1 // Copyright 2014 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
11 "runtime/internal/atomic"
12 "runtime/internal/sys"
16 // set using cmd/go/internal/modload.ModInfoProg
19 // Goroutine scheduler
20 // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
22 // The main concepts are:
24 // M - worker thread, or machine.
25 // P - processor, a resource that is required to execute Go code.
26 // M must have an associated P to execute Go code, however it can be
27 // blocked or in a syscall w/o an associated P.
29 // Design doc at https://golang.org/s/go11sched.
31 // Worker thread parking/unparking.
32 // We need to balance between keeping enough running worker threads to utilize
33 // available hardware parallelism and parking excessive running worker threads
34 // to conserve CPU resources and power. This is not simple for two reasons:
35 // (1) scheduler state is intentionally distributed (in particular, per-P work
36 // queues), so it is not possible to compute global predicates on fast paths;
37 // (2) for optimal thread management we would need to know the future (don't park
38 // a worker thread when a new goroutine will be readied in near future).
40 // Three rejected approaches that would work badly:
41 // 1. Centralize all scheduler state (would inhibit scalability).
42 // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
43 // is a spare P, unpark a thread and handoff it the thread and the goroutine.
44 // This would lead to thread state thrashing, as the thread that readied the
45 // goroutine can be out of work the very next moment, we will need to park it.
46 // Also, it would destroy locality of computation as we want to preserve
47 // dependent goroutines on the same thread; and introduce additional latency.
48 // 3. Unpark an additional thread whenever we ready a goroutine and there is an
49 // idle P, but don't do handoff. This would lead to excessive thread parking/
50 // unparking as the additional threads will instantly park without discovering
53 // The current approach:
55 // This approach applies to three primary sources of potential work: readying a
56 // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
57 // additional details.
59 // We unpark an additional thread when we submit work if (this is wakep()):
60 // 1. There is an idle P, and
61 // 2. There are no "spinning" worker threads.
63 // A worker thread is considered spinning if it is out of local work and did
64 // not find work in the global run queue or netpoller; the spinning state is
65 // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
66 // also considered spinning; we don't do goroutine handoff so such threads are
67 // out of work initially. Spinning threads spin on looking for work in per-P
68 // run queues and timer heaps or from the GC before parking. If a spinning
69 // thread finds work it takes itself out of the spinning state and proceeds to
70 // execution. If it does not find work it takes itself out of the spinning
71 // state and then parks.
73 // If there is at least one spinning thread (sched.nmspinning>1), we don't
74 // unpark new threads when submitting work. To compensate for that, if the last
75 // spinning thread finds work and stops spinning, it must unpark a new spinning
76 // thread. This approach smooths out unjustified spikes of thread unparking,
77 // but at the same time guarantees eventual maximal CPU parallelism
80 // The main implementation complication is that we need to be very careful
81 // during spinning->non-spinning thread transition. This transition can race
82 // with submission of new work, and either one part or another needs to unpark
83 // another worker thread. If they both fail to do that, we can end up with
84 // semi-persistent CPU underutilization.
86 // The general pattern for submission is:
87 // 1. Submit work to the local run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
120 //go:linkname runtime_inittask runtime..inittask
121 var runtime_inittask initTask
123 //go:linkname main_inittask main..inittask
124 var main_inittask initTask
126 // main_init_done is a signal used by cgocallbackg that initialization
127 // has been completed. It is made before _cgo_notify_runtime_init_done,
128 // so all cgo calls can rely on it existing. When main_init is complete,
129 // it is closed, meaning cgocallbackg can reliably receive from it.
130 var main_init_done chan bool
132 //go:linkname main_main main.main
135 // mainStarted indicates that the main M has started.
138 // runtimeInitTime is the nanotime() at which the runtime started.
139 var runtimeInitTime int64
141 // Value to use for signal mask for newly created M's.
142 var initSigmask sigset
144 // The main goroutine.
148 // Racectx of m0->g0 is used only as the parent of the main goroutine.
149 // It must not be used for anything else.
152 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
153 // Using decimal instead of binary GB and MB because
154 // they look nicer in the stack overflow failure message.
155 if goarch.PtrSize == 8 {
156 maxstacksize = 1000000000
158 maxstacksize = 250000000
161 // An upper limit for max stack size. Used to avoid random crashes
162 // after calling SetMaxStack and trying to allocate a stack that is too big,
163 // since stackalloc works with 32-bit sizes.
164 maxstackceiling = 2 * maxstacksize
166 // Allow newproc to start new Ms.
169 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
171 newm(sysmon, nil, -1)
175 // Lock the main goroutine onto this, the main OS thread,
176 // during initialization. Most programs won't care, but a few
177 // do require certain calls to be made by the main thread.
178 // Those can arrange for main.main to run in the main thread
179 // by calling runtime.LockOSThread during initialization
180 // to preserve the lock.
184 throw("runtime.main not on m0")
187 // Record when the world started.
188 // Must be before doInit for tracing init.
189 runtimeInitTime = nanotime()
190 if runtimeInitTime == 0 {
191 throw("nanotime returning zero")
194 if debug.inittrace != 0 {
195 inittrace.id = getg().goid
196 inittrace.active = true
199 doInit(&runtime_inittask) // Must be before defer.
201 // Defer unlock so that runtime.Goexit during init does the unlock too.
211 main_init_done = make(chan bool)
213 if _cgo_pthread_key_created == nil {
214 throw("_cgo_pthread_key_created missing")
217 if _cgo_thread_start == nil {
218 throw("_cgo_thread_start missing")
220 if GOOS != "windows" {
221 if _cgo_setenv == nil {
222 throw("_cgo_setenv missing")
224 if _cgo_unsetenv == nil {
225 throw("_cgo_unsetenv missing")
228 if _cgo_notify_runtime_init_done == nil {
229 throw("_cgo_notify_runtime_init_done missing")
232 // Set the x_crosscall2_ptr C function pointer variable point to crosscall2.
233 if set_crosscall2 == nil {
234 throw("set_crosscall2 missing")
238 // Start the template thread in case we enter Go from
239 // a C-created thread and need to create a new thread.
240 startTemplateThread()
241 cgocall(_cgo_notify_runtime_init_done, nil)
244 doInit(&main_inittask)
246 // Disable init tracing after main init done to avoid overhead
247 // of collecting statistics in malloc and newproc
248 inittrace.active = false
250 close(main_init_done)
255 if isarchive || islibrary {
256 // A program compiled with -buildmode=c-archive or c-shared
257 // has a main, but it is not executed.
260 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
263 runExitHooks(0) // run hooks now, since racefini does not return
267 // Make racy client program work: if panicking on
268 // another goroutine at the same time as main returns,
269 // let the other goroutine finish printing the panic trace.
270 // Once it does, it will exit. See issues 3934 and 20018.
271 if runningPanicDefers.Load() != 0 {
272 // Running deferred functions should not take long.
273 for c := 0; c < 1000; c++ {
274 if runningPanicDefers.Load() == 0 {
280 if panicking.Load() != 0 {
281 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
292 // os_beforeExit is called from os.Exit(0).
294 //go:linkname os_beforeExit os.runtime_beforeExit
295 func os_beforeExit(exitCode int) {
296 runExitHooks(exitCode)
297 if exitCode == 0 && raceenabled {
302 // start forcegc helper goroutine
307 func forcegchelper() {
309 lockInit(&forcegc.lock, lockRankForcegc)
312 if forcegc.idle.Load() {
313 throw("forcegc: phase error")
315 forcegc.idle.Store(true)
316 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
317 // this goroutine is explicitly resumed by sysmon
318 if debug.gctrace > 0 {
321 // Time-triggered, fully concurrent.
322 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
326 // Gosched yields the processor, allowing other goroutines to run. It does not
327 // suspend the current goroutine, so execution resumes automatically.
335 // goschedguarded yields the processor like gosched, but also checks
336 // for forbidden states and opts out of the yield in those cases.
339 func goschedguarded() {
340 mcall(goschedguarded_m)
343 // goschedIfBusy yields the processor like gosched, but only does so if
344 // there are no idle Ps or if we're on the only P and there's nothing in
345 // the run queue. In both cases, there is freely available idle time.
348 func goschedIfBusy() {
350 // Call gosched if gp.preempt is set; we may be in a tight loop that
351 // doesn't otherwise yield.
352 if !gp.preempt && sched.npidle.Load() > 0 {
358 // Puts the current goroutine into a waiting state and calls unlockf on the
361 // If unlockf returns false, the goroutine is resumed.
363 // unlockf must not access this G's stack, as it may be moved between
364 // the call to gopark and the call to unlockf.
366 // Note that because unlockf is called after putting the G into a waiting
367 // state, the G may have already been readied by the time unlockf is called
368 // unless there is external synchronization preventing the G from being
369 // readied. If unlockf returns false, it must guarantee that the G cannot be
370 // externally readied.
372 // Reason explains why the goroutine has been parked. It is displayed in stack
373 // traces and heap dumps. Reasons should be unique and descriptive. Do not
374 // re-use reasons, add new ones.
375 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
376 if reason != waitReasonSleep {
377 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
381 status := readgstatus(gp)
382 if status != _Grunning && status != _Gscanrunning {
383 throw("gopark: bad g status")
386 mp.waitunlockf = unlockf
387 gp.waitreason = reason
388 mp.waittraceev = traceEv
389 mp.waittraceskip = traceskip
391 // can't do anything that might move the G between Ms here.
395 // Puts the current goroutine into a waiting state and unlocks the lock.
396 // The goroutine can be made runnable again by calling goready(gp).
397 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
398 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
401 func goready(gp *g, traceskip int) {
403 ready(gp, traceskip, true)
408 func acquireSudog() *sudog {
409 // Delicate dance: the semaphore implementation calls
410 // acquireSudog, acquireSudog calls new(sudog),
411 // new calls malloc, malloc can call the garbage collector,
412 // and the garbage collector calls the semaphore implementation
414 // Break the cycle by doing acquirem/releasem around new(sudog).
415 // The acquirem/releasem increments m.locks during new(sudog),
416 // which keeps the garbage collector from being invoked.
419 if len(pp.sudogcache) == 0 {
420 lock(&sched.sudoglock)
421 // First, try to grab a batch from central cache.
422 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
423 s := sched.sudogcache
424 sched.sudogcache = s.next
426 pp.sudogcache = append(pp.sudogcache, s)
428 unlock(&sched.sudoglock)
429 // If the central cache is empty, allocate a new one.
430 if len(pp.sudogcache) == 0 {
431 pp.sudogcache = append(pp.sudogcache, new(sudog))
434 n := len(pp.sudogcache)
435 s := pp.sudogcache[n-1]
436 pp.sudogcache[n-1] = nil
437 pp.sudogcache = pp.sudogcache[:n-1]
439 throw("acquireSudog: found s.elem != nil in cache")
446 func releaseSudog(s *sudog) {
448 throw("runtime: sudog with non-nil elem")
451 throw("runtime: sudog with non-false isSelect")
454 throw("runtime: sudog with non-nil next")
457 throw("runtime: sudog with non-nil prev")
459 if s.waitlink != nil {
460 throw("runtime: sudog with non-nil waitlink")
463 throw("runtime: sudog with non-nil c")
467 throw("runtime: releaseSudog with non-nil gp.param")
469 mp := acquirem() // avoid rescheduling to another P
471 if len(pp.sudogcache) == cap(pp.sudogcache) {
472 // Transfer half of local cache to the central cache.
473 var first, last *sudog
474 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
475 n := len(pp.sudogcache)
476 p := pp.sudogcache[n-1]
477 pp.sudogcache[n-1] = nil
478 pp.sudogcache = pp.sudogcache[:n-1]
486 lock(&sched.sudoglock)
487 last.next = sched.sudogcache
488 sched.sudogcache = first
489 unlock(&sched.sudoglock)
491 pp.sudogcache = append(pp.sudogcache, s)
495 // called from assembly.
496 func badmcall(fn func(*g)) {
497 throw("runtime: mcall called on m->g0 stack")
500 func badmcall2(fn func(*g)) {
501 throw("runtime: mcall function returned")
504 func badreflectcall() {
505 panic(plainError("arg size to reflect.call more than 1GB"))
509 //go:nowritebarrierrec
510 func badmorestackg0() {
511 writeErrStr("fatal: morestack on g0\n")
515 //go:nowritebarrierrec
516 func badmorestackgsignal() {
517 writeErrStr("fatal: morestack on gsignal\n")
525 func lockedOSThread() bool {
527 return gp.lockedm != 0 && gp.m.lockedg != 0
531 // allgs contains all Gs ever created (including dead Gs), and thus
534 // Access via the slice is protected by allglock or stop-the-world.
535 // Readers that cannot take the lock may (carefully!) use the atomic
540 // allglen and allgptr are atomic variables that contain len(allgs) and
541 // &allgs[0] respectively. Proper ordering depends on totally-ordered
542 // loads and stores. Writes are protected by allglock.
544 // allgptr is updated before allglen. Readers should read allglen
545 // before allgptr to ensure that allglen is always <= len(allgptr). New
546 // Gs appended during the race can be missed. For a consistent view of
547 // all Gs, allglock must be held.
549 // allgptr copies should always be stored as a concrete type or
550 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
551 // even if it points to a stale array.
556 func allgadd(gp *g) {
557 if readgstatus(gp) == _Gidle {
558 throw("allgadd: bad status Gidle")
562 allgs = append(allgs, gp)
563 if &allgs[0] != allgptr {
564 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
566 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
570 // allGsSnapshot returns a snapshot of the slice of all Gs.
572 // The world must be stopped or allglock must be held.
573 func allGsSnapshot() []*g {
574 assertWorldStoppedOrLockHeld(&allglock)
576 // Because the world is stopped or allglock is held, allgadd
577 // cannot happen concurrently with this. allgs grows
578 // monotonically and existing entries never change, so we can
579 // simply return a copy of the slice header. For added safety,
580 // we trim everything past len because that can still change.
581 return allgs[:len(allgs):len(allgs)]
584 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
585 func atomicAllG() (**g, uintptr) {
586 length := atomic.Loaduintptr(&allglen)
587 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
591 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
592 func atomicAllGIndex(ptr **g, i uintptr) *g {
593 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
596 // forEachG calls fn on every G from allgs.
598 // forEachG takes a lock to exclude concurrent addition of new Gs.
599 func forEachG(fn func(gp *g)) {
601 for _, gp := range allgs {
607 // forEachGRace calls fn on every G from allgs.
609 // forEachGRace avoids locking, but does not exclude addition of new Gs during
610 // execution, which may be missed.
611 func forEachGRace(fn func(gp *g)) {
612 ptr, length := atomicAllG()
613 for i := uintptr(0); i < length; i++ {
614 gp := atomicAllGIndex(ptr, i)
621 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
622 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
626 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
627 // value of the GODEBUG environment variable.
628 func cpuinit(env string) {
630 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
631 cpu.DebugOptions = true
635 // Support cpu feature variables are used in code generated by the compiler
636 // to guard execution of instructions that can not be assumed to be always supported.
639 x86HasPOPCNT = cpu.X86.HasPOPCNT
640 x86HasSSE41 = cpu.X86.HasSSE41
641 x86HasFMA = cpu.X86.HasFMA
644 armHasVFPv4 = cpu.ARM.HasVFPv4
647 arm64HasATOMICS = cpu.ARM64.HasATOMICS
651 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
652 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
653 // early before much of the runtime is initialized.
654 func getGodebugEarly() string {
655 const prefix = "GODEBUG="
658 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
659 // Similar to goenv_unix but extracts the environment value for
661 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
663 for argv_index(argv, argc+1+n) != nil {
667 for i := int32(0); i < n; i++ {
668 p := argv_index(argv, argc+1+i)
669 s := unsafe.String(p, findnull(p))
671 if hasPrefix(s, prefix) {
672 env = gostring(p)[len(prefix):]
680 // The bootstrap sequence is:
684 // make & queue new G
685 // call runtime·mstart
687 // The new G calls runtime·main.
689 lockInit(&sched.lock, lockRankSched)
690 lockInit(&sched.sysmonlock, lockRankSysmon)
691 lockInit(&sched.deferlock, lockRankDefer)
692 lockInit(&sched.sudoglock, lockRankSudog)
693 lockInit(&deadlock, lockRankDeadlock)
694 lockInit(&paniclk, lockRankPanic)
695 lockInit(&allglock, lockRankAllg)
696 lockInit(&allpLock, lockRankAllp)
697 lockInit(&reflectOffs.lock, lockRankReflectOffs)
698 lockInit(&finlock, lockRankFin)
699 lockInit(&trace.bufLock, lockRankTraceBuf)
700 lockInit(&trace.stringsLock, lockRankTraceStrings)
701 lockInit(&trace.lock, lockRankTrace)
702 lockInit(&cpuprof.lock, lockRankCpuprof)
703 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
704 // Enforce that this lock is always a leaf lock.
705 // All of this lock's critical sections should be
707 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
709 // raceinit must be the first call to race detector.
710 // In particular, it must be done before mallocinit below calls racemapshadow.
713 gp.racectx, raceprocctx0 = raceinit()
716 sched.maxmcount = 10000
718 // The world starts stopped.
724 godebug := getGodebugEarly()
725 initPageTrace(godebug) // must run after mallocinit but before anything allocates
726 cpuinit(godebug) // must run before alginit
727 alginit() // maps, hash, fastrand must not be used before this call
728 fastrandinit() // must run before mcommoninit
729 mcommoninit(gp.m, -1)
730 modulesinit() // provides activeModules
731 typelinksinit() // uses maps, activeModules
732 itabsinit() // uses activeModules
733 stkobjinit() // must run before GC starts
735 sigsave(&gp.m.sigmask)
736 initSigmask = gp.m.sigmask
743 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
744 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
745 // set to true by the linker, it means that nothing is consuming the profile, it is
746 // safe to set MemProfileRate to 0.
747 if disableMemoryProfiling {
752 sched.lastpoll.Store(nanotime())
754 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
757 if procresize(procs) != nil {
758 throw("unknown runnable goroutine during bootstrap")
762 // World is effectively started now, as P's can run.
765 if buildVersion == "" {
766 // Condition should never trigger. This code just serves
767 // to ensure runtime·buildVersion is kept in the resulting binary.
768 buildVersion = "unknown"
770 if len(modinfo) == 1 {
771 // Condition should never trigger. This code just serves
772 // to ensure runtime·modinfo is kept in the resulting binary.
