565 // Ownership transfer of pp committed by wakeup. Preemption is now
2570 // Hands off P from syscall or locked M.
2571 // Always runs without a P, so write barriers are not allowed.
2573 //go:nowritebarrierrec
2574 func handoffp(pp *p) {
2575 // handoffp must start an M in any situation where
2576 // findrunnable would return a G to run on pp.
2578 // if it has local work, start it straight away
2579 if !runqempty(pp) || sched.runqsize != 0 {
2580 startm(pp, false, false)
2583 // if there's trace work to do, start it straight away
2584 if (traceEnabled() || traceShuttingDown()) && traceReaderAvailable() != nil {
2585 startm(pp, false, false)
2588 // if it has GC work, start it straight away
2589 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2590 startm(pp, false, false)
2593 // no local work, check that there are no spinning/idle M's,
2594 // otherwise our help is not required
2595 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2596 sched.needspinning.Store(0)
2597 startm(pp, true, false)
2601 if sched.gcwaiting.Load() {
2602 pp.status = _Pgcstop
2604 if sched.stopwait == 0 {
2605 notewakeup(&sched.stopnote)
2610 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2611 sched.safePointFn(pp)
2612 sched.safePointWait--
2613 if sched.safePointWait == 0 {
2614 notewakeup(&sched.safePointNote)
2617 if sched.runqsize != 0 {
2619 startm(pp, false, false)
2622 // If this is the last running P and nobody is polling network,
2623 // need to wakeup another M to poll network.
2624 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2626 startm(pp, false, false)
2630 // The scheduler lock cannot be held when calling wakeNetPoller below
2631 // because wakeNetPoller may call wakep which may call startm.
2632 when := nobarrierWakeTime(pp)
2641 // Tries to add one more P to execute G's.
2642 // Called when a G is made runnable (newproc, ready).
2643 // Must be called with a P.
2645 // Be conservative about spinning threads, only start one if none exist
2647 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2651 // Disable preemption until ownership of pp transfers to the next M in
2652 // startm. Otherwise preemption here would leave pp stuck waiting to
2655 // See preemption comment on acquirem in startm for more details.
2660 pp, _ = pidlegetSpinning(0)
2662 if sched.nmspinning.Add(-1) < 0 {
2663 throw("wakep: negative nmspinning")
2669 // Since we always have a P, the race in the "No M is available"
2670 // comment in startm doesn't apply during the small window between the
2671 // unlock here and lock in startm. A checkdead in between will always
2672 // see at least one running M (ours).
2675 startm(pp, true, false)
2680 // Stops execution of the current m that is locked to a g until the g is runnable again.
2681 // Returns with acquired P.
2682 func stoplockedm() {
2685 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2686 throw("stoplockedm: inconsistent locking")
2689 // Schedule another M to run this p.
2694 // Wait until another thread schedules lockedg again.
2696 status := readgstatus(gp.m.lockedg.ptr())
2697 if status&^_Gscan != _Grunnable {
2698 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2699 dumpgstatus(gp.m.lockedg.ptr())
2700 throw("stoplockedm: not runnable")
2702 acquirep(gp.m.nextp.ptr())
2706 // Schedules the locked m to run the locked gp.
2707 // May run during STW, so write barriers are not allowed.
2709 //go:nowritebarrierrec
2710 func startlockedm(gp *g) {
2711 mp := gp.lockedm.ptr()
2713 throw("startlockedm: locked to me")
2716 throw("startlockedm: m has p")
2718 // directly handoff current P to the locked m
2722 notewakeup(&mp.park)
2726 // Stops the current m for stopTheWorld.
2727 // Returns when the world is restarted.
2731 if !sched.gcwaiting.Load() {
2732 throw("gcstopm: not waiting for gc")
2735 gp.m.spinning = false
2736 // OK to just drop nmspinning here,
2737 // startTheWorld will unpark threads as necessary.
2738 if sched.nmspinning.Add(-1) < 0 {
2739 throw("gcstopm: negative nmspinning")
2744 pp.status = _Pgcstop
2746 if sched.stopwait == 0 {
2747 notewakeup(&sched.stopnote)
2753 // Schedules gp to run on the current M.
2754 // If inheritTime is true, gp inherits the remaining time in the
2755 // current time slice. Otherwise, it starts a new time slice.
2758 // Write barriers are allowed because this is called immediately after
2759 // acquiring a P in several places.
2761 //go:yeswritebarrierrec
2762 func execute(gp *g, inheritTime bool) {
2765 if goroutineProfile.active {
2766 // Make sure that gp has had its stack written out to the goroutine
2767 // profile, exactly as it was when the goroutine profiler first stopped
2769 tryRecordGoroutineProfile(gp, osyield)
2772 // Assign gp.m before entering _Grunning so running Gs have an
2776 casgstatus(gp, _Grunnable, _Grunning)
2779 gp.stackguard0 = gp.stack.lo + stackGuard
2781 mp.p.ptr().schedtick++
2784 // Check whether the profiler needs to be turned on or off.
2785 hz := sched.profilehz
2786 if mp.profilehz != hz {
2787 setThreadCPUProfiler(hz)
2791 // GoSysExit has to happen when we have a P, but before GoStart.
2792 // So we emit it here.
2793 if gp.syscallsp != 0 {
2802 // Finds a runnable goroutine to execute.
2803 // Tries to steal from other P's, get g from local or global queue, poll network.
2804 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2805 // reader) so the caller should try to wake a P.
2806 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2809 // The conditions here and in handoffp must agree: if
2810 // findrunnable would return a G to run, handoffp must start
2815 if sched.gcwaiting.Load() {
2819 if pp.runSafePointFn != 0 {
2823 // now and pollUntil are saved for work stealing later,
2824 // which may steal timers. It's important that between now
2825 // and then, nothing blocks, so these numbers remain mostly
2827 now, pollUntil, _ := checkTimers(pp, 0)
2829 // Try to schedule the trace reader.
2830 if traceEnabled() || traceShuttingDown() {
2833 casgstatus(gp, _Gwaiting, _Grunnable)
2834 traceGoUnpark(gp, 0)
2835 return gp, false, true
2839 // Try to schedule a GC worker.
2840 if gcBlackenEnabled != 0 {
2841 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2843 return gp, false, true
2848 // Check the global runnable queue once in a while to ensure fairness.
2849 // Otherwise two goroutines can completely occupy the local runqueue
2850 // by constantly respawning each other.
2851 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2853 gp := globrunqget(pp, 1)
2856 return gp, false, false
2860 // Wake up the finalizer G.
2861 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2862 if gp := wakefing(); gp != nil {
2866 if *cgo_yield != nil {
2867 asmcgocall(*cgo_yield, nil)
2871 if gp, inheritTime := runqget(pp); gp != nil {
2872 return gp, inheritTime, false
2876 if sched.runqsize != 0 {
2878 gp := globrunqget(pp, 0)
2881 return gp, false, false
2886 // This netpoll is only an optimization before we resort to stealing.
2887 // We can safely skip it if there are no waiters or a thread is blocked
2888 // in netpoll already. If there is any kind of logical race with that
2889 // blocked thread (e.g. it has already returned from netpoll, but does
2890 // not set lastpoll yet), this thread will do blocking netpoll below
2892 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2893 if list := netpoll(0); !list.empty() { // non-blocking
2896 casgstatus(gp, _Gwaiting, _Grunnable)
2898 traceGoUnpark(gp, 0)
2900 return gp, false, false
2904 // Spinning Ms: steal work from other Ps.
2906 // Limit the number of spinning Ms to half the number of busy Ps.
2907 // This is necessary to prevent excessive CPU consumption when
2908 // GOMAXPROCS>>1 but the program parallelism is low.
2909 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2914 gp, inheritTime, tnow, w, newWork := stealWork(now)
2916 // Successfully stole.
2917 return gp, inheritTime, false
2920 // There may be new timer or GC work; restart to
2926 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2927 // Earlier timer to wait for.
2932 // We have nothing to do.
2934 // If we're in the GC mark phase, can safely scan and blacken objects,
2935 // and have work to do, run idle-time marking rather than give up the P.
2936 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2937 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2939 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2941 casgstatus(gp, _Gwaiting, _Grunnable)
2943 traceGoUnpark(gp, 0)
2945 return gp, false, false
2947 gcController.removeIdleMarkWorker()
2951 // If a callback returned and no other goroutine is awake,
2952 // then wake event handler goroutine which pauses execution
2953 // until a callback was triggered.
2954 gp, otherReady := beforeIdle(now, pollUntil)
2956 casgstatus(gp, _Gwaiting, _Grunnable)
2958 traceGoUnpark(gp, 0)
2960 return gp, false, false
2966 // Before we drop our P, make a snapshot of the allp slice,
2967 // which can change underfoot once we no longer block
2968 // safe-points. We don't need to snapshot the contents because
2969 // everything up to cap(allp) is immutable.
2970 allpSnapshot := allp
2971 // Also snapshot masks. Value changes are OK, but we can't allow
2972 // len to change out from under us.
2973 idlepMaskSnapshot := idlepMask
2974 timerpMaskSnapshot := timerpMask
2976 // return P and block
2978 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2982 if sched.runqsize != 0 {
2983 gp := globrunqget(pp, 0)
2985 return gp, false, false
2987 if !mp.spinning && sched.needspinning.Load() == 1 {
2988 // See "Delicate dance" comment below.
2993 if releasep() != pp {
2994 throw("findrunnable: wrong p")
2996 now = pidleput(pp, now)
2999 // Delicate dance: thread transitions from spinning to non-spinning
3000 // state, potentially concurrently with submission of new work. We must
3001 // drop nmspinning first and then check all sources again (with
3002 // #StoreLoad memory barrier in between). If we do it the other way
3003 // around, another thread can submit work after we've checked all
3004 // sources but before we drop nmspinning; as a result nobody will
3005 // unpark a thread to run the work.
3007 // This applies to the following sources of work:
3009 // * Goroutines added to a per-P run queue.
3010 // * New/modified-earlier timers on a per-P timer heap.
3011 // * Idle-priority GC work (barring golang.org/issue/19112).
3013 // If we discover new work below, we need to restore m.spinning as a
3014 // signal for resetspinning to unpark a new worker thread (because
3015 // there can be more than one starving goroutine).
3017 // However, if after discovering new work we also observe no idle Ps
3018 // (either here or in resetspinning), we have a problem. We may be
3019 // racing with a non-spinning M in the block above, having found no
3020 // work and preparing to release its P and park. Allowing that P to go
3021 // idle will result in loss of work conservation (idle P while there is
3022 // runnable work). This could result in complete deadlock in the
3023 // unlikely event that we discover new work (from netpoll) right as we
3024 // are racing with _all_ other Ps going idle.
3026 // We use sched.needspinning to synchronize with non-spinning Ms going
3027 // idle. If needspinning is set when they are about to drop their P,
3028 // they abort the drop and instead become a new spinning M on our
3029 // behalf. If we are not racing and the system is truly fully loaded
3030 // then no spinning threads are required, and the next thread to
3031 // naturally become spinning will clear the flag.
