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 or global run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
121 // This slice records the initializing tasks that need to be
122 // done to start up the runtime. It is built by the linker.
123 var runtime_inittasks []*initTask
125 // main_init_done is a signal used by cgocallbackg that initialization
126 // has been completed. It is made before _cgo_notify_runtime_init_done,
127 // so all cgo calls can rely on it existing. When main_init is complete,
128 // it is closed, meaning cgocallbackg can reliably receive from it.
129 var main_init_done chan bool
131 //go:linkname main_main main.main
134 // mainStarted indicates that the main M has started.
137 // runtimeInitTime is the nanotime() at which the runtime started.
138 var runtimeInitTime int64
140 // Value to use for signal mask for newly created M's.
141 var initSigmask sigset
143 // The main goroutine.
147 // Racectx of m0->g0 is used only as the parent of the main goroutine.
148 // It must not be used for anything else.
151 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
152 // Using decimal instead of binary GB and MB because
153 // they look nicer in the stack overflow failure message.
154 if goarch.PtrSize == 8 {
155 maxstacksize = 1000000000
157 maxstacksize = 250000000
160 // An upper limit for max stack size. Used to avoid random crashes
161 // after calling SetMaxStack and trying to allocate a stack that is too big,
162 // since stackalloc works with 32-bit sizes.
163 maxstackceiling = 2 * maxstacksize
165 // Allow newproc to start new Ms.
168 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
170 newm(sysmon, nil, -1)
174 // Lock the main goroutine onto this, the main OS thread,
175 // during initialization. Most programs won't care, but a few
176 // do require certain calls to be made by the main thread.
177 // Those can arrange for main.main to run in the main thread
178 // by calling runtime.LockOSThread during initialization
179 // to preserve the lock.
183 throw("runtime.main not on m0")
186 // Record when the world started.
187 // Must be before doInit for tracing init.
188 runtimeInitTime = nanotime()
189 if runtimeInitTime == 0 {
190 throw("nanotime returning zero")
193 if debug.inittrace != 0 {
194 inittrace.id = getg().goid
195 inittrace.active = true
198 doInit(runtime_inittasks) // Must be before defer.
200 // Defer unlock so that runtime.Goexit during init does the unlock too.
210 main_init_done = make(chan bool)
212 if _cgo_pthread_key_created == nil {
213 throw("_cgo_pthread_key_created missing")
216 if _cgo_thread_start == nil {
217 throw("_cgo_thread_start missing")
219 if GOOS != "windows" {
220 if _cgo_setenv == nil {
221 throw("_cgo_setenv missing")
223 if _cgo_unsetenv == nil {
224 throw("_cgo_unsetenv missing")
227 if _cgo_notify_runtime_init_done == nil {
228 throw("_cgo_notify_runtime_init_done missing")
231 // Set the x_crosscall2_ptr C function pointer variable point to crosscall2.
232 if set_crosscall2 == nil {
233 throw("set_crosscall2 missing")
237 // Start the template thread in case we enter Go from
238 // a C-created thread and need to create a new thread.
239 startTemplateThread()
240 cgocall(_cgo_notify_runtime_init_done, nil)
243 // Run the initializing tasks. Depending on build mode this
244 // list can arrive a few different ways, but it will always
245 // contain the init tasks computed by the linker for all the
246 // packages in the program (excluding those added at runtime
247 // by package plugin). Run through the modules in dependency
248 // order (the order they are initialized by the dynamic
249 // loader, i.e. they are added to the moduledata linked list).
250 for m := &firstmoduledata; m != nil; m = m.next {
254 // Disable init tracing after main init done to avoid overhead
255 // of collecting statistics in malloc and newproc
256 inittrace.active = false
258 close(main_init_done)
263 if isarchive || islibrary {
264 // A program compiled with -buildmode=c-archive or c-shared
265 // has a main, but it is not executed.
268 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
271 runExitHooks(0) // run hooks now, since racefini does not return
275 // Make racy client program work: if panicking on
276 // another goroutine at the same time as main returns,
277 // let the other goroutine finish printing the panic trace.
278 // Once it does, it will exit. See issues 3934 and 20018.
279 if runningPanicDefers.Load() != 0 {
280 // Running deferred functions should not take long.
281 for c := 0; c < 1000; c++ {
282 if runningPanicDefers.Load() == 0 {
288 if panicking.Load() != 0 {
289 gopark(nil, nil, waitReasonPanicWait, traceBlockForever, 1)
300 // os_beforeExit is called from os.Exit(0).
302 //go:linkname os_beforeExit os.runtime_beforeExit
303 func os_beforeExit(exitCode int) {
304 runExitHooks(exitCode)
305 if exitCode == 0 && raceenabled {
310 // start forcegc helper goroutine
315 func forcegchelper() {
317 lockInit(&forcegc.lock, lockRankForcegc)
320 if forcegc.idle.Load() {
321 throw("forcegc: phase error")
323 forcegc.idle.Store(true)
324 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceBlockSystemGoroutine, 1)
325 // this goroutine is explicitly resumed by sysmon
326 if debug.gctrace > 0 {
329 // Time-triggered, fully concurrent.
330 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
334 // Gosched yields the processor, allowing other goroutines to run. It does not
335 // suspend the current goroutine, so execution resumes automatically.
343 // goschedguarded yields the processor like gosched, but also checks
344 // for forbidden states and opts out of the yield in those cases.
347 func goschedguarded() {
348 mcall(goschedguarded_m)
351 // goschedIfBusy yields the processor like gosched, but only does so if
352 // there are no idle Ps or if we're on the only P and there's nothing in
353 // the run queue. In both cases, there is freely available idle time.
356 func goschedIfBusy() {
358 // Call gosched if gp.preempt is set; we may be in a tight loop that
359 // doesn't otherwise yield.
360 if !gp.preempt && sched.npidle.Load() > 0 {
366 // Puts the current goroutine into a waiting state and calls unlockf on the
369 // If unlockf returns false, the goroutine is resumed.
371 // unlockf must not access this G's stack, as it may be moved between
372 // the call to gopark and the call to unlockf.
374 // Note that because unlockf is called after putting the G into a waiting
375 // state, the G may have already been readied by the time unlockf is called
376 // unless there is external synchronization preventing the G from being
377 // readied. If unlockf returns false, it must guarantee that the G cannot be
378 // externally readied.
380 // Reason explains why the goroutine has been parked. It is displayed in stack
381 // traces and heap dumps. Reasons should be unique and descriptive. Do not
382 // re-use reasons, add new ones.
383 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceReason traceBlockReason, traceskip int) {
384 if reason != waitReasonSleep {
385 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
389 status := readgstatus(gp)
390 if status != _Grunning && status != _Gscanrunning {
391 throw("gopark: bad g status")
394 mp.waitunlockf = unlockf
395 gp.waitreason = reason
396 mp.waitTraceBlockReason = traceReason
397 mp.waitTraceSkip = traceskip
399 // can't do anything that might move the G between Ms here.
403 // Puts the current goroutine into a waiting state and unlocks the lock.
404 // The goroutine can be made runnable again by calling goready(gp).
405 func goparkunlock(lock *mutex, reason waitReason, traceReason traceBlockReason, traceskip int) {
406 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceReason, traceskip)
409 func goready(gp *g, traceskip int) {
411 ready(gp, traceskip, true)
416 func acquireSudog() *sudog {
417 // Delicate dance: the semaphore implementation calls
418 // acquireSudog, acquireSudog calls new(sudog),
419 // new calls malloc, malloc can call the garbage collector,
420 // and the garbage collector calls the semaphore implementation
422 // Break the cycle by doing acquirem/releasem around new(sudog).
423 // The acquirem/releasem increments m.locks during new(sudog),
424 // which keeps the garbage collector from being invoked.
427 if len(pp.sudogcache) == 0 {
428 lock(&sched.sudoglock)
429 // First, try to grab a batch from central cache.
430 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
431 s := sched.sudogcache
432 sched.sudogcache = s.next
434 pp.sudogcache = append(pp.sudogcache, s)
436 unlock(&sched.sudoglock)
437 // If the central cache is empty, allocate a new one.
438 if len(pp.sudogcache) == 0 {
439 pp.sudogcache = append(pp.sudogcache, new(sudog))
442 n := len(pp.sudogcache)
443 s := pp.sudogcache[n-1]
444 pp.sudogcache[n-1] = nil
445 pp.sudogcache = pp.sudogcache[:n-1]
447 throw("acquireSudog: found s.elem != nil in cache")
454 func releaseSudog(s *sudog) {
456 throw("runtime: sudog with non-nil elem")
459 throw("runtime: sudog with non-false isSelect")
462 throw("runtime: sudog with non-nil next")
465 throw("runtime: sudog with non-nil prev")
467 if s.waitlink != nil {
468 throw("runtime: sudog with non-nil waitlink")
471 throw("runtime: sudog with non-nil c")
475 throw("runtime: releaseSudog with non-nil gp.param")
477 mp := acquirem() // avoid rescheduling to another P
479 if len(pp.sudogcache) == cap(pp.sudogcache) {
480 // Transfer half of local cache to the central cache.
481 var first, last *sudog
482 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
483 n := len(pp.sudogcache)
484 p := pp.sudogcache[n-1]
485 pp.sudogcache[n-1] = nil
486 pp.sudogcache = pp.sudogcache[:n-1]
494 lock(&sched.sudoglock)
495 last.next = sched.sudogcache
496 sched.sudogcache = first
497 unlock(&sched.sudoglock)
499 pp.sudogcache = append(pp.sudogcache, s)
503 // called from assembly.
504 func badmcall(fn func(*g)) {
505 throw("runtime: mcall called on m->g0 stack")
508 func badmcall2(fn func(*g)) {
509 throw("runtime: mcall function returned")
512 func badreflectcall() {
513 panic(plainError("arg size to reflect.call more than 1GB"))
517 //go:nowritebarrierrec
518 func badmorestackg0() {
519 writeErrStr("fatal: morestack on g0\n")
523 //go:nowritebarrierrec
524 func badmorestackgsignal() {
525 writeErrStr("fatal: morestack on gsignal\n")
533 func lockedOSThread() bool {
535 return gp.lockedm != 0 && gp.m.lockedg != 0
539 // allgs contains all Gs ever created (including dead Gs), and thus
542 // Access via the slice is protected by allglock or stop-the-world.
543 // Readers that cannot take the lock may (carefully!) use the atomic
548 // allglen and allgptr are atomic variables that contain len(allgs) and
549 // &allgs[0] respectively. Proper ordering depends on totally-ordered
550 // loads and stores. Writes are protected by allglock.
552 // allgptr is updated before allglen. Readers should read allglen
553 // before allgptr to ensure that allglen is always <= len(allgptr). New
554 // Gs appended during the race can be missed. For a consistent view of
555 // all Gs, allglock must be held.
557 // allgptr copies should always be stored as a concrete type or
558 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
559 // even if it points to a stale array.
564 func allgadd(gp *g) {
565 if readgstatus(gp) == _Gidle {
566 throw("allgadd: bad status Gidle")
570 allgs = append(allgs, gp)
571 if &allgs[0] != allgptr {
572 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
574 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
578 // allGsSnapshot returns a snapshot of the slice of all Gs.
580 // The world must be stopped or allglock must be held.
581 func allGsSnapshot() []*g {
582 assertWorldStoppedOrLockHeld(&allglock)
584 // Because the world is stopped or allglock is held, allgadd
585 // cannot happen concurrently with this. allgs grows
586 // monotonically and existing entries never change, so we can
587 // simply return a copy of the slice header. For added safety,
588 // we trim everything past len because that can still change.
589 return allgs[:len(allgs):len(allgs)]
592 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
593 func atomicAllG() (**g, uintptr) {
594 length := atomic.Loaduintptr(&allglen)
595 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
599 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
600 func atomicAllGIndex(ptr **g, i uintptr) *g {
601 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
604 // forEachG calls fn on every G from allgs.
606 // forEachG takes a lock to exclude concurrent addition of new Gs.
607 func forEachG(fn func(gp *g)) {
609 for _, gp := range allgs {
615 // forEachGRace calls fn on every G from allgs.
617 // forEachGRace avoids locking, but does not exclude addition of new Gs during
618 // execution, which may be missed.
619 func forEachGRace(fn func(gp *g)) {
620 ptr, length := atomicAllG()
621 for i := uintptr(0); i < length; i++ {
622 gp := atomicAllGIndex(ptr, i)
629 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
630 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
634 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
635 // value of the GODEBUG environment variable.
636 func cpuinit(env string) {
638 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
639 cpu.DebugOptions = true
643 // Support cpu feature variables are used in code generated by the compiler
644 // to guard execution of instructions that can not be assumed to be always supported.
647 x86HasPOPCNT = cpu.X86.HasPOPCNT
648 x86HasSSE41 = cpu.X86.HasSSE41
649 x86HasFMA = cpu.X86.HasFMA
652 armHasVFPv4 = cpu.ARM.HasVFPv4
655 arm64HasATOMICS = cpu.ARM64.HasATOMICS
659 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
660 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
661 // early before much of the runtime is initialized.
662 func getGodebugEarly() string {
663 const prefix = "GODEBUG="
666 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
667 // Similar to goenv_unix but extracts the environment value for
669 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
671 for argv_index(argv, argc+1+n) != nil {
675 for i := int32(0); i < n; i++ {
676 p := argv_index(argv, argc+1+i)
677 s := unsafe.String(p, findnull(p))
679 if hasPrefix(s, prefix) {
680 env = gostring(p)[len(prefix):]
688 // The bootstrap sequence is:
692 // make & queue new G
693 // call runtime·mstart
695 // The new G calls runtime·main.
697 lockInit(&sched.lock, lockRankSched)
698 lockInit(&sched.sysmonlock, lockRankSysmon)
699 lockInit(&sched.deferlock, lockRankDefer)
700 lockInit(&sched.sudoglock, lockRankSudog)
701 lockInit(&deadlock, lockRankDeadlock)
702 lockInit(&paniclk, lockRankPanic)
703 lockInit(&allglock, lockRankAllg)
704 lockInit(&allpLock, lockRankAllp)
705 lockInit(&reflectOffs.lock, lockRankReflectOffs)
706 lockInit(&finlock, lockRankFin)
707 lockInit(&cpuprof.lock, lockRankCpuprof)
709 // Enforce that this lock is always a leaf lock.
710 // All of this lock's critical sections should be
712 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
714 // raceinit must be the first call to race detector.
715 // In particular, it must be done before mallocinit below calls racemapshadow.
718 gp.racectx, raceprocctx0 = raceinit()
721 sched.maxmcount = 10000
723 // The world starts stopped.
729 godebug := getGodebugEarly()
730 initPageTrace(godebug) // must run after mallocinit but before anything allocates
731 cpuinit(godebug) // must run before alginit
732 alginit() // maps, hash, fastrand must not be used before this call
733 fastrandinit() // must run before mcommoninit
734 mcommoninit(gp.m, -1)
735 modulesinit() // provides activeModules
736 typelinksinit() // uses maps, activeModules
737 itabsinit() // uses activeModules
738 stkobjinit() // must run before GC starts
740 sigsave(&gp.m.sigmask)
741 initSigmask = gp.m.sigmask
750 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
751 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
752 // set to true by the linker, it means that nothing is consuming the profile, it is
753 // safe to set MemProfileRate to 0.
754 if disableMemoryProfiling {
759 sched.lastpoll.Store(nanotime())
761 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
764 if procresize(procs) != nil {
765 throw("unknown runnable goroutine during bootstrap")
769 // World is effectively started now, as P's can run.
772 if buildVersion == "" {
773 // Condition should never trigger. This code just serves
774 // to ensure runtime·buildVersion is kept in the resulting binary.
775 buildVersion = "unknown"
777 if len(modinfo) == 1 {
778 // Condition should never trigger. This code just serves
779 // to ensure runtime·modinfo is kept in the resulting binary.