777 func dumpgstatus(gp *g) {
779 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
780 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
783 // sched.lock must be held.
785 assertLockHeld(&sched.lock)
787 if mcount() > sched.maxmcount {
788 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
789 throw("thread exhaustion")
793 // mReserveID returns the next ID to use for a new m. This new m is immediately
794 // considered 'running' by checkdead.
796 // sched.lock must be held.
797 func mReserveID() int64 {
798 assertLockHeld(&sched.lock)
800 if sched.mnext+1 < sched.mnext {
801 throw("runtime: thread ID overflow")
809 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
810 func mcommoninit(mp *m, id int64) {
813 // g0 stack won't make sense for user (and is not necessary unwindable).
815 callers(1, mp.createstack[:])
826 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
827 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
831 // Same behavior as for 1.17.
832 // TODO: Simplify this.
833 if goarch.BigEndian {
834 mp.fastrand = uint64(lo)<<32 | uint64(hi)
836 mp.fastrand = uint64(hi)<<32 | uint64(lo)
840 if mp.gsignal != nil {
841 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
844 // Add to allm so garbage collector doesn't free g->m
845 // when it is just in a register or thread-local storage.
848 // NumCgoCall() iterates over allm w/o schedlock,
849 // so we need to publish it safely.
850 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
853 // Allocate memory to hold a cgo traceback if the cgo call crashes.
854 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
855 mp.cgoCallers = new(cgoCallers)
859 func (mp *m) becomeSpinning() {
861 sched.nmspinning.Add(1)
862 sched.needspinning.Store(0)
865 func (mp *m) incgocallback() bool {
866 return (!mp.incgo && mp.ncgo > 0) || mp.isextra
869 var fastrandseed uintptr
871 func fastrandinit() {
872 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
876 // Mark gp ready to run.
877 func ready(gp *g, traceskip int, next bool) {
879 traceGoUnpark(gp, traceskip)
882 status := readgstatus(gp)
885 mp := acquirem() // disable preemption because it can be holding p in a local var
886 if status&^_Gscan != _Gwaiting {
888 throw("bad g->status in ready")
891 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
892 casgstatus(gp, _Gwaiting, _Grunnable)
893 runqput(mp.p.ptr(), gp, next)
898 // freezeStopWait is a large value that freezetheworld sets
899 // sched.stopwait to in order to request that all Gs permanently stop.
900 const freezeStopWait = 0x7fffffff
902 // freezing is set to non-zero if the runtime is trying to freeze the
904 var freezing atomic.Bool
906 // Similar to stopTheWorld but best-effort and can be called several times.
907 // There is no reverse operation, used during crashing.
908 // This function must not lock any mutexes.
909 func freezetheworld() {
911 // stopwait and preemption requests can be lost
912 // due to races with concurrently executing threads,
913 // so try several times
914 for i := 0; i < 5; i++ {
915 // this should tell the scheduler to not start any new goroutines
916 sched.stopwait = freezeStopWait
917 sched.gcwaiting.Store(true)
918 // this should stop running goroutines
920 break // no running goroutines
930 // All reads and writes of g's status go through readgstatus, casgstatus
931 // castogscanstatus, casfrom_Gscanstatus.
934 func readgstatus(gp *g) uint32 {
935 return gp.atomicstatus.Load()
938 // The Gscanstatuses are acting like locks and this releases them.
939 // If it proves to be a performance hit we should be able to make these
940 // simple atomic stores but for now we are going to throw if
941 // we see an inconsistent state.
942 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
945 // Check that transition is valid.
948 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
950 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
956 if newval == oldval&^_Gscan {
957 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
961 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
963 throw("casfrom_Gscanstatus: gp->status is not in scan state")
965 releaseLockRank(lockRankGscan)
968 // This will return false if the gp is not in the expected status and the cas fails.
969 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
970 func castogscanstatus(gp *g, oldval, newval uint32) bool {
976 if newval == oldval|_Gscan {
977 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
979 acquireLockRank(lockRankGscan)
985 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
986 throw("castogscanstatus")
990 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
991 // various latencies on every transition instead of sampling them.
992 var casgstatusAlwaysTrack = false
994 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
995 // and casfrom_Gscanstatus instead.
996 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
997 // put it in the Gscan state is finished.
1000 func casgstatus(gp *g, oldval, newval uint32) {
1001 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
1002 systemstack(func() {
1003 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
1004 throw("casgstatus: bad incoming values")
1008 acquireLockRank(lockRankGscan)
1009 releaseLockRank(lockRankGscan)
1011 // See https://golang.org/cl/21503 for justification of the yield delay.
1012 const yieldDelay = 5 * 1000
1015 // loop if gp->atomicstatus is in a scan state giving
1016 // GC time to finish and change the state to oldval.
1017 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1018 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1019 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1022 nextYield = nanotime() + yieldDelay
1024 if nanotime() < nextYield {
1025 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1030 nextYield = nanotime() + yieldDelay/2
1034 if oldval == _Grunning {
1035 // Track every gTrackingPeriod time a goroutine transitions out of running.
1036 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1045 // Handle various kinds of tracking.
1048 // - Time spent in runnable.
1049 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1052 // We transitioned out of runnable, so measure how much
1053 // time we spent in this state and add it to
1056 gp.runnableTime += now - gp.trackingStamp
1057 gp.trackingStamp = 0
1059 if !gp.waitreason.isMutexWait() {
1060 // Not blocking on a lock.
1063 // Blocking on a lock, measure it. Note that because we're
1064 // sampling, we have to multiply by our sampling period to get
1065 // a more representative estimate of the absolute value.
1066 // gTrackingPeriod also represents an accurate sampling period
1067 // because we can only enter this state from _Grunning.
1069 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1070 gp.trackingStamp = 0
1074 if !gp.waitreason.isMutexWait() {
1075 // Not blocking on a lock.
1078 // Blocking on a lock. Write down the timestamp.
1080 gp.trackingStamp = now
1082 // We just transitioned into runnable, so record what
1083 // time that happened.
1085 gp.trackingStamp = now
1087 // We're transitioning into running, so turn off
1088 // tracking and record how much time we spent in
1091 sched.timeToRun.record(gp.runnableTime)
1096 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1098 // Use this over casgstatus when possible to ensure that a waitreason is set.
1099 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1100 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1101 gp.waitreason = reason
1102 casgstatus(gp, old, _Gwaiting)
1105 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1106 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1107 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1108 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1109 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1112 func casgcopystack(gp *g) uint32 {
1114 oldstatus := readgstatus(gp) &^ _Gscan
1115 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1116 throw("copystack: bad status, not Gwaiting or Grunnable")
1118 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1124 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1126 // TODO(austin): This is the only status operation that both changes
1127 // the status and locks the _Gscan bit. Rethink this.
1128 func casGToPreemptScan(gp *g, old, new uint32) {
1129 if old != _Grunning || new != _Gscan|_Gpreempted {
1130 throw("bad g transition")
1132 acquireLockRank(lockRankGscan)
1133 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1137 // casGFromPreempted attempts to transition gp from _Gpreempted to
1138 // _Gwaiting. If successful, the caller is responsible for
1139 // re-scheduling gp.
1140 func casGFromPreempted(gp *g, old, new uint32) bool {
1141 if old != _Gpreempted || new != _Gwaiting {
1142 throw("bad g transition")
1144 gp.waitreason = waitReasonPreempted
1145 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1148 // stopTheWorld stops all P's from executing goroutines, interrupting
1149 // all goroutines at GC safe points and records reason as the reason
1150 // for the stop. On return, only the current goroutine's P is running.
1151 // stopTheWorld must not be called from a system stack and the caller
1152 // must not hold worldsema. The caller must call startTheWorld when
1153 // other P's should resume execution.
1155 // stopTheWorld is safe for multiple goroutines to call at the
1156 // same time. Each will execute its own stop, and the stops will
1159 // This is also used by routines that do stack dumps. If the system is
1160 // in panic or being exited, this may not reliably stop all
1162 func stopTheWorld(reason string) {
1163 semacquire(&worldsema)
1165 gp.m.preemptoff = reason
1166 systemstack(func() {
1167 // Mark the goroutine which called stopTheWorld preemptible so its
1168 // stack may be scanned.
1169 // This lets a mark worker scan us while we try to stop the world
1170 // since otherwise we could get in a mutual preemption deadlock.
1171 // We must not modify anything on the G stack because a stack shrink
1172 // may occur. A stack shrink is otherwise OK though because in order
1173 // to return from this function (and to leave the system stack) we
1174 // must have preempted all goroutines, including any attempting
1175 // to scan our stack, in which case, any stack shrinking will
1176 // have already completed by the time we exit.
1177 // Don't provide a wait reason because we're still executing.
1178 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1179 stopTheWorldWithSema()
1180 casgstatus(gp, _Gwaiting, _Grunning)
1184 // startTheWorld undoes the effects of stopTheWorld.
1185 func startTheWorld() {
1186 systemstack(func() { startTheWorldWithSema(false) })
1188 // worldsema must be held over startTheWorldWithSema to ensure
1189 // gomaxprocs cannot change while worldsema is held.
1191 // Release worldsema with direct handoff to the next waiter, but
1192 // acquirem so that semrelease1 doesn't try to yield our time.
1194 // Otherwise if e.g. ReadMemStats is being called in a loop,
1195 // it might stomp on other attempts to stop the world, such as
1196 // for starting or ending GC. The operation this blocks is
1197 // so heavy-weight that we should just try to be as fair as
1200 // We don't want to just allow us to get preempted between now
1201 // and releasing the semaphore because then we keep everyone
1202 // (including, for example, GCs) waiting longer.
1205 semrelease1(&worldsema, true, 0)
1209 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1210 // until the GC is not running. It also blocks a GC from starting
1211 // until startTheWorldGC is called.
1212 func stopTheWorldGC(reason string) {
1214 stopTheWorld(reason)
1217 // startTheWorldGC undoes the effects of stopTheWorldGC.
1218 func startTheWorldGC() {
1223 // Holding worldsema grants an M the right to try to stop the world.
1224 var worldsema uint32 = 1
1226 // Holding gcsema grants the M the right to block a GC, and blocks
1227 // until the current GC is done. In particular, it prevents gomaxprocs
1228 // from changing concurrently.
1230 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1231 // being changed/enabled during a GC, remove this.
1232 var gcsema uint32 = 1
1234 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1235 // The caller is responsible for acquiring worldsema and disabling
1236 // preemption first and then should stopTheWorldWithSema on the system
1239 // semacquire(&worldsema, 0)
1240 // m.preemptoff = "reason"
1241 // systemstack(stopTheWorldWithSema)
1243 // When finished, the caller must either call startTheWorld or undo
1244 // these three operations separately:
1246 // m.preemptoff = ""
1247 // systemstack(startTheWorldWithSema)
1248 // semrelease(&worldsema)
1250 // It is allowed to acquire worldsema once and then execute multiple
1251 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1252 // Other P's are able to execute between successive calls to
1253 // startTheWorldWithSema and stopTheWorldWithSema.
1254 // Holding worldsema causes any other goroutines invoking
1255 // stopTheWorld to block.
1256 func stopTheWorldWithSema() {
1259 // If we hold a lock, then we won't be able to stop another M
1260 // that is blocked trying to acquire the lock.
1262 throw("stopTheWorld: holding locks")
1266 sched.stopwait = gomaxprocs
1267 sched.gcwaiting.Store(true)
1270 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1272 // try to retake all P's in Psyscall status
1273 for _, pp := range allp {
1275 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1287 pp, _ := pidleget(now)
1291 pp.status = _Pgcstop
1294 wait := sched.stopwait > 0
1297 // wait for remaining P's to stop voluntarily
1300 // wait for 100us, then try to re-preempt in case of any races
1301 if notetsleep(&sched.stopnote, 100*1000) {
1302 noteclear(&sched.stopnote)
1311 if sched.stopwait != 0 {
1312 bad = "stopTheWorld: not stopped (stopwait != 0)"
1314 for _, pp := range allp {
1315 if pp.status != _Pgcstop {
1316 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1320 if freezing.Load() {
1321 // Some other thread is panicking. This can cause the
1322 // sanity checks above to fail if the panic happens in
1323 // the signal handler on a stopped thread. Either way,
1324 // we should halt this thread.
1335 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1336 assertWorldStopped()
1338 mp := acquirem() // disable preemption because it can be holding p in a local var
1339 if netpollinited() {
1340 list := netpoll(0) // non-blocking
1350 p1 := procresize(procs)
1351 sched.gcwaiting.Store(false)
1352 if sched.sysmonwait.Load() {
1353 sched.sysmonwait.Store(false)
1354 notewakeup(&sched.sysmonnote)
1367 throw("startTheWorld: inconsistent mp->nextp")
1370 notewakeup(&mp.park)
1372 // Start M to run P. Do not start another M below.
1377 // Capture start-the-world time before doing clean-up tasks.
1378 startTime := nanotime()
1383 // Wakeup an additional proc in case we have excessive runnable goroutines
1384 // in local queues or in the global queue. If we don't, the proc will park itself.
1385 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1393 // usesLibcall indicates whether this runtime performs system calls
1395 func usesLibcall() bool {
1397 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1400 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1405 // mStackIsSystemAllocated indicates whether this runtime starts on a
1406 // system-allocated stack.
1407 func mStackIsSystemAllocated() bool {
1409 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1413 case "386", "amd64", "arm", "arm64":
1420 // mstart is the entry-point for new Ms.
1421 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1424 // mstart0 is the Go entry-point for new Ms.
1425 // This must not split the stack because we may not even have stack
1426 // bounds set up yet.
1428 // May run during STW (because it doesn't have a P yet), so write
1429 // barriers are not allowed.
1432 //go:nowritebarrierrec
1436 osStack := gp.stack.lo == 0
1438 // Initialize stack bounds from system stack.
1439 // Cgo may have left stack size in stack.hi.
1440 // minit may update the stack bounds.
1442 // Note: these bounds may not be very accurate.
1443 // We set hi to &size, but there are things above
1444 // it. The 1024 is supposed to compensate this,
1445 // but is somewhat arbitrary.
1448 size = 8192 * sys.StackGuardMultiplier
1450 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1451 gp.stack.lo = gp.stack.hi - size + 1024
1453 // Initialize stack guard so that we can start calling regular
1455 gp.stackguard0 = gp.stack.lo + _StackGuard
1456 // This is the g0, so we can also call go:systemstack
1457 // functions, which check stackguard1.
1458 gp.stackguard1 = gp.stackguard0
1461 // Exit this thread.
1462 if mStackIsSystemAllocated() {
1463 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1464 // the stack, but put it in gp.stack before mstart,
1465 // so the logic above hasn't set osStack yet.
1471 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1472 // so that we can set up g0.sched to return to the call of mstart1 above.
1479 throw("bad runtime·mstart")
1482 // Set up m.g0.sched as a label returning to just
1483 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1484 // We're never coming back to mstart1 after we call schedule,
1485 // so other calls can reuse the current frame.
1486 // And goexit0 does a gogo that needs to return from mstart1
1487 // and let mstart0 exit the thread.
1488 gp.sched.g = guintptr(unsafe.Pointer(gp))
1489 gp.sched.pc = getcallerpc()
1490 gp.sched.sp = getcallersp()
1495 // Install signal handlers; after minit so that minit can
1496 // prepare the thread to be able to handle the signals.
1501 if fn := gp.m.mstartfn; fn != nil {
1506 acquirep(gp.m.nextp.ptr())
1512 // mstartm0 implements part of mstart1 that only runs on the m0.
1514 // Write barriers are allowed here because we know the GC can't be
1515 // running yet, so they'll be no-ops.
1517 //go:yeswritebarrierrec
1519 // Create an extra M for callbacks on threads not created by Go.
1520 // An extra M is also needed on Windows for callbacks created by
1521 // syscall.NewCallback. See issue #6751 for details.
1522 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1529 // mPark causes a thread to park itself, returning once woken.
1534 notesleep(&gp.m.park)
1535 noteclear(&gp.m.park)
1538 // mexit tears down and exits the current thread.
1540 // Don't call this directly to exit the thread, since it must run at
1541 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1542 // unwind the stack to the point that exits the thread.
1544 // It is entered with m.p != nil, so write barriers are allowed. It
1545 // will release the P before exiting.
1547 //go:yeswritebarrierrec
1548 func mexit(osStack bool) {
1552 // This is the main thread. Just wedge it.
1554 // On Linux, exiting the main thread puts the process
1555 // into a non-waitable zombie state. On Plan 9,
1556 // exiting the main thread unblocks wait even though
1557 // other threads are still running. On Solaris we can
1558 // neither exitThread nor return from mstart. Other
1559 // bad things probably happen on other platforms.
1561 // We could try to clean up this M more before wedging
1562 // it, but that complicates signal handling.
1563 handoffp(releasep())
1569 throw("locked m0 woke up")
1575 // Free the gsignal stack.
1576 if mp.gsignal != nil {
1577 stackfree(mp.gsignal.stack)
1578 // On some platforms, when calling into VDSO (e.g. nanotime)
1579 // we store our g on the gsignal stack, if there is one.
1580 // Now the stack is freed, unlink it from the m, so we
1581 // won't write to it when calling VDSO code.
1585 // Remove m from allm.
1587 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1593 throw("m not found in allm")
1595 // Delay reaping m until it's done with the stack.
1597 // Put mp on the free list, though it will not be reaped while freeWait
1598 // is freeMWait. mp is no longer reachable via allm, so even if it is
1599 // on an OS stack, we must keep a reference to mp alive so that the GC
1600 // doesn't free mp while we are still using it.
1602 // Note that the free list must not be linked through alllink because
1603 // some functions walk allm without locking, so may be using alllink.
1604 mp.freeWait.Store(freeMWait)
1605 mp.freelink = sched.freem
1609 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1612 handoffp(releasep())
1613 // After this point we must not have write barriers.
1615 // Invoke the deadlock detector. This must happen after
1616 // handoffp because it may have started a new M to take our
1623 if GOOS == "darwin" || GOOS == "ios" {
1624 // Make sure pendingPreemptSignals is correct when an M exits.
1626 if mp.signalPending.Load() != 0 {
1627 pendingPreemptSignals.Add(-1)
1631 // Destroy all allocated resources. After this is called, we may no
1632 // longer take any locks.
1636 // No more uses of mp, so it is safe to drop the reference.
1637 mp.freeWait.Store(freeMRef)
1639 // Return from mstart and let the system thread
1640 // library free the g0 stack and terminate the thread.