3033 // Also see "Worker thread parking/unparking" comment at the top of the
3035 wasSpinning := mp.spinning
3038 if sched.nmspinning.Add(-1) < 0 {
3039 throw("findrunnable: negative nmspinning")
3042 // Note the for correctness, only the last M transitioning from
3043 // spinning to non-spinning must perform these rechecks to
3044 // ensure no missed work. However, the runtime has some cases
3045 // of transient increments of nmspinning that are decremented
3046 // without going through this path, so we must be conservative
3047 // and perform the check on all spinning Ms.
3049 // See https://go.dev/issue/43997.
3051 // Check all runqueues once again.
3052 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
3059 // Check for idle-priority GC work again.
3060 pp, gp := checkIdleGCNoP()
3065 // Run the idle worker.
3066 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
3067 casgstatus(gp, _Gwaiting, _Grunnable)
3069 traceGoUnpark(gp, 0)
3071 return gp, false, false
3074 // Finally, check for timer creation or expiry concurrently with
3075 // transitioning from spinning to non-spinning.
3077 // Note that we cannot use checkTimers here because it calls
3078 // adjusttimers which may need to allocate memory, and that isn't
3079 // allowed when we don't have an active P.
3080 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
3083 // Poll network until next timer.
3084 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
3085 sched.pollUntil.Store(pollUntil)
3087 throw("findrunnable: netpoll with p")
3090 throw("findrunnable: netpoll with spinning")
3096 delay = pollUntil - now
3102 // When using fake time, just poll.
3105 list := netpoll(delay) // block until new work is available
3106 sched.pollUntil.Store(0)
3107 sched.lastpoll.Store(now)
3108 if faketime != 0 && list.empty() {
3109 // Using fake time and nothing is ready; stop M.
3110 // When all M's stop, checkdead will call timejump.
3115 pp, _ := pidleget(now)
3124 casgstatus(gp, _Gwaiting, _Grunnable)
3126 traceGoUnpark(gp, 0)
3128 return gp, false, false
3135 } else if pollUntil != 0 && netpollinited() {
3136 pollerPollUntil := sched.pollUntil.Load()
3137 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3145 // pollWork reports whether there is non-background work this P could
3146 // be doing. This is a fairly lightweight check to be used for
3147 // background work loops, like idle GC. It checks a subset of the
3148 // conditions checked by the actual scheduler.
3149 func pollWork() bool {
3150 if sched.runqsize != 0 {
3153 p := getg().m.p.ptr()
3157 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3158 if list := netpoll(0); !list.empty() {
3166 // stealWork attempts to steal a runnable goroutine or timer from any P.
3168 // If newWork is true, new work may have been readied.
3170 // If now is not 0 it is the current time. stealWork returns the passed time or
3171 // the current time if now was passed as 0.
3172 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3173 pp := getg().m.p.ptr()
3177 const stealTries = 4
3178 for i := 0; i < stealTries; i++ {
3179 stealTimersOrRunNextG := i == stealTries-1
3181 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3182 if sched.gcwaiting.Load() {
3183 // GC work may be available.
3184 return nil, false, now, pollUntil, true
3186 p2 := allp[enum.position()]
3191 // Steal timers from p2. This call to checkTimers is the only place
3192 // where we might hold a lock on a different P's timers. We do this
3193 // once on the last pass before checking runnext because stealing
3194 // from the other P's runnext should be the last resort, so if there
3195 // are timers to steal do that first.
3197 // We only check timers on one of the stealing iterations because
3198 // the time stored in now doesn't change in this loop and checking
3199 // the timers for each P more than once with the same value of now
3200 // is probably a waste of time.
3202 // timerpMask tells us whether the P may have timers at all. If it
3203 // can't, no need to check at all.
3204 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3205 tnow, w, ran := checkTimers(p2, now)
3207 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3211 // Running the timers may have
3212 // made an arbitrary number of G's
3213 // ready and added them to this P's
3214 // local run queue. That invalidates
3215 // the assumption of runqsteal
3216 // that it always has room to add
3217 // stolen G's. So check now if there
3218 // is a local G to run.
3219 if gp, inheritTime := runqget(pp); gp != nil {
3220 return gp, inheritTime, now, pollUntil, ranTimer
3226 // Don't bother to attempt to steal if p2 is idle.
3227 if !idlepMask.read(enum.position()) {
3228 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3229 return gp, false, now, pollUntil, ranTimer
3235 // No goroutines found to steal. Regardless, running a timer may have
3236 // made some goroutine ready that we missed. Indicate the next timer to
3238 return nil, false, now, pollUntil, ranTimer
3241 // Check all Ps for a runnable G to steal.
3243 // On entry we have no P. If a G is available to steal and a P is available,
3244 // the P is returned which the caller should acquire and attempt to steal the
3246 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3247 for id, p2 := range allpSnapshot {
3248 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3250 pp, _ := pidlegetSpinning(0)
3252 // Can't get a P, don't bother checking remaining Ps.
3261 // No work available.
3265 // Check all Ps for a timer expiring sooner than pollUntil.
3267 // Returns updated pollUntil value.
3268 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3269 for id, p2 := range allpSnapshot {
3270 if timerpMaskSnapshot.read(uint32(id)) {
3271 w := nobarrierWakeTime(p2)
3272 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3281 // Check for idle-priority GC, without a P on entry.
3283 // If some GC work, a P, and a worker G are all available, the P and G will be
3284 // returned. The returned P has not been wired yet.
3285 func checkIdleGCNoP() (*p, *g) {
3286 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3287 // must check again after acquiring a P. As an optimization, we also check
3288 // if an idle mark worker is needed at all. This is OK here, because if we
3289 // observe that one isn't needed, at least one is currently running. Even if
3290 // it stops running, its own journey into the scheduler should schedule it
3291 // again, if need be (at which point, this check will pass, if relevant).
3292 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3295 if !gcMarkWorkAvailable(nil) {
3299 // Work is available; we can start an idle GC worker only if there is
3300 // an available P and available worker G.
3302 // We can attempt to acquire these in either order, though both have
3303 // synchronization concerns (see below). Workers are almost always
3304 // available (see comment in findRunnableGCWorker for the one case
3305 // there may be none). Since we're slightly less likely to find a P,
3306 // check for that first.
3308 // Synchronization: note that we must hold sched.lock until we are
3309 // committed to keeping it. Otherwise we cannot put the unnecessary P
3310 // back in sched.pidle without performing the full set of idle
3311 // transition checks.
3313 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3314 // the assumption in gcControllerState.findRunnableGCWorker that an
3315 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3317 pp, now := pidlegetSpinning(0)
3323 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3324 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3330 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3334 gcController.removeIdleMarkWorker()
3340 return pp, node.gp.ptr()
3343 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3344 // going to wake up before the when argument; or it wakes an idle P to service
3345 // timers and the network poller if there isn't one already.
3346 func wakeNetPoller(when int64) {
3347 if sched.lastpoll.Load() == 0 {
3348 // In findrunnable we ensure that when polling the pollUntil
3349 // field is either zero or the time to which the current
3350 // poll is expected to run. This can have a spurious wakeup
3351 // but should never miss a wakeup.
3352 pollerPollUntil := sched.pollUntil.Load()
3353 if pollerPollUntil == 0 || pollerPollUntil > when {
3357 // There are no threads in the network poller, try to get
3358 // one there so it can handle new timers.
3359 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3365 func resetspinning() {
3368 throw("resetspinning: not a spinning m")
3370 gp.m.spinning = false
3371 nmspinning := sched.nmspinning.Add(-1)
3373 throw("findrunnable: negative nmspinning")
3375 // M wakeup policy is deliberately somewhat conservative, so check if we
3376 // need to wakeup another P here. See "Worker thread parking/unparking"
3377 // comment at the top of the file for details.
3381 // injectglist adds each runnable G on the list to some run queue,
3382 // and clears glist. If there is no current P, they are added to the
3383 // global queue, and up to npidle M's are started to run them.
3384 // Otherwise, for each idle P, this adds a G to the global queue
3385 // and starts an M. Any remaining G's are added to the current P's
3387 // This may temporarily acquire sched.lock.
3388 // Can run concurrently with GC.
3389 func injectglist(glist *gList) {
3394 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3395 traceGoUnpark(gp, 0)
3399 // Mark all the goroutines as runnable before we put them
3400 // on the run queues.
3401 head := glist.head.ptr()
3404 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3407 casgstatus(gp, _Gwaiting, _Grunnable)
3410 // Turn the gList into a gQueue.
3416 startIdle := func(n int) {
3417 for i := 0; i < n; i++ {
3418 mp := acquirem() // See comment in startm.
3421 pp, _ := pidlegetSpinning(0)
3428 startm(pp, false, true)
3434 pp := getg().m.p.ptr()
3437 globrunqputbatch(&q, int32(qsize))
3443 npidle := int(sched.npidle.Load())
3446 for n = 0; n < npidle && !q.empty(); n++ {
3452 globrunqputbatch(&globq, int32(n))
3459 runqputbatch(pp, &q, qsize)
3463 // One round of scheduler: find a runnable goroutine and execute it.
3469 throw("schedule: holding locks")
3472 if mp.lockedg != 0 {
3474 execute(mp.lockedg.ptr(), false) // Never returns.
3477 // We should not schedule away from a g that is executing a cgo call,
3478 // since the cgo call is using the m's g0 stack.
3480 throw("schedule: in cgo")
3487 // Safety check: if we are spinning, the run queue should be empty.
3488 // Check this before calling checkTimers, as that might call
3489 // goready to put a ready goroutine on the local run queue.
3490 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3491 throw("schedule: spinning with local work")
3494 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3496 // This thread is going to run a goroutine and is not spinning anymore,
3497 // so if it was marked as spinning we need to reset it now and potentially
3498 // start a new spinning M.
3503 if sched.disable.user && !schedEnabled(gp) {
3504 // Scheduling of this goroutine is disabled. Put it on
3505 // the list of pending runnable goroutines for when we
3506 // re-enable user scheduling and look again.
3508 if schedEnabled(gp) {
3509 // Something re-enabled scheduling while we
3510 // were acquiring the lock.
3513 sched.disable.runnable.pushBack(gp)
3520 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3521 // wake a P if there is one.
3525 if gp.lockedm != 0 {
3526 // Hands off own p to the locked m,
3527 // then blocks waiting for a new p.
3532 execute(gp, inheritTime)
3535 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3536 // Typically a caller sets gp's status away from Grunning and then
3537 // immediately calls dropg to finish the job. The caller is also responsible
3538 // for arranging that gp will be restarted using ready at an
3539 // appropriate time. After calling dropg and arranging for gp to be
3540 // readied later, the caller can do other work but eventually should
3541 // call schedule to restart the scheduling of goroutines on this m.
3545 setMNoWB(&gp.m.curg.m, nil)
3546 setGNoWB(&gp.m.curg, nil)
3549 // checkTimers runs any timers for the P that are ready.
3550 // If now is not 0 it is the current time.
3551 // It returns the passed time or the current time if now was passed as 0.
3552 // and the time when the next timer should run or 0 if there is no next timer,
3553 // and reports whether it ran any timers.
3554 // If the time when the next timer should run is not 0,
3555 // it is always larger than the returned time.
3556 // We pass now in and out to avoid extra calls of nanotime.