784 func dumpgstatus(gp *g) {
786 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
787 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
790 // sched.lock must be held.
792 assertLockHeld(&sched.lock)
794 // Exclude extra M's, which are used for cgocallback from threads
797 // The purpose of the SetMaxThreads limit is to avoid accidental fork
798 // bomb from something like millions of goroutines blocking on system
799 // calls, causing the runtime to create millions of threads. By
800 // definition, this isn't a problem for threads created in C, so we
801 // exclude them from the limit. See https://go.dev/issue/60004.
802 count := mcount() - int32(extraMInUse.Load()) - int32(extraMLength.Load())
803 if count > sched.maxmcount {
804 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
805 throw("thread exhaustion")
809 // mReserveID returns the next ID to use for a new m. This new m is immediately
810 // considered 'running' by checkdead.
812 // sched.lock must be held.
813 func mReserveID() int64 {
814 assertLockHeld(&sched.lock)
816 if sched.mnext+1 < sched.mnext {
817 throw("runtime: thread ID overflow")
825 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
826 func mcommoninit(mp *m, id int64) {
829 // g0 stack won't make sense for user (and is not necessary unwindable).
831 callers(1, mp.createstack[:])
842 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
843 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
847 // Same behavior as for 1.17.
848 // TODO: Simplify this.
849 if goarch.BigEndian {
850 mp.fastrand = uint64(lo)<<32 | uint64(hi)
852 mp.fastrand = uint64(hi)<<32 | uint64(lo)
856 if mp.gsignal != nil {
857 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
860 // Add to allm so garbage collector doesn't free g->m
861 // when it is just in a register or thread-local storage.
864 // NumCgoCall() iterates over allm w/o schedlock,
865 // so we need to publish it safely.
866 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
869 // Allocate memory to hold a cgo traceback if the cgo call crashes.
870 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
871 mp.cgoCallers = new(cgoCallers)
875 func (mp *m) becomeSpinning() {
877 sched.nmspinning.Add(1)
878 sched.needspinning.Store(0)
881 func (mp *m) hasCgoOnStack() bool {
882 return mp.ncgo > 0 || mp.isextra
885 var fastrandseed uintptr
887 func fastrandinit() {
888 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
892 // Mark gp ready to run.
893 func ready(gp *g, traceskip int, next bool) {
895 traceGoUnpark(gp, traceskip)
898 status := readgstatus(gp)
901 mp := acquirem() // disable preemption because it can be holding p in a local var
902 if status&^_Gscan != _Gwaiting {
904 throw("bad g->status in ready")
907 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
908 casgstatus(gp, _Gwaiting, _Grunnable)
909 runqput(mp.p.ptr(), gp, next)
914 // freezeStopWait is a large value that freezetheworld sets
915 // sched.stopwait to in order to request that all Gs permanently stop.
916 const freezeStopWait = 0x7fffffff
918 // freezing is set to non-zero if the runtime is trying to freeze the
920 var freezing atomic.Bool
922 // Similar to stopTheWorld but best-effort and can be called several times.
923 // There is no reverse operation, used during crashing.
924 // This function must not lock any mutexes.
925 func freezetheworld() {
927 if debug.dontfreezetheworld > 0 {
928 // Don't prempt Ps to stop goroutines. That will perturb
929 // scheduler state, making debugging more difficult. Instead,
930 // allow goroutines to continue execution.
932 // fatalpanic will tracebackothers to trace all goroutines. It
933 // is unsafe to trace a running goroutine, so tracebackothers
934 // will skip running goroutines. That is OK and expected, we
935 // expect users of dontfreezetheworld to use core files anyway.
937 // However, allowing the scheduler to continue running free
938 // introduces a race: a goroutine may be stopped when
939 // tracebackothers checks its status, and then start running
940 // later when we are in the middle of traceback, potentially
943 // To mitigate this, when an M naturally enters the scheduler,
944 // schedule checks if freezing is set and if so stops
945 // execution. This guarantees that while Gs can transition from
946 // running to stopped, they can never transition from stopped
949 // The sleep here allows racing Ms that missed freezing and are
950 // about to run a G to complete the transition to running
951 // before we start traceback.
956 // stopwait and preemption requests can be lost
957 // due to races with concurrently executing threads,
958 // so try several times
959 for i := 0; i < 5; i++ {
960 // this should tell the scheduler to not start any new goroutines
961 sched.stopwait = freezeStopWait
962 sched.gcwaiting.Store(true)
963 // this should stop running goroutines
965 break // no running goroutines
975 // All reads and writes of g's status go through readgstatus, casgstatus
976 // castogscanstatus, casfrom_Gscanstatus.
979 func readgstatus(gp *g) uint32 {
980 return gp.atomicstatus.Load()
983 // The Gscanstatuses are acting like locks and this releases them.
984 // If it proves to be a performance hit we should be able to make these
985 // simple atomic stores but for now we are going to throw if
986 // we see an inconsistent state.
987 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
990 // Check that transition is valid.
993 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
995 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
1001 if newval == oldval&^_Gscan {
1002 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
1006 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
1008 throw("casfrom_Gscanstatus: gp->status is not in scan state")
1010 releaseLockRank(lockRankGscan)
1013 // This will return false if the gp is not in the expected status and the cas fails.
1014 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
1015 func castogscanstatus(gp *g, oldval, newval uint32) bool {
1021 if newval == oldval|_Gscan {
1022 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
1024 acquireLockRank(lockRankGscan)
1030 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
1031 throw("castogscanstatus")
1032 panic("not reached")
1035 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
1036 // various latencies on every transition instead of sampling them.
1037 var casgstatusAlwaysTrack = false
1039 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
1040 // and casfrom_Gscanstatus instead.
1041 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
1042 // put it in the Gscan state is finished.
1045 func casgstatus(gp *g, oldval, newval uint32) {
1046 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
1047 systemstack(func() {
1048 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
1049 throw("casgstatus: bad incoming values")
1053 acquireLockRank(lockRankGscan)
1054 releaseLockRank(lockRankGscan)
1056 // See https://golang.org/cl/21503 for justification of the yield delay.
1057 const yieldDelay = 5 * 1000
1060 // loop if gp->atomicstatus is in a scan state giving
1061 // GC time to finish and change the state to oldval.
1062 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1063 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1064 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1067 nextYield = nanotime() + yieldDelay
1069 if nanotime() < nextYield {
1070 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1075 nextYield = nanotime() + yieldDelay/2
1079 if oldval == _Grunning {
1080 // Track every gTrackingPeriod time a goroutine transitions out of running.
1081 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1090 // Handle various kinds of tracking.
1093 // - Time spent in runnable.
1094 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1097 // We transitioned out of runnable, so measure how much
1098 // time we spent in this state and add it to
1101 gp.runnableTime += now - gp.trackingStamp
1102 gp.trackingStamp = 0
1104 if !gp.waitreason.isMutexWait() {
1105 // Not blocking on a lock.
1108 // Blocking on a lock, measure it. Note that because we're
1109 // sampling, we have to multiply by our sampling period to get
1110 // a more representative estimate of the absolute value.
1111 // gTrackingPeriod also represents an accurate sampling period
1112 // because we can only enter this state from _Grunning.
1114 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1115 gp.trackingStamp = 0
1119 if !gp.waitreason.isMutexWait() {
1120 // Not blocking on a lock.
1123 // Blocking on a lock. Write down the timestamp.
1125 gp.trackingStamp = now
1127 // We just transitioned into runnable, so record what
1128 // time that happened.
1130 gp.trackingStamp = now
1132 // We're transitioning into running, so turn off
1133 // tracking and record how much time we spent in
1136 sched.timeToRun.record(gp.runnableTime)
1141 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1143 // Use this over casgstatus when possible to ensure that a waitreason is set.
1144 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1145 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1146 gp.waitreason = reason
1147 casgstatus(gp, old, _Gwaiting)
1150 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1151 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1152 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1153 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1154 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1157 func casgcopystack(gp *g) uint32 {
1159 oldstatus := readgstatus(gp) &^ _Gscan
1160 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1161 throw("copystack: bad status, not Gwaiting or Grunnable")
1163 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1169 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1171 // TODO(austin): This is the only status operation that both changes
1172 // the status and locks the _Gscan bit. Rethink this.
1173 func casGToPreemptScan(gp *g, old, new uint32) {
1174 if old != _Grunning || new != _Gscan|_Gpreempted {
1175 throw("bad g transition")
1177 acquireLockRank(lockRankGscan)
1178 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1182 // casGFromPreempted attempts to transition gp from _Gpreempted to
1183 // _Gwaiting. If successful, the caller is responsible for
1184 // re-scheduling gp.
1185 func casGFromPreempted(gp *g, old, new uint32) bool {
1186 if old != _Gpreempted || new != _Gwaiting {
1187 throw("bad g transition")
1189 gp.waitreason = waitReasonPreempted
1190 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1193 // stwReason is an enumeration of reasons the world is stopping.
1194 type stwReason uint8
1196 // Reasons to stop-the-world.
1198 // Avoid reusing reasons and add new ones instead.
1200 stwUnknown stwReason = iota // "unknown"
1201 stwGCMarkTerm // "GC mark termination"
1202 stwGCSweepTerm // "GC sweep termination"
1203 stwWriteHeapDump // "write heap dump"
1204 stwGoroutineProfile // "goroutine profile"
1205 stwGoroutineProfileCleanup // "goroutine profile cleanup"
1206 stwAllGoroutinesStack // "all goroutines stack trace"
1207 stwReadMemStats // "read mem stats"
1208 stwAllThreadsSyscall // "AllThreadsSyscall"
1209 stwGOMAXPROCS // "GOMAXPROCS"
1210 stwStartTrace // "start trace"
1211 stwStopTrace // "stop trace"
1212 stwForTestCountPagesInUse // "CountPagesInUse (test)"
1213 stwForTestReadMetricsSlow // "ReadMetricsSlow (test)"
1214 stwForTestReadMemStatsSlow // "ReadMemStatsSlow (test)"
1215 stwForTestPageCachePagesLeaked // "PageCachePagesLeaked (test)"
1216 stwForTestResetDebugLog // "ResetDebugLog (test)"
1219 func (r stwReason) String() string {
1220 return stwReasonStrings[r]
1223 // If you add to this list, also add it to src/internal/trace/parser.go.
1224 // If you change the values of any of the stw* constants, bump the trace
1225 // version number and make a copy of this.
1226 var stwReasonStrings = [...]string{
1227 stwUnknown: "unknown",
1228 stwGCMarkTerm: "GC mark termination",
1229 stwGCSweepTerm: "GC sweep termination",
1230 stwWriteHeapDump: "write heap dump",
1231 stwGoroutineProfile: "goroutine profile",
1232 stwGoroutineProfileCleanup: "goroutine profile cleanup",
1233 stwAllGoroutinesStack: "all goroutines stack trace",
1234 stwReadMemStats: "read mem stats",
1235 stwAllThreadsSyscall: "AllThreadsSyscall",
1236 stwGOMAXPROCS: "GOMAXPROCS",
1237 stwStartTrace: "start trace",
1238 stwStopTrace: "stop trace",
1239 stwForTestCountPagesInUse: "CountPagesInUse (test)",
1240 stwForTestReadMetricsSlow: "ReadMetricsSlow (test)",
1241 stwForTestReadMemStatsSlow: "ReadMemStatsSlow (test)",
1242 stwForTestPageCachePagesLeaked: "PageCachePagesLeaked (test)",
1243 stwForTestResetDebugLog: "ResetDebugLog (test)",
1246 // stopTheWorld stops all P's from executing goroutines, interrupting
1247 // all goroutines at GC safe points and records reason as the reason
1248 // for the stop. On return, only the current goroutine's P is running.
1249 // stopTheWorld must not be called from a system stack and the caller
1250 // must not hold worldsema. The caller must call startTheWorld when
1251 // other P's should resume execution.
1253 // stopTheWorld is safe for multiple goroutines to call at the
1254 // same time. Each will execute its own stop, and the stops will
1257 // This is also used by routines that do stack dumps. If the system is
1258 // in panic or being exited, this may not reliably stop all
1260 func stopTheWorld(reason stwReason) {
1261 semacquire(&worldsema)
1263 gp.m.preemptoff = reason.String()
1264 systemstack(func() {
1265 // Mark the goroutine which called stopTheWorld preemptible so its
1266 // stack may be scanned.
1267 // This lets a mark worker scan us while we try to stop the world
1268 // since otherwise we could get in a mutual preemption deadlock.
1269 // We must not modify anything on the G stack because a stack shrink
1270 // may occur. A stack shrink is otherwise OK though because in order
1271 // to return from this function (and to leave the system stack) we
1272 // must have preempted all goroutines, including any attempting
1273 // to scan our stack, in which case, any stack shrinking will
1274 // have already completed by the time we exit.
1275 // Don't provide a wait reason because we're still executing.
1276 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1277 stopTheWorldWithSema(reason)
1278 casgstatus(gp, _Gwaiting, _Grunning)
1282 // startTheWorld undoes the effects of stopTheWorld.
1283 func startTheWorld() {
1284 systemstack(func() { startTheWorldWithSema() })
1286 // worldsema must be held over startTheWorldWithSema to ensure
1287 // gomaxprocs cannot change while worldsema is held.
1289 // Release worldsema with direct handoff to the next waiter, but
1290 // acquirem so that semrelease1 doesn't try to yield our time.
1292 // Otherwise if e.g. ReadMemStats is being called in a loop,
1293 // it might stomp on other attempts to stop the world, such as
1294 // for starting or ending GC. The operation this blocks is
1295 // so heavy-weight that we should just try to be as fair as
1298 // We don't want to just allow us to get preempted between now
1299 // and releasing the semaphore because then we keep everyone
1300 // (including, for example, GCs) waiting longer.
1303 semrelease1(&worldsema, true, 0)
1307 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1308 // until the GC is not running. It also blocks a GC from starting
1309 // until startTheWorldGC is called.
1310 func stopTheWorldGC(reason stwReason) {
1312 stopTheWorld(reason)
1315 // startTheWorldGC undoes the effects of stopTheWorldGC.
1316 func startTheWorldGC() {
1321 // Holding worldsema grants an M the right to try to stop the world.
1322 var worldsema uint32 = 1
1324 // Holding gcsema grants the M the right to block a GC, and blocks
1325 // until the current GC is done. In particular, it prevents gomaxprocs
1326 // from changing concurrently.
1328 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1329 // being changed/enabled during a GC, remove this.
1330 var gcsema uint32 = 1
1332 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1333 // The caller is responsible for acquiring worldsema and disabling
1334 // preemption first and then should stopTheWorldWithSema on the system
1337 // semacquire(&worldsema, 0)
1338 // m.preemptoff = "reason"
1339 // systemstack(stopTheWorldWithSema)
1341 // When finished, the caller must either call startTheWorld or undo
1342 // these three operations separately:
1344 // m.preemptoff = ""
1345 // systemstack(startTheWorldWithSema)
1346 // semrelease(&worldsema)
1348 // It is allowed to acquire worldsema once and then execute multiple
1349 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1350 // Other P's are able to execute between successive calls to
1351 // startTheWorldWithSema and stopTheWorldWithSema.
1352 // Holding worldsema causes any other goroutines invoking
1353 // stopTheWorld to block.
1354 func stopTheWorldWithSema(reason stwReason) {
1356 traceSTWStart(reason)
1360 // If we hold a lock, then we won't be able to stop another M
1361 // that is blocked trying to acquire the lock.