1644 // mstart is the thread's entry point, so there's nothing to
1645 // return to. Exit the thread directly. exitThread will clear
1646 // m.freeWait when it's done with the stack and the m can be
1648 exitThread(&mp.freeWait)
1651 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1652 // If a P is currently executing code, this will bring the P to a GC
1653 // safe point and execute fn on that P. If the P is not executing code
1654 // (it is idle or in a syscall), this will call fn(p) directly while
1655 // preventing the P from exiting its state. This does not ensure that
1656 // fn will run on every CPU executing Go code, but it acts as a global
1657 // memory barrier. GC uses this as a "ragged barrier."
1659 // The caller must hold worldsema.
1662 func forEachP(fn func(*p)) {
1664 pp := getg().m.p.ptr()
1667 if sched.safePointWait != 0 {
1668 throw("forEachP: sched.safePointWait != 0")
1670 sched.safePointWait = gomaxprocs - 1
1671 sched.safePointFn = fn
1673 // Ask all Ps to run the safe point function.
1674 for _, p2 := range allp {
1676 atomic.Store(&p2.runSafePointFn, 1)
1681 // Any P entering _Pidle or _Psyscall from now on will observe
1682 // p.runSafePointFn == 1 and will call runSafePointFn when
1683 // changing its status to _Pidle/_Psyscall.
1685 // Run safe point function for all idle Ps. sched.pidle will
1686 // not change because we hold sched.lock.
1687 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1688 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1690 sched.safePointWait--
1694 wait := sched.safePointWait > 0
1697 // Run fn for the current P.
1700 // Force Ps currently in _Psyscall into _Pidle and hand them
1701 // off to induce safe point function execution.
1702 for _, p2 := range allp {
1704 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1714 // Wait for remaining Ps to run fn.
1717 // Wait for 100us, then try to re-preempt in
1718 // case of any races.
1720 // Requires system stack.
1721 if notetsleep(&sched.safePointNote, 100*1000) {
1722 noteclear(&sched.safePointNote)
1728 if sched.safePointWait != 0 {
1729 throw("forEachP: not done")
1731 for _, p2 := range allp {
1732 if p2.runSafePointFn != 0 {
1733 throw("forEachP: P did not run fn")
1738 sched.safePointFn = nil
1743 // runSafePointFn runs the safe point function, if any, for this P.
1744 // This should be called like
1746 // if getg().m.p.runSafePointFn != 0 {
1750 // runSafePointFn must be checked on any transition in to _Pidle or
1751 // _Psyscall to avoid a race where forEachP sees that the P is running
1752 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1753 // nor the P run the safe-point function.
1754 func runSafePointFn() {
1755 p := getg().m.p.ptr()
1756 // Resolve the race between forEachP running the safe-point
1757 // function on this P's behalf and this P running the
1758 // safe-point function directly.
1759 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1762 sched.safePointFn(p)
1764 sched.safePointWait--
1765 if sched.safePointWait == 0 {
1766 notewakeup(&sched.safePointNote)
1771 // When running with cgo, we call _cgo_thread_start
1772 // to start threads for us so that we can play nicely with
1774 var cgoThreadStart unsafe.Pointer
1776 type cgothreadstart struct {
1782 // Allocate a new m unassociated with any thread.
1783 // Can use p for allocation context if needed.
1784 // fn is recorded as the new m's m.mstartfn.
1785 // id is optional pre-allocated m ID. Omit by passing -1.
1787 // This function is allowed to have write barriers even if the caller
1788 // isn't because it borrows pp.
1790 //go:yeswritebarrierrec
1791 func allocm(pp *p, fn func(), id int64) *m {
1794 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1795 // disable preemption to ensure it is not stolen, which would make the
1796 // caller lose ownership.
1801 acquirep(pp) // temporarily borrow p for mallocs in this function
1804 // Release the free M list. We need to do this somewhere and
1805 // this may free up a stack we can use.
1806 if sched.freem != nil {
1809 for freem := sched.freem; freem != nil; {
1810 wait := freem.freeWait.Load()
1811 if wait == freeMWait {
1812 next := freem.freelink
1813 freem.freelink = newList
1818 // Free the stack if needed. For freeMRef, there is
1819 // nothing to do except drop freem from the sched.freem
1821 if wait == freeMStack {
1822 // stackfree must be on the system stack, but allocm is
1823 // reachable off the system stack transitively from
1825 systemstack(func() {
1826 stackfree(freem.g0.stack)
1829 freem = freem.freelink
1831 sched.freem = newList
1839 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1840 // Windows and Plan 9 will layout sched stack on OS stack.
1841 if iscgo || mStackIsSystemAllocated() {
1844 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1848 if pp == gp.m.p.ptr() {
1853 allocmLock.runlock()
1857 // needm is called when a cgo callback happens on a
1858 // thread without an m (a thread not created by Go).
1859 // In this case, needm is expected to find an m to use
1860 // and return with m, g initialized correctly.
1861 // Since m and g are not set now (likely nil, but see below)
1862 // needm is limited in what routines it can call. In particular
1863 // it can only call nosplit functions (textflag 7) and cannot
1864 // do any scheduling that requires an m.
1866 // In order to avoid needing heavy lifting here, we adopt
1867 // the following strategy: there is a stack of available m's
1868 // that can be stolen. Using compare-and-swap
1869 // to pop from the stack has ABA races, so we simulate
1870 // a lock by doing an exchange (via Casuintptr) to steal the stack
1871 // head and replace the top pointer with MLOCKED (1).
1872 // This serves as a simple spin lock that we can use even
1873 // without an m. The thread that locks the stack in this way
1874 // unlocks the stack by storing a valid stack head pointer.
1876 // In order to make sure that there is always an m structure
1877 // available to be stolen, we maintain the invariant that there
1878 // is always one more than needed. At the beginning of the
1879 // program (if cgo is in use) the list is seeded with a single m.
1880 // If needm finds that it has taken the last m off the list, its job
1881 // is - once it has installed its own m so that it can do things like
1882 // allocate memory - to create a spare m and put it on the list.
1884 // Each of these extra m's also has a g0 and a curg that are
1885 // pressed into service as the scheduling stack and current
1886 // goroutine for the duration of the cgo callback.
1888 // It calls dropm to put the m back on the list,
1889 // 1. when the callback is done with the m in non-pthread platforms,
1890 // 2. or when the C thread exiting on pthread platforms.
1892 // The signal argument indicates whether we're called from a signal
1896 func needm(signal bool) {
1897 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1898 // Can happen if C/C++ code calls Go from a global ctor.
1899 // Can also happen on Windows if a global ctor uses a
1900 // callback created by syscall.NewCallback. See issue #6751
1903 // Can not throw, because scheduler is not initialized yet.
1904 writeErrStr("fatal error: cgo callback before cgo call\n")
1908 // Save and block signals before getting an M.
1909 // The signal handler may call needm itself,
1910 // and we must avoid a deadlock. Also, once g is installed,
1911 // any incoming signals will try to execute,
1912 // but we won't have the sigaltstack settings and other data
1913 // set up appropriately until the end of minit, which will
1914 // unblock the signals. This is the same dance as when
1915 // starting a new m to run Go code via newosproc.
1920 // Lock extra list, take head, unlock popped list.
1921 // nilokay=false is safe here because of the invariant above,
1922 // that the extra list always contains or will soon contain
1924 mp := lockextra(false)
1926 // Set needextram when we've just emptied the list,
1927 // so that the eventual call into cgocallbackg will
1928 // allocate a new m for the extra list. We delay the
1929 // allocation until then so that it can be done
1930 // after exitsyscall makes sure it is okay to be
1931 // running at all (that is, there's no garbage collection
1932 // running right now).
1933 mp.needextram = mp.schedlink == 0
1935 unlockextra(mp.schedlink.ptr())
1937 // Store the original signal mask for use by minit.
1938 mp.sigmask = sigmask
1940 // Install TLS on some platforms (previously setg
1941 // would do this if necessary).
1944 // Install g (= m->g0) and set the stack bounds
1945 // to match the current stack. If we don't actually know
1946 // how big the stack is, like we don't know how big any
1947 // scheduling stack is, but we assume there's at least 32 kB.
1948 // If we can get a more accurate stack bound from pthread,
1952 gp.stack.hi = getcallersp() + 1024
1953 gp.stack.lo = getcallersp() - 32*1024
1954 if !signal && _cgo_getstackbound != nil {
1955 // Don't adjust if called from the signal handler.
1956 // We are on the signal stack, not the pthread stack.
1957 // (We could get the stack bounds from sigaltstack, but
1958 // we're getting out of the signal handler very soon
1959 // anyway. Not worth it.)
1960 asmcgocall(_cgo_getstackbound, unsafe.Pointer(gp))
1962 gp.stackguard0 = gp.stack.lo + _StackGuard
1964 // Should mark we are already in Go now.
1965 // Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
1966 // which means the extram list may be empty, that will cause a deadlock.
1967 mp.isExtraInC = false
1969 // Initialize this thread to use the m.
1973 // mp.curg is now a real goroutine.
1974 casgstatus(mp.curg, _Gdead, _Gsyscall)
1978 // Acquire an extra m and bind it to the C thread when a pthread key has been created.
1981 func needAndBindM() {
1984 if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
1989 // newextram allocates m's and puts them on the extra list.
1990 // It is called with a working local m, so that it can do things
1991 // like call schedlock and allocate.
1993 c := extraMWaiters.Swap(0)
1995 for i := uint32(0); i < c; i++ {
1999 // Make sure there is at least one extra M.
2000 mp := lockextra(true)
2008 // oneNewExtraM allocates an m and puts it on the extra list.
2009 func oneNewExtraM() {
2010 // Create extra goroutine locked to extra m.
2011 // The goroutine is the context in which the cgo callback will run.
2012 // The sched.pc will never be returned to, but setting it to
2013 // goexit makes clear to the traceback routines where
2014 // the goroutine stack ends.
2015 mp := allocm(nil, nil, -1)
2017 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
2018 gp.sched.sp = gp.stack.hi
2019 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
2021 gp.sched.g = guintptr(unsafe.Pointer(gp))
2022 gp.syscallpc = gp.sched.pc
2023 gp.syscallsp = gp.sched.sp
2024 gp.stktopsp = gp.sched.sp
2025 // malg returns status as _Gidle. Change to _Gdead before
2026 // adding to allg where GC can see it. We use _Gdead to hide
2027 // this from tracebacks and stack scans since it isn't a
2028 // "real" goroutine until needm grabs it.
2029 casgstatus(gp, _Gidle, _Gdead)
2033 // mark we are in C by default.
2034 mp.isExtraInC = true
2038 gp.goid = sched.goidgen.Add(1)
2039 gp.sysblocktraced = true
2041 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
2044 // Trigger two trace events for the locked g in the extra m,
2045 // since the next event of the g will be traceEvGoSysExit in exitsyscall,
2046 // while calling from C thread to Go.
2047 traceGoCreate(gp, 0) // no start pc
2049 traceEvent(traceEvGoInSyscall, -1, gp.goid)
2051 // put on allg for garbage collector
2054 // gp is now on the allg list, but we don't want it to be
2055 // counted by gcount. It would be more "proper" to increment
2056 // sched.ngfree, but that requires locking. Incrementing ngsys
2057 // has the same effect.
2060 // Add m to the extra list.
2061 mnext := lockextra(true)
2062 mp.schedlink.set(mnext)
2067 // dropm puts the current m back onto the extra list.
2069 // 1. On systems without pthreads, like Windows
2070 // dropm is called when a cgo callback has called needm but is now
2071 // done with the callback and returning back into the non-Go thread.
2073 // The main expense here is the call to signalstack to release the
2074 // m's signal stack, and then the call to needm on the next callback
2075 // from this thread. It is tempting to try to save the m for next time,
2076 // which would eliminate both these costs, but there might not be
2077 // a next time: the current thread (which Go does not control) might exit.
2078 // If we saved the m for that thread, there would be an m leak each time
2079 // such a thread exited. Instead, we acquire and release an m on each
2080 // call. These should typically not be scheduling operations, just a few
2081 // atomics, so the cost should be small.
2083 // 2. On systems with pthreads
2084 // dropm is called while a non-Go thread is exiting.
2085 // We allocate a pthread per-thread variable using pthread_key_create,
2086 // to register a thread-exit-time destructor.
2087 // And store the g into a thread-specific value associated with the pthread key,
2088 // when first return back to C.
2089 // So that the destructor would invoke dropm while the non-Go thread is exiting.
2090 // This is much faster since it avoids expensive signal-related syscalls.
2092 // NOTE: this always runs without a P, so, nowritebarrierrec required.
2094 //go:nowritebarrierrec
2096 // Clear m and g, and return m to the extra list.
2097 // After the call to setg we can only call nosplit functions
2098 // with no pointer manipulation.
2101 // Return mp.curg to dead state.
2102 casgstatus(mp.curg, _Gsyscall, _Gdead)
2103 mp.curg.preemptStop = false
2106 // Block signals before unminit.
2107 // Unminit unregisters the signal handling stack (but needs g on some systems).
2108 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2109 // It's important not to try to handle a signal between those two steps.
2110 sigmask := mp.sigmask
2114 mnext := lockextra(true)
2116 mp.schedlink.set(mnext)
2120 // Commit the release of mp.
2123 msigrestore(sigmask)
2126 // bindm store the g0 of the current m into a thread-specific value.
2128 // We allocate a pthread per-thread variable using pthread_key_create,
2129 // to register a thread-exit-time destructor.
2130 // We are here setting the thread-specific value of the pthread key, to enable the destructor.
2131 // So that the pthread_key_destructor would dropm while the C thread is exiting.
2133 // And the saved g will be used in pthread_key_destructor,
2134 // since the g stored in the TLS by Go might be cleared in some platforms,
2135 // before the destructor invoked, so, we restore g by the stored g, before dropm.
2137 // We store g0 instead of m, to make the assembly code simpler,
2138 // since we need to restore g0 in runtime.cgocallback.
2140 // On systems without pthreads, like Windows, bindm shouldn't be used.
2142 // NOTE: this always runs without a P, so, nowritebarrierrec required.
2145 //go:nowritebarrierrec
2147 if GOOS == "windows" || GOOS == "plan9" {
2148 fatal("bindm in unexpected GOOS")
2152 fatal("the current g is not g0")
2154 if _cgo_bindm != nil {
2155 asmcgocall(_cgo_bindm, unsafe.Pointer(g))
2159 // A helper function for EnsureDropM.
2160 func getm() uintptr {
2161 return uintptr(unsafe.Pointer(getg().m))
2164 var extram atomic.Uintptr
2165 var extraMCount uint32 // Protected by lockextra
2166 var extraMWaiters atomic.Uint32
2168 // lockextra locks the extra list and returns the list head.
2169 // The caller must unlock the list by storing a new list head
2170 // to extram. If nilokay is true, then lockextra will
2171 // return a nil list head if that's what it finds. If nilokay is false,
2172 // lockextra will keep waiting until the list head is no longer nil.
2175 func lockextra(nilokay bool) *m {
2180 old := extram.Load()
2185 if old == 0 && !nilokay {
2187 // Add 1 to the number of threads
2188 // waiting for an M.
2189 // This is cleared by newextram.
2190 extraMWaiters.Add(1)
2196 if extram.CompareAndSwap(old, locked) {
2197 return (*m)(unsafe.Pointer(old))
2205 func unlockextra(mp *m) {
2206 extram.Store(uintptr(unsafe.Pointer(mp)))
2210 // allocmLock is locked for read when creating new Ms in allocm and their
2211 // addition to allm. Thus acquiring this lock for write blocks the
2212 // creation of new Ms.
2215 // execLock serializes exec and clone to avoid bugs or unspecified
2216 // behaviour around exec'ing while creating/destroying threads. See
2221 // These errors are reported (via writeErrStr) by some OS-specific
2222 // versions of newosproc and newosproc0.
2224 failthreadcreate = "runtime: failed to create new OS thread\n"
2225 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2228 // newmHandoff contains a list of m structures that need new OS threads.
2229 // This is used by newm in situations where newm itself can't safely
2230 // start an OS thread.
2231 var newmHandoff struct {
2234 // newm points to a list of M structures that need new OS
2235 // threads. The list is linked through m.schedlink.
2238 // waiting indicates that wake needs to be notified when an m
2239 // is put on the list.
2243 // haveTemplateThread indicates that the templateThread has
2244 // been started. This is not protected by lock. Use cas to set
2246 haveTemplateThread uint32
2249 // Create a new m. It will start off with a call to fn, or else the scheduler.
2250 // fn needs to be static and not a heap allocated closure.
2251 // May run with m.p==nil, so write barriers are not allowed.
2253 // id is optional pre-allocated m ID. Omit by passing -1.
2255 //go:nowritebarrierrec
2256 func newm(fn func(), pp *p, id int64) {
2257 // allocm adds a new M to allm, but they do not start until created by
2258 // the OS in newm1 or the template thread.
2260 // doAllThreadsSyscall requires that every M in allm will eventually
2261 // start and be signal-able, even with a STW.
2263 // Disable preemption here until we start the thread to ensure that
2264 // newm is not preempted between allocm and starting the new thread,
2265 // ensuring that anything added to allm is guaranteed to eventually
2269 mp := allocm(pp, fn, id)
2271 mp.sigmask = initSigmask
2272 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2273 // We're on a locked M or a thread that may have been
2274 // started by C. The kernel state of this thread may
2275 // be strange (the user may have locked it for that
2276 // purpose). We don't want to clone that into another
2277 // thread. Instead, ask a known-good thread to create
2278 // the thread for us.
2280 // This is disabled on Plan 9. See golang.org/issue/22227.
2282 // TODO: This may be unnecessary on Windows, which
2283 // doesn't model thread creation off fork.
2284 lock(&newmHandoff.lock)
2285 if newmHandoff.haveTemplateThread == 0 {
2286 throw("on a locked thread with no template thread")
2288 mp.schedlink = newmHandoff.newm
2289 newmHandoff.newm.set(mp)
2290 if newmHandoff.waiting {
2291 newmHandoff.waiting = false
2292 notewakeup(&newmHandoff.wake)
2294 unlock(&newmHandoff.lock)
2295 // The M has not started yet, but the template thread does not
2296 // participate in STW, so it will always process queued Ms and
2297 // it is safe to releasem.