3558 //go:yeswritebarrierrec
3559 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3560 // If it's not yet time for the first timer, or the first adjusted
3561 // timer, then there is nothing to do.
3562 next := pp.timer0When.Load()
3563 nextAdj := pp.timerModifiedEarliest.Load()
3564 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3569 // No timers to run or adjust.
3570 return now, 0, false
3577 // Next timer is not ready to run, but keep going
3578 // if we would clear deleted timers.
3579 // This corresponds to the condition below where
3580 // we decide whether to call clearDeletedTimers.
3581 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3582 return now, next, false
3586 lock(&pp.timersLock)
3588 if len(pp.timers) > 0 {
3589 adjusttimers(pp, now)
3590 for len(pp.timers) > 0 {
3591 // Note that runtimer may temporarily unlock
3593 if tw := runtimer(pp, now); tw != 0 {
3603 // If this is the local P, and there are a lot of deleted timers,
3604 // clear them out. We only do this for the local P to reduce
3605 // lock contention on timersLock.
3606 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3607 clearDeletedTimers(pp)
3610 unlock(&pp.timersLock)
3612 return now, pollUntil, ran
3615 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3616 unlock((*mutex)(lock))
3620 // park continuation on g0.
3621 func park_m(gp *g) {
3625 traceGoPark(mp.waittraceev, mp.waittraceskip)
3628 // N.B. Not using casGToWaiting here because the waitreason is
3629 // set by park_m's caller.
3630 casgstatus(gp, _Grunning, _Gwaiting)
3633 if fn := mp.waitunlockf; fn != nil {
3634 ok := fn(gp, mp.waitlock)
3635 mp.waitunlockf = nil
3639 traceGoUnpark(gp, 2)
3641 casgstatus(gp, _Gwaiting, _Grunnable)
3642 execute(gp, true) // Schedule it back, never returns.
3648 func goschedImpl(gp *g) {
3649 status := readgstatus(gp)
3650 if status&^_Gscan != _Grunning {
3652 throw("bad g status")
3654 casgstatus(gp, _Grunning, _Grunnable)
3663 // Gosched continuation on g0.
3664 func gosched_m(gp *g) {
3671 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3672 func goschedguarded_m(gp *g) {
3674 if !canPreemptM(gp.m) {
3675 gogo(&gp.sched) // never return
3684 func gopreempt_m(gp *g) {
3691 // preemptPark parks gp and puts it in _Gpreempted.
3694 func preemptPark(gp *g) {
3696 traceGoPark(traceEvGoBlock, 0)
3698 status := readgstatus(gp)
3699 if status&^_Gscan != _Grunning {
3701 throw("bad g status")
3704 if gp.asyncSafePoint {
3705 // Double-check that async preemption does not
3706 // happen in SPWRITE assembly functions.
3707 // isAsyncSafePoint must exclude this case.
3708 f := findfunc(gp.sched.pc)
3710 throw("preempt at unknown pc")
3712 if f.flag&abi.FuncFlagSPWrite != 0 {
3713 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3714 throw("preempt SPWRITE")
3718 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3719 // be in _Grunning when we dropg because then we'd be running
3720 // without an M, but the moment we're in _Gpreempted,
3721 // something could claim this G before we've fully cleaned it
3722 // up. Hence, we set the scan bit to lock down further
3723 // transitions until we can dropg.
3724 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3726 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3730 // goyield is like Gosched, but it:
3731 // - emits a GoPreempt trace event instead of a GoSched trace event
3732 // - puts the current G on the runq of the current P instead of the globrunq
3738 func goyield_m(gp *g) {
3743 casgstatus(gp, _Grunning, _Grunnable)
3745 runqput(pp, gp, false)
3749 // Finishes execution of the current goroutine.
3760 // goexit continuation on g0.
3761 func goexit0(gp *g) {
3765 casgstatus(gp, _Grunning, _Gdead)
3766 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3767 if isSystemGoroutine(gp, false) {
3771 locked := gp.lockedm != 0
3774 gp.preemptStop = false
3775 gp.paniconfault = false
3776 gp._defer = nil // should be true already but just in case.
3777 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3779 gp.waitreason = waitReasonZero
3784 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3785 // Flush assist credit to the global pool. This gives
3786 // better information to pacing if the application is
3787 // rapidly creating an exiting goroutines.
3788 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3789 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3790 gcController.bgScanCredit.Add(scanCredit)
3791 gp.gcAssistBytes = 0
3796 if GOARCH == "wasm" { // no threads yet on wasm
3798 schedule() // never returns
3801 if mp.lockedInt != 0 {
3802 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3803 throw("internal lockOSThread error")
3807 // The goroutine may have locked this thread because
3808 // it put it in an unusual kernel state. Kill it
3809 // rather than returning it to the thread pool.
3811 // Return to mstart, which will release the P and exit
3813 if GOOS != "plan9" { // See golang.org/issue/22227.
3816 // Clear lockedExt on plan9 since we may end up re-using
3824 // save updates getg().sched to refer to pc and sp so that a following
3825 // gogo will restore pc and sp.
3827 // save must not have write barriers because invoking a write barrier
3828 // can clobber getg().sched.
3831 //go:nowritebarrierrec
3832 func save(pc, sp uintptr) {
3835 if gp == gp.m.g0 || gp == gp.m.gsignal {
3836 // m.g0.sched is special and must describe the context
3837 // for exiting the thread. mstart1 writes to it directly.
3838 // m.gsignal.sched should not be used at all.
3839 // This check makes sure save calls do not accidentally
3840 // run in contexts where they'd write to system g's.
3841 throw("save on system g not allowed")
3848 // We need to ensure ctxt is zero, but can't have a write
3849 // barrier here. However, it should always already be zero.
3851 if gp.sched.ctxt != nil {
3856 // The goroutine g is about to enter a system call.
3857 // Record that it's not using the cpu anymore.
3858 // This is called only from the go syscall library and cgocall,
3859 // not from the low-level system calls used by the runtime.
3861 // Entersyscall cannot split the stack: the save must
3862 // make g->sched refer to the caller's stack segment, because
3863 // entersyscall is going to return immediately after.
3865 // Nothing entersyscall calls can split the stack either.
3866 // We cannot safely move the stack during an active call to syscall,
3867 // because we do not know which of the uintptr arguments are
3868 // really pointers (back into the stack).
3869 // In practice, this means that we make the fast path run through
3870 // entersyscall doing no-split things, and the slow path has to use systemstack
3871 // to run bigger things on the system stack.
3873 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3874 // saved SP and PC are restored. This is needed when exitsyscall will be called
3875 // from a function further up in the call stack than the parent, as g->syscallsp
3876 // must always point to a valid stack frame. entersyscall below is the normal
3877 // entry point for syscalls, which obtains the SP and PC from the caller.
3880 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3881 // If the syscall does not block, that is it, we do not emit any other events.
3882 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3883 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3884 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3885 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3886 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3887 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3888 // and we wait for the increment before emitting traceGoSysExit.
3889 // Note that the increment is done even if tracing is not enabled,
3890 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3893 func reentersyscall(pc, sp uintptr) {
3896 // Disable preemption because during this function g is in Gsyscall status,
3897 // but can have inconsistent g->sched, do not let GC observe it.
3900 // Entersyscall must not call any function that might split/grow the stack.
3901 // (See details in comment above.)
3902 // Catch calls that might, by replacing the stack guard with something that
3903 // will trip any stack check and leaving a flag to tell newstack to die.
3904 gp.stackguard0 = stackPreempt
3905 gp.throwsplit = true
3907 // Leave SP around for GC and traceback.
3911 casgstatus(gp, _Grunning, _Gsyscall)
3912 if staticLockRanking {
3913 // When doing static lock ranking casgstatus can call
3914 // systemstack which clobbers g.sched.
3917 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3918 systemstack(func() {
3919 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3920 throw("entersyscall")
3925 systemstack(traceGoSysCall)
3926 // systemstack itself clobbers g.sched.{pc,sp} and we might
3927 // need them later when the G is genuinely blocked in a
3932 if sched.sysmonwait.Load() {
3933 systemstack(entersyscall_sysmon)
3937 if gp.m.p.ptr().runSafePointFn != 0 {
3938 // runSafePointFn may stack split if run on this stack
3939 systemstack(runSafePointFn)
3943 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3948 atomic.Store(&pp.status, _Psyscall)
3949 if sched.gcwaiting.Load() {
3950 systemstack(entersyscall_gcwait)
3957 // Standard syscall entry used by the go syscall library and normal cgo calls.
3959 // This is exported via linkname to assembly in the syscall package and x/sys.
3962 //go:linkname entersyscall
3963 func entersyscall() {
3964 reentersyscall(getcallerpc(), getcallersp())
3967 func entersyscall_sysmon() {
3969 if sched.sysmonwait.Load() {
3970 sched.sysmonwait.Store(false)
3971 notewakeup(&sched.sysmonnote)
3976 func entersyscall_gcwait() {
3978 pp := gp.m.oldp.ptr()
3981 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3987 if sched.stopwait--; sched.stopwait == 0 {
3988 notewakeup(&sched.stopnote)
3994 // The same as entersyscall(), but with a hint that the syscall is blocking.
3997 func entersyscallblock() {
4000 gp.m.locks++ // see comment in entersyscall
4001 gp.throwsplit = true
4002 gp.stackguard0 = stackPreempt // see comment in entersyscall
4003 gp.m.syscalltick = gp.m.p.ptr().syscalltick
4004 gp.m.p.ptr().syscalltick++
4006 // Leave SP around for GC and traceback.
4010 gp.syscallsp = gp.sched.sp
4011 gp.syscallpc = gp.sched.pc
4012 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
4016 systemstack(func() {
4017 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
4018 throw("entersyscallblock")
4021 casgstatus(gp, _Grunning, _Gsyscall)
4022 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
4023 systemstack(func() {
4024 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
4025 throw("entersyscallblock")
4029 systemstack(entersyscallblock_handoff)
4031 // Resave for traceback during blocked call.
4032 save(getcallerpc(), getcallersp())
4037 func entersyscallblock_handoff() {
4040 traceGoSysBlock(getg().m.p.ptr())
4042 handoffp(releasep())
4045 // The goroutine g exited its system call.
4046 // Arrange for it to run on a cpu again.
4047 // This is called only from the go syscall library, not
4048 // from the low-level system calls used by the runtime.
4050 // Write barriers are not allowed because our P may have been stolen.
4052 // This is exported via linkname to assembly in the syscall package.
4055 //go:nowritebarrierrec
4056 //go:linkname exitsyscall
4057 func exitsyscall() {
4060 gp.m.locks++ // see comment in entersyscall
4061 if getcallersp() > gp.syscallsp {
4062 throw("exitsyscall: syscall frame is no longer valid")
4066 oldp := gp.m.oldp.ptr()
4068 if exitsyscallfast(oldp) {
4069 // When exitsyscallfast returns success, we have a P so can now use
4071 if goroutineProfile.active {
4072 // Make sure that gp has had its stack written out to the goroutine
4073 // profile, exactly as it was when the goroutine profiler first
4074 // stopped the world.