1363 throw("stopTheWorld: holding locks")
1367 sched.stopwait = gomaxprocs
1368 sched.gcwaiting.Store(true)
1371 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1373 // try to retake all P's in Psyscall status
1374 for _, pp := range allp {
1376 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1388 pp, _ := pidleget(now)
1392 pp.status = _Pgcstop
1395 wait := sched.stopwait > 0
1398 // wait for remaining P's to stop voluntarily
1401 // wait for 100us, then try to re-preempt in case of any races
1402 if notetsleep(&sched.stopnote, 100*1000) {
1403 noteclear(&sched.stopnote)
1412 if sched.stopwait != 0 {
1413 bad = "stopTheWorld: not stopped (stopwait != 0)"
1415 for _, pp := range allp {
1416 if pp.status != _Pgcstop {
1417 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1421 if freezing.Load() {
1422 // Some other thread is panicking. This can cause the
1423 // sanity checks above to fail if the panic happens in
1424 // the signal handler on a stopped thread. Either way,
1425 // we should halt this thread.
1436 func startTheWorldWithSema() int64 {
1437 assertWorldStopped()
1439 mp := acquirem() // disable preemption because it can be holding p in a local var
1440 if netpollinited() {
1441 list, delta := netpoll(0) // non-blocking
1443 netpollAdjustWaiters(delta)
1452 p1 := procresize(procs)
1453 sched.gcwaiting.Store(false)
1454 if sched.sysmonwait.Load() {
1455 sched.sysmonwait.Store(false)
1456 notewakeup(&sched.sysmonnote)
1469 throw("startTheWorld: inconsistent mp->nextp")
1472 notewakeup(&mp.park)
1474 // Start M to run P. Do not start another M below.
1479 // Capture start-the-world time before doing clean-up tasks.
1480 startTime := nanotime()
1485 // Wakeup an additional proc in case we have excessive runnable goroutines
1486 // in local queues or in the global queue. If we don't, the proc will park itself.
1487 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1495 // usesLibcall indicates whether this runtime performs system calls
1497 func usesLibcall() bool {
1499 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1502 return GOARCH != "mips64"
1507 // mStackIsSystemAllocated indicates whether this runtime starts on a
1508 // system-allocated stack.
1509 func mStackIsSystemAllocated() bool {
1511 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1514 return GOARCH != "mips64"
1519 // mstart is the entry-point for new Ms.
1520 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1523 // mstart0 is the Go entry-point for new Ms.
1524 // This must not split the stack because we may not even have stack
1525 // bounds set up yet.
1527 // May run during STW (because it doesn't have a P yet), so write
1528 // barriers are not allowed.
1531 //go:nowritebarrierrec
1535 osStack := gp.stack.lo == 0
1537 // Initialize stack bounds from system stack.
1538 // Cgo may have left stack size in stack.hi.
1539 // minit may update the stack bounds.
1541 // Note: these bounds may not be very accurate.
1542 // We set hi to &size, but there are things above
1543 // it. The 1024 is supposed to compensate this,
1544 // but is somewhat arbitrary.
1547 size = 16384 * sys.StackGuardMultiplier
1549 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1550 gp.stack.lo = gp.stack.hi - size + 1024
1552 // Initialize stack guard so that we can start calling regular
1554 gp.stackguard0 = gp.stack.lo + stackGuard
1555 // This is the g0, so we can also call go:systemstack
1556 // functions, which check stackguard1.
1557 gp.stackguard1 = gp.stackguard0
1560 // Exit this thread.
1561 if mStackIsSystemAllocated() {
1562 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1563 // the stack, but put it in gp.stack before mstart,
1564 // so the logic above hasn't set osStack yet.
1570 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1571 // so that we can set up g0.sched to return to the call of mstart1 above.
1578 throw("bad runtime·mstart")
1581 // Set up m.g0.sched as a label returning to just
1582 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1583 // We're never coming back to mstart1 after we call schedule,
1584 // so other calls can reuse the current frame.
1585 // And goexit0 does a gogo that needs to return from mstart1
1586 // and let mstart0 exit the thread.
1587 gp.sched.g = guintptr(unsafe.Pointer(gp))
1588 gp.sched.pc = getcallerpc()
1589 gp.sched.sp = getcallersp()
1594 // Install signal handlers; after minit so that minit can
1595 // prepare the thread to be able to handle the signals.
1600 if fn := gp.m.mstartfn; fn != nil {
1605 acquirep(gp.m.nextp.ptr())
1611 // mstartm0 implements part of mstart1 that only runs on the m0.
1613 // Write barriers are allowed here because we know the GC can't be
1614 // running yet, so they'll be no-ops.
1616 //go:yeswritebarrierrec
1618 // Create an extra M for callbacks on threads not created by Go.
1619 // An extra M is also needed on Windows for callbacks created by
1620 // syscall.NewCallback. See issue #6751 for details.
1621 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1628 // mPark causes a thread to park itself, returning once woken.
1633 notesleep(&gp.m.park)
1634 noteclear(&gp.m.park)
1637 // mexit tears down and exits the current thread.
1639 // Don't call this directly to exit the thread, since it must run at
1640 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1641 // unwind the stack to the point that exits the thread.
1643 // It is entered with m.p != nil, so write barriers are allowed. It
1644 // will release the P before exiting.
1646 //go:yeswritebarrierrec
1647 func mexit(osStack bool) {
1651 // This is the main thread. Just wedge it.
1653 // On Linux, exiting the main thread puts the process
1654 // into a non-waitable zombie state. On Plan 9,
1655 // exiting the main thread unblocks wait even though
1656 // other threads are still running. On Solaris we can
1657 // neither exitThread nor return from mstart. Other
1658 // bad things probably happen on other platforms.
1660 // We could try to clean up this M more before wedging
1661 // it, but that complicates signal handling.
1662 handoffp(releasep())
1668 throw("locked m0 woke up")
1674 // Free the gsignal stack.
1675 if mp.gsignal != nil {
1676 stackfree(mp.gsignal.stack)
1677 // On some platforms, when calling into VDSO (e.g. nanotime)
1678 // we store our g on the gsignal stack, if there is one.
1679 // Now the stack is freed, unlink it from the m, so we
1680 // won't write to it when calling VDSO code.
1684 // Remove m from allm.
1686 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1692 throw("m not found in allm")
1694 // Delay reaping m until it's done with the stack.
1696 // Put mp on the free list, though it will not be reaped while freeWait
1697 // is freeMWait. mp is no longer reachable via allm, so even if it is
1698 // on an OS stack, we must keep a reference to mp alive so that the GC
1699 // doesn't free mp while we are still using it.
1701 // Note that the free list must not be linked through alllink because
1702 // some functions walk allm without locking, so may be using alllink.
1703 mp.freeWait.Store(freeMWait)
1704 mp.freelink = sched.freem
1708 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1711 handoffp(releasep())
1712 // After this point we must not have write barriers.
1714 // Invoke the deadlock detector. This must happen after
1715 // handoffp because it may have started a new M to take our
1722 if GOOS == "darwin" || GOOS == "ios" {
1723 // Make sure pendingPreemptSignals is correct when an M exits.
1725 if mp.signalPending.Load() != 0 {
1726 pendingPreemptSignals.Add(-1)
1730 // Destroy all allocated resources. After this is called, we may no
1731 // longer take any locks.
1735 // No more uses of mp, so it is safe to drop the reference.
1736 mp.freeWait.Store(freeMRef)
1738 // Return from mstart and let the system thread
1739 // library free the g0 stack and terminate the thread.
1743 // mstart is the thread's entry point, so there's nothing to
1744 // return to. Exit the thread directly. exitThread will clear
1745 // m.freeWait when it's done with the stack and the m can be
1747 exitThread(&mp.freeWait)
1750 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1751 // If a P is currently executing code, this will bring the P to a GC
1752 // safe point and execute fn on that P. If the P is not executing code
1753 // (it is idle or in a syscall), this will call fn(p) directly while
1754 // preventing the P from exiting its state. This does not ensure that
1755 // fn will run on every CPU executing Go code, but it acts as a global
1756 // memory barrier. GC uses this as a "ragged barrier."
1758 // The caller must hold worldsema.
1761 func forEachP(fn func(*p)) {
1763 pp := getg().m.p.ptr()
1766 if sched.safePointWait != 0 {
1767 throw("forEachP: sched.safePointWait != 0")
1769 sched.safePointWait = gomaxprocs - 1
1770 sched.safePointFn = fn
1772 // Ask all Ps to run the safe point function.
1773 for _, p2 := range allp {
1775 atomic.Store(&p2.runSafePointFn, 1)
1780 // Any P entering _Pidle or _Psyscall from now on will observe
1781 // p.runSafePointFn == 1 and will call runSafePointFn when
1782 // changing its status to _Pidle/_Psyscall.
1784 // Run safe point function for all idle Ps. sched.pidle will
1785 // not change because we hold sched.lock.
1786 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1787 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1789 sched.safePointWait--
1793 wait := sched.safePointWait > 0
1796 // Run fn for the current P.
1799 // Force Ps currently in _Psyscall into _Pidle and hand them
1800 // off to induce safe point function execution.
1801 for _, p2 := range allp {
1803 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1813 // Wait for remaining Ps to run fn.
1816 // Wait for 100us, then try to re-preempt in
1817 // case of any races.
1819 // Requires system stack.
1820 if notetsleep(&sched.safePointNote, 100*1000) {
1821 noteclear(&sched.safePointNote)
1827 if sched.safePointWait != 0 {
1828 throw("forEachP: not done")
1830 for _, p2 := range allp {
1831 if p2.runSafePointFn != 0 {
1832 throw("forEachP: P did not run fn")
1837 sched.safePointFn = nil
1842 // runSafePointFn runs the safe point function, if any, for this P.
1843 // This should be called like
1845 // if getg().m.p.runSafePointFn != 0 {
1849 // runSafePointFn must be checked on any transition in to _Pidle or
1850 // _Psyscall to avoid a race where forEachP sees that the P is running
1851 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1852 // nor the P run the safe-point function.
1853 func runSafePointFn() {
1854 p := getg().m.p.ptr()
1855 // Resolve the race between forEachP running the safe-point
1856 // function on this P's behalf and this P running the
1857 // safe-point function directly.
1858 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1861 sched.safePointFn(p)
1863 sched.safePointWait--
1864 if sched.safePointWait == 0 {
1865 notewakeup(&sched.safePointNote)
1870 // When running with cgo, we call _cgo_thread_start
1871 // to start threads for us so that we can play nicely with
1873 var cgoThreadStart unsafe.Pointer
1875 type cgothreadstart struct {
1881 // Allocate a new m unassociated with any thread.
1882 // Can use p for allocation context if needed.
1883 // fn is recorded as the new m's m.mstartfn.
1884 // id is optional pre-allocated m ID. Omit by passing -1.
1886 // This function is allowed to have write barriers even if the caller
1887 // isn't because it borrows pp.
1889 //go:yeswritebarrierrec
1890 func allocm(pp *p, fn func(), id int64) *m {
1893 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1894 // disable preemption to ensure it is not stolen, which would make the
1895 // caller lose ownership.
1900 acquirep(pp) // temporarily borrow p for mallocs in this function
1903 // Release the free M list. We need to do this somewhere and
1904 // this may free up a stack we can use.
1905 if sched.freem != nil {
1908 for freem := sched.freem; freem != nil; {
1909 wait := freem.freeWait.Load()
1910 if wait == freeMWait {
1911 next := freem.freelink
1912 freem.freelink = newList
1917 // Free the stack if needed. For freeMRef, there is
1918 // nothing to do except drop freem from the sched.freem
1920 if wait == freeMStack {
1921 // stackfree must be on the system stack, but allocm is
1922 // reachable off the system stack transitively from
1924 systemstack(func() {
1925 stackfree(freem.g0.stack)
1928 freem = freem.freelink
1930 sched.freem = newList
1938 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1939 // Windows and Plan 9 will layout sched stack on OS stack.
1940 if iscgo || mStackIsSystemAllocated() {
1943 mp.g0 = malg(16384 * sys.StackGuardMultiplier)
1947 if pp == gp.m.p.ptr() {
1952 allocmLock.runlock()
1956 // needm is called when a cgo callback happens on a
1957 // thread without an m (a thread not created by Go).
1958 // In this case, needm is expected to find an m to use
1959 // and return with m, g initialized correctly.
1960 // Since m and g are not set now (likely nil, but see below)
1961 // needm is limited in what routines it can call. In particular
1962 // it can only call nosplit functions (textflag 7) and cannot
1963 // do any scheduling that requires an m.
1965 // In order to avoid needing heavy lifting here, we adopt
1966 // the following strategy: there is a stack of available m's
1967 // that can be stolen. Using compare-and-swap
1968 // to pop from the stack has ABA races, so we simulate
1969 // a lock by doing an exchange (via Casuintptr) to steal the stack
1970 // head and replace the top pointer with MLOCKED (1).
1971 // This serves as a simple spin lock that we can use even
1972 // without an m. The thread that locks the stack in this way
1973 // unlocks the stack by storing a valid stack head pointer.
1975 // In order to make sure that there is always an m structure
1976 // available to be stolen, we maintain the invariant that there
1977 // is always one more than needed. At the beginning of the
1978 // program (if cgo is in use) the list is seeded with a single m.
1979 // If needm finds that it has taken the last m off the list, its job
1980 // is - once it has installed its own m so that it can do things like
1981 // allocate memory - to create a spare m and put it on the list.
1983 // Each of these extra m's also has a g0 and a curg that are
1984 // pressed into service as the scheduling stack and current
1985 // goroutine for the duration of the cgo callback.
1987 // It calls dropm to put the m back on the list,
1988 // 1. when the callback is done with the m in non-pthread platforms,
1989 // 2. or when the C thread exiting on pthread platforms.
1991 // The signal argument indicates whether we're called from a signal
1995 func needm(signal bool) {
1996 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1997 // Can happen if C/C++ code calls Go from a global ctor.
1998 // Can also happen on Windows if a global ctor uses a
1999 // callback created by syscall.NewCallback. See issue #6751
2002 // Can not throw, because scheduler is not initialized yet.
2003 writeErrStr("fatal error: cgo callback before cgo call\n")
2007 // Save and block signals before getting an M.
2008 // The signal handler may call needm itself,
2009 // and we must avoid a deadlock. Also, once g is installed,
2010 // any incoming signals will try to execute,
2011 // but we won't have the sigaltstack settings and other data
2012 // set up appropriately until the end of minit, which will
2013 // unblock the signals. This is the same dance as when
2014 // starting a new m to run Go code via newosproc.
2019 // getExtraM is safe here because of the invariant above,
2020 // that the extra list always contains or will soon contain
2022 mp, last := getExtraM()
2024 // Set needextram when we've just emptied the list,
2025 // so that the eventual call into cgocallbackg will
2026 // allocate a new m for the extra list. We delay the
2027 // allocation until then so that it can be done
2028 // after exitsyscall makes sure it is okay to be
2029 // running at all (that is, there's no garbage collection
2030 // running right now).
2031 mp.needextram = last
2033 // Store the original signal mask for use by minit.
2034 mp.sigmask = sigmask
2036 // Install TLS on some platforms (previously setg
2037 // would do this if necessary).
2040 // Install g (= m->g0) and set the stack bounds
2041 // to match the current stack.
2044 callbackUpdateSystemStack(mp, sp, signal)
2046 // Should mark we are already in Go now.
2047 // Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
2048 // which means the extram list may be empty, that will cause a deadlock.
2049 mp.isExtraInC = false
2051 // Initialize this thread to use the m.
2055 // mp.curg is now a real goroutine.
2056 casgstatus(mp.curg, _Gdead, _Gsyscall)
2060 // Acquire an extra m and bind it to the C thread when a pthread key has been created.
2063 func needAndBindM() {
2066 if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
2071 // newextram allocates m's and puts them on the extra list.
2072 // It is called with a working local m, so that it can do things
2073 // like call schedlock and allocate.
2075 c := extraMWaiters.Swap(0)
2077 for i := uint32(0); i < c; i++ {
2080 } else if extraMLength.Load() == 0 {
2081 // Make sure there is at least one extra M.