2307 var ts cgothreadstart
2308 if _cgo_thread_start == nil {
2309 throw("_cgo_thread_start missing")
2312 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2313 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2315 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2318 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2320 execLock.rlock() // Prevent process clone.
2321 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2325 execLock.rlock() // Prevent process clone.
2330 // startTemplateThread starts the template thread if it is not already
2333 // The calling thread must itself be in a known-good state.
2334 func startTemplateThread() {
2335 if GOARCH == "wasm" { // no threads on wasm yet
2339 // Disable preemption to guarantee that the template thread will be
2340 // created before a park once haveTemplateThread is set.
2342 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2346 newm(templateThread, nil, -1)
2350 // templateThread is a thread in a known-good state that exists solely
2351 // to start new threads in known-good states when the calling thread
2352 // may not be in a good state.
2354 // Many programs never need this, so templateThread is started lazily
2355 // when we first enter a state that might lead to running on a thread
2356 // in an unknown state.
2358 // templateThread runs on an M without a P, so it must not have write
2361 //go:nowritebarrierrec
2362 func templateThread() {
2369 lock(&newmHandoff.lock)
2370 for newmHandoff.newm != 0 {
2371 newm := newmHandoff.newm.ptr()
2372 newmHandoff.newm = 0
2373 unlock(&newmHandoff.lock)
2375 next := newm.schedlink.ptr()
2380 lock(&newmHandoff.lock)
2382 newmHandoff.waiting = true
2383 noteclear(&newmHandoff.wake)
2384 unlock(&newmHandoff.lock)
2385 notesleep(&newmHandoff.wake)
2389 // Stops execution of the current m until new work is available.
2390 // Returns with acquired P.
2394 if gp.m.locks != 0 {
2395 throw("stopm holding locks")
2398 throw("stopm holding p")
2401 throw("stopm spinning")
2408 acquirep(gp.m.nextp.ptr())
2413 // startm's caller incremented nmspinning. Set the new M's spinning.
2414 getg().m.spinning = true
2417 // Schedules some M to run the p (creates an M if necessary).
2418 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2419 // May run with m.p==nil, so write barriers are not allowed.
2420 // If spinning is set, the caller has incremented nmspinning and must provide a
2421 // P. startm will set m.spinning in the newly started M.
2423 // Callers passing a non-nil P must call from a non-preemptible context. See
2424 // comment on acquirem below.
2426 // Must not have write barriers because this may be called without a P.
2428 //go:nowritebarrierrec
2429 func startm(pp *p, spinning bool) {
2430 // Disable preemption.
2432 // Every owned P must have an owner that will eventually stop it in the
2433 // event of a GC stop request. startm takes transient ownership of a P
2434 // (either from argument or pidleget below) and transfers ownership to
2435 // a started M, which will be responsible for performing the stop.
2437 // Preemption must be disabled during this transient ownership,
2438 // otherwise the P this is running on may enter GC stop while still
2439 // holding the transient P, leaving that P in limbo and deadlocking the
2442 // Callers passing a non-nil P must already be in non-preemptible
2443 // context, otherwise such preemption could occur on function entry to
2444 // startm. Callers passing a nil P may be preemptible, so we must
2445 // disable preemption before acquiring a P from pidleget below.
2450 // TODO(prattmic): All remaining calls to this function
2451 // with _p_ == nil could be cleaned up to find a P
2452 // before calling startm.
2453 throw("startm: P required for spinning=true")
2464 // No M is available, we must drop sched.lock and call newm.
2465 // However, we already own a P to assign to the M.
2467 // Once sched.lock is released, another G (e.g., in a syscall),
2468 // could find no idle P while checkdead finds a runnable G but
2469 // no running M's because this new M hasn't started yet, thus
2470 // throwing in an apparent deadlock.
2472 // Avoid this situation by pre-allocating the ID for the new M,
2473 // thus marking it as 'running' before we drop sched.lock. This
2474 // new M will eventually run the scheduler to execute any
2481 // The caller incremented nmspinning, so set m.spinning in the new M.
2485 // Ownership transfer of pp committed by start in newm.
2486 // Preemption is now safe.
2492 throw("startm: m is spinning")
2495 throw("startm: m has p")
2497 if spinning && !runqempty(pp) {
2498 throw("startm: p has runnable gs")
2500 // The caller incremented nmspinning, so set m.spinning in the new M.
2501 nmp.spinning = spinning
2503 notewakeup(&nmp.park)
2504 // Ownership transfer of pp committed by wakeup. Preemption is now
2509 // Hands off P from syscall or locked M.
2510 // Always runs without a P, so write barriers are not allowed.
2512 //go:nowritebarrierrec
2513 func handoffp(pp *p) {
2514 // handoffp must start an M in any situation where
2515 // findrunnable would return a G to run on pp.
2517 // if it has local work, start it straight away
2518 if !runqempty(pp) || sched.runqsize != 0 {
2522 // if there's trace work to do, start it straight away
2523 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2527 // if it has GC work, start it straight away
2528 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2532 // no local work, check that there are no spinning/idle M's,
2533 // otherwise our help is not required
2534 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2535 sched.needspinning.Store(0)
2540 if sched.gcwaiting.Load() {
2541 pp.status = _Pgcstop
2543 if sched.stopwait == 0 {
2544 notewakeup(&sched.stopnote)
2549 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2550 sched.safePointFn(pp)
2551 sched.safePointWait--
2552 if sched.safePointWait == 0 {
2553 notewakeup(&sched.safePointNote)
2556 if sched.runqsize != 0 {
2561 // If this is the last running P and nobody is polling network,
2562 // need to wakeup another M to poll network.
2563 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2569 // The scheduler lock cannot be held when calling wakeNetPoller below
2570 // because wakeNetPoller may call wakep which may call startm.
2571 when := nobarrierWakeTime(pp)
2580 // Tries to add one more P to execute G's.
2581 // Called when a G is made runnable (newproc, ready).
2582 // Must be called with a P.
2584 // Be conservative about spinning threads, only start one if none exist
2586 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2590 // Disable preemption until ownership of pp transfers to the next M in
2591 // startm. Otherwise preemption here would leave pp stuck waiting to
2594 // See preemption comment on acquirem in startm for more details.
2599 pp, _ = pidlegetSpinning(0)
2601 if sched.nmspinning.Add(-1) < 0 {
2602 throw("wakep: negative nmspinning")
2608 // Since we always have a P, the race in the "No M is available"
2609 // comment in startm doesn't apply during the small window between the
2610 // unlock here and lock in startm. A checkdead in between will always
2611 // see at least one running M (ours).
2619 // Stops execution of the current m that is locked to a g until the g is runnable again.
2620 // Returns with acquired P.
2621 func stoplockedm() {
2624 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2625 throw("stoplockedm: inconsistent locking")
2628 // Schedule another M to run this p.
2633 // Wait until another thread schedules lockedg again.
2635 status := readgstatus(gp.m.lockedg.ptr())
2636 if status&^_Gscan != _Grunnable {
2637 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2638 dumpgstatus(gp.m.lockedg.ptr())
2639 throw("stoplockedm: not runnable")
2641 acquirep(gp.m.nextp.ptr())
2645 // Schedules the locked m to run the locked gp.
2646 // May run during STW, so write barriers are not allowed.
2648 //go:nowritebarrierrec
2649 func startlockedm(gp *g) {
2650 mp := gp.lockedm.ptr()
2652 throw("startlockedm: locked to me")
2655 throw("startlockedm: m has p")
2657 // directly handoff current P to the locked m
2661 notewakeup(&mp.park)
2665 // Stops the current m for stopTheWorld.
2666 // Returns when the world is restarted.
2670 if !sched.gcwaiting.Load() {
2671 throw("gcstopm: not waiting for gc")
2674 gp.m.spinning = false
2675 // OK to just drop nmspinning here,
2676 // startTheWorld will unpark threads as necessary.
2677 if sched.nmspinning.Add(-1) < 0 {
2678 throw("gcstopm: negative nmspinning")
2683 pp.status = _Pgcstop
2685 if sched.stopwait == 0 {
2686 notewakeup(&sched.stopnote)
2692 // Schedules gp to run on the current M.
2693 // If inheritTime is true, gp inherits the remaining time in the
2694 // current time slice. Otherwise, it starts a new time slice.
2697 // Write barriers are allowed because this is called immediately after
2698 // acquiring a P in several places.
2700 //go:yeswritebarrierrec
2701 func execute(gp *g, inheritTime bool) {
2704 if goroutineProfile.active {
2705 // Make sure that gp has had its stack written out to the goroutine
2706 // profile, exactly as it was when the goroutine profiler first stopped
2708 tryRecordGoroutineProfile(gp, osyield)
2711 // Assign gp.m before entering _Grunning so running Gs have an
2715 casgstatus(gp, _Grunnable, _Grunning)
2718 gp.stackguard0 = gp.stack.lo + _StackGuard
2720 mp.p.ptr().schedtick++
2723 // Check whether the profiler needs to be turned on or off.
2724 hz := sched.profilehz
2725 if mp.profilehz != hz {
2726 setThreadCPUProfiler(hz)
2730 // GoSysExit has to happen when we have a P, but before GoStart.
2731 // So we emit it here.
2732 if gp.syscallsp != 0 && gp.sysblocktraced {
2733 traceGoSysExit(gp.sysexitticks)
2741 // Finds a runnable goroutine to execute.
2742 // Tries to steal from other P's, get g from local or global queue, poll network.
2743 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2744 // reader) so the caller should try to wake a P.
2745 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2748 // The conditions here and in handoffp must agree: if
2749 // findrunnable would return a G to run, handoffp must start
2754 if sched.gcwaiting.Load() {
2758 if pp.runSafePointFn != 0 {
2762 // now and pollUntil are saved for work stealing later,
2763 // which may steal timers. It's important that between now
2764 // and then, nothing blocks, so these numbers remain mostly
2766 now, pollUntil, _ := checkTimers(pp, 0)
2768 // Try to schedule the trace reader.
2769 if trace.enabled || trace.shutdown {
2772 casgstatus(gp, _Gwaiting, _Grunnable)
2773 traceGoUnpark(gp, 0)
2774 return gp, false, true
2778 // Try to schedule a GC worker.
2779 if gcBlackenEnabled != 0 {
2780 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2782 return gp, false, true
2787 // Check the global runnable queue once in a while to ensure fairness.
2788 // Otherwise two goroutines can completely occupy the local runqueue
2789 // by constantly respawning each other.
2790 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2792 gp := globrunqget(pp, 1)
2795 return gp, false, false
2799 // Wake up the finalizer G.
2800 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2801 if gp := wakefing(); gp != nil {
2805 if *cgo_yield != nil {
2806 asmcgocall(*cgo_yield, nil)
2810 if gp, inheritTime := runqget(pp); gp != nil {
2811 return gp, inheritTime, false
2815 if sched.runqsize != 0 {
2817 gp := globrunqget(pp, 0)
2820 return gp, false, false
2825 // This netpoll is only an optimization before we resort to stealing.
2826 // We can safely skip it if there are no waiters or a thread is blocked
2827 // in netpoll already. If there is any kind of logical race with that
2828 // blocked thread (e.g. it has already returned from netpoll, but does
2829 // not set lastpoll yet), this thread will do blocking netpoll below
2831 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2832 if list := netpoll(0); !list.empty() { // non-blocking
2835 casgstatus(gp, _Gwaiting, _Grunnable)
2837 traceGoUnpark(gp, 0)
2839 return gp, false, false
2843 // Spinning Ms: steal work from other Ps.
2845 // Limit the number of spinning Ms to half the number of busy Ps.
2846 // This is necessary to prevent excessive CPU consumption when
2847 // GOMAXPROCS>>1 but the program parallelism is low.
2848 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2853 gp, inheritTime, tnow, w, newWork := stealWork(now)
2855 // Successfully stole.
2856 return gp, inheritTime, false
2859 // There may be new timer or GC work; restart to
2865 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2866 // Earlier timer to wait for.
2871 // We have nothing to do.
2873 // If we're in the GC mark phase, can safely scan and blacken objects,
2874 // and have work to do, run idle-time marking rather than give up the P.
2875 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2876 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2878 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2880 casgstatus(gp, _Gwaiting, _Grunnable)
2882 traceGoUnpark(gp, 0)
2884 return gp, false, false
2886 gcController.removeIdleMarkWorker()
2890 // If a callback returned and no other goroutine is awake,
2891 // then wake event handler goroutine which pauses execution
2892 // until a callback was triggered.
2893 gp, otherReady := beforeIdle(now, pollUntil)
2895 casgstatus(gp, _Gwaiting, _Grunnable)
2897 traceGoUnpark(gp, 0)
2899 return gp, false, false
2905 // Before we drop our P, make a snapshot of the allp slice,
2906 // which can change underfoot once we no longer block
2907 // safe-points. We don't need to snapshot the contents because
2908 // everything up to cap(allp) is immutable.
2909 allpSnapshot := allp
2910 // Also snapshot masks. Value changes are OK, but we can't allow
2911 // len to change out from under us.
2912 idlepMaskSnapshot := idlepMask
2913 timerpMaskSnapshot := timerpMask
2915 // return P and block
2917 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2921 if sched.runqsize != 0 {
2922 gp := globrunqget(pp, 0)
2924 return gp, false, false
2926 if !mp.spinning && sched.needspinning.Load() == 1 {
2927 // See "Delicate dance" comment below.
2932 if releasep() != pp {
2933 throw("findrunnable: wrong p")
2935 now = pidleput(pp, now)
2938 // Delicate dance: thread transitions from spinning to non-spinning
2939 // state, potentially concurrently with submission of new work. We must
2940 // drop nmspinning first and then check all sources again (with
2941 // #StoreLoad memory barrier in between). If we do it the other way
2942 // around, another thread can submit work after we've checked all
2943 // sources but before we drop nmspinning; as a result nobody will
2944 // unpark a thread to run the work.
2946 // This applies to the following sources of work:
2948 // * Goroutines added to a per-P run queue.
2949 // * New/modified-earlier timers on a per-P timer heap.
2950 // * Idle-priority GC work (barring golang.org/issue/19112).
2952 // If we discover new work below, we need to restore m.spinning as a
2953 // signal for resetspinning to unpark a new worker thread (because
2954 // there can be more than one starving goroutine).
2956 // However, if after discovering new work we also observe no idle Ps
2957 // (either here or in resetspinning), we have a problem. We may be
2958 // racing with a non-spinning M in the block above, having found no
2959 // work and preparing to release its P and park. Allowing that P to go
2960 // idle will result in loss of work conservation (idle P while there is
2961 // runnable work). This could result in complete deadlock in the
2962 // unlikely event that we discover new work (from netpoll) right as we
2963 // are racing with _all_ other Ps going idle.
2965 // We use sched.needspinning to synchronize with non-spinning Ms going
2966 // idle. If needspinning is set when they are about to drop their P,
2967 // they abort the drop and instead become a new spinning M on our
2968 // behalf. If we are not racing and the system is truly fully loaded
2969 // then no spinning threads are required, and the next thread to
2970 // naturally become spinning will clear the flag.
2972 // Also see "Worker thread parking/unparking" comment at the top of the
2974 wasSpinning := mp.spinning
2977 if sched.nmspinning.Add(-1) < 0 {
2978 throw("findrunnable: negative nmspinning")
2981 // Note the for correctness, only the last M transitioning from
2982 // spinning to non-spinning must perform these rechecks to
2983 // ensure no missed work. However, the runtime has some cases
2984 // of transient increments of nmspinning that are decremented
2985 // without going through this path, so we must be conservative
2986 // and perform the check on all spinning Ms.
2988 // See https://go.dev/issue/43997.
2990 // Check all runqueues once again.
2991 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2998 // Check for idle-priority GC work again.
2999 pp, gp := checkIdleGCNoP()
3004 // Run the idle worker.
3005 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
3006 casgstatus(gp, _Gwaiting, _Grunnable)
3008 traceGoUnpark(gp, 0)
3010 return gp, false, false
3013 // Finally, check for timer creation or expiry concurrently with
3014 // transitioning from spinning to non-spinning.
3016 // Note that we cannot use checkTimers here because it calls
3017 // adjusttimers which may need to allocate memory, and that isn't
3018 // allowed when we don't have an active P.
3019 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
3022 // Poll network until next timer.
3023 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
3024 sched.pollUntil.Store(pollUntil)
3026 throw("findrunnable: netpoll with p")
3029 throw("findrunnable: netpoll with spinning")
3035 delay = pollUntil - now
3041 // When using fake time, just poll.
3044 list := netpoll(delay) // block until new work is available
3045 sched.pollUntil.Store(0)
3046 sched.lastpoll.Store(now)
3047 if faketime != 0 && list.empty() {
3048 // Using fake time and nothing is ready; stop M.
3049 // When all M's stop, checkdead will call timejump.
3054 pp, _ := pidleget(now)
3063 casgstatus(gp, _Gwaiting, _Grunnable)
3065 traceGoUnpark(gp, 0)
3067 return gp, false, false
3074 } else if pollUntil != 0 && netpollinited() {
3075 pollerPollUntil := sched.pollUntil.Load()
3076 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3084 // pollWork reports whether there is non-background work this P could
3085 // be doing. This is a fairly lightweight check to be used for
3086 // background work loops, like idle GC. It checks a subset of the
3087 // conditions checked by the actual scheduler.
3088 func pollWork() bool {
3089 if sched.runqsize != 0 {
3092 p := getg().m.p.ptr()
3096 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3097 if list := netpoll(0); !list.empty() {
3105 // stealWork attempts to steal a runnable goroutine or timer from any P.
3107 // If newWork is true, new work may have been readied.
3109 // If now is not 0 it is the current time. stealWork returns the passed time or
3110 // the current time if now was passed as 0.
3111 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3112 pp := getg().m.p.ptr()
3116 const stealTries = 4
3117 for i := 0; i < stealTries; i++ {
3118 stealTimersOrRunNextG := i == stealTries-1
3120 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3121 if sched.gcwaiting.Load() {
3122 // GC work may be available.
3123 return nil, false, now, pollUntil, true
3125 p2 := allp[enum.position()]
3130 // Steal timers from p2. This call to checkTimers is the only place
3131 // where we might hold a lock on a different P's timers. We do this
3132 // once on the last pass before checking runnext because stealing
3133 // from the other P's runnext should be the last resort, so if there
3134 // are timers to steal do that first.
3136 // We only check timers on one of the stealing iterations because
3137 // the time stored in now doesn't change in this loop and checking
3138 // the timers for each P more than once with the same value of now
3139 // is probably a waste of time.