4075 systemstack(func() {
4076 tryRecordGoroutineProfileWB(gp)
4080 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4081 systemstack(traceGoStart)
4084 // There's a cpu for us, so we can run.
4085 gp.m.p.ptr().syscalltick++
4086 // We need to cas the status and scan before resuming...
4087 casgstatus(gp, _Gsyscall, _Grunning)
4089 // Garbage collector isn't running (since we are),
4090 // so okay to clear syscallsp.
4094 // restore the preemption request in case we've cleared it in newstack
4095 gp.stackguard0 = stackPreempt
4097 // otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
4098 gp.stackguard0 = gp.stack.lo + stackGuard
4100 gp.throwsplit = false
4102 if sched.disable.user && !schedEnabled(gp) {
4103 // Scheduling of this goroutine is disabled.
4111 // Wait till traceGoSysBlock event is emitted.
4112 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4113 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
4116 // We can't trace syscall exit right now because we don't have a P.
4117 // Tracing code can invoke write barriers that cannot run without a P.
4118 // So instead we remember the syscall exit time and emit the event
4119 // in execute when we have a P.
4120 gp.trace.sysExitTicks = cputicks()
4125 // Call the scheduler.
4128 // Scheduler returned, so we're allowed to run now.
4129 // Delete the syscallsp information that we left for
4130 // the garbage collector during the system call.
4131 // Must wait until now because until gosched returns
4132 // we don't know for sure that the garbage collector
4135 gp.m.p.ptr().syscalltick++
4136 gp.throwsplit = false
4140 func exitsyscallfast(oldp *p) bool {
4143 // Freezetheworld sets stopwait but does not retake P's.
4144 if sched.stopwait == freezeStopWait {
4148 // Try to re-acquire the last P.
4149 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4150 // There's a cpu for us, so we can run.
4152 exitsyscallfast_reacquired()
4156 // Try to get any other idle P.
4157 if sched.pidle != 0 {
4159 systemstack(func() {
4160 ok = exitsyscallfast_pidle()
4161 if ok && traceEnabled() {
4163 // Wait till traceGoSysBlock event is emitted.
4164 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4165 for oldp.syscalltick == gp.m.syscalltick {
4179 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4180 // has successfully reacquired the P it was running on before the
4184 func exitsyscallfast_reacquired() {
4186 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4188 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4189 // traceGoSysBlock for this syscall was already emitted,
4190 // but here we effectively retake the p from the new syscall running on the same p.
4191 systemstack(func() {
4192 // Denote blocking of the new syscall.
4193 traceGoSysBlock(gp.m.p.ptr())
4194 // Denote completion of the current syscall.
4198 gp.m.p.ptr().syscalltick++
4202 func exitsyscallfast_pidle() bool {
4204 pp, _ := pidleget(0)
4205 if pp != nil && sched.sysmonwait.Load() {
4206 sched.sysmonwait.Store(false)
4207 notewakeup(&sched.sysmonnote)
4217 // exitsyscall slow path on g0.
4218 // Failed to acquire P, enqueue gp as runnable.
4220 // Called via mcall, so gp is the calling g from this M.
4222 //go:nowritebarrierrec
4223 func exitsyscall0(gp *g) {
4224 casgstatus(gp, _Gsyscall, _Grunnable)
4228 if schedEnabled(gp) {
4235 // Below, we stoplockedm if gp is locked. globrunqput releases
4236 // ownership of gp, so we must check if gp is locked prior to
4237 // committing the release by unlocking sched.lock, otherwise we
4238 // could race with another M transitioning gp from unlocked to
4240 locked = gp.lockedm != 0
4241 } else if sched.sysmonwait.Load() {
4242 sched.sysmonwait.Store(false)
4243 notewakeup(&sched.sysmonnote)
4248 execute(gp, false) // Never returns.
4251 // Wait until another thread schedules gp and so m again.
4253 // N.B. lockedm must be this M, as this g was running on this M
4254 // before entersyscall.
4256 execute(gp, false) // Never returns.
4259 schedule() // Never returns.
4262 // Called from syscall package before fork.
4264 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4266 func syscall_runtime_BeforeFork() {
4269 // Block signals during a fork, so that the child does not run
4270 // a signal handler before exec if a signal is sent to the process
4271 // group. See issue #18600.
4273 sigsave(&gp.m.sigmask)
4276 // This function is called before fork in syscall package.
4277 // Code between fork and exec must not allocate memory nor even try to grow stack.
4278 // Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
4279 // runtime_AfterFork will undo this in parent process, but not in child.
4280 gp.stackguard0 = stackFork
4283 // Called from syscall package after fork in parent.
4285 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4287 func syscall_runtime_AfterFork() {
4290 // See the comments in beforefork.
4291 gp.stackguard0 = gp.stack.lo + stackGuard
4293 msigrestore(gp.m.sigmask)
4298 // inForkedChild is true while manipulating signals in the child process.
4299 // This is used to avoid calling libc functions in case we are using vfork.
4300 var inForkedChild bool
4302 // Called from syscall package after fork in child.
4303 // It resets non-sigignored signals to the default handler, and
4304 // restores the signal mask in preparation for the exec.
4306 // Because this might be called during a vfork, and therefore may be
4307 // temporarily sharing address space with the parent process, this must
4308 // not change any global variables or calling into C code that may do so.
4310 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4312 //go:nowritebarrierrec
4313 func syscall_runtime_AfterForkInChild() {
4314 // It's OK to change the global variable inForkedChild here
4315 // because we are going to change it back. There is no race here,
4316 // because if we are sharing address space with the parent process,
4317 // then the parent process can not be running concurrently.
4318 inForkedChild = true
4320 clearSignalHandlers()
4322 // When we are the child we are the only thread running,
4323 // so we know that nothing else has changed gp.m.sigmask.
4324 msigrestore(getg().m.sigmask)
4326 inForkedChild = false
4329 // pendingPreemptSignals is the number of preemption signals
4330 // that have been sent but not received. This is only used on Darwin.
4332 var pendingPreemptSignals atomic.Int32
4334 // Called from syscall package before Exec.
4336 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4337 func syscall_runtime_BeforeExec() {
4338 // Prevent thread creation during exec.
4341 // On Darwin, wait for all pending preemption signals to
4342 // be received. See issue #41702.
4343 if GOOS == "darwin" || GOOS == "ios" {
4344 for pendingPreemptSignals.Load() > 0 {
4350 // Called from syscall package after Exec.
4352 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4353 func syscall_runtime_AfterExec() {
4357 // Allocate a new g, with a stack big enough for stacksize bytes.
4358 func malg(stacksize int32) *g {
4361 stacksize = round2(stackSystem + stacksize)
4362 systemstack(func() {
4363 newg.stack = stackalloc(uint32(stacksize))
4365 newg.stackguard0 = newg.stack.lo + stackGuard
4366 newg.stackguard1 = ^uintptr(0)
4367 // Clear the bottom word of the stack. We record g
4368 // there on gsignal stack during VDSO on ARM and ARM64.
4369 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4374 // Create a new g running fn.
4375 // Put it on the queue of g's waiting to run.
4376 // The compiler turns a go statement into a call to this.
4377 func newproc(fn *funcval) {
4380 systemstack(func() {
4381 newg := newproc1(fn, gp, pc)
4383 pp := getg().m.p.ptr()
4384 runqput(pp, newg, true)
4392 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4393 // address of the go statement that created this. The caller is responsible
4394 // for adding the new g to the scheduler.
4395 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4397 fatal("go of nil func value")
4400 mp := acquirem() // disable preemption because we hold M and P in local vars.
4404 newg = malg(stackMin)
4405 casgstatus(newg, _Gidle, _Gdead)
4406 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4408 if newg.stack.hi == 0 {
4409 throw("newproc1: newg missing stack")
4412 if readgstatus(newg) != _Gdead {
4413 throw("newproc1: new g is not Gdead")
4416 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4417 totalSize = alignUp(totalSize, sys.StackAlign)
4418 sp := newg.stack.hi - totalSize
4422 *(*uintptr)(unsafe.Pointer(sp)) = 0
4424 spArg += sys.MinFrameSize
4427 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4430 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4431 newg.sched.g = guintptr(unsafe.Pointer(newg))
4432 gostartcallfn(&newg.sched, fn)
4433 newg.parentGoid = callergp.goid
4434 newg.gopc = callerpc
4435 newg.ancestors = saveAncestors(callergp)
4436 newg.startpc = fn.fn
4437 if isSystemGoroutine(newg, false) {
4440 // Only user goroutines inherit pprof labels.
4442 newg.labels = mp.curg.labels
4444 if goroutineProfile.active {
4445 // A concurrent goroutine profile is running. It should include
4446 // exactly the set of goroutines that were alive when the goroutine
4447 // profiler first stopped the world. That does not include newg, so
4448 // mark it as not needing a profile before transitioning it from
4450 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4453 // Track initial transition?
4454 newg.trackingSeq = uint8(fastrand())
4455 if newg.trackingSeq%gTrackingPeriod == 0 {
4456 newg.tracking = true
4458 casgstatus(newg, _Gdead, _Grunnable)
4459 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4461 if pp.goidcache == pp.goidcacheend {
4462 // Sched.goidgen is the last allocated id,
4463 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4464 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4465 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4466 pp.goidcache -= _GoidCacheBatch - 1
4467 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4469 newg.goid = pp.goidcache
4472 newg.racectx = racegostart(callerpc)
4473 if newg.labels != nil {
4474 // See note in proflabel.go on labelSync's role in synchronizing
4475 // with the reads in the signal handler.
4476 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4480 traceGoCreate(newg, newg.startpc)
4487 // saveAncestors copies previous ancestors of the given caller g and
4488 // includes info for the current caller into a new set of tracebacks for
4489 // a g being created.
4490 func saveAncestors(callergp *g) *[]ancestorInfo {
4491 // Copy all prior info, except for the root goroutine (goid 0).
4492 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4495 var callerAncestors []ancestorInfo
4496 if callergp.ancestors != nil {
4497 callerAncestors = *callergp.ancestors
4499 n := int32(len(callerAncestors)) + 1
4500 if n > debug.tracebackancestors {
4501 n = debug.tracebackancestors
4503 ancestors := make([]ancestorInfo, n)
4504 copy(ancestors[1:], callerAncestors)
4506 var pcs [tracebackInnerFrames]uintptr
4507 npcs := gcallers(callergp, 0, pcs[:])
4508 ipcs := make([]uintptr, npcs)
4510 ancestors[0] = ancestorInfo{
4512 goid: callergp.goid,
4513 gopc: callergp.gopc,
4516 ancestorsp := new([]ancestorInfo)
4517 *ancestorsp = ancestors
4521 // Put on gfree list.
4522 // If local list is too long, transfer a batch to the global list.
4523 func gfput(pp *p, gp *g) {
4524 if readgstatus(gp) != _Gdead {
4525 throw("gfput: bad status (not Gdead)")
4528 stksize := gp.stack.hi - gp.stack.lo
4530 if stksize != uintptr(startingStackSize) {
4531 // non-standard stack size - free it.