2086 // oneNewExtraM allocates an m and puts it on the extra list.
2087 func oneNewExtraM() {
2088 // Create extra goroutine locked to extra m.
2089 // The goroutine is the context in which the cgo callback will run.
2090 // The sched.pc will never be returned to, but setting it to
2091 // goexit makes clear to the traceback routines where
2092 // the goroutine stack ends.
2093 mp := allocm(nil, nil, -1)
2095 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
2096 gp.sched.sp = gp.stack.hi
2097 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
2099 gp.sched.g = guintptr(unsafe.Pointer(gp))
2100 gp.syscallpc = gp.sched.pc
2101 gp.syscallsp = gp.sched.sp
2102 gp.stktopsp = gp.sched.sp
2103 // malg returns status as _Gidle. Change to _Gdead before
2104 // adding to allg where GC can see it. We use _Gdead to hide
2105 // this from tracebacks and stack scans since it isn't a
2106 // "real" goroutine until needm grabs it.
2107 casgstatus(gp, _Gidle, _Gdead)
2111 // mark we are in C by default.
2112 mp.isExtraInC = true
2116 gp.goid = sched.goidgen.Add(1)
2118 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
2121 traceOneNewExtraM(gp)
2123 // put on allg for garbage collector
2126 // gp is now on the allg list, but we don't want it to be
2127 // counted by gcount. It would be more "proper" to increment
2128 // sched.ngfree, but that requires locking. Incrementing ngsys
2129 // has the same effect.
2132 // Add m to the extra list.
2136 // dropm puts the current m back onto the extra list.
2138 // 1. On systems without pthreads, like Windows
2139 // dropm is called when a cgo callback has called needm but is now
2140 // done with the callback and returning back into the non-Go thread.
2142 // The main expense here is the call to signalstack to release the
2143 // m's signal stack, and then the call to needm on the next callback
2144 // from this thread. It is tempting to try to save the m for next time,
2145 // which would eliminate both these costs, but there might not be
2146 // a next time: the current thread (which Go does not control) might exit.
2147 // If we saved the m for that thread, there would be an m leak each time
2148 // such a thread exited. Instead, we acquire and release an m on each
2149 // call. These should typically not be scheduling operations, just a few
2150 // atomics, so the cost should be small.
2152 // 2. On systems with pthreads
2153 // dropm is called while a non-Go thread is exiting.
2154 // We allocate a pthread per-thread variable using pthread_key_create,
2155 // to register a thread-exit-time destructor.
2156 // And store the g into a thread-specific value associated with the pthread key,
2157 // when first return back to C.
2158 // So that the destructor would invoke dropm while the non-Go thread is exiting.
2159 // This is much faster since it avoids expensive signal-related syscalls.
2161 // This always runs without a P, so //go:nowritebarrierrec is required.
2163 // This may run with a different stack than was recorded in g0 (there is no
2164 // call to callbackUpdateSystemStack prior to dropm), so this must be
2165 // //go:nosplit to avoid the stack bounds check.
2167 //go:nowritebarrierrec
2170 // Clear m and g, and return m to the extra list.
2171 // After the call to setg we can only call nosplit functions
2172 // with no pointer manipulation.
2175 // Return mp.curg to dead state.
2176 casgstatus(mp.curg, _Gsyscall, _Gdead)
2177 mp.curg.preemptStop = false
2180 // Block signals before unminit.
2181 // Unminit unregisters the signal handling stack (but needs g on some systems).
2182 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2183 // It's important not to try to handle a signal between those two steps.
2184 sigmask := mp.sigmask
2190 // Clear g0 stack bounds to ensure that needm always refreshes the
2191 // bounds when reusing this M.
2200 msigrestore(sigmask)
2203 // bindm store the g0 of the current m into a thread-specific value.
2205 // We allocate a pthread per-thread variable using pthread_key_create,
2206 // to register a thread-exit-time destructor.
2207 // We are here setting the thread-specific value of the pthread key, to enable the destructor.
2208 // So that the pthread_key_destructor would dropm while the C thread is exiting.
2210 // And the saved g will be used in pthread_key_destructor,
2211 // since the g stored in the TLS by Go might be cleared in some platforms,
2212 // before the destructor invoked, so, we restore g by the stored g, before dropm.
2214 // We store g0 instead of m, to make the assembly code simpler,
2215 // since we need to restore g0 in runtime.cgocallback.
2217 // On systems without pthreads, like Windows, bindm shouldn't be used.
2219 // NOTE: this always runs without a P, so, nowritebarrierrec required.
2222 //go:nowritebarrierrec
2224 if GOOS == "windows" || GOOS == "plan9" {
2225 fatal("bindm in unexpected GOOS")
2229 fatal("the current g is not g0")
2231 if _cgo_bindm != nil {
2232 asmcgocall(_cgo_bindm, unsafe.Pointer(g))
2236 // A helper function for EnsureDropM.
2237 func getm() uintptr {
2238 return uintptr(unsafe.Pointer(getg().m))
2242 // Locking linked list of extra M's, via mp.schedlink. Must be accessed
2243 // only via lockextra/unlockextra.
2245 // Can't be atomic.Pointer[m] because we use an invalid pointer as a
2246 // "locked" sentinel value. M's on this list remain visible to the GC
2247 // because their mp.curg is on allgs.
2248 extraM atomic.Uintptr
2249 // Number of M's in the extraM list.
2250 extraMLength atomic.Uint32
2251 // Number of waiters in lockextra.
2252 extraMWaiters atomic.Uint32
2254 // Number of extra M's in use by threads.
2255 extraMInUse atomic.Uint32
2258 // lockextra locks the extra list and returns the list head.
2259 // The caller must unlock the list by storing a new list head
2260 // to extram. If nilokay is true, then lockextra will
2261 // return a nil list head if that's what it finds. If nilokay is false,
2262 // lockextra will keep waiting until the list head is no longer nil.
2265 func lockextra(nilokay bool) *m {
2270 old := extraM.Load()
2275 if old == 0 && !nilokay {
2277 // Add 1 to the number of threads
2278 // waiting for an M.
2279 // This is cleared by newextram.
2280 extraMWaiters.Add(1)
2286 if extraM.CompareAndSwap(old, locked) {
2287 return (*m)(unsafe.Pointer(old))
2295 func unlockextra(mp *m, delta int32) {
2296 extraMLength.Add(delta)
2297 extraM.Store(uintptr(unsafe.Pointer(mp)))
2300 // Return an M from the extra M list. Returns last == true if the list becomes
2301 // empty because of this call.
2303 // Spins waiting for an extra M, so caller must ensure that the list always
2304 // contains or will soon contain at least one M.
2307 func getExtraM() (mp *m, last bool) {
2308 mp = lockextra(false)
2310 unlockextra(mp.schedlink.ptr(), -1)
2311 return mp, mp.schedlink.ptr() == nil
2314 // Returns an extra M back to the list. mp must be from getExtraM. Newly
2315 // allocated M's should use addExtraM.
2318 func putExtraM(mp *m) {
2323 // Adds a newly allocated M to the extra M list.
2326 func addExtraM(mp *m) {
2327 mnext := lockextra(true)
2328 mp.schedlink.set(mnext)
2333 // allocmLock is locked for read when creating new Ms in allocm and their
2334 // addition to allm. Thus acquiring this lock for write blocks the
2335 // creation of new Ms.
2338 // execLock serializes exec and clone to avoid bugs or unspecified
2339 // behaviour around exec'ing while creating/destroying threads. See
2344 // These errors are reported (via writeErrStr) by some OS-specific
2345 // versions of newosproc and newosproc0.
2347 failthreadcreate = "runtime: failed to create new OS thread\n"
2348 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2351 // newmHandoff contains a list of m structures that need new OS threads.
2352 // This is used by newm in situations where newm itself can't safely
2353 // start an OS thread.
2354 var newmHandoff struct {
2357 // newm points to a list of M structures that need new OS
2358 // threads. The list is linked through m.schedlink.
2361 // waiting indicates that wake needs to be notified when an m
2362 // is put on the list.
2366 // haveTemplateThread indicates that the templateThread has
2367 // been started. This is not protected by lock. Use cas to set
2369 haveTemplateThread uint32
2372 // Create a new m. It will start off with a call to fn, or else the scheduler.
2373 // fn needs to be static and not a heap allocated closure.
2374 // May run with m.p==nil, so write barriers are not allowed.
2376 // id is optional pre-allocated m ID. Omit by passing -1.
2378 //go:nowritebarrierrec
2379 func newm(fn func(), pp *p, id int64) {
2380 // allocm adds a new M to allm, but they do not start until created by
2381 // the OS in newm1 or the template thread.
2383 // doAllThreadsSyscall requires that every M in allm will eventually
2384 // start and be signal-able, even with a STW.
2386 // Disable preemption here until we start the thread to ensure that
2387 // newm is not preempted between allocm and starting the new thread,
2388 // ensuring that anything added to allm is guaranteed to eventually
2392 mp := allocm(pp, fn, id)
2394 mp.sigmask = initSigmask
2395 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2396 // We're on a locked M or a thread that may have been
2397 // started by C. The kernel state of this thread may
2398 // be strange (the user may have locked it for that
2399 // purpose). We don't want to clone that into another
2400 // thread. Instead, ask a known-good thread to create
2401 // the thread for us.
2403 // This is disabled on Plan 9. See golang.org/issue/22227.
2405 // TODO: This may be unnecessary on Windows, which
2406 // doesn't model thread creation off fork.
2407 lock(&newmHandoff.lock)
2408 if newmHandoff.haveTemplateThread == 0 {
2409 throw("on a locked thread with no template thread")
2411 mp.schedlink = newmHandoff.newm
2412 newmHandoff.newm.set(mp)
2413 if newmHandoff.waiting {
2414 newmHandoff.waiting = false
2415 notewakeup(&newmHandoff.wake)
2417 unlock(&newmHandoff.lock)
2418 // The M has not started yet, but the template thread does not
2419 // participate in STW, so it will always process queued Ms and
2420 // it is safe to releasem.
2430 var ts cgothreadstart
2431 if _cgo_thread_start == nil {
2432 throw("_cgo_thread_start missing")
2435 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2436 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2438 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2441 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2443 execLock.rlock() // Prevent process clone.
2444 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2448 execLock.rlock() // Prevent process clone.
2453 // startTemplateThread starts the template thread if it is not already
2456 // The calling thread must itself be in a known-good state.
2457 func startTemplateThread() {
2458 if GOARCH == "wasm" { // no threads on wasm yet
2462 // Disable preemption to guarantee that the template thread will be
2463 // created before a park once haveTemplateThread is set.
2465 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2469 newm(templateThread, nil, -1)
2473 // templateThread is a thread in a known-good state that exists solely
2474 // to start new threads in known-good states when the calling thread
2475 // may not be in a good state.
2477 // Many programs never need this, so templateThread is started lazily
2478 // when we first enter a state that might lead to running on a thread
2479 // in an unknown state.
2481 // templateThread runs on an M without a P, so it must not have write
2484 //go:nowritebarrierrec
2485 func templateThread() {
2492 lock(&newmHandoff.lock)
2493 for newmHandoff.newm != 0 {
2494 newm := newmHandoff.newm.ptr()
2495 newmHandoff.newm = 0
2496 unlock(&newmHandoff.lock)
2498 next := newm.schedlink.ptr()
2503 lock(&newmHandoff.lock)
2505 newmHandoff.waiting = true
2506 noteclear(&newmHandoff.wake)
2507 unlock(&newmHandoff.lock)
2508 notesleep(&newmHandoff.wake)
2512 // Stops execution of the current m until new work is available.
2513 // Returns with acquired P.
2517 if gp.m.locks != 0 {
2518 throw("stopm holding locks")
2521 throw("stopm holding p")
2524 throw("stopm spinning")
2531 acquirep(gp.m.nextp.ptr())
2536 // startm's caller incremented nmspinning. Set the new M's spinning.
2537 getg().m.spinning = true
2540 // Schedules some M to run the p (creates an M if necessary).
2541 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2542 // May run with m.p==nil, so write barriers are not allowed.
2543 // If spinning is set, the caller has incremented nmspinning and must provide a
2544 // P. startm will set m.spinning in the newly started M.
2546 // Callers passing a non-nil P must call from a non-preemptible context. See
2547 // comment on acquirem below.
2549 // Argument lockheld indicates whether the caller already acquired the
2550 // scheduler lock. Callers holding the lock when making the call must pass
2551 // true. The lock might be temporarily dropped, but will be reacquired before
2554 // Must not have write barriers because this may be called without a P.
2556 //go:nowritebarrierrec
2557 func startm(pp *p, spinning, lockheld bool) {
2558 // Disable preemption.
2560 // Every owned P must have an owner that will eventually stop it in the
2561 // event of a GC stop request. startm takes transient ownership of a P
2562 // (either from argument or pidleget below) and transfers ownership to
2563 // a started M, which will be responsible for performing the stop.
2565 // Preemption must be disabled during this transient ownership,
2566 // otherwise the P this is running on may enter GC stop while still
2567 // holding the transient P, leaving that P in limbo and deadlocking the
2570 // Callers passing a non-nil P must already be in non-preemptible
2571 // context, otherwise such preemption could occur on function entry to
2572 // startm. Callers passing a nil P may be preemptible, so we must
2573 // disable preemption before acquiring a P from pidleget below.
2580 // TODO(prattmic): All remaining calls to this function
2581 // with _p_ == nil could be cleaned up to find a P
2582 // before calling startm.
2583 throw("startm: P required for spinning=true")
2596 // No M is available, we must drop sched.lock and call newm.
2597 // However, we already own a P to assign to the M.
2599 // Once sched.lock is released, another G (e.g., in a syscall),
2600 // could find no idle P while checkdead finds a runnable G but
2601 // no running M's because this new M hasn't started yet, thus
2602 // throwing in an apparent deadlock.
2603 // This apparent deadlock is possible when startm is called
2604 // from sysmon, which doesn't count as a running M.
2606 // Avoid this situation by pre-allocating the ID for the new M,
2607 // thus marking it as 'running' before we drop sched.lock. This
2608 // new M will eventually run the scheduler to execute any
2615 // The caller incremented nmspinning, so set m.spinning in the new M.
2623 // Ownership transfer of pp committed by start in newm.
2624 // Preemption is now safe.
2632 throw("startm: m is spinning")
2635 throw("startm: m has p")
2637 if spinning && !runqempty(pp) {
2638 throw("startm: p has runnable gs")
2640 // The caller incremented nmspinning, so set m.spinning in the new M.
2641 nmp.spinning = spinning
2643 notewakeup(&nmp.park)
2644 // Ownership transfer of pp committed by wakeup. Preemption is now
2649 // Hands off P from syscall or locked M.
2650 // Always runs without a P, so write barriers are not allowed.
2652 //go:nowritebarrierrec
2653 func handoffp(pp *p) {
2654 // handoffp must start an M in any situation where
2655 // findrunnable would return a G to run on pp.
2657 // if it has local work, start it straight away
2658 if !runqempty(pp) || sched.runqsize != 0 {
2659 startm(pp, false, false)
2662 // if there's trace work to do, start it straight away
2663 if (traceEnabled() || traceShuttingDown()) && traceReaderAvailable() != nil {
2664 startm(pp, false, false)
2667 // if it has GC work, start it straight away
2668 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2669 startm(pp, false, false)
2672 // no local work, check that there are no spinning/idle M's,
2673 // otherwise our help is not required
2674 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2675 sched.needspinning.Store(0)
2676 startm(pp, true, false)
2680 if sched.gcwaiting.Load() {
2681 pp.status = _Pgcstop
2683 if sched.stopwait == 0 {
2684 notewakeup(&sched.stopnote)
2689 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2690 sched.safePointFn(pp)
2691 sched.safePointWait--
2692 if sched.safePointWait == 0 {
2693 notewakeup(&sched.safePointNote)
2696 if sched.runqsize != 0 {
2698 startm(pp, false, false)
2701 // If this is the last running P and nobody is polling network,
2702 // need to wakeup another M to poll network.