3141 // timerpMask tells us whether the P may have timers at all. If it
3142 // can't, no need to check at all.
3143 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3144 tnow, w, ran := checkTimers(p2, now)
3146 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3150 // Running the timers may have
3151 // made an arbitrary number of G's
3152 // ready and added them to this P's
3153 // local run queue. That invalidates
3154 // the assumption of runqsteal
3155 // that it always has room to add
3156 // stolen G's. So check now if there
3157 // is a local G to run.
3158 if gp, inheritTime := runqget(pp); gp != nil {
3159 return gp, inheritTime, now, pollUntil, ranTimer
3165 // Don't bother to attempt to steal if p2 is idle.
3166 if !idlepMask.read(enum.position()) {
3167 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3168 return gp, false, now, pollUntil, ranTimer
3174 // No goroutines found to steal. Regardless, running a timer may have
3175 // made some goroutine ready that we missed. Indicate the next timer to
3177 return nil, false, now, pollUntil, ranTimer
3180 // Check all Ps for a runnable G to steal.
3182 // On entry we have no P. If a G is available to steal and a P is available,
3183 // the P is returned which the caller should acquire and attempt to steal the
3185 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3186 for id, p2 := range allpSnapshot {
3187 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3189 pp, _ := pidlegetSpinning(0)
3191 // Can't get a P, don't bother checking remaining Ps.
3200 // No work available.
3204 // Check all Ps for a timer expiring sooner than pollUntil.
3206 // Returns updated pollUntil value.
3207 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3208 for id, p2 := range allpSnapshot {
3209 if timerpMaskSnapshot.read(uint32(id)) {
3210 w := nobarrierWakeTime(p2)
3211 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3220 // Check for idle-priority GC, without a P on entry.
3222 // If some GC work, a P, and a worker G are all available, the P and G will be
3223 // returned. The returned P has not been wired yet.
3224 func checkIdleGCNoP() (*p, *g) {
3225 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3226 // must check again after acquiring a P. As an optimization, we also check
3227 // if an idle mark worker is needed at all. This is OK here, because if we
3228 // observe that one isn't needed, at least one is currently running. Even if
3229 // it stops running, its own journey into the scheduler should schedule it
3230 // again, if need be (at which point, this check will pass, if relevant).
3231 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3234 if !gcMarkWorkAvailable(nil) {
3238 // Work is available; we can start an idle GC worker only if there is
3239 // an available P and available worker G.
3241 // We can attempt to acquire these in either order, though both have
3242 // synchronization concerns (see below). Workers are almost always
3243 // available (see comment in findRunnableGCWorker for the one case
3244 // there may be none). Since we're slightly less likely to find a P,
3245 // check for that first.
3247 // Synchronization: note that we must hold sched.lock until we are
3248 // committed to keeping it. Otherwise we cannot put the unnecessary P
3249 // back in sched.pidle without performing the full set of idle
3250 // transition checks.
3252 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3253 // the assumption in gcControllerState.findRunnableGCWorker that an
3254 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3256 pp, now := pidlegetSpinning(0)
3262 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3263 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3269 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3273 gcController.removeIdleMarkWorker()
3279 return pp, node.gp.ptr()
3282 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3283 // going to wake up before the when argument; or it wakes an idle P to service
3284 // timers and the network poller if there isn't one already.
3285 func wakeNetPoller(when int64) {
3286 if sched.lastpoll.Load() == 0 {
3287 // In findrunnable we ensure that when polling the pollUntil
3288 // field is either zero or the time to which the current
3289 // poll is expected to run. This can have a spurious wakeup
3290 // but should never miss a wakeup.
3291 pollerPollUntil := sched.pollUntil.Load()
3292 if pollerPollUntil == 0 || pollerPollUntil > when {
3296 // There are no threads in the network poller, try to get
3297 // one there so it can handle new timers.
3298 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3304 func resetspinning() {
3307 throw("resetspinning: not a spinning m")
3309 gp.m.spinning = false
3310 nmspinning := sched.nmspinning.Add(-1)
3312 throw("findrunnable: negative nmspinning")
3314 // M wakeup policy is deliberately somewhat conservative, so check if we
3315 // need to wakeup another P here. See "Worker thread parking/unparking"
3316 // comment at the top of the file for details.
3320 // injectglist adds each runnable G on the list to some run queue,
3321 // and clears glist. If there is no current P, they are added to the
3322 // global queue, and up to npidle M's are started to run them.
3323 // Otherwise, for each idle P, this adds a G to the global queue
3324 // and starts an M. Any remaining G's are added to the current P's
3326 // This may temporarily acquire sched.lock.
3327 // Can run concurrently with GC.
3328 func injectglist(glist *gList) {
3333 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3334 traceGoUnpark(gp, 0)
3338 // Mark all the goroutines as runnable before we put them
3339 // on the run queues.
3340 head := glist.head.ptr()
3343 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3346 casgstatus(gp, _Gwaiting, _Grunnable)
3349 // Turn the gList into a gQueue.
3355 startIdle := func(n int) {
3356 for i := 0; i < n; i++ {
3357 mp := acquirem() // See comment in startm.
3360 pp, _ := pidlegetSpinning(0)
3373 pp := getg().m.p.ptr()
3376 globrunqputbatch(&q, int32(qsize))
3382 npidle := int(sched.npidle.Load())
3385 for n = 0; n < npidle && !q.empty(); n++ {
3391 globrunqputbatch(&globq, int32(n))
3398 runqputbatch(pp, &q, qsize)
3402 // One round of scheduler: find a runnable goroutine and execute it.
3408 throw("schedule: holding locks")
3411 if mp.lockedg != 0 {
3413 execute(mp.lockedg.ptr(), false) // Never returns.
3416 // We should not schedule away from a g that is executing a cgo call,
3417 // since the cgo call is using the m's g0 stack.
3419 throw("schedule: in cgo")
3426 // Safety check: if we are spinning, the run queue should be empty.
3427 // Check this before calling checkTimers, as that might call
3428 // goready to put a ready goroutine on the local run queue.
3429 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3430 throw("schedule: spinning with local work")
3433 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3435 // This thread is going to run a goroutine and is not spinning anymore,
3436 // so if it was marked as spinning we need to reset it now and potentially
3437 // start a new spinning M.
3442 if sched.disable.user && !schedEnabled(gp) {
3443 // Scheduling of this goroutine is disabled. Put it on
3444 // the list of pending runnable goroutines for when we
3445 // re-enable user scheduling and look again.
3447 if schedEnabled(gp) {
3448 // Something re-enabled scheduling while we
3449 // were acquiring the lock.
3452 sched.disable.runnable.pushBack(gp)
3459 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3460 // wake a P if there is one.
3464 if gp.lockedm != 0 {
3465 // Hands off own p to the locked m,
3466 // then blocks waiting for a new p.
3471 execute(gp, inheritTime)
3474 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3475 // Typically a caller sets gp's status away from Grunning and then
3476 // immediately calls dropg to finish the job. The caller is also responsible
3477 // for arranging that gp will be restarted using ready at an
3478 // appropriate time. After calling dropg and arranging for gp to be
3479 // readied later, the caller can do other work but eventually should
3480 // call schedule to restart the scheduling of goroutines on this m.
3484 setMNoWB(&gp.m.curg.m, nil)
3485 setGNoWB(&gp.m.curg, nil)
3488 // checkTimers runs any timers for the P that are ready.
3489 // If now is not 0 it is the current time.
3490 // It returns the passed time or the current time if now was passed as 0.
3491 // and the time when the next timer should run or 0 if there is no next timer,
3492 // and reports whether it ran any timers.
3493 // If the time when the next timer should run is not 0,
3494 // it is always larger than the returned time.
3495 // We pass now in and out to avoid extra calls of nanotime.
3497 //go:yeswritebarrierrec
3498 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3499 // If it's not yet time for the first timer, or the first adjusted
3500 // timer, then there is nothing to do.
3501 next := pp.timer0When.Load()
3502 nextAdj := pp.timerModifiedEarliest.Load()
3503 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3508 // No timers to run or adjust.
3509 return now, 0, false
3516 // Next timer is not ready to run, but keep going
3517 // if we would clear deleted timers.
3518 // This corresponds to the condition below where
3519 // we decide whether to call clearDeletedTimers.
3520 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3521 return now, next, false
3525 lock(&pp.timersLock)
3527 if len(pp.timers) > 0 {
3528 adjusttimers(pp, now)
3529 for len(pp.timers) > 0 {
3530 // Note that runtimer may temporarily unlock
3532 if tw := runtimer(pp, now); tw != 0 {
3542 // If this is the local P, and there are a lot of deleted timers,
3543 // clear them out. We only do this for the local P to reduce
3544 // lock contention on timersLock.
3545 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3546 clearDeletedTimers(pp)
3549 unlock(&pp.timersLock)
3551 return now, pollUntil, ran
3554 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3555 unlock((*mutex)(lock))
3559 // park continuation on g0.
3560 func park_m(gp *g) {
3564 traceGoPark(mp.waittraceev, mp.waittraceskip)
3567 // N.B. Not using casGToWaiting here because the waitreason is
3568 // set by park_m's caller.
3569 casgstatus(gp, _Grunning, _Gwaiting)
3572 if fn := mp.waitunlockf; fn != nil {
3573 ok := fn(gp, mp.waitlock)
3574 mp.waitunlockf = nil
3578 traceGoUnpark(gp, 2)
3580 casgstatus(gp, _Gwaiting, _Grunnable)
3581 execute(gp, true) // Schedule it back, never returns.
3587 func goschedImpl(gp *g) {
3588 status := readgstatus(gp)
3589 if status&^_Gscan != _Grunning {
3591 throw("bad g status")
3593 casgstatus(gp, _Grunning, _Grunnable)
3602 // Gosched continuation on g0.
3603 func gosched_m(gp *g) {
3610 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3611 func goschedguarded_m(gp *g) {
3613 if !canPreemptM(gp.m) {
3614 gogo(&gp.sched) // never return
3623 func gopreempt_m(gp *g) {
3630 // preemptPark parks gp and puts it in _Gpreempted.
3633 func preemptPark(gp *g) {
3635 traceGoPark(traceEvGoBlock, 0)
3637 status := readgstatus(gp)
3638 if status&^_Gscan != _Grunning {
3640 throw("bad g status")
3643 if gp.asyncSafePoint {
3644 // Double-check that async preemption does not
3645 // happen in SPWRITE assembly functions.
3646 // isAsyncSafePoint must exclude this case.
3647 f := findfunc(gp.sched.pc)
3649 throw("preempt at unknown pc")
3651 if f.flag&funcFlag_SPWRITE != 0 {
3652 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3653 throw("preempt SPWRITE")
3657 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3658 // be in _Grunning when we dropg because then we'd be running
3659 // without an M, but the moment we're in _Gpreempted,
3660 // something could claim this G before we've fully cleaned it
3661 // up. Hence, we set the scan bit to lock down further
3662 // transitions until we can dropg.
3663 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3665 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3669 // goyield is like Gosched, but it:
3670 // - emits a GoPreempt trace event instead of a GoSched trace event
3671 // - puts the current G on the runq of the current P instead of the globrunq
3677 func goyield_m(gp *g) {
3682 casgstatus(gp, _Grunning, _Grunnable)
3684 runqput(pp, gp, false)
3688 // Finishes execution of the current goroutine.
3699 // goexit continuation on g0.
3700 func goexit0(gp *g) {
3704 casgstatus(gp, _Grunning, _Gdead)
3705 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3706 if isSystemGoroutine(gp, false) {
3710 locked := gp.lockedm != 0
3713 gp.preemptStop = false
3714 gp.paniconfault = false
3715 gp._defer = nil // should be true already but just in case.
3716 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3718 gp.waitreason = waitReasonZero
3723 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3724 // Flush assist credit to the global pool. This gives
3725 // better information to pacing if the application is
3726 // rapidly creating an exiting goroutines.
3727 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3728 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3729 gcController.bgScanCredit.Add(scanCredit)
3730 gp.gcAssistBytes = 0
3735 if GOARCH == "wasm" { // no threads yet on wasm
3737 schedule() // never returns
3740 if mp.lockedInt != 0 {
3741 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3742 throw("internal lockOSThread error")
3746 // The goroutine may have locked this thread because
3747 // it put it in an unusual kernel state. Kill it
3748 // rather than returning it to the thread pool.
3750 // Return to mstart, which will release the P and exit
3752 if GOOS != "plan9" { // See golang.org/issue/22227.
3755 // Clear lockedExt on plan9 since we may end up re-using
3763 // save updates getg().sched to refer to pc and sp so that a following
3764 // gogo will restore pc and sp.
3766 // save must not have write barriers because invoking a write barrier
3767 // can clobber getg().sched.
3770 //go:nowritebarrierrec
3771 func save(pc, sp uintptr) {
3774 if gp == gp.m.g0 || gp == gp.m.gsignal {
3775 // m.g0.sched is special and must describe the context
3776 // for exiting the thread. mstart1 writes to it directly.
3777 // m.gsignal.sched should not be used at all.
3778 // This check makes sure save calls do not accidentally
3779 // run in contexts where they'd write to system g's.
3780 throw("save on system g not allowed")
3787 // We need to ensure ctxt is zero, but can't have a write
3788 // barrier here. However, it should always already be zero.
3790 if gp.sched.ctxt != nil {
3795 // The goroutine g is about to enter a system call.
3796 // Record that it's not using the cpu anymore.
3797 // This is called only from the go syscall library and cgocall,
3798 // not from the low-level system calls used by the runtime.
3800 // Entersyscall cannot split the stack: the save must
3801 // make g->sched refer to the caller's stack segment, because
3802 // entersyscall is going to return immediately after.
3804 // Nothing entersyscall calls can split the stack either.
3805 // We cannot safely move the stack during an active call to syscall,
3806 // because we do not know which of the uintptr arguments are
3807 // really pointers (back into the stack).
3808 // In practice, this means that we make the fast path run through
3809 // entersyscall doing no-split things, and the slow path has to use systemstack
3810 // to run bigger things on the system stack.
3812 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3813 // saved SP and PC are restored. This is needed when exitsyscall will be called
3814 // from a function further up in the call stack than the parent, as g->syscallsp
3815 // must always point to a valid stack frame. entersyscall below is the normal
3816 // entry point for syscalls, which obtains the SP and PC from the caller.
3819 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3820 // If the syscall does not block, that is it, we do not emit any other events.
3821 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3822 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3823 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3824 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3825 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3826 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3827 // and we wait for the increment before emitting traceGoSysExit.
3828 // Note that the increment is done even if tracing is not enabled,
3829 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3832 func reentersyscall(pc, sp uintptr) {
3835 // Disable preemption because during this function g is in Gsyscall status,
3836 // but can have inconsistent g->sched, do not let GC observe it.
3839 // Entersyscall must not call any function that might split/grow the stack.
3840 // (See details in comment above.)
3841 // Catch calls that might, by replacing the stack guard with something that
3842 // will trip any stack check and leaving a flag to tell newstack to die.
3843 gp.stackguard0 = stackPreempt
3844 gp.throwsplit = true
3846 // Leave SP around for GC and traceback.
3850 casgstatus(gp, _Grunning, _Gsyscall)
3851 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3852 systemstack(func() {
3853 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3854 throw("entersyscall")
3859 systemstack(traceGoSysCall)
3860 // systemstack itself clobbers g.sched.{pc,sp} and we might
3861 // need them later when the G is genuinely blocked in a
3866 if sched.sysmonwait.Load() {
3867 systemstack(entersyscall_sysmon)
3871 if gp.m.p.ptr().runSafePointFn != 0 {
3872 // runSafePointFn may stack split if run on this stack
3873 systemstack(runSafePointFn)
3877 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3878 gp.sysblocktraced = true
3883 atomic.Store(&pp.status, _Psyscall)
3884 if sched.gcwaiting.Load() {
3885 systemstack(entersyscall_gcwait)
3892 // Standard syscall entry used by the go syscall library and normal cgo calls.
3894 // This is exported via linkname to assembly in the syscall package and x/sys.
3897 //go:linkname entersyscall
3898 func entersyscall() {
3899 reentersyscall(getcallerpc(), getcallersp())
3902 func entersyscall_sysmon() {
3904 if sched.sysmonwait.Load() {
3905 sched.sysmonwait.Store(false)
3906 notewakeup(&sched.sysmonnote)
3911 func entersyscall_gcwait() {
3913 pp := gp.m.oldp.ptr()
3916 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3922 if sched.stopwait--; sched.stopwait == 0 {
3923 notewakeup(&sched.stopnote)
3929 // The same as entersyscall(), but with a hint that the syscall is blocking.
3932 func entersyscallblock() {
3935 gp.m.locks++ // see comment in entersyscall
3936 gp.throwsplit = true
3937 gp.stackguard0 = stackPreempt // see comment in entersyscall
3938 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3939 gp.sysblocktraced = true
3940 gp.m.p.ptr().syscalltick++
3942 // Leave SP around for GC and traceback.
3946 gp.syscallsp = gp.sched.sp
3947 gp.syscallpc = gp.sched.pc
3948 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3952 systemstack(func() {
3953 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3954 throw("entersyscallblock")
3957 casgstatus(gp, _Grunning, _Gsyscall)
3958 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3959 systemstack(func() {
3960 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3961 throw("entersyscallblock")
3965 systemstack(entersyscallblock_handoff)
3967 // Resave for traceback during blocked call.
3968 save(getcallerpc(), getcallersp())
3973 func entersyscallblock_handoff() {
3976 traceGoSysBlock(getg().m.p.ptr())
3978 handoffp(releasep())
3981 // The goroutine g exited its system call.
3982 // Arrange for it to run on a cpu again.
3983 // This is called only from the go syscall library, not
3984 // from the low-level system calls used by the runtime.
3986 // Write barriers are not allowed because our P may have been stolen.
3988 // This is exported via linkname to assembly in the syscall package.
3991 //go:nowritebarrierrec
3992 //go:linkname exitsyscall
3993 func exitsyscall() {
3996 gp.m.locks++ // see comment in entersyscall
3997 if getcallersp() > gp.syscallsp {
3998 throw("exitsyscall: syscall frame is no longer valid")
4002 oldp := gp.m.oldp.ptr()
4004 if exitsyscallfast(oldp) {
4005 // When exitsyscallfast returns success, we have a P so can now use
4007 if goroutineProfile.active {
4008 // Make sure that gp has had its stack written out to the goroutine
4009 // profile, exactly as it was when the goroutine profiler first
4010 // stopped the world.