4540 if pp.gFree.n >= 64 {
4546 for pp.gFree.n >= 32 {
4547 gp := pp.gFree.pop()
4549 if gp.stack.lo == 0 {
4556 lock(&sched.gFree.lock)
4557 sched.gFree.noStack.pushAll(noStackQ)
4558 sched.gFree.stack.pushAll(stackQ)
4559 sched.gFree.n += inc
4560 unlock(&sched.gFree.lock)
4564 // Get from gfree list.
4565 // If local list is empty, grab a batch from global list.
4566 func gfget(pp *p) *g {
4568 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4569 lock(&sched.gFree.lock)
4570 // Move a batch of free Gs to the P.
4571 for pp.gFree.n < 32 {
4572 // Prefer Gs with stacks.
4573 gp := sched.gFree.stack.pop()
4575 gp = sched.gFree.noStack.pop()
4584 unlock(&sched.gFree.lock)
4587 gp := pp.gFree.pop()
4592 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4593 // Deallocate old stack. We kept it in gfput because it was the
4594 // right size when the goroutine was put on the free list, but
4595 // the right size has changed since then.
4596 systemstack(func() {
4603 if gp.stack.lo == 0 {
4604 // Stack was deallocated in gfput or just above. Allocate a new one.
4605 systemstack(func() {
4606 gp.stack = stackalloc(startingStackSize)
4608 gp.stackguard0 = gp.stack.lo + stackGuard
4611 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4614 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4617 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4623 // Purge all cached G's from gfree list to the global list.
4624 func gfpurge(pp *p) {
4630 for !pp.gFree.empty() {
4631 gp := pp.gFree.pop()
4633 if gp.stack.lo == 0 {
4640 lock(&sched.gFree.lock)
4641 sched.gFree.noStack.pushAll(noStackQ)
4642 sched.gFree.stack.pushAll(stackQ)
4643 sched.gFree.n += inc
4644 unlock(&sched.gFree.lock)
4647 // Breakpoint executes a breakpoint trap.
4652 // dolockOSThread is called by LockOSThread and lockOSThread below
4653 // after they modify m.locked. Do not allow preemption during this call,
4654 // or else the m might be different in this function than in the caller.
4657 func dolockOSThread() {
4658 if GOARCH == "wasm" {
4659 return // no threads on wasm yet
4662 gp.m.lockedg.set(gp)
4663 gp.lockedm.set(gp.m)
4666 // LockOSThread wires the calling goroutine to its current operating system thread.
4667 // The calling goroutine will always execute in that thread,
4668 // and no other goroutine will execute in it,
4669 // until the calling goroutine has made as many calls to
4670 // UnlockOSThread as to LockOSThread.
4671 // If the calling goroutine exits without unlocking the thread,
4672 // the thread will be terminated.
4674 // All init functions are run on the startup thread. Calling LockOSThread
4675 // from an init function will cause the main function to be invoked on
4678 // A goroutine should call LockOSThread before calling OS services or
4679 // non-Go library functions that depend on per-thread state.
4682 func LockOSThread() {
4683 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4684 // If we need to start a new thread from the locked
4685 // thread, we need the template thread. Start it now
4686 // while we're in a known-good state.
4687 startTemplateThread()
4691 if gp.m.lockedExt == 0 {
4693 panic("LockOSThread nesting overflow")
4699 func lockOSThread() {
4700 getg().m.lockedInt++
4704 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4705 // after they update m->locked. Do not allow preemption during this call,
4706 // or else the m might be in different in this function than in the caller.
4709 func dounlockOSThread() {
4710 if GOARCH == "wasm" {
4711 return // no threads on wasm yet
4714 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4721 // UnlockOSThread undoes an earlier call to LockOSThread.
4722 // If this drops the number of active LockOSThread calls on the
4723 // calling goroutine to zero, it unwires the calling goroutine from
4724 // its fixed operating system thread.
4725 // If there are no active LockOSThread calls, this is a no-op.
4727 // Before calling UnlockOSThread, the caller must ensure that the OS
4728 // thread is suitable for running other goroutines. If the caller made
4729 // any permanent changes to the state of the thread that would affect
4730 // other goroutines, it should not call this function and thus leave
4731 // the goroutine locked to the OS thread until the goroutine (and
4732 // hence the thread) exits.
4735 func UnlockOSThread() {
4737 if gp.m.lockedExt == 0 {
4745 func unlockOSThread() {
4747 if gp.m.lockedInt == 0 {
4748 systemstack(badunlockosthread)
4754 func badunlockosthread() {
4755 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4758 func gcount() int32 {
4759 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4760 for _, pp := range allp {
4764 // All these variables can be changed concurrently, so the result can be inconsistent.
4765 // But at least the current goroutine is running.
4772 func mcount() int32 {
4773 return int32(sched.mnext - sched.nmfreed)
4777 signalLock atomic.Uint32
4779 // Must hold signalLock to write. Reads may be lock-free, but
4780 // signalLock should be taken to synchronize with changes.
4784 func _System() { _System() }
4785 func _ExternalCode() { _ExternalCode() }
4786 func _LostExternalCode() { _LostExternalCode() }
4787 func _GC() { _GC() }
4788 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4789 func _VDSO() { _VDSO() }
4791 // Called if we receive a SIGPROF signal.
4792 // Called by the signal handler, may run during STW.
4794 //go:nowritebarrierrec
4795 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4796 if prof.hz.Load() == 0 {
4800 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4801 // We must check this to avoid a deadlock between setcpuprofilerate
4802 // and the call to cpuprof.add, below.
4803 if mp != nil && mp.profilehz == 0 {
4807 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4808 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4809 // the critical section, it creates a deadlock (when writing the sample).
4810 // As a workaround, create a counter of SIGPROFs while in critical section
4811 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4812 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4813 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4814 if f := findfunc(pc); f.valid() {
4815 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4816 cpuprof.lostAtomic++
4820 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4821 // runtime/internal/atomic functions call into kernel
4822 // helpers on arm < 7. See
4823 // runtime/internal/atomic/sys_linux_arm.s.
4824 cpuprof.lostAtomic++
4829 // Profiling runs concurrently with GC, so it must not allocate.
4830 // Set a trap in case the code does allocate.
4831 // Note that on windows, one thread takes profiles of all the
4832 // other threads, so mp is usually not getg().m.
4833 // In fact mp may not even be stopped.
4834 // See golang.org/issue/17165.
4835 getg().m.mallocing++
4838 var stk [maxCPUProfStack]uintptr
4840 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4842 // Check cgoCallersUse to make sure that we are not
4843 // interrupting other code that is fiddling with
4844 // cgoCallers. We are running in a signal handler
4845 // with all signals blocked, so we don't have to worry
4846 // about any other code interrupting us.
4847 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4848 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4851 n += copy(stk[:], mp.cgoCallers[:cgoOff])
4852 mp.cgoCallers[0] = 0
4855 // Collect Go stack that leads to the cgo call.
4856 u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
4857 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4858 // Libcall, i.e. runtime syscall on windows.
4859 // Collect Go stack that leads to the call.
4860 u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
4861 } else if mp != nil && mp.vdsoSP != 0 {
4862 // VDSO call, e.g. nanotime1 on Linux.
4863 // Collect Go stack that leads to the call.
4864 u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
4866 u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
4868 n += tracebackPCs(&u, 0, stk[n:])
4871 // Normal traceback is impossible or has failed.
4872 // Account it against abstract "System" or "GC".
4875 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4876 } else if pc > firstmoduledata.etext {
4877 // "ExternalCode" is better than "etext".
4878 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4881 if mp.preemptoff != "" {
4882 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4884 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4888 if prof.hz.Load() != 0 {
4889 // Note: it can happen on Windows that we interrupted a system thread
4890 // with no g, so gp could nil. The other nil checks are done out of
4891 // caution, but not expected to be nil in practice.
4892 var tagPtr *unsafe.Pointer
4893 if gp != nil && gp.m != nil && gp.m.curg != nil {
4894 tagPtr = &gp.m.curg.labels
4896 cpuprof.add(tagPtr, stk[:n])
4900 if gp != nil && gp.m != nil {
4901 if gp.m.curg != nil {
4906 traceCPUSample(gprof, pp, stk[:n])
4908 getg().m.mallocing--
4911 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4912 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4913 func setcpuprofilerate(hz int32) {
4914 // Force sane arguments.
4919 // Disable preemption, otherwise we can be rescheduled to another thread
4920 // that has profiling enabled.
4924 // Stop profiler on this thread so that it is safe to lock prof.
4925 // if a profiling signal came in while we had prof locked,
4926 // it would deadlock.
4927 setThreadCPUProfiler(0)
4929 for !prof.signalLock.CompareAndSwap(0, 1) {
4932 if prof.hz.Load() != hz {
4933 setProcessCPUProfiler(hz)
4936 prof.signalLock.Store(0)
4939 sched.profilehz = hz
4943 setThreadCPUProfiler(hz)
4949 // init initializes pp, which may be a freshly allocated p or a
4950 // previously destroyed p, and transitions it to status _Pgcstop.
4951 func (pp *p) init(id int32) {
4953 pp.status = _Pgcstop
4954 pp.sudogcache = pp.sudogbuf[:0]
4955 pp.deferpool = pp.deferpoolbuf[:0]
4957 if pp.mcache == nil {
4960 throw("missing mcache?")
4962 // Use the bootstrap mcache0. Only one P will get
4963 // mcache0: the one with ID 0.
4966 pp.mcache = allocmcache()
4969 if raceenabled && pp.raceprocctx == 0 {
4971 pp.raceprocctx = raceprocctx0
4972 raceprocctx0 = 0 // bootstrap
4974 pp.raceprocctx = raceproccreate()
4977 lockInit(&pp.timersLock, lockRankTimers)
4979 // This P may get timers when it starts running. Set the mask here
4980 // since the P may not go through pidleget (notably P 0 on startup).
4982 // Similarly, we may not go through pidleget before this P starts
4983 // running if it is P 0 on startup.
4987 // destroy releases all of the resources associated with pp and
4988 // transitions it to status _Pdead.
4990 // sched.lock must be held and the world must be stopped.
4991 func (pp *p) destroy() {
4992 assertLockHeld(&sched.lock)
4993 assertWorldStopped()
4995 // Move all runnable goroutines to the global queue
4996 for pp.runqhead != pp.runqtail {
4997 // Pop from tail of local queue
4999 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
5000 // Push onto head of global queue
5003 if pp.runnext != 0 {
5004 globrunqputhead(pp.runnext.ptr())
5007 if len(pp.timers) > 0 {
5008 plocal := getg().m.p.ptr()
5009 // The world is stopped, but we acquire timersLock to
5010 // protect against sysmon calling timeSleepUntil.
5011 // This is the only case where we hold the timersLock of
5012 // more than one P, so there are no deadlock concerns.
5013 lock(&plocal.timersLock)
5014 lock(&pp.timersLock)
5015 moveTimers(plocal, pp.timers)
5017 pp.numTimers.Store(0)
5018 pp.deletedTimers.Store(0)
5019 pp.timer0When.Store(0)
5020 unlock(&pp.timersLock)
5021 unlock(&plocal.timersLock)
5023 // Flush p's write barrier buffer.