2703 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2705 startm(pp, false, false)
2709 // The scheduler lock cannot be held when calling wakeNetPoller below
2710 // because wakeNetPoller may call wakep which may call startm.
2711 when := nobarrierWakeTime(pp)
2720 // Tries to add one more P to execute G's.
2721 // Called when a G is made runnable (newproc, ready).
2722 // Must be called with a P.
2724 // Be conservative about spinning threads, only start one if none exist
2726 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2730 // Disable preemption until ownership of pp transfers to the next M in
2731 // startm. Otherwise preemption here would leave pp stuck waiting to
2734 // See preemption comment on acquirem in startm for more details.
2739 pp, _ = pidlegetSpinning(0)
2741 if sched.nmspinning.Add(-1) < 0 {
2742 throw("wakep: negative nmspinning")
2748 // Since we always have a P, the race in the "No M is available"
2749 // comment in startm doesn't apply during the small window between the
2750 // unlock here and lock in startm. A checkdead in between will always
2751 // see at least one running M (ours).
2754 startm(pp, true, false)
2759 // Stops execution of the current m that is locked to a g until the g is runnable again.
2760 // Returns with acquired P.
2761 func stoplockedm() {
2764 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2765 throw("stoplockedm: inconsistent locking")
2768 // Schedule another M to run this p.
2773 // Wait until another thread schedules lockedg again.
2775 status := readgstatus(gp.m.lockedg.ptr())
2776 if status&^_Gscan != _Grunnable {
2777 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2778 dumpgstatus(gp.m.lockedg.ptr())
2779 throw("stoplockedm: not runnable")
2781 acquirep(gp.m.nextp.ptr())
2785 // Schedules the locked m to run the locked gp.
2786 // May run during STW, so write barriers are not allowed.
2788 //go:nowritebarrierrec
2789 func startlockedm(gp *g) {
2790 mp := gp.lockedm.ptr()
2792 throw("startlockedm: locked to me")
2795 throw("startlockedm: m has p")
2797 // directly handoff current P to the locked m
2801 notewakeup(&mp.park)
2805 // Stops the current m for stopTheWorld.
2806 // Returns when the world is restarted.
2810 if !sched.gcwaiting.Load() {
2811 throw("gcstopm: not waiting for gc")
2814 gp.m.spinning = false
2815 // OK to just drop nmspinning here,
2816 // startTheWorld will unpark threads as necessary.
2817 if sched.nmspinning.Add(-1) < 0 {
2818 throw("gcstopm: negative nmspinning")
2823 pp.status = _Pgcstop
2825 if sched.stopwait == 0 {
2826 notewakeup(&sched.stopnote)
2832 // Schedules gp to run on the current M.
2833 // If inheritTime is true, gp inherits the remaining time in the
2834 // current time slice. Otherwise, it starts a new time slice.
2837 // Write barriers are allowed because this is called immediately after
2838 // acquiring a P in several places.
2840 //go:yeswritebarrierrec
2841 func execute(gp *g, inheritTime bool) {
2844 if goroutineProfile.active {
2845 // Make sure that gp has had its stack written out to the goroutine
2846 // profile, exactly as it was when the goroutine profiler first stopped
2848 tryRecordGoroutineProfile(gp, osyield)
2851 // Assign gp.m before entering _Grunning so running Gs have an
2855 casgstatus(gp, _Grunnable, _Grunning)
2858 gp.stackguard0 = gp.stack.lo + stackGuard
2860 mp.p.ptr().schedtick++
2863 // Check whether the profiler needs to be turned on or off.
2864 hz := sched.profilehz
2865 if mp.profilehz != hz {
2866 setThreadCPUProfiler(hz)
2870 // GoSysExit has to happen when we have a P, but before GoStart.
2871 // So we emit it here.
2872 if gp.syscallsp != 0 {
2881 // Finds a runnable goroutine to execute.
2882 // Tries to steal from other P's, get g from local or global queue, poll network.
2883 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2884 // reader) so the caller should try to wake a P.
2885 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2888 // The conditions here and in handoffp must agree: if
2889 // findrunnable would return a G to run, handoffp must start
2894 if sched.gcwaiting.Load() {
2898 if pp.runSafePointFn != 0 {
2902 // now and pollUntil are saved for work stealing later,
2903 // which may steal timers. It's important that between now
2904 // and then, nothing blocks, so these numbers remain mostly
2906 now, pollUntil, _ := checkTimers(pp, 0)
2908 // Try to schedule the trace reader.
2909 if traceEnabled() || traceShuttingDown() {
2912 casgstatus(gp, _Gwaiting, _Grunnable)
2913 traceGoUnpark(gp, 0)
2914 return gp, false, true
2918 // Try to schedule a GC worker.
2919 if gcBlackenEnabled != 0 {
2920 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2922 return gp, false, true
2927 // Check the global runnable queue once in a while to ensure fairness.
2928 // Otherwise two goroutines can completely occupy the local runqueue
2929 // by constantly respawning each other.
2930 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2932 gp := globrunqget(pp, 1)
2935 return gp, false, false
2939 // Wake up the finalizer G.
2940 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2941 if gp := wakefing(); gp != nil {
2945 if *cgo_yield != nil {
2946 asmcgocall(*cgo_yield, nil)
2950 if gp, inheritTime := runqget(pp); gp != nil {
2951 return gp, inheritTime, false
2955 if sched.runqsize != 0 {
2957 gp := globrunqget(pp, 0)
2960 return gp, false, false
2965 // This netpoll is only an optimization before we resort to stealing.
2966 // We can safely skip it if there are no waiters or a thread is blocked
2967 // in netpoll already. If there is any kind of logical race with that
2968 // blocked thread (e.g. it has already returned from netpoll, but does
2969 // not set lastpoll yet), this thread will do blocking netpoll below
2971 if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 {
2972 if list, delta := netpoll(0); !list.empty() { // non-blocking
2975 netpollAdjustWaiters(delta)
2976 casgstatus(gp, _Gwaiting, _Grunnable)
2978 traceGoUnpark(gp, 0)
2980 return gp, false, false
2984 // Spinning Ms: steal work from other Ps.
2986 // Limit the number of spinning Ms to half the number of busy Ps.
2987 // This is necessary to prevent excessive CPU consumption when
2988 // GOMAXPROCS>>1 but the program parallelism is low.
2989 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2994 gp, inheritTime, tnow, w, newWork := stealWork(now)
2996 // Successfully stole.
2997 return gp, inheritTime, false
3000 // There may be new timer or GC work; restart to
3006 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3007 // Earlier timer to wait for.
3012 // We have nothing to do.
3014 // If we're in the GC mark phase, can safely scan and blacken objects,
3015 // and have work to do, run idle-time marking rather than give up the P.
3016 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
3017 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3019 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
3021 casgstatus(gp, _Gwaiting, _Grunnable)
3023 traceGoUnpark(gp, 0)
3025 return gp, false, false
3027 gcController.removeIdleMarkWorker()
3031 // If a callback returned and no other goroutine is awake,
3032 // then wake event handler goroutine which pauses execution
3033 // until a callback was triggered.
3034 gp, otherReady := beforeIdle(now, pollUntil)
3036 casgstatus(gp, _Gwaiting, _Grunnable)
3038 traceGoUnpark(gp, 0)
3040 return gp, false, false
3046 // Before we drop our P, make a snapshot of the allp slice,
3047 // which can change underfoot once we no longer block
3048 // safe-points. We don't need to snapshot the contents because
3049 // everything up to cap(allp) is immutable.
3050 allpSnapshot := allp
3051 // Also snapshot masks. Value changes are OK, but we can't allow
3052 // len to change out from under us.
3053 idlepMaskSnapshot := idlepMask
3054 timerpMaskSnapshot := timerpMask
3056 // return P and block
3058 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
3062 if sched.runqsize != 0 {
3063 gp := globrunqget(pp, 0)
3065 return gp, false, false
3067 if !mp.spinning && sched.needspinning.Load() == 1 {
3068 // See "Delicate dance" comment below.
3073 if releasep() != pp {
3074 throw("findrunnable: wrong p")
3076 now = pidleput(pp, now)
3079 // Delicate dance: thread transitions from spinning to non-spinning
3080 // state, potentially concurrently with submission of new work. We must
3081 // drop nmspinning first and then check all sources again (with
3082 // #StoreLoad memory barrier in between). If we do it the other way
3083 // around, another thread can submit work after we've checked all
3084 // sources but before we drop nmspinning; as a result nobody will
3085 // unpark a thread to run the work.
3087 // This applies to the following sources of work:
3089 // * Goroutines added to the global or a per-P run queue.
3090 // * New/modified-earlier timers on a per-P timer heap.
3091 // * Idle-priority GC work (barring golang.org/issue/19112).
3093 // If we discover new work below, we need to restore m.spinning as a
3094 // signal for resetspinning to unpark a new worker thread (because
3095 // there can be more than one starving goroutine).
3097 // However, if after discovering new work we also observe no idle Ps
3098 // (either here or in resetspinning), we have a problem. We may be
3099 // racing with a non-spinning M in the block above, having found no
3100 // work and preparing to release its P and park. Allowing that P to go
3101 // idle will result in loss of work conservation (idle P while there is
3102 // runnable work). This could result in complete deadlock in the
3103 // unlikely event that we discover new work (from netpoll) right as we
3104 // are racing with _all_ other Ps going idle.
3106 // We use sched.needspinning to synchronize with non-spinning Ms going
3107 // idle. If needspinning is set when they are about to drop their P,
3108 // they abort the drop and instead become a new spinning M on our
3109 // behalf. If we are not racing and the system is truly fully loaded
3110 // then no spinning threads are required, and the next thread to
3111 // naturally become spinning will clear the flag.
3113 // Also see "Worker thread parking/unparking" comment at the top of the
3115 wasSpinning := mp.spinning
3118 if sched.nmspinning.Add(-1) < 0 {
3119 throw("findrunnable: negative nmspinning")
3122 // Note the for correctness, only the last M transitioning from
3123 // spinning to non-spinning must perform these rechecks to
3124 // ensure no missed work. However, the runtime has some cases
3125 // of transient increments of nmspinning that are decremented
3126 // without going through this path, so we must be conservative
3127 // and perform the check on all spinning Ms.
3129 // See https://go.dev/issue/43997.
3131 // Check global and P runqueues again.
3134 if sched.runqsize != 0 {
3135 pp, _ := pidlegetSpinning(0)
3137 gp := globrunqget(pp, 0)
3139 throw("global runq empty with non-zero runqsize")
3144 return gp, false, false
3149 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
3156 // Check for idle-priority GC work again.
3157 pp, gp := checkIdleGCNoP()
3162 // Run the idle worker.
3163 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
3164 casgstatus(gp, _Gwaiting, _Grunnable)
3166 traceGoUnpark(gp, 0)
3168 return gp, false, false
3171 // Finally, check for timer creation or expiry concurrently with
3172 // transitioning from spinning to non-spinning.
3174 // Note that we cannot use checkTimers here because it calls
3175 // adjusttimers which may need to allocate memory, and that isn't
3176 // allowed when we don't have an active P.
3177 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
3180 // Poll network until next timer.
3181 if netpollinited() && (netpollAnyWaiters() || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
3182 sched.pollUntil.Store(pollUntil)
3184 throw("findrunnable: netpoll with p")
3187 throw("findrunnable: netpoll with spinning")
3194 delay = pollUntil - now
3200 // When using fake time, just poll.
3203 list, delta := netpoll(delay) // block until new work is available
3204 // Refresh now again, after potentially blocking.
3206 sched.pollUntil.Store(0)
3207 sched.lastpoll.Store(now)
3208 if faketime != 0 && list.empty() {
3209 // Using fake time and nothing is ready; stop M.
3210 // When all M's stop, checkdead will call timejump.
3215 pp, _ := pidleget(now)
3219 netpollAdjustWaiters(delta)
3225 netpollAdjustWaiters(delta)
3226 casgstatus(gp, _Gwaiting, _Grunnable)
3228 traceGoUnpark(gp, 0)
3230 return gp, false, false
3237 } else if pollUntil != 0 && netpollinited() {
3238 pollerPollUntil := sched.pollUntil.Load()
3239 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3247 // pollWork reports whether there is non-background work this P could
3248 // be doing. This is a fairly lightweight check to be used for
3249 // background work loops, like idle GC. It checks a subset of the
3250 // conditions checked by the actual scheduler.
3251 func pollWork() bool {
3252 if sched.runqsize != 0 {
3255 p := getg().m.p.ptr()
3259 if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 {
3260 if list, delta := netpoll(0); !list.empty() {
3262 netpollAdjustWaiters(delta)
3269 // stealWork attempts to steal a runnable goroutine or timer from any P.
3271 // If newWork is true, new work may have been readied.
3273 // If now is not 0 it is the current time. stealWork returns the passed time or
3274 // the current time if now was passed as 0.
3275 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3276 pp := getg().m.p.ptr()
3280 const stealTries = 4
3281 for i := 0; i < stealTries; i++ {
3282 stealTimersOrRunNextG := i == stealTries-1
3284 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3285 if sched.gcwaiting.Load() {
3286 // GC work may be available.
3287 return nil, false, now, pollUntil, true
3289 p2 := allp[enum.position()]
3294 // Steal timers from p2. This call to checkTimers is the only place
3295 // where we might hold a lock on a different P's timers. We do this
3296 // once on the last pass before checking runnext because stealing
3297 // from the other P's runnext should be the last resort, so if there
3298 // are timers to steal do that first.
3300 // We only check timers on one of the stealing iterations because
3301 // the time stored in now doesn't change in this loop and checking
3302 // the timers for each P more than once with the same value of now
3303 // is probably a waste of time.
3305 // timerpMask tells us whether the P may have timers at all. If it
3306 // can't, no need to check at all.
3307 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3308 tnow, w, ran := checkTimers(p2, now)
3310 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3314 // Running the timers may have
3315 // made an arbitrary number of G's
3316 // ready and added them to this P's
3317 // local run queue. That invalidates
3318 // the assumption of runqsteal
3319 // that it always has room to add
3320 // stolen G's. So check now if there
3321 // is a local G to run.
3322 if gp, inheritTime := runqget(pp); gp != nil {
3323 return gp, inheritTime, now, pollUntil, ranTimer
3329 // Don't bother to attempt to steal if p2 is idle.
3330 if !idlepMask.read(enum.position()) {
3331 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3332 return gp, false, now, pollUntil, ranTimer
3338 // No goroutines found to steal. Regardless, running a timer may have
3339 // made some goroutine ready that we missed. Indicate the next timer to
3341 return nil, false, now, pollUntil, ranTimer
3344 // Check all Ps for a runnable G to steal.
3346 // On entry we have no P. If a G is available to steal and a P is available,
3347 // the P is returned which the caller should acquire and attempt to steal the
3349 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3350 for id, p2 := range allpSnapshot {
3351 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3353 pp, _ := pidlegetSpinning(0)
3355 // Can't get a P, don't bother checking remaining Ps.
3364 // No work available.
3368 // Check all Ps for a timer expiring sooner than pollUntil.
3370 // Returns updated pollUntil value.
3371 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3372 for id, p2 := range allpSnapshot {
3373 if timerpMaskSnapshot.read(uint32(id)) {
3374 w := nobarrierWakeTime(p2)
3375 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3384 // Check for idle-priority GC, without a P on entry.
3386 // If some GC work, a P, and a worker G are all available, the P and G will be
3387 // returned. The returned P has not been wired yet.