4011 systemstack(func() {
4012 tryRecordGoroutineProfileWB(gp)
4016 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4017 systemstack(traceGoStart)
4020 // There's a cpu for us, so we can run.
4021 gp.m.p.ptr().syscalltick++
4022 // We need to cas the status and scan before resuming...
4023 casgstatus(gp, _Gsyscall, _Grunning)
4025 // Garbage collector isn't running (since we are),
4026 // so okay to clear syscallsp.
4030 // restore the preemption request in case we've cleared it in newstack
4031 gp.stackguard0 = stackPreempt
4033 // otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
4034 gp.stackguard0 = gp.stack.lo + _StackGuard
4036 gp.throwsplit = false
4038 if sched.disable.user && !schedEnabled(gp) {
4039 // Scheduling of this goroutine is disabled.
4048 // Wait till traceGoSysBlock event is emitted.
4049 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4050 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
4053 // We can't trace syscall exit right now because we don't have a P.
4054 // Tracing code can invoke write barriers that cannot run without a P.
4055 // So instead we remember the syscall exit time and emit the event
4056 // in execute when we have a P.
4057 gp.sysexitticks = cputicks()
4062 // Call the scheduler.
4065 // Scheduler returned, so we're allowed to run now.
4066 // Delete the syscallsp information that we left for
4067 // the garbage collector during the system call.
4068 // Must wait until now because until gosched returns
4069 // we don't know for sure that the garbage collector
4072 gp.m.p.ptr().syscalltick++
4073 gp.throwsplit = false
4077 func exitsyscallfast(oldp *p) bool {
4080 // Freezetheworld sets stopwait but does not retake P's.
4081 if sched.stopwait == freezeStopWait {
4085 // Try to re-acquire the last P.
4086 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4087 // There's a cpu for us, so we can run.
4089 exitsyscallfast_reacquired()
4093 // Try to get any other idle P.
4094 if sched.pidle != 0 {
4096 systemstack(func() {
4097 ok = exitsyscallfast_pidle()
4098 if ok && trace.enabled {
4100 // Wait till traceGoSysBlock event is emitted.
4101 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4102 for oldp.syscalltick == gp.m.syscalltick {
4116 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4117 // has successfully reacquired the P it was running on before the
4121 func exitsyscallfast_reacquired() {
4123 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4125 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4126 // traceGoSysBlock for this syscall was already emitted,
4127 // but here we effectively retake the p from the new syscall running on the same p.
4128 systemstack(func() {
4129 // Denote blocking of the new syscall.
4130 traceGoSysBlock(gp.m.p.ptr())
4131 // Denote completion of the current syscall.
4135 gp.m.p.ptr().syscalltick++
4139 func exitsyscallfast_pidle() bool {
4141 pp, _ := pidleget(0)
4142 if pp != nil && sched.sysmonwait.Load() {
4143 sched.sysmonwait.Store(false)
4144 notewakeup(&sched.sysmonnote)
4154 // exitsyscall slow path on g0.
4155 // Failed to acquire P, enqueue gp as runnable.
4157 // Called via mcall, so gp is the calling g from this M.
4159 //go:nowritebarrierrec
4160 func exitsyscall0(gp *g) {
4161 casgstatus(gp, _Gsyscall, _Grunnable)
4165 if schedEnabled(gp) {
4172 // Below, we stoplockedm if gp is locked. globrunqput releases
4173 // ownership of gp, so we must check if gp is locked prior to
4174 // committing the release by unlocking sched.lock, otherwise we
4175 // could race with another M transitioning gp from unlocked to
4177 locked = gp.lockedm != 0
4178 } else if sched.sysmonwait.Load() {
4179 sched.sysmonwait.Store(false)
4180 notewakeup(&sched.sysmonnote)
4185 execute(gp, false) // Never returns.
4188 // Wait until another thread schedules gp and so m again.
4190 // N.B. lockedm must be this M, as this g was running on this M
4191 // before entersyscall.
4193 execute(gp, false) // Never returns.
4196 schedule() // Never returns.
4199 // Called from syscall package before fork.
4201 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4203 func syscall_runtime_BeforeFork() {
4206 // Block signals during a fork, so that the child does not run
4207 // a signal handler before exec if a signal is sent to the process
4208 // group. See issue #18600.
4210 sigsave(&gp.m.sigmask)
4213 // This function is called before fork in syscall package.
4214 // Code between fork and exec must not allocate memory nor even try to grow stack.
4215 // Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
4216 // runtime_AfterFork will undo this in parent process, but not in child.
4217 gp.stackguard0 = stackFork
4220 // Called from syscall package after fork in parent.
4222 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4224 func syscall_runtime_AfterFork() {
4227 // See the comments in beforefork.
4228 gp.stackguard0 = gp.stack.lo + _StackGuard
4230 msigrestore(gp.m.sigmask)
4235 // inForkedChild is true while manipulating signals in the child process.
4236 // This is used to avoid calling libc functions in case we are using vfork.
4237 var inForkedChild bool
4239 // Called from syscall package after fork in child.
4240 // It resets non-sigignored signals to the default handler, and
4241 // restores the signal mask in preparation for the exec.
4243 // Because this might be called during a vfork, and therefore may be
4244 // temporarily sharing address space with the parent process, this must
4245 // not change any global variables or calling into C code that may do so.
4247 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4249 //go:nowritebarrierrec
4250 func syscall_runtime_AfterForkInChild() {
4251 // It's OK to change the global variable inForkedChild here
4252 // because we are going to change it back. There is no race here,
4253 // because if we are sharing address space with the parent process,
4254 // then the parent process can not be running concurrently.
4255 inForkedChild = true
4257 clearSignalHandlers()
4259 // When we are the child we are the only thread running,
4260 // so we know that nothing else has changed gp.m.sigmask.
4261 msigrestore(getg().m.sigmask)
4263 inForkedChild = false
4266 // pendingPreemptSignals is the number of preemption signals
4267 // that have been sent but not received. This is only used on Darwin.
4269 var pendingPreemptSignals atomic.Int32
4271 // Called from syscall package before Exec.
4273 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4274 func syscall_runtime_BeforeExec() {
4275 // Prevent thread creation during exec.
4278 // On Darwin, wait for all pending preemption signals to
4279 // be received. See issue #41702.
4280 if GOOS == "darwin" || GOOS == "ios" {
4281 for pendingPreemptSignals.Load() > 0 {
4287 // Called from syscall package after Exec.
4289 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4290 func syscall_runtime_AfterExec() {
4294 // Allocate a new g, with a stack big enough for stacksize bytes.
4295 func malg(stacksize int32) *g {
4298 stacksize = round2(_StackSystem + stacksize)
4299 systemstack(func() {
4300 newg.stack = stackalloc(uint32(stacksize))
4302 newg.stackguard0 = newg.stack.lo + _StackGuard
4303 newg.stackguard1 = ^uintptr(0)
4304 // Clear the bottom word of the stack. We record g
4305 // there on gsignal stack during VDSO on ARM and ARM64.
4306 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4311 // Create a new g running fn.
4312 // Put it on the queue of g's waiting to run.
4313 // The compiler turns a go statement into a call to this.
4314 func newproc(fn *funcval) {
4317 systemstack(func() {
4318 newg := newproc1(fn, gp, pc)
4320 pp := getg().m.p.ptr()
4321 runqput(pp, newg, true)
4329 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4330 // address of the go statement that created this. The caller is responsible
4331 // for adding the new g to the scheduler.
4332 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4334 fatal("go of nil func value")
4337 mp := acquirem() // disable preemption because we hold M and P in local vars.
4341 newg = malg(_StackMin)
4342 casgstatus(newg, _Gidle, _Gdead)
4343 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4345 if newg.stack.hi == 0 {
4346 throw("newproc1: newg missing stack")
4349 if readgstatus(newg) != _Gdead {
4350 throw("newproc1: new g is not Gdead")
4353 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4354 totalSize = alignUp(totalSize, sys.StackAlign)
4355 sp := newg.stack.hi - totalSize
4359 *(*uintptr)(unsafe.Pointer(sp)) = 0
4361 spArg += sys.MinFrameSize
4364 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4367 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4368 newg.sched.g = guintptr(unsafe.Pointer(newg))
4369 gostartcallfn(&newg.sched, fn)
4370 newg.parentGoid = callergp.goid
4371 newg.gopc = callerpc
4372 newg.ancestors = saveAncestors(callergp)
4373 newg.startpc = fn.fn
4374 if isSystemGoroutine(newg, false) {
4377 // Only user goroutines inherit pprof labels.
4379 newg.labels = mp.curg.labels
4381 if goroutineProfile.active {
4382 // A concurrent goroutine profile is running. It should include
4383 // exactly the set of goroutines that were alive when the goroutine
4384 // profiler first stopped the world. That does not include newg, so
4385 // mark it as not needing a profile before transitioning it from
4387 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4390 // Track initial transition?
4391 newg.trackingSeq = uint8(fastrand())
4392 if newg.trackingSeq%gTrackingPeriod == 0 {
4393 newg.tracking = true
4395 casgstatus(newg, _Gdead, _Grunnable)
4396 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4398 if pp.goidcache == pp.goidcacheend {
4399 // Sched.goidgen is the last allocated id,
4400 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4401 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4402 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4403 pp.goidcache -= _GoidCacheBatch - 1
4404 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4406 newg.goid = pp.goidcache
4409 newg.racectx = racegostart(callerpc)
4410 if newg.labels != nil {
4411 // See note in proflabel.go on labelSync's role in synchronizing
4412 // with the reads in the signal handler.
4413 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4417 traceGoCreate(newg, newg.startpc)
4424 // saveAncestors copies previous ancestors of the given caller g and
4425 // includes info for the current caller into a new set of tracebacks for
4426 // a g being created.
4427 func saveAncestors(callergp *g) *[]ancestorInfo {
4428 // Copy all prior info, except for the root goroutine (goid 0).
4429 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4432 var callerAncestors []ancestorInfo
4433 if callergp.ancestors != nil {
4434 callerAncestors = *callergp.ancestors
4436 n := int32(len(callerAncestors)) + 1
4437 if n > debug.tracebackancestors {
4438 n = debug.tracebackancestors
4440 ancestors := make([]ancestorInfo, n)
4441 copy(ancestors[1:], callerAncestors)
4443 var pcs [tracebackInnerFrames]uintptr
4444 npcs := gcallers(callergp, 0, pcs[:])
4445 ipcs := make([]uintptr, npcs)
4447 ancestors[0] = ancestorInfo{
4449 goid: callergp.goid,
4450 gopc: callergp.gopc,
4453 ancestorsp := new([]ancestorInfo)
4454 *ancestorsp = ancestors
4458 // Put on gfree list.
4459 // If local list is too long, transfer a batch to the global list.
4460 func gfput(pp *p, gp *g) {
4461 if readgstatus(gp) != _Gdead {
4462 throw("gfput: bad status (not Gdead)")
4465 stksize := gp.stack.hi - gp.stack.lo
4467 if stksize != uintptr(startingStackSize) {
4468 // non-standard stack size - free it.
4477 if pp.gFree.n >= 64 {
4483 for pp.gFree.n >= 32 {
4484 gp := pp.gFree.pop()
4486 if gp.stack.lo == 0 {
4493 lock(&sched.gFree.lock)
4494 sched.gFree.noStack.pushAll(noStackQ)
4495 sched.gFree.stack.pushAll(stackQ)
4496 sched.gFree.n += inc
4497 unlock(&sched.gFree.lock)
4501 // Get from gfree list.
4502 // If local list is empty, grab a batch from global list.
4503 func gfget(pp *p) *g {
4505 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4506 lock(&sched.gFree.lock)
4507 // Move a batch of free Gs to the P.
4508 for pp.gFree.n < 32 {
4509 // Prefer Gs with stacks.
4510 gp := sched.gFree.stack.pop()
4512 gp = sched.gFree.noStack.pop()
4521 unlock(&sched.gFree.lock)
4524 gp := pp.gFree.pop()
4529 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4530 // Deallocate old stack. We kept it in gfput because it was the
4531 // right size when the goroutine was put on the free list, but
4532 // the right size has changed since then.
4533 systemstack(func() {
4540 if gp.stack.lo == 0 {
4541 // Stack was deallocated in gfput or just above. Allocate a new one.
4542 systemstack(func() {
4543 gp.stack = stackalloc(startingStackSize)
4545 gp.stackguard0 = gp.stack.lo + _StackGuard
4548 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4551 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4554 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4560 // Purge all cached G's from gfree list to the global list.
4561 func gfpurge(pp *p) {
4567 for !pp.gFree.empty() {
4568 gp := pp.gFree.pop()
4570 if gp.stack.lo == 0 {
4577 lock(&sched.gFree.lock)
4578 sched.gFree.noStack.pushAll(noStackQ)
4579 sched.gFree.stack.pushAll(stackQ)
4580 sched.gFree.n += inc
4581 unlock(&sched.gFree.lock)
4584 // Breakpoint executes a breakpoint trap.
4589 // dolockOSThread is called by LockOSThread and lockOSThread below
4590 // after they modify m.locked. Do not allow preemption during this call,
4591 // or else the m might be different in this function than in the caller.
4594 func dolockOSThread() {
4595 if GOARCH == "wasm" {
4596 return // no threads on wasm yet
4599 gp.m.lockedg.set(gp)
4600 gp.lockedm.set(gp.m)
4603 // LockOSThread wires the calling goroutine to its current operating system thread.
4604 // The calling goroutine will always execute in that thread,
4605 // and no other goroutine will execute in it,
4606 // until the calling goroutine has made as many calls to
4607 // UnlockOSThread as to LockOSThread.
4608 // If the calling goroutine exits without unlocking the thread,
4609 // the thread will be terminated.
4611 // All init functions are run on the startup thread. Calling LockOSThread
4612 // from an init function will cause the main function to be invoked on
4615 // A goroutine should call LockOSThread before calling OS services or
4616 // non-Go library functions that depend on per-thread state.
4619 func LockOSThread() {
4620 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4621 // If we need to start a new thread from the locked
4622 // thread, we need the template thread. Start it now
4623 // while we're in a known-good state.
4624 startTemplateThread()
4628 if gp.m.lockedExt == 0 {
4630 panic("LockOSThread nesting overflow")
4636 func lockOSThread() {
4637 getg().m.lockedInt++
4641 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4642 // after they update m->locked. Do not allow preemption during this call,
4643 // or else the m might be in different in this function than in the caller.
4646 func dounlockOSThread() {
4647 if GOARCH == "wasm" {
4648 return // no threads on wasm yet
4651 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4658 // UnlockOSThread undoes an earlier call to LockOSThread.
4659 // If this drops the number of active LockOSThread calls on the
4660 // calling goroutine to zero, it unwires the calling goroutine from
4661 // its fixed operating system thread.
4662 // If there are no active LockOSThread calls, this is a no-op.
4664 // Before calling UnlockOSThread, the caller must ensure that the OS
4665 // thread is suitable for running other goroutines. If the caller made
4666 // any permanent changes to the state of the thread that would affect
4667 // other goroutines, it should not call this function and thus leave
4668 // the goroutine locked to the OS thread until the goroutine (and
4669 // hence the thread) exits.
4672 func UnlockOSThread() {
4674 if gp.m.lockedExt == 0 {
4682 func unlockOSThread() {
4684 if gp.m.lockedInt == 0 {
4685 systemstack(badunlockosthread)
4691 func badunlockosthread() {
4692 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4695 func gcount() int32 {
4696 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4697 for _, pp := range allp {
4701 // All these variables can be changed concurrently, so the result can be inconsistent.
4702 // But at least the current goroutine is running.
4709 func mcount() int32 {
4710 return int32(sched.mnext - sched.nmfreed)
4714 signalLock atomic.Uint32
4716 // Must hold signalLock to write. Reads may be lock-free, but
4717 // signalLock should be taken to synchronize with changes.
4721 func _System() { _System() }
4722 func _ExternalCode() { _ExternalCode() }
4723 func _LostExternalCode() { _LostExternalCode() }
4724 func _GC() { _GC() }
4725 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4726 func _VDSO() { _VDSO() }
4728 // Called if we receive a SIGPROF signal.
4729 // Called by the signal handler, may run during STW.
4731 //go:nowritebarrierrec
4732 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4733 if prof.hz.Load() == 0 {
4737 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4738 // We must check this to avoid a deadlock between setcpuprofilerate
4739 // and the call to cpuprof.add, below.
4740 if mp != nil && mp.profilehz == 0 {
4744 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4745 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4746 // the critical section, it creates a deadlock (when writing the sample).
4747 // As a workaround, create a counter of SIGPROFs while in critical section
4748 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4749 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4750 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4751 if f := findfunc(pc); f.valid() {
4752 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4753 cpuprof.lostAtomic++
4757 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4758 // runtime/internal/atomic functions call into kernel
4759 // helpers on arm < 7. See
4760 // runtime/internal/atomic/sys_linux_arm.s.
4761 cpuprof.lostAtomic++
4766 // Profiling runs concurrently with GC, so it must not allocate.
4767 // Set a trap in case the code does allocate.
4768 // Note that on windows, one thread takes profiles of all the
4769 // other threads, so mp is usually not getg().m.
4770 // In fact mp may not even be stopped.
4771 // See golang.org/issue/17165.
4772 getg().m.mallocing++
4775 var stk [maxCPUProfStack]uintptr
4777 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4779 // Check cgoCallersUse to make sure that we are not
4780 // interrupting other code that is fiddling with
4781 // cgoCallers. We are running in a signal handler
4782 // with all signals blocked, so we don't have to worry
4783 // about any other code interrupting us.
4784 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4785 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4788 n += copy(stk[:], mp.cgoCallers[:cgoOff])
4789 mp.cgoCallers[0] = 0
4792 // Collect Go stack that leads to the cgo call.
4793 u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
4794 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4795 // Libcall, i.e. runtime syscall on windows.
4796 // Collect Go stack that leads to the call.
4797 u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
4798 } else if mp != nil && mp.vdsoSP != 0 {
4799 // VDSO call, e.g. nanotime1 on Linux.
4800 // Collect Go stack that leads to the call.
4801 u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
4803 u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
4805 n += tracebackPCs(&u, 0, stk[n:])
4808 // Normal traceback is impossible or has failed.
4809 // Account it against abstract "System" or "GC".