5024 if gcphase != _GCoff {
5028 for i := range pp.sudogbuf {
5029 pp.sudogbuf[i] = nil
5031 pp.sudogcache = pp.sudogbuf[:0]
5032 for j := range pp.deferpoolbuf {
5033 pp.deferpoolbuf[j] = nil
5035 pp.deferpool = pp.deferpoolbuf[:0]
5036 systemstack(func() {
5037 for i := 0; i < pp.mspancache.len; i++ {
5038 // Safe to call since the world is stopped.
5039 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
5041 pp.mspancache.len = 0
5043 pp.pcache.flush(&mheap_.pages)
5044 unlock(&mheap_.lock)
5046 freemcache(pp.mcache)
5051 if pp.timerRaceCtx != 0 {
5052 // The race detector code uses a callback to fetch
5053 // the proc context, so arrange for that callback
5054 // to see the right thing.
5055 // This hack only works because we are the only
5061 racectxend(pp.timerRaceCtx)
5066 raceprocdestroy(pp.raceprocctx)
5073 // Change number of processors.
5075 // sched.lock must be held, and the world must be stopped.
5077 // gcworkbufs must not be being modified by either the GC or the write barrier
5078 // code, so the GC must not be running if the number of Ps actually changes.
5080 // Returns list of Ps with local work, they need to be scheduled by the caller.
5081 func procresize(nprocs int32) *p {
5082 assertLockHeld(&sched.lock)
5083 assertWorldStopped()
5086 if old < 0 || nprocs <= 0 {
5087 throw("procresize: invalid arg")
5090 traceGomaxprocs(nprocs)
5093 // update statistics
5095 if sched.procresizetime != 0 {
5096 sched.totaltime += int64(old) * (now - sched.procresizetime)
5098 sched.procresizetime = now
5100 maskWords := (nprocs + 31) / 32
5102 // Grow allp if necessary.
5103 if nprocs > int32(len(allp)) {
5104 // Synchronize with retake, which could be running
5105 // concurrently since it doesn't run on a P.
5107 if nprocs <= int32(cap(allp)) {
5108 allp = allp[:nprocs]
5110 nallp := make([]*p, nprocs)
5111 // Copy everything up to allp's cap so we
5112 // never lose old allocated Ps.
5113 copy(nallp, allp[:cap(allp)])
5117 if maskWords <= int32(cap(idlepMask)) {
5118 idlepMask = idlepMask[:maskWords]
5119 timerpMask = timerpMask[:maskWords]
5121 nidlepMask := make([]uint32, maskWords)
5122 // No need to copy beyond len, old Ps are irrelevant.
5123 copy(nidlepMask, idlepMask)
5124 idlepMask = nidlepMask
5126 ntimerpMask := make([]uint32, maskWords)
5127 copy(ntimerpMask, timerpMask)
5128 timerpMask = ntimerpMask
5133 // initialize new P's
5134 for i := old; i < nprocs; i++ {
5140 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5144 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5145 // continue to use the current P
5146 gp.m.p.ptr().status = _Prunning
5147 gp.m.p.ptr().mcache.prepareForSweep()
5149 // release the current P and acquire allp[0].
5151 // We must do this before destroying our current P
5152 // because p.destroy itself has write barriers, so we
5153 // need to do that from a valid P.
5156 // Pretend that we were descheduled
5157 // and then scheduled again to keep
5160 traceProcStop(gp.m.p.ptr())
5174 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5177 // release resources from unused P's
5178 for i := nprocs; i < old; i++ {
5181 // can't free P itself because it can be referenced by an M in syscall
5185 if int32(len(allp)) != nprocs {
5187 allp = allp[:nprocs]
5188 idlepMask = idlepMask[:maskWords]
5189 timerpMask = timerpMask[:maskWords]
5194 for i := nprocs - 1; i >= 0; i-- {
5196 if gp.m.p.ptr() == pp {
5204 pp.link.set(runnablePs)
5208 stealOrder.reset(uint32(nprocs))
5209 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5210 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5212 // Notify the limiter that the amount of procs has changed.
5213 gcCPULimiter.resetCapacity(now, nprocs)
5218 // Associate p and the current m.
5220 // This function is allowed to have write barriers even if the caller
5221 // isn't because it immediately acquires pp.
5223 //go:yeswritebarrierrec
5224 func acquirep(pp *p) {
5225 // Do the part that isn't allowed to have write barriers.
5228 // Have p; write barriers now allowed.
5230 // Perform deferred mcache flush before this P can allocate
5231 // from a potentially stale mcache.
5232 pp.mcache.prepareForSweep()
5239 // wirep is the first step of acquirep, which actually associates the
5240 // current M to pp. This is broken out so we can disallow write
5241 // barriers for this part, since we don't yet have a P.
5243 //go:nowritebarrierrec
5249 throw("wirep: already in go")
5251 if pp.m != 0 || pp.status != _Pidle {
5256 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5257 throw("wirep: invalid p state")
5261 pp.status = _Prunning
5264 // Disassociate p and the current m.
5265 func releasep() *p {
5269 throw("releasep: invalid arg")
5272 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5273 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5274 throw("releasep: invalid p state")
5277 traceProcStop(gp.m.p.ptr())
5285 func incidlelocked(v int32) {
5287 sched.nmidlelocked += v
5294 // Check for deadlock situation.
5295 // The check is based on number of running M's, if 0 -> deadlock.
5296 // sched.lock must be held.
5298 assertLockHeld(&sched.lock)
5300 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5301 // there are no running goroutines. The calling program is
5302 // assumed to be running.
5303 if islibrary || isarchive {
5307 // If we are dying because of a signal caught on an already idle thread,
5308 // freezetheworld will cause all running threads to block.
5309 // And runtime will essentially enter into deadlock state,
5310 // except that there is a thread that will call exit soon.
5311 if panicking.Load() > 0 {
5315 // If we are not running under cgo, but we have an extra M then account
5316 // for it. (It is possible to have an extra M on Windows without cgo to
5317 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5320 if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
5324 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5329 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5331 throw("checkdead: inconsistent counts")
5335 forEachG(func(gp *g) {
5336 if isSystemGoroutine(gp, false) {
5339 s := readgstatus(gp)
5340 switch s &^ _Gscan {
5347 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5349 throw("checkdead: runnable g")
5352 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5353 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5354 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5357 // Maybe jump time forward for playground.
5359 if when := timeSleepUntil(); when < maxWhen {
5362 // Start an M to steal the timer.
5363 pp, _ := pidleget(faketime)
5365 // There should always be a free P since
5366 // nothing is running.
5368 throw("checkdead: no p for timer")
5372 // There should always be a free M since
5373 // nothing is running.
5375 throw("checkdead: no m for timer")
5377 // M must be spinning to steal. We set this to be
5378 // explicit, but since this is the only M it would
5379 // become spinning on its own anyways.
5380 sched.nmspinning.Add(1)
5383 notewakeup(&mp.park)
5388 // There are no goroutines running, so we can look at the P's.
5389 for _, pp := range allp {
5390 if len(pp.timers) > 0 {
5395 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5396 fatal("all goroutines are asleep - deadlock!")
5399 // forcegcperiod is the maximum time in nanoseconds between garbage
5400 // collections. If we go this long without a garbage collection, one
5401 // is forced to run.
5403 // This is a variable for testing purposes. It normally doesn't change.
5404 var forcegcperiod int64 = 2 * 60 * 1e9
5406 // needSysmonWorkaround is true if the workaround for
5407 // golang.org/issue/42515 is needed on NetBSD.
5408 var needSysmonWorkaround bool = false
5410 // Always runs without a P, so write barriers are not allowed.
5412 //go:nowritebarrierrec
5419 lasttrace := int64(0)
5420 idle := 0 // how many cycles in succession we had not wokeup somebody
5424 if idle == 0 { // start with 20us sleep...
5426 } else if idle > 50 { // start doubling the sleep after 1ms...
5429 if delay > 10*1000 { // up to 10ms
5434 // sysmon should not enter deep sleep if schedtrace is enabled so that
5435 // it can print that information at the right time.
5437 // It should also not enter deep sleep if there are any active P's so
5438 // that it can retake P's from syscalls, preempt long running G's, and
5439 // poll the network if all P's are busy for long stretches.
5441 // It should wakeup from deep sleep if any P's become active either due
5442 // to exiting a syscall or waking up due to a timer expiring so that it
5443 // can resume performing those duties. If it wakes from a syscall it
5444 // resets idle and delay as a bet that since it had retaken a P from a
5445 // syscall before, it may need to do it again shortly after the
5446 // application starts work again. It does not reset idle when waking
5447 // from a timer to avoid adding system load to applications that spend
5448 // most of their time sleeping.
5450 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5452 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5453 syscallWake := false
5454 next := timeSleepUntil()
5456 sched.sysmonwait.Store(true)
5458 // Make wake-up period small enough
5459 // for the sampling to be correct.
5460 sleep := forcegcperiod / 2
5461 if next-now < sleep {
5464 shouldRelax := sleep >= osRelaxMinNS
5468 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5473 sched.sysmonwait.Store(false)
5474 noteclear(&sched.sysmonnote)
5484 lock(&sched.sysmonlock)
5485 // Update now in case we blocked on sysmonnote or spent a long time
5486 // blocked on schedlock or sysmonlock above.
5489 // trigger libc interceptors if needed
5490 if *cgo_yield != nil {
5491 asmcgocall(*cgo_yield, nil)
5493 // poll network if not polled for more than 10ms
5494 lastpoll := sched.lastpoll.Load()
5495 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5496 sched.lastpoll.CompareAndSwap(lastpoll, now)
5497 list := netpoll(0) // non-blocking - returns list of goroutines
5499 // Need to decrement number of idle locked M's
5500 // (pretending that one more is running) before injectglist.
5501 // Otherwise it can lead to the following situation:
5502 // injectglist grabs all P's but before it starts M's to run the P's,
5503 // another M returns from syscall, finishes running its G,
5504 // observes that there is no work to do and no other running M's
5505 // and reports deadlock.
5511 if GOOS == "netbsd" && needSysmonWorkaround {
5512 // netpoll is responsible for waiting for timer
5513 // expiration, so we typically don't have to worry
5514 // about starting an M to service timers. (Note that
5515 // sleep for timeSleepUntil above simply ensures sysmon
5516 // starts running again when that timer expiration may
5517 // cause Go code to run again).
5519 // However, netbsd has a kernel bug that sometimes
5520 // misses netpollBreak wake-ups, which can lead to
5521 // unbounded delays servicing timers. If we detect this
5522 // overrun, then startm to get something to handle the
5525 // See issue 42515 and
5526 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5527 if next := timeSleepUntil(); next < now {
5528 startm(nil, false, false)
5531 if scavenger.sysmonWake.Load() != 0 {
5532 // Kick the scavenger awake if someone requested it.
5535 // retake P's blocked in syscalls
5536 // and preempt long running G's
5537 if retake(now) != 0 {
5542 // check if we need to force a GC
5543 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5545 forcegc.idle.Store(false)
5547 list.push(forcegc.g)
5549 unlock(&forcegc.lock)
5551 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5553 schedtrace(debug.scheddetail > 0)
5555 unlock(&sched.sysmonlock)
5559 type sysmontick struct {
5566 // forcePreemptNS is the time slice given to a G before it is
5568 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5570 func retake(now int64) uint32 {
5572 // Prevent allp slice changes. This lock will be completely
5573 // uncontended unless we're already stopping the world.