3388 func checkIdleGCNoP() (*p, *g) {
3389 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3390 // must check again after acquiring a P. As an optimization, we also check
3391 // if an idle mark worker is needed at all. This is OK here, because if we
3392 // observe that one isn't needed, at least one is currently running. Even if
3393 // it stops running, its own journey into the scheduler should schedule it
3394 // again, if need be (at which point, this check will pass, if relevant).
3395 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3398 if !gcMarkWorkAvailable(nil) {
3402 // Work is available; we can start an idle GC worker only if there is
3403 // an available P and available worker G.
3405 // We can attempt to acquire these in either order, though both have
3406 // synchronization concerns (see below). Workers are almost always
3407 // available (see comment in findRunnableGCWorker for the one case
3408 // there may be none). Since we're slightly less likely to find a P,
3409 // check for that first.
3411 // Synchronization: note that we must hold sched.lock until we are
3412 // committed to keeping it. Otherwise we cannot put the unnecessary P
3413 // back in sched.pidle without performing the full set of idle
3414 // transition checks.
3416 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3417 // the assumption in gcControllerState.findRunnableGCWorker that an
3418 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3420 pp, now := pidlegetSpinning(0)
3426 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3427 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3433 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3437 gcController.removeIdleMarkWorker()
3443 return pp, node.gp.ptr()
3446 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3447 // going to wake up before the when argument; or it wakes an idle P to service
3448 // timers and the network poller if there isn't one already.
3449 func wakeNetPoller(when int64) {
3450 if sched.lastpoll.Load() == 0 {
3451 // In findrunnable we ensure that when polling the pollUntil
3452 // field is either zero or the time to which the current
3453 // poll is expected to run. This can have a spurious wakeup
3454 // but should never miss a wakeup.
3455 pollerPollUntil := sched.pollUntil.Load()
3456 if pollerPollUntil == 0 || pollerPollUntil > when {
3460 // There are no threads in the network poller, try to get
3461 // one there so it can handle new timers.
3462 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3468 func resetspinning() {
3471 throw("resetspinning: not a spinning m")
3473 gp.m.spinning = false
3474 nmspinning := sched.nmspinning.Add(-1)
3476 throw("findrunnable: negative nmspinning")
3478 // M wakeup policy is deliberately somewhat conservative, so check if we
3479 // need to wakeup another P here. See "Worker thread parking/unparking"
3480 // comment at the top of the file for details.
3484 // injectglist adds each runnable G on the list to some run queue,
3485 // and clears glist. If there is no current P, they are added to the
3486 // global queue, and up to npidle M's are started to run them.
3487 // Otherwise, for each idle P, this adds a G to the global queue
3488 // and starts an M. Any remaining G's are added to the current P's
3490 // This may temporarily acquire sched.lock.
3491 // Can run concurrently with GC.
3492 func injectglist(glist *gList) {
3497 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3498 traceGoUnpark(gp, 0)
3502 // Mark all the goroutines as runnable before we put them
3503 // on the run queues.
3504 head := glist.head.ptr()
3507 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3510 casgstatus(gp, _Gwaiting, _Grunnable)
3513 // Turn the gList into a gQueue.
3519 startIdle := func(n int) {
3520 for i := 0; i < n; i++ {
3521 mp := acquirem() // See comment in startm.
3524 pp, _ := pidlegetSpinning(0)
3531 startm(pp, false, true)
3537 pp := getg().m.p.ptr()
3540 globrunqputbatch(&q, int32(qsize))
3546 npidle := int(sched.npidle.Load())
3549 for n = 0; n < npidle && !q.empty(); n++ {
3555 globrunqputbatch(&globq, int32(n))
3562 runqputbatch(pp, &q, qsize)
3566 // One round of scheduler: find a runnable goroutine and execute it.
3572 throw("schedule: holding locks")
3575 if mp.lockedg != 0 {
3577 execute(mp.lockedg.ptr(), false) // Never returns.
3580 // We should not schedule away from a g that is executing a cgo call,
3581 // since the cgo call is using the m's g0 stack.
3583 throw("schedule: in cgo")
3590 // Safety check: if we are spinning, the run queue should be empty.
3591 // Check this before calling checkTimers, as that might call
3592 // goready to put a ready goroutine on the local run queue.
3593 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3594 throw("schedule: spinning with local work")
3597 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3599 if debug.dontfreezetheworld > 0 && freezing.Load() {
3600 // See comment in freezetheworld. We don't want to perturb
3601 // scheduler state, so we didn't gcstopm in findRunnable, but
3602 // also don't want to allow new goroutines to run.
3604 // Deadlock here rather than in the findRunnable loop so if
3605 // findRunnable is stuck in a loop we don't perturb that
3611 // This thread is going to run a goroutine and is not spinning anymore,
3612 // so if it was marked as spinning we need to reset it now and potentially
3613 // start a new spinning M.
3618 if sched.disable.user && !schedEnabled(gp) {
3619 // Scheduling of this goroutine is disabled. Put it on
3620 // the list of pending runnable goroutines for when we
3621 // re-enable user scheduling and look again.
3623 if schedEnabled(gp) {
3624 // Something re-enabled scheduling while we
3625 // were acquiring the lock.
3628 sched.disable.runnable.pushBack(gp)
3635 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3636 // wake a P if there is one.
3640 if gp.lockedm != 0 {
3641 // Hands off own p to the locked m,
3642 // then blocks waiting for a new p.
3647 execute(gp, inheritTime)
3650 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3651 // Typically a caller sets gp's status away from Grunning and then
3652 // immediately calls dropg to finish the job. The caller is also responsible
3653 // for arranging that gp will be restarted using ready at an
3654 // appropriate time. After calling dropg and arranging for gp to be
3655 // readied later, the caller can do other work but eventually should
3656 // call schedule to restart the scheduling of goroutines on this m.
3660 setMNoWB(&gp.m.curg.m, nil)
3661 setGNoWB(&gp.m.curg, nil)
3664 // checkTimers runs any timers for the P that are ready.
3665 // If now is not 0 it is the current time.
3666 // It returns the passed time or the current time if now was passed as 0.
3667 // and the time when the next timer should run or 0 if there is no next timer,
3668 // and reports whether it ran any timers.
3669 // If the time when the next timer should run is not 0,
3670 // it is always larger than the returned time.
3671 // We pass now in and out to avoid extra calls of nanotime.
3673 //go:yeswritebarrierrec
3674 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3675 // If it's not yet time for the first timer, or the first adjusted
3676 // timer, then there is nothing to do.
3677 next := pp.timer0When.Load()
3678 nextAdj := pp.timerModifiedEarliest.Load()
3679 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3684 // No timers to run or adjust.
3685 return now, 0, false
3692 // Next timer is not ready to run, but keep going
3693 // if we would clear deleted timers.
3694 // This corresponds to the condition below where
3695 // we decide whether to call clearDeletedTimers.
3696 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3697 return now, next, false
3701 lock(&pp.timersLock)
3703 if len(pp.timers) > 0 {
3704 adjusttimers(pp, now)
3705 for len(pp.timers) > 0 {
3706 // Note that runtimer may temporarily unlock
3708 if tw := runtimer(pp, now); tw != 0 {
3718 // If this is the local P, and there are a lot of deleted timers,
3719 // clear them out. We only do this for the local P to reduce
3720 // lock contention on timersLock.
3721 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3722 clearDeletedTimers(pp)
3725 unlock(&pp.timersLock)
3727 return now, pollUntil, ran
3730 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3731 unlock((*mutex)(lock))
3735 // park continuation on g0.
3736 func park_m(gp *g) {
3740 traceGoPark(mp.waitTraceBlockReason, mp.waitTraceSkip)
3743 // N.B. Not using casGToWaiting here because the waitreason is
3744 // set by park_m's caller.
3745 casgstatus(gp, _Grunning, _Gwaiting)
3748 if fn := mp.waitunlockf; fn != nil {
3749 ok := fn(gp, mp.waitlock)
3750 mp.waitunlockf = nil
3754 traceGoUnpark(gp, 2)
3756 casgstatus(gp, _Gwaiting, _Grunnable)
3757 execute(gp, true) // Schedule it back, never returns.
3763 func goschedImpl(gp *g) {
3764 status := readgstatus(gp)
3765 if status&^_Gscan != _Grunning {
3767 throw("bad g status")
3769 casgstatus(gp, _Grunning, _Grunnable)
3782 // Gosched continuation on g0.
3783 func gosched_m(gp *g) {
3790 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3791 func goschedguarded_m(gp *g) {
3793 if !canPreemptM(gp.m) {
3794 gogo(&gp.sched) // never return
3803 func gopreempt_m(gp *g) {
3810 // preemptPark parks gp and puts it in _Gpreempted.
3813 func preemptPark(gp *g) {
3815 traceGoPark(traceBlockPreempted, 0)
3817 status := readgstatus(gp)
3818 if status&^_Gscan != _Grunning {
3820 throw("bad g status")
3823 if gp.asyncSafePoint {
3824 // Double-check that async preemption does not
3825 // happen in SPWRITE assembly functions.
3826 // isAsyncSafePoint must exclude this case.
3827 f := findfunc(gp.sched.pc)
3829 throw("preempt at unknown pc")
3831 if f.flag&abi.FuncFlagSPWrite != 0 {
3832 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3833 throw("preempt SPWRITE")
3837 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3838 // be in _Grunning when we dropg because then we'd be running
3839 // without an M, but the moment we're in _Gpreempted,
3840 // something could claim this G before we've fully cleaned it
3841 // up. Hence, we set the scan bit to lock down further
3842 // transitions until we can dropg.
3843 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3845 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3849 // goyield is like Gosched, but it:
3850 // - emits a GoPreempt trace event instead of a GoSched trace event
3851 // - puts the current G on the runq of the current P instead of the globrunq
3857 func goyield_m(gp *g) {
3862 casgstatus(gp, _Grunning, _Grunnable)
3864 runqput(pp, gp, false)
3868 // Finishes execution of the current goroutine.
3879 // goexit continuation on g0.
3880 func goexit0(gp *g) {
3884 casgstatus(gp, _Grunning, _Gdead)
3885 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3886 if isSystemGoroutine(gp, false) {
3890 locked := gp.lockedm != 0
3893 gp.preemptStop = false
3894 gp.paniconfault = false
3895 gp._defer = nil // should be true already but just in case.
3896 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3898 gp.waitreason = waitReasonZero
3903 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3904 // Flush assist credit to the global pool. This gives
3905 // better information to pacing if the application is
3906 // rapidly creating an exiting goroutines.
3907 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3908 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3909 gcController.bgScanCredit.Add(scanCredit)
3910 gp.gcAssistBytes = 0
3915 if GOARCH == "wasm" { // no threads yet on wasm
3917 schedule() // never returns
3920 if mp.lockedInt != 0 {
3921 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3922 throw("internal lockOSThread error")
3926 // The goroutine may have locked this thread because
3927 // it put it in an unusual kernel state. Kill it
3928 // rather than returning it to the thread pool.
3930 // Return to mstart, which will release the P and exit
3932 if GOOS != "plan9" { // See golang.org/issue/22227.
3935 // Clear lockedExt on plan9 since we may end up re-using
3943 // save updates getg().sched to refer to pc and sp so that a following
3944 // gogo will restore pc and sp.
3946 // save must not have write barriers because invoking a write barrier
3947 // can clobber getg().sched.
3950 //go:nowritebarrierrec
3951 func save(pc, sp uintptr) {
3954 if gp == gp.m.g0 || gp == gp.m.gsignal {
3955 // m.g0.sched is special and must describe the context
3956 // for exiting the thread. mstart1 writes to it directly.
3957 // m.gsignal.sched should not be used at all.
3958 // This check makes sure save calls do not accidentally
3959 // run in contexts where they'd write to system g's.
3960 throw("save on system g not allowed")
3967 // We need to ensure ctxt is zero, but can't have a write
3968 // barrier here. However, it should always already be zero.
3970 if gp.sched.ctxt != nil {
3975 // The goroutine g is about to enter a system call.
3976 // Record that it's not using the cpu anymore.
3977 // This is called only from the go syscall library and cgocall,
3978 // not from the low-level system calls used by the runtime.
3980 // Entersyscall cannot split the stack: the save must
3981 // make g->sched refer to the caller's stack segment, because
3982 // entersyscall is going to return immediately after.
3984 // Nothing entersyscall calls can split the stack either.
3985 // We cannot safely move the stack during an active call to syscall,
3986 // because we do not know which of the uintptr arguments are
3987 // really pointers (back into the stack).
3988 // In practice, this means that we make the fast path run through
3989 // entersyscall doing no-split things, and the slow path has to use systemstack
3990 // to run bigger things on the system stack.
3992 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3993 // saved SP and PC are restored. This is needed when exitsyscall will be called
3994 // from a function further up in the call stack than the parent, as g->syscallsp
3995 // must always point to a valid stack frame. entersyscall below is the normal
3996 // entry point for syscalls, which obtains the SP and PC from the caller.
3999 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
4000 // If the syscall does not block, that is it, we do not emit any other events.
4001 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
4002 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
4003 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
4004 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
4005 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
4006 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
4007 // and we wait for the increment before emitting traceGoSysExit.
4008 // Note that the increment is done even if tracing is not enabled,
4009 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
4012 func reentersyscall(pc, sp uintptr) {
4015 // Disable preemption because during this function g is in Gsyscall status,
4016 // but can have inconsistent g->sched, do not let GC observe it.
4019 // Entersyscall must not call any function that might split/grow the stack.
4020 // (See details in comment above.)
4021 // Catch calls that might, by replacing the stack guard with something that
4022 // will trip any stack check and leaving a flag to tell newstack to die.
4023 gp.stackguard0 = stackPreempt
4024 gp.throwsplit = true
4026 // Leave SP around for GC and traceback.
4030 casgstatus(gp, _Grunning, _Gsyscall)
4031 if staticLockRanking {
4032 // When doing static lock ranking casgstatus can call
4033 // systemstack which clobbers g.sched.
4036 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
4037 systemstack(func() {
4038 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
4039 throw("entersyscall")
4044 systemstack(traceGoSysCall)
4045 // systemstack itself clobbers g.sched.{pc,sp} and we might
4046 // need them later when the G is genuinely blocked in a
4051 if sched.sysmonwait.Load() {
4052 systemstack(entersyscall_sysmon)
4056 if gp.m.p.ptr().runSafePointFn != 0 {
4057 // runSafePointFn may stack split if run on this stack
4058 systemstack(runSafePointFn)
4062 gp.m.syscalltick = gp.m.p.ptr().syscalltick
4067 atomic.Store(&pp.status, _Psyscall)
4068 if sched.gcwaiting.Load() {
4069 systemstack(entersyscall_gcwait)
4076 // Standard syscall entry used by the go syscall library and normal cgo calls.
4078 // This is exported via linkname to assembly in the syscall package and x/sys.
4081 //go:linkname entersyscall
4082 func entersyscall() {
4083 reentersyscall(getcallerpc(), getcallersp())
4086 func entersyscall_sysmon() {
4088 if sched.sysmonwait.Load() {
4089 sched.sysmonwait.Store(false)
4090 notewakeup(&sched.sysmonnote)
4095 func entersyscall_gcwait() {
4097 pp := gp.m.oldp.ptr()
4100 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
4106 if sched.stopwait--; sched.stopwait == 0 {
4107 notewakeup(&sched.stopnote)
4113 // The same as entersyscall(), but with a hint that the syscall is blocking.
4116 func entersyscallblock() {
4119 gp.m.locks++ // see comment in entersyscall
4120 gp.throwsplit = true
4121 gp.stackguard0 = stackPreempt // see comment in entersyscall
4122 gp.m.syscalltick = gp.m.p.ptr().syscalltick
4123 gp.m.p.ptr().syscalltick++
4125 // Leave SP around for GC and traceback.