4812 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4813 } else if pc > firstmoduledata.etext {
4814 // "ExternalCode" is better than "etext".
4815 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4818 if mp.preemptoff != "" {
4819 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4821 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4825 if prof.hz.Load() != 0 {
4826 // Note: it can happen on Windows that we interrupted a system thread
4827 // with no g, so gp could nil. The other nil checks are done out of
4828 // caution, but not expected to be nil in practice.
4829 var tagPtr *unsafe.Pointer
4830 if gp != nil && gp.m != nil && gp.m.curg != nil {
4831 tagPtr = &gp.m.curg.labels
4833 cpuprof.add(tagPtr, stk[:n])
4837 if gp != nil && gp.m != nil {
4838 if gp.m.curg != nil {
4843 traceCPUSample(gprof, pp, stk[:n])
4845 getg().m.mallocing--
4848 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4849 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4850 func setcpuprofilerate(hz int32) {
4851 // Force sane arguments.
4856 // Disable preemption, otherwise we can be rescheduled to another thread
4857 // that has profiling enabled.
4861 // Stop profiler on this thread so that it is safe to lock prof.
4862 // if a profiling signal came in while we had prof locked,
4863 // it would deadlock.
4864 setThreadCPUProfiler(0)
4866 for !prof.signalLock.CompareAndSwap(0, 1) {
4869 if prof.hz.Load() != hz {
4870 setProcessCPUProfiler(hz)
4873 prof.signalLock.Store(0)
4876 sched.profilehz = hz
4880 setThreadCPUProfiler(hz)
4886 // init initializes pp, which may be a freshly allocated p or a
4887 // previously destroyed p, and transitions it to status _Pgcstop.
4888 func (pp *p) init(id int32) {
4890 pp.status = _Pgcstop
4891 pp.sudogcache = pp.sudogbuf[:0]
4892 pp.deferpool = pp.deferpoolbuf[:0]
4894 if pp.mcache == nil {
4897 throw("missing mcache?")
4899 // Use the bootstrap mcache0. Only one P will get
4900 // mcache0: the one with ID 0.
4903 pp.mcache = allocmcache()
4906 if raceenabled && pp.raceprocctx == 0 {
4908 pp.raceprocctx = raceprocctx0
4909 raceprocctx0 = 0 // bootstrap
4911 pp.raceprocctx = raceproccreate()
4914 lockInit(&pp.timersLock, lockRankTimers)
4916 // This P may get timers when it starts running. Set the mask here
4917 // since the P may not go through pidleget (notably P 0 on startup).
4919 // Similarly, we may not go through pidleget before this P starts
4920 // running if it is P 0 on startup.
4924 // destroy releases all of the resources associated with pp and
4925 // transitions it to status _Pdead.
4927 // sched.lock must be held and the world must be stopped.
4928 func (pp *p) destroy() {
4929 assertLockHeld(&sched.lock)
4930 assertWorldStopped()
4932 // Move all runnable goroutines to the global queue
4933 for pp.runqhead != pp.runqtail {
4934 // Pop from tail of local queue
4936 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4937 // Push onto head of global queue
4940 if pp.runnext != 0 {
4941 globrunqputhead(pp.runnext.ptr())
4944 if len(pp.timers) > 0 {
4945 plocal := getg().m.p.ptr()
4946 // The world is stopped, but we acquire timersLock to
4947 // protect against sysmon calling timeSleepUntil.
4948 // This is the only case where we hold the timersLock of
4949 // more than one P, so there are no deadlock concerns.
4950 lock(&plocal.timersLock)
4951 lock(&pp.timersLock)
4952 moveTimers(plocal, pp.timers)
4954 pp.numTimers.Store(0)
4955 pp.deletedTimers.Store(0)
4956 pp.timer0When.Store(0)
4957 unlock(&pp.timersLock)
4958 unlock(&plocal.timersLock)
4960 // Flush p's write barrier buffer.
4961 if gcphase != _GCoff {
4965 for i := range pp.sudogbuf {
4966 pp.sudogbuf[i] = nil
4968 pp.sudogcache = pp.sudogbuf[:0]
4969 for j := range pp.deferpoolbuf {
4970 pp.deferpoolbuf[j] = nil
4972 pp.deferpool = pp.deferpoolbuf[:0]
4973 systemstack(func() {
4974 for i := 0; i < pp.mspancache.len; i++ {
4975 // Safe to call since the world is stopped.
4976 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4978 pp.mspancache.len = 0
4980 pp.pcache.flush(&mheap_.pages)
4981 unlock(&mheap_.lock)
4983 freemcache(pp.mcache)
4988 if pp.timerRaceCtx != 0 {
4989 // The race detector code uses a callback to fetch
4990 // the proc context, so arrange for that callback
4991 // to see the right thing.
4992 // This hack only works because we are the only
4998 racectxend(pp.timerRaceCtx)
5003 raceprocdestroy(pp.raceprocctx)
5010 // Change number of processors.
5012 // sched.lock must be held, and the world must be stopped.
5014 // gcworkbufs must not be being modified by either the GC or the write barrier
5015 // code, so the GC must not be running if the number of Ps actually changes.
5017 // Returns list of Ps with local work, they need to be scheduled by the caller.
5018 func procresize(nprocs int32) *p {
5019 assertLockHeld(&sched.lock)
5020 assertWorldStopped()
5023 if old < 0 || nprocs <= 0 {
5024 throw("procresize: invalid arg")
5027 traceGomaxprocs(nprocs)
5030 // update statistics
5032 if sched.procresizetime != 0 {
5033 sched.totaltime += int64(old) * (now - sched.procresizetime)
5035 sched.procresizetime = now
5037 maskWords := (nprocs + 31) / 32
5039 // Grow allp if necessary.
5040 if nprocs > int32(len(allp)) {
5041 // Synchronize with retake, which could be running
5042 // concurrently since it doesn't run on a P.
5044 if nprocs <= int32(cap(allp)) {
5045 allp = allp[:nprocs]
5047 nallp := make([]*p, nprocs)
5048 // Copy everything up to allp's cap so we
5049 // never lose old allocated Ps.
5050 copy(nallp, allp[:cap(allp)])
5054 if maskWords <= int32(cap(idlepMask)) {
5055 idlepMask = idlepMask[:maskWords]
5056 timerpMask = timerpMask[:maskWords]
5058 nidlepMask := make([]uint32, maskWords)
5059 // No need to copy beyond len, old Ps are irrelevant.
5060 copy(nidlepMask, idlepMask)
5061 idlepMask = nidlepMask
5063 ntimerpMask := make([]uint32, maskWords)
5064 copy(ntimerpMask, timerpMask)
5065 timerpMask = ntimerpMask
5070 // initialize new P's
5071 for i := old; i < nprocs; i++ {
5077 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5081 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5082 // continue to use the current P
5083 gp.m.p.ptr().status = _Prunning
5084 gp.m.p.ptr().mcache.prepareForSweep()
5086 // release the current P and acquire allp[0].
5088 // We must do this before destroying our current P
5089 // because p.destroy itself has write barriers, so we
5090 // need to do that from a valid P.
5093 // Pretend that we were descheduled
5094 // and then scheduled again to keep
5097 traceProcStop(gp.m.p.ptr())
5111 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5114 // release resources from unused P's
5115 for i := nprocs; i < old; i++ {
5118 // can't free P itself because it can be referenced by an M in syscall
5122 if int32(len(allp)) != nprocs {
5124 allp = allp[:nprocs]
5125 idlepMask = idlepMask[:maskWords]
5126 timerpMask = timerpMask[:maskWords]
5131 for i := nprocs - 1; i >= 0; i-- {
5133 if gp.m.p.ptr() == pp {
5141 pp.link.set(runnablePs)
5145 stealOrder.reset(uint32(nprocs))
5146 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5147 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5149 // Notify the limiter that the amount of procs has changed.
5150 gcCPULimiter.resetCapacity(now, nprocs)
5155 // Associate p and the current m.
5157 // This function is allowed to have write barriers even if the caller
5158 // isn't because it immediately acquires pp.
5160 //go:yeswritebarrierrec
5161 func acquirep(pp *p) {
5162 // Do the part that isn't allowed to have write barriers.
5165 // Have p; write barriers now allowed.
5167 // Perform deferred mcache flush before this P can allocate
5168 // from a potentially stale mcache.
5169 pp.mcache.prepareForSweep()
5176 // wirep is the first step of acquirep, which actually associates the
5177 // current M to pp. This is broken out so we can disallow write
5178 // barriers for this part, since we don't yet have a P.
5180 //go:nowritebarrierrec
5186 throw("wirep: already in go")
5188 if pp.m != 0 || pp.status != _Pidle {
5193 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5194 throw("wirep: invalid p state")
5198 pp.status = _Prunning
5201 // Disassociate p and the current m.
5202 func releasep() *p {
5206 throw("releasep: invalid arg")
5209 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5210 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5211 throw("releasep: invalid p state")
5214 traceProcStop(gp.m.p.ptr())
5222 func incidlelocked(v int32) {
5224 sched.nmidlelocked += v
5231 // Check for deadlock situation.
5232 // The check is based on number of running M's, if 0 -> deadlock.
5233 // sched.lock must be held.
5235 assertLockHeld(&sched.lock)
5237 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5238 // there are no running goroutines. The calling program is
5239 // assumed to be running.
5240 if islibrary || isarchive {
5244 // If we are dying because of a signal caught on an already idle thread,
5245 // freezetheworld will cause all running threads to block.
5246 // And runtime will essentially enter into deadlock state,
5247 // except that there is a thread that will call exit soon.
5248 if panicking.Load() > 0 {
5252 // If we are not running under cgo, but we have an extra M then account
5253 // for it. (It is possible to have an extra M on Windows without cgo to
5254 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5257 if !iscgo && cgoHasExtraM {
5258 mp := lockextra(true)
5259 haveExtraM := extraMCount > 0
5266 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5271 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5272 throw("checkdead: inconsistent counts")
5276 forEachG(func(gp *g) {
5277 if isSystemGoroutine(gp, false) {
5280 s := readgstatus(gp)
5281 switch s &^ _Gscan {
5288 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5289 throw("checkdead: runnable g")
5292 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5293 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5294 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5297 // Maybe jump time forward for playground.
5299 if when := timeSleepUntil(); when < maxWhen {
5302 // Start an M to steal the timer.
5303 pp, _ := pidleget(faketime)
5305 // There should always be a free P since
5306 // nothing is running.
5307 throw("checkdead: no p for timer")
5311 // There should always be a free M since
5312 // nothing is running.
5313 throw("checkdead: no m for timer")
5315 // M must be spinning to steal. We set this to be
5316 // explicit, but since this is the only M it would
5317 // become spinning on its own anyways.
5318 sched.nmspinning.Add(1)
5321 notewakeup(&mp.park)
5326 // There are no goroutines running, so we can look at the P's.
5327 for _, pp := range allp {
5328 if len(pp.timers) > 0 {
5333 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5334 fatal("all goroutines are asleep - deadlock!")
5337 // forcegcperiod is the maximum time in nanoseconds between garbage
5338 // collections. If we go this long without a garbage collection, one
5339 // is forced to run.
5341 // This is a variable for testing purposes. It normally doesn't change.
5342 var forcegcperiod int64 = 2 * 60 * 1e9
5344 // needSysmonWorkaround is true if the workaround for
5345 // golang.org/issue/42515 is needed on NetBSD.
5346 var needSysmonWorkaround bool = false
5348 // Always runs without a P, so write barriers are not allowed.
5350 //go:nowritebarrierrec
5357 lasttrace := int64(0)
5358 idle := 0 // how many cycles in succession we had not wokeup somebody
5362 if idle == 0 { // start with 20us sleep...
5364 } else if idle > 50 { // start doubling the sleep after 1ms...
5367 if delay > 10*1000 { // up to 10ms
5372 // sysmon should not enter deep sleep if schedtrace is enabled so that
5373 // it can print that information at the right time.
5375 // It should also not enter deep sleep if there are any active P's so
5376 // that it can retake P's from syscalls, preempt long running G's, and
5377 // poll the network if all P's are busy for long stretches.
5379 // It should wakeup from deep sleep if any P's become active either due
5380 // to exiting a syscall or waking up due to a timer expiring so that it
5381 // can resume performing those duties. If it wakes from a syscall it
5382 // resets idle and delay as a bet that since it had retaken a P from a
5383 // syscall before, it may need to do it again shortly after the
5384 // application starts work again. It does not reset idle when waking
5385 // from a timer to avoid adding system load to applications that spend
5386 // most of their time sleeping.
5388 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5390 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5391 syscallWake := false
5392 next := timeSleepUntil()
5394 sched.sysmonwait.Store(true)
5396 // Make wake-up period small enough
5397 // for the sampling to be correct.
5398 sleep := forcegcperiod / 2
5399 if next-now < sleep {
5402 shouldRelax := sleep >= osRelaxMinNS
5406 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5411 sched.sysmonwait.Store(false)
5412 noteclear(&sched.sysmonnote)
5422 lock(&sched.sysmonlock)
5423 // Update now in case we blocked on sysmonnote or spent a long time
5424 // blocked on schedlock or sysmonlock above.
5427 // trigger libc interceptors if needed
5428 if *cgo_yield != nil {
5429 asmcgocall(*cgo_yield, nil)
5431 // poll network if not polled for more than 10ms
5432 lastpoll := sched.lastpoll.Load()
5433 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5434 sched.lastpoll.CompareAndSwap(lastpoll, now)
5435 list := netpoll(0) // non-blocking - returns list of goroutines
5437 // Need to decrement number of idle locked M's
5438 // (pretending that one more is running) before injectglist.
5439 // Otherwise it can lead to the following situation:
5440 // injectglist grabs all P's but before it starts M's to run the P's,
5441 // another M returns from syscall, finishes running its G,
5442 // observes that there is no work to do and no other running M's
5443 // and reports deadlock.
5449 if GOOS == "netbsd" && needSysmonWorkaround {
5450 // netpoll is responsible for waiting for timer
5451 // expiration, so we typically don't have to worry
5452 // about starting an M to service timers. (Note that
5453 // sleep for timeSleepUntil above simply ensures sysmon
5454 // starts running again when that timer expiration may
5455 // cause Go code to run again).
5457 // However, netbsd has a kernel bug that sometimes
5458 // misses netpollBreak wake-ups, which can lead to
5459 // unbounded delays servicing timers. If we detect this
5460 // overrun, then startm to get something to handle the
5463 // See issue 42515 and
5464 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5465 if next := timeSleepUntil(); next < now {
5469 if scavenger.sysmonWake.Load() != 0 {
5470 // Kick the scavenger awake if someone requested it.
5473 // retake P's blocked in syscalls
5474 // and preempt long running G's
5475 if retake(now) != 0 {
5480 // check if we need to force a GC
5481 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5483 forcegc.idle.Store(false)
5485 list.push(forcegc.g)
5487 unlock(&forcegc.lock)
5489 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5491 schedtrace(debug.scheddetail > 0)
5493 unlock(&sched.sysmonlock)
5497 type sysmontick struct {
5504 // forcePreemptNS is the time slice given to a G before it is
5506 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5508 func retake(now int64) uint32 {
5510 // Prevent allp slice changes. This lock will be completely
5511 // uncontended unless we're already stopping the world.
5513 // We can't use a range loop over allp because we may
5514 // temporarily drop the allpLock. Hence, we need to re-fetch
5515 // allp each time around the loop.
5516 for i := 0; i < len(allp); i++ {
5519 // This can happen if procresize has grown
5520 // allp but not yet created new Ps.
5523 pd := &pp.sysmontick
5526 if s == _Prunning || s == _Psyscall {
5527 // Preempt G if it's running for too long.
5528 t := int64(pp.schedtick)
5529 if int64(pd.schedtick) != t {
5530 pd.schedtick = uint32(t)
5532 } else if pd.schedwhen+forcePreemptNS <= now {
5534 // In case of syscall, preemptone() doesn't
5535 // work, because there is no M wired to P.
5540 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5541 t := int64(pp.syscalltick)
5542 if !sysretake && int64(pd.syscalltick) != t {
5543 pd.syscalltick = uint32(t)
5544 pd.syscallwhen = now
5547 // On the one hand we don't want to retake Ps if there is no other work to do,
5548 // but on the other hand we want to retake them eventually
5549 // because they can prevent the sysmon thread from deep sleep.
5550 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5553 // Drop allpLock so we can take sched.lock.
5555 // Need to decrement number of idle locked M's
5556 // (pretending that one more is running) before the CAS.
5557 // Otherwise the M from which we retake can exit the syscall,
5558 // increment nmidle and report deadlock.
5560 if atomic.Cas(&pp.status, s, _Pidle) {
5577 // Tell all goroutines that they have been preempted and they should stop.
5578 // This function is purely best-effort. It can fail to inform a goroutine if a
5579 // processor just started running it.
5580 // No locks need to be held.
5581 // Returns true if preemption request was issued to at least one goroutine.
5582 func preemptall() bool {
5584 for _, pp := range allp {
5585 if pp.status != _Prunning {
5595 // Tell the goroutine running on processor P to stop.
5596 // This function is purely best-effort. It can incorrectly fail to inform the
5597 // goroutine. It can inform the wrong goroutine. Even if it informs the
5598 // correct goroutine, that goroutine might ignore the request if it is
5599 // simultaneously executing newstack.
5600 // No lock needs to be held.
5601 // Returns true if preemption request was issued.
5602 // The actual preemption will happen at some point in the future
5603 // and will be indicated by the gp->status no longer being
5605 func preemptone(pp *p) bool {
5607 if mp == nil || mp == getg().m {
5611 if gp == nil || gp == mp.g0 {
5617 // Every call in a goroutine checks for stack overflow by
5618 // comparing the current stack pointer to gp->stackguard0.
5619 // Setting gp->stackguard0 to StackPreempt folds
5620 // preemption into the normal stack overflow check.
5621 gp.stackguard0 = stackPreempt
5623 // Request an async preemption of this P.
5624 if preemptMSupported && debug.asyncpreemptoff == 0 {
5634 func schedtrace(detailed bool) {
5641 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)
5643 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5645 // We must be careful while reading data from P's, M's and G's.
5646 // Even if we hold schedlock, most data can be changed concurrently.