5575 // We can't use a range loop over allp because we may
5576 // temporarily drop the allpLock. Hence, we need to re-fetch
5577 // allp each time around the loop.
5578 for i := 0; i < len(allp); i++ {
5581 // This can happen if procresize has grown
5582 // allp but not yet created new Ps.
5585 pd := &pp.sysmontick
5588 if s == _Prunning || s == _Psyscall {
5589 // Preempt G if it's running for too long.
5590 t := int64(pp.schedtick)
5591 if int64(pd.schedtick) != t {
5592 pd.schedtick = uint32(t)
5594 } else if pd.schedwhen+forcePreemptNS <= now {
5596 // In case of syscall, preemptone() doesn't
5597 // work, because there is no M wired to P.
5602 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5603 t := int64(pp.syscalltick)
5604 if !sysretake && int64(pd.syscalltick) != t {
5605 pd.syscalltick = uint32(t)
5606 pd.syscallwhen = now
5609 // On the one hand we don't want to retake Ps if there is no other work to do,
5610 // but on the other hand we want to retake them eventually
5611 // because they can prevent the sysmon thread from deep sleep.
5612 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5615 // Drop allpLock so we can take sched.lock.
5617 // Need to decrement number of idle locked M's
5618 // (pretending that one more is running) before the CAS.
5619 // Otherwise the M from which we retake can exit the syscall,
5620 // increment nmidle and report deadlock.
5622 if atomic.Cas(&pp.status, s, _Pidle) {
5639 // Tell all goroutines that they have been preempted and they should stop.
5640 // This function is purely best-effort. It can fail to inform a goroutine if a
5641 // processor just started running it.
5642 // No locks need to be held.
5643 // Returns true if preemption request was issued to at least one goroutine.
5644 func preemptall() bool {
5646 for _, pp := range allp {
5647 if pp.status != _Prunning {
5657 // Tell the goroutine running on processor P to stop.
5658 // This function is purely best-effort. It can incorrectly fail to inform the
5659 // goroutine. It can inform the wrong goroutine. Even if it informs the
5660 // correct goroutine, that goroutine might ignore the request if it is
5661 // simultaneously executing newstack.
5662 // No lock needs to be held.
5663 // Returns true if preemption request was issued.
5664 // The actual preemption will happen at some point in the future
5665 // and will be indicated by the gp->status no longer being
5667 func preemptone(pp *p) bool {
5669 if mp == nil || mp == getg().m {
5673 if gp == nil || gp == mp.g0 {
5679 // Every call in a goroutine checks for stack overflow by
5680 // comparing the current stack pointer to gp->stackguard0.
5681 // Setting gp->stackguard0 to StackPreempt folds
5682 // preemption into the normal stack overflow check.
5683 gp.stackguard0 = stackPreempt
5685 // Request an async preemption of this P.
5686 if preemptMSupported && debug.asyncpreemptoff == 0 {
5696 func schedtrace(detailed bool) {
5703 print("SCHED ", (now-starttime)/1e6, "ms: gomaxprocs=", gomaxprocs, " idleprocs=", sched.npidle.Load(), " threads=", mcount(), " spinningthreads=", sched.nmspinning.Load(), " needspinning=", sched.needspinning.Load(), " idlethreads=", sched.nmidle, " runqueue=", sched.runqsize)
5705 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5707 // We must be careful while reading data from P's, M's and G's.
5708 // Even if we hold schedlock, most data can be changed concurrently.
5709 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5710 for i, pp := range allp {
5712 h := atomic.Load(&pp.runqhead)
5713 t := atomic.Load(&pp.runqtail)
5715 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5721 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5723 // In non-detailed mode format lengths of per-P run queues as:
5724 // [len1 len2 len3 len4]
5730 if i == len(allp)-1 {
5741 for mp := allm; mp != nil; mp = mp.alllink {
5743 print(" M", mp.id, ": p=")
5755 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5756 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5764 forEachG(func(gp *g) {
5765 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5772 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5782 // schedEnableUser enables or disables the scheduling of user
5785 // This does not stop already running user goroutines, so the caller
5786 // should first stop the world when disabling user goroutines.
5787 func schedEnableUser(enable bool) {
5789 if sched.disable.user == !enable {
5793 sched.disable.user = !enable
5795 n := sched.disable.n
5797 globrunqputbatch(&sched.disable.runnable, n)
5799 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5800 startm(nil, false, false)
5807 // schedEnabled reports whether gp should be scheduled. It returns
5808 // false is scheduling of gp is disabled.
5810 // sched.lock must be held.
5811 func schedEnabled(gp *g) bool {
5812 assertLockHeld(&sched.lock)
5814 if sched.disable.user {
5815 return isSystemGoroutine(gp, true)
5820 // Put mp on midle list.
5821 // sched.lock must be held.
5822 // May run during STW, so write barriers are not allowed.
5824 //go:nowritebarrierrec
5826 assertLockHeld(&sched.lock)
5828 mp.schedlink = sched.midle
5834 // Try to get an m from midle list.
5835 // sched.lock must be held.
5836 // May run during STW, so write barriers are not allowed.
5838 //go:nowritebarrierrec
5840 assertLockHeld(&sched.lock)
5842 mp := sched.midle.ptr()
5844 sched.midle = mp.schedlink
5850 // Put gp on the global runnable queue.
5851 // sched.lock must be held.
5852 // May run during STW, so write barriers are not allowed.
5854 //go:nowritebarrierrec
5855 func globrunqput(gp *g) {
5856 assertLockHeld(&sched.lock)
5858 sched.runq.pushBack(gp)
5862 // Put gp at the head of the global runnable queue.
5863 // sched.lock must be held.
5864 // May run during STW, so write barriers are not allowed.
5866 //go:nowritebarrierrec
5867 func globrunqputhead(gp *g) {
5868 assertLockHeld(&sched.lock)
5874 // Put a batch of runnable goroutines on the global runnable queue.
5875 // This clears *batch.
5876 // sched.lock must be held.
5877 // May run during STW, so write barriers are not allowed.
5879 //go:nowritebarrierrec
5880 func globrunqputbatch(batch *gQueue, n int32) {
5881 assertLockHeld(&sched.lock)
5883 sched.runq.pushBackAll(*batch)
5888 // Try get a batch of G's from the global runnable queue.
5889 // sched.lock must be held.
5890 func globrunqget(pp *p, max int32) *g {
5891 assertLockHeld(&sched.lock)
5893 if sched.runqsize == 0 {
5897 n := sched.runqsize/gomaxprocs + 1
5898 if n > sched.runqsize {
5901 if max > 0 && n > max {
5904 if n > int32(len(pp.runq))/2 {
5905 n = int32(len(pp.runq)) / 2
5910 gp := sched.runq.pop()
5913 gp1 := sched.runq.pop()
5914 runqput(pp, gp1, false)
5919 // pMask is an atomic bitstring with one bit per P.
5922 // read returns true if P id's bit is set.
5923 func (p pMask) read(id uint32) bool {
5925 mask := uint32(1) << (id % 32)
5926 return (atomic.Load(&p[word]) & mask) != 0
5929 // set sets P id's bit.
5930 func (p pMask) set(id int32) {
5932 mask := uint32(1) << (id % 32)
5933 atomic.Or(&p[word], mask)
5936 // clear clears P id's bit.
5937 func (p pMask) clear(id int32) {
5939 mask := uint32(1) << (id % 32)
5940 atomic.And(&p[word], ^mask)
5943 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5945 // Ideally, the timer mask would be kept immediately consistent on any timer
5946 // operations. Unfortunately, updating a shared global data structure in the
5947 // timer hot path adds too much overhead in applications frequently switching
5948 // between no timers and some timers.
5950 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5951 // running P (returned by pidleget) may add a timer at any time, so its mask
5952 // must be set. An idle P (passed to pidleput) cannot add new timers while
5953 // idle, so if it has no timers at that time, its mask may be cleared.
5955 // Thus, we get the following effects on timer-stealing in findrunnable:
5957 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5958 // (for work- or timer-stealing; this is the ideal case).
5959 // - Running Ps must always be checked.
5960 // - Idle Ps whose timers are stolen must continue to be checked until they run
5961 // again, even after timer expiration.
5963 // When the P starts running again, the mask should be set, as a timer may be
5964 // added at any time.
5966 // TODO(prattmic): Additional targeted updates may improve the above cases.
5967 // e.g., updating the mask when stealing a timer.
5968 func updateTimerPMask(pp *p) {
5969 if pp.numTimers.Load() > 0 {
5973 // Looks like there are no timers, however another P may transiently
5974 // decrement numTimers when handling a timerModified timer in
5975 // checkTimers. We must take timersLock to serialize with these changes.
5976 lock(&pp.timersLock)
5977 if pp.numTimers.Load() == 0 {
5978 timerpMask.clear(pp.id)
5980 unlock(&pp.timersLock)
5983 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5984 // to nanotime or zero. Returns now or the current time if now was zero.
5986 // This releases ownership of p. Once sched.lock is released it is no longer
5989 // sched.lock must be held.
5991 // May run during STW, so write barriers are not allowed.
5993 //go:nowritebarrierrec
5994 func pidleput(pp *p, now int64) int64 {
5995 assertLockHeld(&sched.lock)
5998 throw("pidleput: P has non-empty run queue")
6003 updateTimerPMask(pp) // clear if there are no timers.
6004 idlepMask.set(pp.id)
6005 pp.link = sched.pidle
6008 if !pp.limiterEvent.start(limiterEventIdle, now) {
6009 throw("must be able to track idle limiter event")
6014 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
6016 // sched.lock must be held.
6018 // May run during STW, so write barriers are not allowed.
6020 //go:nowritebarrierrec
6021 func pidleget(now int64) (*p, int64) {
6022 assertLockHeld(&sched.lock)
6024 pp := sched.pidle.ptr()
6026 // Timer may get added at any time now.
6030 timerpMask.set(pp.id)
6031 idlepMask.clear(pp.id)
6032 sched.pidle = pp.link
6033 sched.npidle.Add(-1)
6034 pp.limiterEvent.stop(limiterEventIdle, now)
6039 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
6040 // This is called by spinning Ms (or callers than need a spinning M) that have
6041 // found work. If no P is available, this must synchronized with non-spinning
6042 // Ms that may be preparing to drop their P without discovering this work.
6044 // sched.lock must be held.
6046 // May run during STW, so write barriers are not allowed.
6048 //go:nowritebarrierrec
6049 func pidlegetSpinning(now int64) (*p, int64) {
6050 assertLockHeld(&sched.lock)
6052 pp, now := pidleget(now)
6054 // See "Delicate dance" comment in findrunnable. We found work
6055 // that we cannot take, we must synchronize with non-spinning
6056 // Ms that may be preparing to drop their P.
6057 sched.needspinning.Store(1)
6064 // runqempty reports whether pp has no Gs on its local run queue.
6065 // It never returns true spuriously.
6066 func runqempty(pp *p) bool {
6067 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
6068 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
6069 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
6070 // does not mean the queue is empty.