4129 gp.syscallsp = gp.sched.sp
4130 gp.syscallpc = gp.sched.pc
4131 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
4135 systemstack(func() {
4136 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
4137 throw("entersyscallblock")
4140 casgstatus(gp, _Grunning, _Gsyscall)
4141 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
4142 systemstack(func() {
4143 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
4144 throw("entersyscallblock")
4148 systemstack(entersyscallblock_handoff)
4150 // Resave for traceback during blocked call.
4151 save(getcallerpc(), getcallersp())
4156 func entersyscallblock_handoff() {
4159 traceGoSysBlock(getg().m.p.ptr())
4161 handoffp(releasep())
4164 // The goroutine g exited its system call.
4165 // Arrange for it to run on a cpu again.
4166 // This is called only from the go syscall library, not
4167 // from the low-level system calls used by the runtime.
4169 // Write barriers are not allowed because our P may have been stolen.
4171 // This is exported via linkname to assembly in the syscall package.
4174 //go:nowritebarrierrec
4175 //go:linkname exitsyscall
4176 func exitsyscall() {
4179 gp.m.locks++ // see comment in entersyscall
4180 if getcallersp() > gp.syscallsp {
4181 throw("exitsyscall: syscall frame is no longer valid")
4185 oldp := gp.m.oldp.ptr()
4187 if exitsyscallfast(oldp) {
4188 // When exitsyscallfast returns success, we have a P so can now use
4190 if goroutineProfile.active {
4191 // Make sure that gp has had its stack written out to the goroutine
4192 // profile, exactly as it was when the goroutine profiler first
4193 // stopped the world.
4194 systemstack(func() {
4195 tryRecordGoroutineProfileWB(gp)
4199 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4200 systemstack(traceGoStart)
4203 // There's a cpu for us, so we can run.
4204 gp.m.p.ptr().syscalltick++
4205 // We need to cas the status and scan before resuming...
4206 casgstatus(gp, _Gsyscall, _Grunning)
4208 // Garbage collector isn't running (since we are),
4209 // so okay to clear syscallsp.
4213 // restore the preemption request in case we've cleared it in newstack
4214 gp.stackguard0 = stackPreempt
4216 // otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
4217 gp.stackguard0 = gp.stack.lo + stackGuard
4219 gp.throwsplit = false
4221 if sched.disable.user && !schedEnabled(gp) {
4222 // Scheduling of this goroutine is disabled.
4230 // Wait till traceGoSysBlock event is emitted.
4231 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4232 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
4235 // We can't trace syscall exit right now because we don't have a P.
4236 // Tracing code can invoke write barriers that cannot run without a P.
4237 // So instead we remember the syscall exit time and emit the event
4238 // in execute when we have a P.
4239 gp.trace.sysExitTime = traceClockNow()
4244 // Call the scheduler.
4247 // Scheduler returned, so we're allowed to run now.
4248 // Delete the syscallsp information that we left for
4249 // the garbage collector during the system call.
4250 // Must wait until now because until gosched returns
4251 // we don't know for sure that the garbage collector
4254 gp.m.p.ptr().syscalltick++
4255 gp.throwsplit = false
4259 func exitsyscallfast(oldp *p) bool {
4262 // Freezetheworld sets stopwait but does not retake P's.
4263 if sched.stopwait == freezeStopWait {
4267 // Try to re-acquire the last P.
4268 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4269 // There's a cpu for us, so we can run.
4271 exitsyscallfast_reacquired()
4275 // Try to get any other idle P.
4276 if sched.pidle != 0 {
4278 systemstack(func() {
4279 ok = exitsyscallfast_pidle()
4280 if ok && traceEnabled() {
4282 // Wait till traceGoSysBlock event is emitted.
4283 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4284 for oldp.syscalltick == gp.m.syscalltick {
4298 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4299 // has successfully reacquired the P it was running on before the
4303 func exitsyscallfast_reacquired() {
4305 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4307 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4308 // traceGoSysBlock for this syscall was already emitted,
4309 // but here we effectively retake the p from the new syscall running on the same p.
4310 systemstack(func() {
4311 // Denote blocking of the new syscall.
4312 traceGoSysBlock(gp.m.p.ptr())
4313 // Denote completion of the current syscall.
4317 gp.m.p.ptr().syscalltick++
4321 func exitsyscallfast_pidle() bool {
4323 pp, _ := pidleget(0)
4324 if pp != nil && sched.sysmonwait.Load() {
4325 sched.sysmonwait.Store(false)
4326 notewakeup(&sched.sysmonnote)
4336 // exitsyscall slow path on g0.
4337 // Failed to acquire P, enqueue gp as runnable.
4339 // Called via mcall, so gp is the calling g from this M.
4341 //go:nowritebarrierrec
4342 func exitsyscall0(gp *g) {
4343 casgstatus(gp, _Gsyscall, _Grunnable)
4347 if schedEnabled(gp) {
4354 // Below, we stoplockedm if gp is locked. globrunqput releases
4355 // ownership of gp, so we must check if gp is locked prior to
4356 // committing the release by unlocking sched.lock, otherwise we
4357 // could race with another M transitioning gp from unlocked to
4359 locked = gp.lockedm != 0
4360 } else if sched.sysmonwait.Load() {
4361 sched.sysmonwait.Store(false)
4362 notewakeup(&sched.sysmonnote)
4367 execute(gp, false) // Never returns.
4370 // Wait until another thread schedules gp and so m again.
4372 // N.B. lockedm must be this M, as this g was running on this M
4373 // before entersyscall.
4375 execute(gp, false) // Never returns.
4378 schedule() // Never returns.
4381 // Called from syscall package before fork.
4383 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4385 func syscall_runtime_BeforeFork() {
4388 // Block signals during a fork, so that the child does not run
4389 // a signal handler before exec if a signal is sent to the process
4390 // group. See issue #18600.
4392 sigsave(&gp.m.sigmask)
4395 // This function is called before fork in syscall package.
4396 // Code between fork and exec must not allocate memory nor even try to grow stack.
4397 // Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
4398 // runtime_AfterFork will undo this in parent process, but not in child.
4399 gp.stackguard0 = stackFork
4402 // Called from syscall package after fork in parent.
4404 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4406 func syscall_runtime_AfterFork() {
4409 // See the comments in beforefork.
4410 gp.stackguard0 = gp.stack.lo + stackGuard
4412 msigrestore(gp.m.sigmask)
4417 // inForkedChild is true while manipulating signals in the child process.
4418 // This is used to avoid calling libc functions in case we are using vfork.
4419 var inForkedChild bool
4421 // Called from syscall package after fork in child.
4422 // It resets non-sigignored signals to the default handler, and
4423 // restores the signal mask in preparation for the exec.
4425 // Because this might be called during a vfork, and therefore may be
4426 // temporarily sharing address space with the parent process, this must
4427 // not change any global variables or calling into C code that may do so.
4429 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4431 //go:nowritebarrierrec
4432 func syscall_runtime_AfterForkInChild() {
4433 // It's OK to change the global variable inForkedChild here
4434 // because we are going to change it back. There is no race here,
4435 // because if we are sharing address space with the parent process,
4436 // then the parent process can not be running concurrently.
4437 inForkedChild = true
4439 clearSignalHandlers()
4441 // When we are the child we are the only thread running,
4442 // so we know that nothing else has changed gp.m.sigmask.
4443 msigrestore(getg().m.sigmask)
4445 inForkedChild = false
4448 // pendingPreemptSignals is the number of preemption signals
4449 // that have been sent but not received. This is only used on Darwin.
4451 var pendingPreemptSignals atomic.Int32
4453 // Called from syscall package before Exec.
4455 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4456 func syscall_runtime_BeforeExec() {
4457 // Prevent thread creation during exec.
4460 // On Darwin, wait for all pending preemption signals to
4461 // be received. See issue #41702.
4462 if GOOS == "darwin" || GOOS == "ios" {
4463 for pendingPreemptSignals.Load() > 0 {
4469 // Called from syscall package after Exec.
4471 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4472 func syscall_runtime_AfterExec() {
4476 // Allocate a new g, with a stack big enough for stacksize bytes.
4477 func malg(stacksize int32) *g {
4480 stacksize = round2(stackSystem + stacksize)
4481 systemstack(func() {
4482 newg.stack = stackalloc(uint32(stacksize))
4484 newg.stackguard0 = newg.stack.lo + stackGuard
4485 newg.stackguard1 = ^uintptr(0)
4486 // Clear the bottom word of the stack. We record g
4487 // there on gsignal stack during VDSO on ARM and ARM64.
4488 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4493 // Create a new g running fn.
4494 // Put it on the queue of g's waiting to run.
4495 // The compiler turns a go statement into a call to this.
4496 func newproc(fn *funcval) {
4499 systemstack(func() {
4500 newg := newproc1(fn, gp, pc)
4502 pp := getg().m.p.ptr()
4503 runqput(pp, newg, true)
4511 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4512 // address of the go statement that created this. The caller is responsible
4513 // for adding the new g to the scheduler.
4514 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4516 fatal("go of nil func value")
4519 mp := acquirem() // disable preemption because we hold M and P in local vars.
4523 newg = malg(stackMin)
4524 casgstatus(newg, _Gidle, _Gdead)
4525 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4527 if newg.stack.hi == 0 {
4528 throw("newproc1: newg missing stack")
4531 if readgstatus(newg) != _Gdead {
4532 throw("newproc1: new g is not Gdead")
4535 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4536 totalSize = alignUp(totalSize, sys.StackAlign)
4537 sp := newg.stack.hi - totalSize
4540 *(*uintptr)(unsafe.Pointer(sp)) = 0
4543 if GOARCH == "arm64" {
4545 *(*uintptr)(unsafe.Pointer(sp - goarch.PtrSize)) = 0
4548 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4551 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4552 newg.sched.g = guintptr(unsafe.Pointer(newg))
4553 gostartcallfn(&newg.sched, fn)
4554 newg.parentGoid = callergp.goid
4555 newg.gopc = callerpc
4556 newg.ancestors = saveAncestors(callergp)
4557 newg.startpc = fn.fn
4558 if isSystemGoroutine(newg, false) {
4561 // Only user goroutines inherit pprof labels.
4563 newg.labels = mp.curg.labels
4565 if goroutineProfile.active {
4566 // A concurrent goroutine profile is running. It should include
4567 // exactly the set of goroutines that were alive when the goroutine
4568 // profiler first stopped the world. That does not include newg, so
4569 // mark it as not needing a profile before transitioning it from
4571 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4574 // Track initial transition?
4575 newg.trackingSeq = uint8(fastrand())
4576 if newg.trackingSeq%gTrackingPeriod == 0 {
4577 newg.tracking = true
4579 casgstatus(newg, _Gdead, _Grunnable)
4580 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4582 if pp.goidcache == pp.goidcacheend {
4583 // Sched.goidgen is the last allocated id,
4584 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4585 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4586 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4587 pp.goidcache -= _GoidCacheBatch - 1
4588 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4590 newg.goid = pp.goidcache
4593 newg.racectx = racegostart(callerpc)
4595 if newg.labels != nil {
4596 // See note in proflabel.go on labelSync's role in synchronizing
4597 // with the reads in the signal handler.
4598 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4602 traceGoCreate(newg, newg.startpc)
4609 // saveAncestors copies previous ancestors of the given caller g and
4610 // includes info for the current caller into a new set of tracebacks for
4611 // a g being created.
4612 func saveAncestors(callergp *g) *[]ancestorInfo {
4613 // Copy all prior info, except for the root goroutine (goid 0).
4614 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4617 var callerAncestors []ancestorInfo
4618 if callergp.ancestors != nil {
4619 callerAncestors = *callergp.ancestors
4621 n := int32(len(callerAncestors)) + 1
4622 if n > debug.tracebackancestors {
4623 n = debug.tracebackancestors
4625 ancestors := make([]ancestorInfo, n)
4626 copy(ancestors[1:], callerAncestors)
4628 var pcs [tracebackInnerFrames]uintptr
4629 npcs := gcallers(callergp, 0, pcs[:])
4630 ipcs := make([]uintptr, npcs)
4632 ancestors[0] = ancestorInfo{
4634 goid: callergp.goid,
4635 gopc: callergp.gopc,
4638 ancestorsp := new([]ancestorInfo)
4639 *ancestorsp = ancestors
4643 // Put on gfree list.
4644 // If local list is too long, transfer a batch to the global list.
4645 func gfput(pp *p, gp *g) {
4646 if readgstatus(gp) != _Gdead {
4647 throw("gfput: bad status (not Gdead)")
4650 stksize := gp.stack.hi - gp.stack.lo
4652 if stksize != uintptr(startingStackSize) {
4653 // non-standard stack size - free it.
4662 if pp.gFree.n >= 64 {
4668 for pp.gFree.n >= 32 {
4669 gp := pp.gFree.pop()
4671 if gp.stack.lo == 0 {
4678 lock(&sched.gFree.lock)
4679 sched.gFree.noStack.pushAll(noStackQ)
4680 sched.gFree.stack.pushAll(stackQ)
4681 sched.gFree.n += inc
4682 unlock(&sched.gFree.lock)
4686 // Get from gfree list.
4687 // If local list is empty, grab a batch from global list.
4688 func gfget(pp *p) *g {
4690 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4691 lock(&sched.gFree.lock)
4692 // Move a batch of free Gs to the P.
4693 for pp.gFree.n < 32 {
4694 // Prefer Gs with stacks.
4695 gp := sched.gFree.stack.pop()
4697 gp = sched.gFree.noStack.pop()
4706 unlock(&sched.gFree.lock)
4709 gp := pp.gFree.pop()
4714 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4715 // Deallocate old stack. We kept it in gfput because it was the
4716 // right size when the goroutine was put on the free list, but
4717 // the right size has changed since then.
4718 systemstack(func() {
4725 if gp.stack.lo == 0 {
4726 // Stack was deallocated in gfput or just above. Allocate a new one.
4727 systemstack(func() {
4728 gp.stack = stackalloc(startingStackSize)
4730 gp.stackguard0 = gp.stack.lo + stackGuard
4733 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4736 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4739 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4745 // Purge all cached G's from gfree list to the global list.
4746 func gfpurge(pp *p) {
4752 for !pp.gFree.empty() {
4753 gp := pp.gFree.pop()
4755 if gp.stack.lo == 0 {
4762 lock(&sched.gFree.lock)
4763 sched.gFree.noStack.pushAll(noStackQ)
4764 sched.gFree.stack.pushAll(stackQ)
4765 sched.gFree.n += inc
4766 unlock(&sched.gFree.lock)
4769 // Breakpoint executes a breakpoint trap.
4774 // dolockOSThread is called by LockOSThread and lockOSThread below
4775 // after they modify m.locked. Do not allow preemption during this call,
4776 // or else the m might be different in this function than in the caller.
4779 func dolockOSThread() {
4780 if GOARCH == "wasm" {
4781 return // no threads on wasm yet
4784 gp.m.lockedg.set(gp)
4785 gp.lockedm.set(gp.m)
4788 // LockOSThread wires the calling goroutine to its current operating system thread.
4789 // The calling goroutine will always execute in that thread,
4790 // and no other goroutine will execute in it,
4791 // until the calling goroutine has made as many calls to
4792 // UnlockOSThread as to LockOSThread.
4793 // If the calling goroutine exits without unlocking the thread,
4794 // the thread will be terminated.
4796 // All init functions are run on the startup thread. Calling LockOSThread
4797 // from an init function will cause the main function to be invoked on
4800 // A goroutine should call LockOSThread before calling OS services or
4801 // non-Go library functions that depend on per-thread state.
4804 func LockOSThread() {
4805 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4806 // If we need to start a new thread from the locked
4807 // thread, we need the template thread. Start it now
4808 // while we're in a known-good state.
4809 startTemplateThread()
4813 if gp.m.lockedExt == 0 {
4815 panic("LockOSThread nesting overflow")
4821 func lockOSThread() {
4822 getg().m.lockedInt++
4826 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4827 // after they update m->locked. Do not allow preemption during this call,
4828 // or else the m might be in different in this function than in the caller.