5647 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5648 for i, pp := range allp {
5650 h := atomic.Load(&pp.runqhead)
5651 t := atomic.Load(&pp.runqtail)
5653 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5659 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5661 // In non-detailed mode format lengths of per-P run queues as:
5662 // [len1 len2 len3 len4]
5668 if i == len(allp)-1 {
5679 for mp := allm; mp != nil; mp = mp.alllink {
5681 print(" M", mp.id, ": p=")
5693 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5694 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5702 forEachG(func(gp *g) {
5703 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5710 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5720 // schedEnableUser enables or disables the scheduling of user
5723 // This does not stop already running user goroutines, so the caller
5724 // should first stop the world when disabling user goroutines.
5725 func schedEnableUser(enable bool) {
5727 if sched.disable.user == !enable {
5731 sched.disable.user = !enable
5733 n := sched.disable.n
5735 globrunqputbatch(&sched.disable.runnable, n)
5737 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5745 // schedEnabled reports whether gp should be scheduled. It returns
5746 // false is scheduling of gp is disabled.
5748 // sched.lock must be held.
5749 func schedEnabled(gp *g) bool {
5750 assertLockHeld(&sched.lock)
5752 if sched.disable.user {
5753 return isSystemGoroutine(gp, true)
5758 // Put mp on midle list.
5759 // sched.lock must be held.
5760 // May run during STW, so write barriers are not allowed.
5762 //go:nowritebarrierrec
5764 assertLockHeld(&sched.lock)
5766 mp.schedlink = sched.midle
5772 // Try to get an m from midle list.
5773 // sched.lock must be held.
5774 // May run during STW, so write barriers are not allowed.
5776 //go:nowritebarrierrec
5778 assertLockHeld(&sched.lock)
5780 mp := sched.midle.ptr()
5782 sched.midle = mp.schedlink
5788 // Put gp on the global runnable queue.
5789 // sched.lock must be held.
5790 // May run during STW, so write barriers are not allowed.
5792 //go:nowritebarrierrec
5793 func globrunqput(gp *g) {
5794 assertLockHeld(&sched.lock)
5796 sched.runq.pushBack(gp)
5800 // Put gp at the head of the global runnable queue.
5801 // sched.lock must be held.
5802 // May run during STW, so write barriers are not allowed.
5804 //go:nowritebarrierrec
5805 func globrunqputhead(gp *g) {
5806 assertLockHeld(&sched.lock)
5812 // Put a batch of runnable goroutines on the global runnable queue.
5813 // This clears *batch.
5814 // sched.lock must be held.
5815 // May run during STW, so write barriers are not allowed.
5817 //go:nowritebarrierrec
5818 func globrunqputbatch(batch *gQueue, n int32) {
5819 assertLockHeld(&sched.lock)
5821 sched.runq.pushBackAll(*batch)
5826 // Try get a batch of G's from the global runnable queue.
5827 // sched.lock must be held.
5828 func globrunqget(pp *p, max int32) *g {
5829 assertLockHeld(&sched.lock)
5831 if sched.runqsize == 0 {
5835 n := sched.runqsize/gomaxprocs + 1
5836 if n > sched.runqsize {
5839 if max > 0 && n > max {
5842 if n > int32(len(pp.runq))/2 {
5843 n = int32(len(pp.runq)) / 2
5848 gp := sched.runq.pop()
5851 gp1 := sched.runq.pop()
5852 runqput(pp, gp1, false)
5857 // pMask is an atomic bitstring with one bit per P.
5860 // read returns true if P id's bit is set.
5861 func (p pMask) read(id uint32) bool {
5863 mask := uint32(1) << (id % 32)
5864 return (atomic.Load(&p[word]) & mask) != 0
5867 // set sets P id's bit.
5868 func (p pMask) set(id int32) {
5870 mask := uint32(1) << (id % 32)
5871 atomic.Or(&p[word], mask)
5874 // clear clears P id's bit.
5875 func (p pMask) clear(id int32) {
5877 mask := uint32(1) << (id % 32)
5878 atomic.And(&p[word], ^mask)
5881 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5883 // Ideally, the timer mask would be kept immediately consistent on any timer
5884 // operations. Unfortunately, updating a shared global data structure in the
5885 // timer hot path adds too much overhead in applications frequently switching
5886 // between no timers and some timers.
5888 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5889 // running P (returned by pidleget) may add a timer at any time, so its mask
5890 // must be set. An idle P (passed to pidleput) cannot add new timers while
5891 // idle, so if it has no timers at that time, its mask may be cleared.
5893 // Thus, we get the following effects on timer-stealing in findrunnable:
5895 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5896 // (for work- or timer-stealing; this is the ideal case).
5897 // - Running Ps must always be checked.
5898 // - Idle Ps whose timers are stolen must continue to be checked until they run
5899 // again, even after timer expiration.
5901 // When the P starts running again, the mask should be set, as a timer may be
5902 // added at any time.
5904 // TODO(prattmic): Additional targeted updates may improve the above cases.
5905 // e.g., updating the mask when stealing a timer.
5906 func updateTimerPMask(pp *p) {
5907 if pp.numTimers.Load() > 0 {
5911 // Looks like there are no timers, however another P may transiently
5912 // decrement numTimers when handling a timerModified timer in
5913 // checkTimers. We must take timersLock to serialize with these changes.
5914 lock(&pp.timersLock)
5915 if pp.numTimers.Load() == 0 {
5916 timerpMask.clear(pp.id)
5918 unlock(&pp.timersLock)
5921 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5922 // to nanotime or zero. Returns now or the current time if now was zero.
5924 // This releases ownership of p. Once sched.lock is released it is no longer
5927 // sched.lock must be held.
5929 // May run during STW, so write barriers are not allowed.
5931 //go:nowritebarrierrec
5932 func pidleput(pp *p, now int64) int64 {
5933 assertLockHeld(&sched.lock)
5936 throw("pidleput: P has non-empty run queue")
5941 updateTimerPMask(pp) // clear if there are no timers.
5942 idlepMask.set(pp.id)
5943 pp.link = sched.pidle
5946 if !pp.limiterEvent.start(limiterEventIdle, now) {
5947 throw("must be able to track idle limiter event")
5952 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5954 // sched.lock must be held.
5956 // May run during STW, so write barriers are not allowed.
5958 //go:nowritebarrierrec
5959 func pidleget(now int64) (*p, int64) {
5960 assertLockHeld(&sched.lock)
5962 pp := sched.pidle.ptr()
5964 // Timer may get added at any time now.
5968 timerpMask.set(pp.id)
5969 idlepMask.clear(pp.id)
5970 sched.pidle = pp.link
5971 sched.npidle.Add(-1)
5972 pp.limiterEvent.stop(limiterEventIdle, now)
5977 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5978 // This is called by spinning Ms (or callers than need a spinning M) that have
5979 // found work. If no P is available, this must synchronized with non-spinning
5980 // Ms that may be preparing to drop their P without discovering this work.
5982 // sched.lock must be held.
5984 // May run during STW, so write barriers are not allowed.
5986 //go:nowritebarrierrec
5987 func pidlegetSpinning(now int64) (*p, int64) {
5988 assertLockHeld(&sched.lock)
5990 pp, now := pidleget(now)
5992 // See "Delicate dance" comment in findrunnable. We found work
5993 // that we cannot take, we must synchronize with non-spinning
5994 // Ms that may be preparing to drop their P.
5995 sched.needspinning.Store(1)
6002 // runqempty reports whether pp has no Gs on its local run queue.
6003 // It never returns true spuriously.
6004 func runqempty(pp *p) bool {
6005 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
6006 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
6007 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
6008 // does not mean the queue is empty.
6010 head := atomic.Load(&pp.runqhead)
6011 tail := atomic.Load(&pp.runqtail)
6012 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
6013 if tail == atomic.Load(&pp.runqtail) {
6014 return head == tail && runnext == 0
6019 // To shake out latent assumptions about scheduling order,
6020 // we introduce some randomness into scheduling decisions
6021 // when running with the race detector.
6022 // The need for this was made obvious by changing the
6023 // (deterministic) scheduling order in Go 1.5 and breaking
6024 // many poorly-written tests.
6025 // With the randomness here, as long as the tests pass
6026 // consistently with -race, they shouldn't have latent scheduling
6028 const randomizeScheduler = raceenabled
6030 // runqput tries to put g on the local runnable queue.
6031 // If next is false, runqput adds g to the tail of the runnable queue.
6032 // If next is true, runqput puts g in the pp.runnext slot.
6033 // If the run queue is full, runnext puts g on the global queue.
6034 // Executed only by the owner P.
6035 func runqput(pp *p, gp *g, next bool) {
6036 if randomizeScheduler && next && fastrandn(2) == 0 {
6042 oldnext := pp.runnext
6043 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
6049 // Kick the old runnext out to the regular run queue.
6054 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6056 if t-h < uint32(len(pp.runq)) {
6057 pp.runq[t%uint32(len(pp.runq))].set(gp)
6058 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
6061 if runqputslow(pp, gp, h, t) {
6064 // the queue is not full, now the put above must succeed
6068 // Put g and a batch of work from local runnable queue on global queue.
6069 // Executed only by the owner P.
6070 func runqputslow(pp *p, gp *g, h, t uint32) bool {
6071 var batch [len(pp.runq)/2 + 1]*g
6073 // First, grab a batch from local queue.
6076 if n != uint32(len(pp.runq)/2) {
6077 throw("runqputslow: queue is not full")
6079 for i := uint32(0); i < n; i++ {
6080 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6082 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6087 if randomizeScheduler {
6088 for i := uint32(1); i <= n; i++ {
6089 j := fastrandn(i + 1)
6090 batch[i], batch[j] = batch[j], batch[i]
6094 // Link the goroutines.
6095 for i := uint32(0); i < n; i++ {
6096 batch[i].schedlink.set(batch[i+1])
6099 q.head.set(batch[0])
6100 q.tail.set(batch[n])
6102 // Now put the batch on global queue.
6104 globrunqputbatch(&q, int32(n+1))
6109 // runqputbatch tries to put all the G's on q on the local runnable queue.
6110 // If the queue is full, they are put on the global queue; in that case
6111 // this will temporarily acquire the scheduler lock.
6112 // Executed only by the owner P.
6113 func runqputbatch(pp *p, q *gQueue, qsize int) {
6114 h := atomic.LoadAcq(&pp.runqhead)
6117 for !q.empty() && t-h < uint32(len(pp.runq)) {
6119 pp.runq[t%uint32(len(pp.runq))].set(gp)
6125 if randomizeScheduler {
6126 off := func(o uint32) uint32 {
6127 return (pp.runqtail + o) % uint32(len(pp.runq))
6129 for i := uint32(1); i < n; i++ {
6130 j := fastrandn(i + 1)
6131 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6135 atomic.StoreRel(&pp.runqtail, t)
6138 globrunqputbatch(q, int32(qsize))
6143 // Get g from local runnable queue.
6144 // If inheritTime is true, gp should inherit the remaining time in the
6145 // current time slice. Otherwise, it should start a new time slice.
6146 // Executed only by the owner P.
6147 func runqget(pp *p) (gp *g, inheritTime bool) {
6148 // If there's a runnext, it's the next G to run.
6150 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6151 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6152 // Hence, there's no need to retry this CAS if it fails.
6153 if next != 0 && pp.runnext.cas(next, 0) {
6154 return next.ptr(), true
6158 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6163 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6164 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6170 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6171 // Executed only by the owner P.
6172 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6173 oldNext := pp.runnext
6174 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6175 drainQ.pushBack(oldNext.ptr())
6180 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6186 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6190 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6194 // We've inverted the order in which it gets G's from the local P's runnable queue
6195 // and then advances the head pointer because we don't want to mess up the statuses of G's
6196 // while runqdrain() and runqsteal() are running in parallel.
6197 // Thus we should advance the head pointer before draining the local P into a gQueue,
6198 // so that we can update any gp.schedlink only after we take the full ownership of G,
6199 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6200 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6201 for i := uint32(0); i < qn; i++ {
6202 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6209 // Grabs a batch of goroutines from pp's runnable queue into batch.
6210 // Batch is a ring buffer starting at batchHead.
6211 // Returns number of grabbed goroutines.
6212 // Can be executed by any P.
6213 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6215 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6216 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6221 // Try to steal from pp.runnext.
6222 if next := pp.runnext; next != 0 {
6223 if pp.status == _Prunning {
6224 // Sleep to ensure that pp isn't about to run the g
6225 // we are about to steal.
6226 // The important use case here is when the g running
6227 // on pp ready()s another g and then almost
6228 // immediately blocks. Instead of stealing runnext
6229 // in this window, back off to give pp a chance to
6230 // schedule runnext. This will avoid thrashing gs
6231 // between different Ps.
6232 // A sync chan send/recv takes ~50ns as of time of
6233 // writing, so 3us gives ~50x overshoot.
6234 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6237 // On some platforms system timer granularity is
6238 // 1-15ms, which is way too much for this
6239 // optimization. So just yield.
6243 if !pp.runnext.cas(next, 0) {
6246 batch[batchHead%uint32(len(batch))] = next
6252 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6255 for i := uint32(0); i < n; i++ {
6256 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6257 batch[(batchHead+i)%uint32(len(batch))] = g
6259 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6265 // Steal half of elements from local runnable queue of p2
6266 // and put onto local runnable queue of p.
6267 // Returns one of the stolen elements (or nil if failed).
6268 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6270 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6275 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6279 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6280 if t-h+n >= uint32(len(pp.runq)) {
6281 throw("runqsteal: runq overflow")
6283 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6287 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6288 // be on one gQueue or gList at a time.
6289 type gQueue struct {
6294 // empty reports whether q is empty.
6295 func (q *gQueue) empty() bool {
6299 // push adds gp to the head of q.
6300 func (q *gQueue) push(gp *g) {
6301 gp.schedlink = q.head
6308 // pushBack adds gp to the tail of q.
6309 func (q *gQueue) pushBack(gp *g) {
6312 q.tail.ptr().schedlink.set(gp)
6319 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6321 func (q *gQueue) pushBackAll(q2 gQueue) {
6325 q2.tail.ptr().schedlink = 0
6327 q.tail.ptr().schedlink = q2.head
6334 // pop removes and returns the head of queue q. It returns nil if
6336 func (q *gQueue) pop() *g {
6339 q.head = gp.schedlink
6347 // popList takes all Gs in q and returns them as a gList.
6348 func (q *gQueue) popList() gList {
6349 stack := gList{q.head}
6354 // A gList is a list of Gs linked through g.schedlink. A G can only be
6355 // on one gQueue or gList at a time.
6360 // empty reports whether l is empty.
6361 func (l *gList) empty() bool {
6365 // push adds gp to the head of l.
6366 func (l *gList) push(gp *g) {
6367 gp.schedlink = l.head
6371 // pushAll prepends all Gs in q to l.
6372 func (l *gList) pushAll(q gQueue) {
6374 q.tail.ptr().schedlink = l.head
6379 // pop removes and returns the head of l. If l is empty, it returns nil.
6380 func (l *gList) pop() *g {
6383 l.head = gp.schedlink
6388 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6389 func setMaxThreads(in int) (out int) {
6391 out = int(sched.maxmcount)
6392 if in > 0x7fffffff { // MaxInt32
6393 sched.maxmcount = 0x7fffffff
6395 sched.maxmcount = int32(in)
6403 func procPin() int {
6408 return int(mp.p.ptr().id)
6417 //go:linkname sync_runtime_procPin sync.runtime_procPin
6419 func sync_runtime_procPin() int {
6423 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6425 func sync_runtime_procUnpin() {
6429 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6431 func sync_atomic_runtime_procPin() int {
6435 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6437 func sync_atomic_runtime_procUnpin() {
6441 // Active spinning for sync.Mutex.
6443 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6445 func sync_runtime_canSpin(i int) bool {
6446 // sync.Mutex is cooperative, so we are conservative with spinning.
6447 // Spin only few times and only if running on a multicore machine and
6448 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6449 // As opposed to runtime mutex we don't do passive spinning here,
6450 // because there can be work on global runq or on other Ps.
6451 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6454 if p := getg().m.p.ptr(); !runqempty(p) {
6460 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6462 func sync_runtime_doSpin() {
6463 procyield(active_spin_cnt)
6466 var stealOrder randomOrder
6468 // randomOrder/randomEnum are helper types for randomized work stealing.
6469 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6470 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6471 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6472 type randomOrder struct {
6477 type randomEnum struct {
6484 func (ord *randomOrder) reset(count uint32) {
6486 ord.coprimes = ord.coprimes[:0]
6487 for i := uint32(1); i <= count; i++ {
6488 if gcd(i, count) == 1 {
6489 ord.coprimes = append(ord.coprimes, i)
6494 func (ord *randomOrder) start(i uint32) randomEnum {
6498 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6502 func (enum *randomEnum) done() bool {
6503 return enum.i == enum.count
6506 func (enum *randomEnum) next() {
6508 enum.pos = (enum.pos + enum.inc) % enum.count
6511 func (enum *randomEnum) position() uint32 {
6515 func gcd(a, b uint32) uint32 {
6522 // An initTask represents the set of initializations that need to be done for a package.
6523 // Keep in sync with ../../test/initempty.go:initTask
6524 type initTask struct {
6525 // TODO: pack the first 3 fields more tightly?
6526 state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
6529 // followed by ndeps instances of an *initTask, one per package depended on
6530 // followed by nfns pcs, one per init function to run
6533 // inittrace stores statistics for init functions which are
6534 // updated by malloc and newproc when active is true.
6535 var inittrace tracestat
6537 type tracestat struct {
6538 active bool // init tracing activation status
6539 id uint64 // init goroutine id
6540 allocs uint64 // heap allocations
6541 bytes uint64 // heap allocated bytes
6544 func doInit(t *initTask) {
6546 case 2: // fully initialized
6548 case 1: // initialization in progress
6549 throw("recursive call during initialization - linker skew")
6550 default: // not initialized yet
6551 t.state = 1 // initialization in progress
6553 for i := uintptr(0); i < t.ndeps; i++ {
6554 p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
6555 t2 := *(**initTask)(p)
6560 t.state = 2 // initialization done
6569 if inittrace.active {
6571 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6575 firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
6576 for i := uintptr(0); i < t.nfns; i++ {
6577 p := add(firstFunc, i*goarch.PtrSize)
6578 f := *(*func())(unsafe.Pointer(&p))
6582 if inittrace.active {
6584 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6587 f := *(*func())(unsafe.Pointer(&firstFunc))
6588 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6591 print("init ", pkg, " @")
6592 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6593 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6594 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6595 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6599 t.state = 2 // initialization done