6072 head := atomic.Load(&pp.runqhead)
6073 tail := atomic.Load(&pp.runqtail)
6074 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
6075 if tail == atomic.Load(&pp.runqtail) {
6076 return head == tail && runnext == 0
6081 // To shake out latent assumptions about scheduling order,
6082 // we introduce some randomness into scheduling decisions
6083 // when running with the race detector.
6084 // The need for this was made obvious by changing the
6085 // (deterministic) scheduling order in Go 1.5 and breaking
6086 // many poorly-written tests.
6087 // With the randomness here, as long as the tests pass
6088 // consistently with -race, they shouldn't have latent scheduling
6090 const randomizeScheduler = raceenabled
6092 // runqput tries to put g on the local runnable queue.
6093 // If next is false, runqput adds g to the tail of the runnable queue.
6094 // If next is true, runqput puts g in the pp.runnext slot.
6095 // If the run queue is full, runnext puts g on the global queue.
6096 // Executed only by the owner P.
6097 func runqput(pp *p, gp *g, next bool) {
6098 if randomizeScheduler && next && fastrandn(2) == 0 {
6104 oldnext := pp.runnext
6105 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
6111 // Kick the old runnext out to the regular run queue.
6116 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6118 if t-h < uint32(len(pp.runq)) {
6119 pp.runq[t%uint32(len(pp.runq))].set(gp)
6120 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
6123 if runqputslow(pp, gp, h, t) {
6126 // the queue is not full, now the put above must succeed
6130 // Put g and a batch of work from local runnable queue on global queue.
6131 // Executed only by the owner P.
6132 func runqputslow(pp *p, gp *g, h, t uint32) bool {
6133 var batch [len(pp.runq)/2 + 1]*g
6135 // First, grab a batch from local queue.
6138 if n != uint32(len(pp.runq)/2) {
6139 throw("runqputslow: queue is not full")
6141 for i := uint32(0); i < n; i++ {
6142 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6144 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6149 if randomizeScheduler {
6150 for i := uint32(1); i <= n; i++ {
6151 j := fastrandn(i + 1)
6152 batch[i], batch[j] = batch[j], batch[i]
6156 // Link the goroutines.
6157 for i := uint32(0); i < n; i++ {
6158 batch[i].schedlink.set(batch[i+1])
6161 q.head.set(batch[0])
6162 q.tail.set(batch[n])
6164 // Now put the batch on global queue.
6166 globrunqputbatch(&q, int32(n+1))
6171 // runqputbatch tries to put all the G's on q on the local runnable queue.
6172 // If the queue is full, they are put on the global queue; in that case
6173 // this will temporarily acquire the scheduler lock.
6174 // Executed only by the owner P.
6175 func runqputbatch(pp *p, q *gQueue, qsize int) {
6176 h := atomic.LoadAcq(&pp.runqhead)
6179 for !q.empty() && t-h < uint32(len(pp.runq)) {
6181 pp.runq[t%uint32(len(pp.runq))].set(gp)
6187 if randomizeScheduler {
6188 off := func(o uint32) uint32 {
6189 return (pp.runqtail + o) % uint32(len(pp.runq))
6191 for i := uint32(1); i < n; i++ {
6192 j := fastrandn(i + 1)
6193 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6197 atomic.StoreRel(&pp.runqtail, t)
6200 globrunqputbatch(q, int32(qsize))
6205 // Get g from local runnable queue.
6206 // If inheritTime is true, gp should inherit the remaining time in the
6207 // current time slice. Otherwise, it should start a new time slice.
6208 // Executed only by the owner P.
6209 func runqget(pp *p) (gp *g, inheritTime bool) {
6210 // If there's a runnext, it's the next G to run.
6212 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6213 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6214 // Hence, there's no need to retry this CAS if it fails.
6215 if next != 0 && pp.runnext.cas(next, 0) {
6216 return next.ptr(), true
6220 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6225 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6226 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6232 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6233 // Executed only by the owner P.
6234 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6235 oldNext := pp.runnext
6236 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6237 drainQ.pushBack(oldNext.ptr())
6242 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6248 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6252 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6256 // We've inverted the order in which it gets G's from the local P's runnable queue
6257 // and then advances the head pointer because we don't want to mess up the statuses of G's
6258 // while runqdrain() and runqsteal() are running in parallel.
6259 // Thus we should advance the head pointer before draining the local P into a gQueue,
6260 // so that we can update any gp.schedlink only after we take the full ownership of G,
6261 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6262 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6263 for i := uint32(0); i < qn; i++ {
6264 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6271 // Grabs a batch of goroutines from pp's runnable queue into batch.
6272 // Batch is a ring buffer starting at batchHead.
6273 // Returns number of grabbed goroutines.
6274 // Can be executed by any P.
6275 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6277 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6278 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6283 // Try to steal from pp.runnext.
6284 if next := pp.runnext; next != 0 {
6285 if pp.status == _Prunning {
6286 // Sleep to ensure that pp isn't about to run the g
6287 // we are about to steal.
6288 // The important use case here is when the g running
6289 // on pp ready()s another g and then almost
6290 // immediately blocks. Instead of stealing runnext
6291 // in this window, back off to give pp a chance to
6292 // schedule runnext. This will avoid thrashing gs
6293 // between different Ps.
6294 // A sync chan send/recv takes ~50ns as of time of
6295 // writing, so 3us gives ~50x overshoot.
6296 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6299 // On some platforms system timer granularity is
6300 // 1-15ms, which is way too much for this
6301 // optimization. So just yield.
6305 if !pp.runnext.cas(next, 0) {
6308 batch[batchHead%uint32(len(batch))] = next
6314 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6317 for i := uint32(0); i < n; i++ {
6318 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6319 batch[(batchHead+i)%uint32(len(batch))] = g
6321 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6327 // Steal half of elements from local runnable queue of p2
6328 // and put onto local runnable queue of p.
6329 // Returns one of the stolen elements (or nil if failed).
6330 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6332 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6337 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6341 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6342 if t-h+n >= uint32(len(pp.runq)) {
6343 throw("runqsteal: runq overflow")
6345 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6349 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6350 // be on one gQueue or gList at a time.
6351 type gQueue struct {
6356 // empty reports whether q is empty.
6357 func (q *gQueue) empty() bool {
6361 // push adds gp to the head of q.
6362 func (q *gQueue) push(gp *g) {
6363 gp.schedlink = q.head
6370 // pushBack adds gp to the tail of q.
6371 func (q *gQueue) pushBack(gp *g) {
6374 q.tail.ptr().schedlink.set(gp)
6381 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6383 func (q *gQueue) pushBackAll(q2 gQueue) {
6387 q2.tail.ptr().schedlink = 0
6389 q.tail.ptr().schedlink = q2.head
6396 // pop removes and returns the head of queue q. It returns nil if
6398 func (q *gQueue) pop() *g {
6401 q.head = gp.schedlink
6409 // popList takes all Gs in q and returns them as a gList.
6410 func (q *gQueue) popList() gList {
6411 stack := gList{q.head}
6416 // A gList is a list of Gs linked through g.schedlink. A G can only be
6417 // on one gQueue or gList at a time.
6422 // empty reports whether l is empty.
6423 func (l *gList) empty() bool {
6427 // push adds gp to the head of l.
6428 func (l *gList) push(gp *g) {
6429 gp.schedlink = l.head
6433 // pushAll prepends all Gs in q to l.
6434 func (l *gList) pushAll(q gQueue) {
6436 q.tail.ptr().schedlink = l.head
6441 // pop removes and returns the head of l. If l is empty, it returns nil.
6442 func (l *gList) pop() *g {
6445 l.head = gp.schedlink
6450 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6451 func setMaxThreads(in int) (out int) {
6453 out = int(sched.maxmcount)
6454 if in > 0x7fffffff { // MaxInt32
6455 sched.maxmcount = 0x7fffffff
6457 sched.maxmcount = int32(in)
6465 func procPin() int {
6470 return int(mp.p.ptr().id)
6479 //go:linkname sync_runtime_procPin sync.runtime_procPin
6481 func sync_runtime_procPin() int {
6485 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6487 func sync_runtime_procUnpin() {
6491 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6493 func sync_atomic_runtime_procPin() int {
6497 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6499 func sync_atomic_runtime_procUnpin() {
6503 // Active spinning for sync.Mutex.
6505 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6507 func sync_runtime_canSpin(i int) bool {
6508 // sync.Mutex is cooperative, so we are conservative with spinning.
6509 // Spin only few times and only if running on a multicore machine and
6510 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6511 // As opposed to runtime mutex we don't do passive spinning here,
6512 // because there can be work on global runq or on other Ps.
6513 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6516 if p := getg().m.p.ptr(); !runqempty(p) {
6522 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6524 func sync_runtime_doSpin() {
6525 procyield(active_spin_cnt)
6528 var stealOrder randomOrder
6530 // randomOrder/randomEnum are helper types for randomized work stealing.
6531 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6532 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6533 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6534 type randomOrder struct {
6539 type randomEnum struct {
6546 func (ord *randomOrder) reset(count uint32) {
6548 ord.coprimes = ord.coprimes[:0]
6549 for i := uint32(1); i <= count; i++ {
6550 if gcd(i, count) == 1 {
6551 ord.coprimes = append(ord.coprimes, i)
6556 func (ord *randomOrder) start(i uint32) randomEnum {
6560 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6564 func (enum *randomEnum) done() bool {
6565 return enum.i == enum.count
6568 func (enum *randomEnum) next() {
6570 enum.pos = (enum.pos + enum.inc) % enum.count
6573 func (enum *randomEnum) position() uint32 {
6577 func gcd(a, b uint32) uint32 {
6584 // An initTask represents the set of initializations that need to be done for a package.
6585 // Keep in sync with ../../test/noinit.go:initTask
6586 type initTask struct {
6587 state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
6589 // followed by nfns pcs, uintptr sized, one per init function to run
6592 // inittrace stores statistics for init functions which are
6593 // updated by malloc and newproc when active is true.
6594 var inittrace tracestat
6596 type tracestat struct {
6597 active bool // init tracing activation status
6598 id uint64 // init goroutine id
6599 allocs uint64 // heap allocations
6600 bytes uint64 // heap allocated bytes
6603 func doInit(ts []*initTask) {
6604 for _, t := range ts {
6609 func doInit1(t *initTask) {
6611 case 2: // fully initialized
6613 case 1: // initialization in progress
6614 throw("recursive call during initialization - linker skew")
6615 default: // not initialized yet
6616 t.state = 1 // initialization in progress
6623 if inittrace.active {
6625 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6630 // We should have pruned all of these in the linker.
6631 throw("inittask with no functions")
6634 firstFunc := add(unsafe.Pointer(t), 8)
6635 for i := uint32(0); i < t.nfns; i++ {
6636 p := add(firstFunc, uintptr(i)*goarch.PtrSize)
6637 f := *(*func())(unsafe.Pointer(&p))
6641 if inittrace.active {
6643 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6646 f := *(*func())(unsafe.Pointer(&firstFunc))
6647 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6650 print("init ", pkg, " @")
6651 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6652 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6653 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6654 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6658 t.state = 2 // initialization done