4831 func dounlockOSThread() {
4832 if GOARCH == "wasm" {
4833 return // no threads on wasm yet
4836 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4843 // UnlockOSThread undoes an earlier call to LockOSThread.
4844 // If this drops the number of active LockOSThread calls on the
4845 // calling goroutine to zero, it unwires the calling goroutine from
4846 // its fixed operating system thread.
4847 // If there are no active LockOSThread calls, this is a no-op.
4849 // Before calling UnlockOSThread, the caller must ensure that the OS
4850 // thread is suitable for running other goroutines. If the caller made
4851 // any permanent changes to the state of the thread that would affect
4852 // other goroutines, it should not call this function and thus leave
4853 // the goroutine locked to the OS thread until the goroutine (and
4854 // hence the thread) exits.
4857 func UnlockOSThread() {
4859 if gp.m.lockedExt == 0 {
4867 func unlockOSThread() {
4869 if gp.m.lockedInt == 0 {
4870 systemstack(badunlockosthread)
4876 func badunlockosthread() {
4877 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4880 func gcount() int32 {
4881 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4882 for _, pp := range allp {
4886 // All these variables can be changed concurrently, so the result can be inconsistent.
4887 // But at least the current goroutine is running.
4894 func mcount() int32 {
4895 return int32(sched.mnext - sched.nmfreed)
4899 signalLock atomic.Uint32
4901 // Must hold signalLock to write. Reads may be lock-free, but
4902 // signalLock should be taken to synchronize with changes.
4906 func _System() { _System() }
4907 func _ExternalCode() { _ExternalCode() }
4908 func _LostExternalCode() { _LostExternalCode() }
4909 func _GC() { _GC() }
4910 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4911 func _VDSO() { _VDSO() }
4913 // Called if we receive a SIGPROF signal.
4914 // Called by the signal handler, may run during STW.
4916 //go:nowritebarrierrec
4917 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4918 if prof.hz.Load() == 0 {
4922 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4923 // We must check this to avoid a deadlock between setcpuprofilerate
4924 // and the call to cpuprof.add, below.
4925 if mp != nil && mp.profilehz == 0 {
4929 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4930 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4931 // the critical section, it creates a deadlock (when writing the sample).
4932 // As a workaround, create a counter of SIGPROFs while in critical section
4933 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4934 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4935 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4936 if f := findfunc(pc); f.valid() {
4937 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4938 cpuprof.lostAtomic++
4942 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4943 // runtime/internal/atomic functions call into kernel
4944 // helpers on arm < 7. See
4945 // runtime/internal/atomic/sys_linux_arm.s.
4946 cpuprof.lostAtomic++
4951 // Profiling runs concurrently with GC, so it must not allocate.
4952 // Set a trap in case the code does allocate.
4953 // Note that on windows, one thread takes profiles of all the
4954 // other threads, so mp is usually not getg().m.
4955 // In fact mp may not even be stopped.
4956 // See golang.org/issue/17165.
4957 getg().m.mallocing++
4960 var stk [maxCPUProfStack]uintptr
4962 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4964 // Check cgoCallersUse to make sure that we are not
4965 // interrupting other code that is fiddling with
4966 // cgoCallers. We are running in a signal handler
4967 // with all signals blocked, so we don't have to worry
4968 // about any other code interrupting us.
4969 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4970 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4973 n += copy(stk[:], mp.cgoCallers[:cgoOff])
4974 mp.cgoCallers[0] = 0
4977 // Collect Go stack that leads to the cgo call.
4978 u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
4979 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4980 // Libcall, i.e. runtime syscall on windows.
4981 // Collect Go stack that leads to the call.
4982 u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
4983 } else if mp != nil && mp.vdsoSP != 0 {
4984 // VDSO call, e.g. nanotime1 on Linux.
4985 // Collect Go stack that leads to the call.
4986 u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
4988 u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
4990 n += tracebackPCs(&u, 0, stk[n:])
4993 // Normal traceback is impossible or has failed.
4994 // Account it against abstract "System" or "GC".
4997 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4998 } else if pc > firstmoduledata.etext {
4999 // "ExternalCode" is better than "etext".
5000 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
5003 if mp.preemptoff != "" {
5004 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
5006 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
5010 if prof.hz.Load() != 0 {
5011 // Note: it can happen on Windows that we interrupted a system thread
5012 // with no g, so gp could nil. The other nil checks are done out of
5013 // caution, but not expected to be nil in practice.
5014 var tagPtr *unsafe.Pointer
5015 if gp != nil && gp.m != nil && gp.m.curg != nil {
5016 tagPtr = &gp.m.curg.labels
5018 cpuprof.add(tagPtr, stk[:n])
5022 if gp != nil && gp.m != nil {
5023 if gp.m.curg != nil {
5028 traceCPUSample(gprof, pp, stk[:n])
5030 getg().m.mallocing--
5033 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
5034 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
5035 func setcpuprofilerate(hz int32) {
5036 // Force sane arguments.
5041 // Disable preemption, otherwise we can be rescheduled to another thread
5042 // that has profiling enabled.
5046 // Stop profiler on this thread so that it is safe to lock prof.
5047 // if a profiling signal came in while we had prof locked,
5048 // it would deadlock.
5049 setThreadCPUProfiler(0)
5051 for !prof.signalLock.CompareAndSwap(0, 1) {
5054 if prof.hz.Load() != hz {
5055 setProcessCPUProfiler(hz)
5058 prof.signalLock.Store(0)
5061 sched.profilehz = hz
5065 setThreadCPUProfiler(hz)
5071 // init initializes pp, which may be a freshly allocated p or a
5072 // previously destroyed p, and transitions it to status _Pgcstop.
5073 func (pp *p) init(id int32) {
5075 pp.status = _Pgcstop
5076 pp.sudogcache = pp.sudogbuf[:0]
5077 pp.deferpool = pp.deferpoolbuf[:0]
5079 if pp.mcache == nil {
5082 throw("missing mcache?")
5084 // Use the bootstrap mcache0. Only one P will get
5085 // mcache0: the one with ID 0.
5088 pp.mcache = allocmcache()
5091 if raceenabled && pp.raceprocctx == 0 {
5093 pp.raceprocctx = raceprocctx0
5094 raceprocctx0 = 0 // bootstrap
5096 pp.raceprocctx = raceproccreate()
5099 lockInit(&pp.timersLock, lockRankTimers)
5101 // This P may get timers when it starts running. Set the mask here
5102 // since the P may not go through pidleget (notably P 0 on startup).
5104 // Similarly, we may not go through pidleget before this P starts
5105 // running if it is P 0 on startup.
5109 // destroy releases all of the resources associated with pp and
5110 // transitions it to status _Pdead.
5112 // sched.lock must be held and the world must be stopped.
5113 func (pp *p) destroy() {
5114 assertLockHeld(&sched.lock)
5115 assertWorldStopped()
5117 // Move all runnable goroutines to the global queue
5118 for pp.runqhead != pp.runqtail {
5119 // Pop from tail of local queue
5121 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
5122 // Push onto head of global queue
5125 if pp.runnext != 0 {
5126 globrunqputhead(pp.runnext.ptr())
5129 if len(pp.timers) > 0 {
5130 plocal := getg().m.p.ptr()
5131 // The world is stopped, but we acquire timersLock to
5132 // protect against sysmon calling timeSleepUntil.
5133 // This is the only case where we hold the timersLock of
5134 // more than one P, so there are no deadlock concerns.
5135 lock(&plocal.timersLock)
5136 lock(&pp.timersLock)
5137 moveTimers(plocal, pp.timers)
5139 pp.numTimers.Store(0)
5140 pp.deletedTimers.Store(0)
5141 pp.timer0When.Store(0)
5142 unlock(&pp.timersLock)
5143 unlock(&plocal.timersLock)
5145 // Flush p's write barrier buffer.
5146 if gcphase != _GCoff {
5150 for i := range pp.sudogbuf {
5151 pp.sudogbuf[i] = nil
5153 pp.sudogcache = pp.sudogbuf[:0]
5154 pp.pinnerCache = nil
5155 for j := range pp.deferpoolbuf {
5156 pp.deferpoolbuf[j] = nil
5158 pp.deferpool = pp.deferpoolbuf[:0]
5159 systemstack(func() {
5160 for i := 0; i < pp.mspancache.len; i++ {
5161 // Safe to call since the world is stopped.
5162 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
5164 pp.mspancache.len = 0
5166 pp.pcache.flush(&mheap_.pages)
5167 unlock(&mheap_.lock)
5169 freemcache(pp.mcache)
5174 if pp.timerRaceCtx != 0 {
5175 // The race detector code uses a callback to fetch
5176 // the proc context, so arrange for that callback
5177 // to see the right thing.
5178 // This hack only works because we are the only
5184 racectxend(pp.timerRaceCtx)
5189 raceprocdestroy(pp.raceprocctx)
5196 // Change number of processors.
5198 // sched.lock must be held, and the world must be stopped.
5200 // gcworkbufs must not be being modified by either the GC or the write barrier
5201 // code, so the GC must not be running if the number of Ps actually changes.
5203 // Returns list of Ps with local work, they need to be scheduled by the caller.
5204 func procresize(nprocs int32) *p {
5205 assertLockHeld(&sched.lock)
5206 assertWorldStopped()
5209 if old < 0 || nprocs <= 0 {
5210 throw("procresize: invalid arg")
5213 traceGomaxprocs(nprocs)
5216 // update statistics
5218 if sched.procresizetime != 0 {
5219 sched.totaltime += int64(old) * (now - sched.procresizetime)
5221 sched.procresizetime = now
5223 maskWords := (nprocs + 31) / 32
5225 // Grow allp if necessary.
5226 if nprocs > int32(len(allp)) {
5227 // Synchronize with retake, which could be running
5228 // concurrently since it doesn't run on a P.
5230 if nprocs <= int32(cap(allp)) {
5231 allp = allp[:nprocs]
5233 nallp := make([]*p, nprocs)
5234 // Copy everything up to allp's cap so we
5235 // never lose old allocated Ps.
5236 copy(nallp, allp[:cap(allp)])
5240 if maskWords <= int32(cap(idlepMask)) {
5241 idlepMask = idlepMask[:maskWords]
5242 timerpMask = timerpMask[:maskWords]
5244 nidlepMask := make([]uint32, maskWords)
5245 // No need to copy beyond len, old Ps are irrelevant.
5246 copy(nidlepMask, idlepMask)
5247 idlepMask = nidlepMask
5249 ntimerpMask := make([]uint32, maskWords)
5250 copy(ntimerpMask, timerpMask)
5251 timerpMask = ntimerpMask
5256 // initialize new P's
5257 for i := old; i < nprocs; i++ {
5263 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5267 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5268 // continue to use the current P
5269 gp.m.p.ptr().status = _Prunning
5270 gp.m.p.ptr().mcache.prepareForSweep()
5272 // release the current P and acquire allp[0].
5274 // We must do this before destroying our current P
5275 // because p.destroy itself has write barriers, so we
5276 // need to do that from a valid P.
5279 // Pretend that we were descheduled
5280 // and then scheduled again to keep
5283 traceProcStop(gp.m.p.ptr())
5297 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5300 // release resources from unused P's
5301 for i := nprocs; i < old; i++ {
5304 // can't free P itself because it can be referenced by an M in syscall
5308 if int32(len(allp)) != nprocs {
5310 allp = allp[:nprocs]
5311 idlepMask = idlepMask[:maskWords]
5312 timerpMask = timerpMask[:maskWords]
5317 for i := nprocs - 1; i >= 0; i-- {
5319 if gp.m.p.ptr() == pp {
5327 pp.link.set(runnablePs)
5331 stealOrder.reset(uint32(nprocs))
5332 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5333 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5335 // Notify the limiter that the amount of procs has changed.
5336 gcCPULimiter.resetCapacity(now, nprocs)
5341 // Associate p and the current m.
5343 // This function is allowed to have write barriers even if the caller
5344 // isn't because it immediately acquires pp.
5346 //go:yeswritebarrierrec
5347 func acquirep(pp *p) {
5348 // Do the part that isn't allowed to have write barriers.
5351 // Have p; write barriers now allowed.
5353 // Perform deferred mcache flush before this P can allocate
5354 // from a potentially stale mcache.
5355 pp.mcache.prepareForSweep()
5362 // wirep is the first step of acquirep, which actually associates the
5363 // current M to pp. This is broken out so we can disallow write
5364 // barriers for this part, since we don't yet have a P.
5366 //go:nowritebarrierrec
5372 throw("wirep: already in go")
5374 if pp.m != 0 || pp.status != _Pidle {
5379 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5380 throw("wirep: invalid p state")
5384 pp.status = _Prunning
5387 // Disassociate p and the current m.
5388 func releasep() *p {
5392 throw("releasep: invalid arg")
5395 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5396 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5397 throw("releasep: invalid p state")
5400 traceProcStop(gp.m.p.ptr())
5408 func incidlelocked(v int32) {
5410 sched.nmidlelocked += v
5417 // Check for deadlock situation.
5418 // The check is based on number of running M's, if 0 -> deadlock.
5419 // sched.lock must be held.
5421 assertLockHeld(&sched.lock)
5423 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5424 // there are no running goroutines. The calling program is
5425 // assumed to be running.
5426 if islibrary || isarchive {
5430 // If we are dying because of a signal caught on an already idle thread,
5431 // freezetheworld will cause all running threads to block.
5432 // And runtime will essentially enter into deadlock state,
5433 // except that there is a thread that will call exit soon.
5434 if panicking.Load() > 0 {
5438 // If we are not running under cgo, but we have an extra M then account
5439 // for it. (It is possible to have an extra M on Windows without cgo to
5440 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5443 if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
5447 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5452 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5454 throw("checkdead: inconsistent counts")
5458 forEachG(func(gp *g) {
5459 if isSystemGoroutine(gp, false) {
5462 s := readgstatus(gp)
5463 switch s &^ _Gscan {
5470 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5472 throw("checkdead: runnable g")
5475 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5476 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5477 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5480 // Maybe jump time forward for playground.
5482 if when := timeSleepUntil(); when < maxWhen {
5485 // Start an M to steal the timer.
5486 pp, _ := pidleget(faketime)
5488 // There should always be a free P since
5489 // nothing is running.
5491 throw("checkdead: no p for timer")
5495 // There should always be a free M since
5496 // nothing is running.
5498 throw("checkdead: no m for timer")
5500 // M must be spinning to steal. We set this to be
5501 // explicit, but since this is the only M it would
5502 // become spinning on its own anyways.
5503 sched.nmspinning.Add(1)
5506 notewakeup(&mp.park)
5511 // There are no goroutines running, so we can look at the P's.
5512 for _, pp := range allp {
5513 if len(pp.timers) > 0 {
5518 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5519 fatal("all goroutines are asleep - deadlock!")
5522 // forcegcperiod is the maximum time in nanoseconds between garbage
5523 // collections. If we go this long without a garbage collection, one
5524 // is forced to run.
5526 // This is a variable for testing purposes. It normally doesn't change.
5527 var forcegcperiod int64 = 2 * 60 * 1e9
5529 // needSysmonWorkaround is true if the workaround for
5530 // golang.org/issue/42515 is needed on NetBSD.
5531 var needSysmonWorkaround bool = false
5533 // Always runs without a P, so write barriers are not allowed.
5535 //go:nowritebarrierrec
5542 lasttrace := int64(0)
5543 idle := 0 // how many cycles in succession we had not wokeup somebody
5547 if idle == 0 { // start with 20us sleep...
5549 } else if idle > 50 { // start doubling the sleep after 1ms...
5552 if delay > 10*1000 { // up to 10ms
5557 // sysmon should not enter deep sleep if schedtrace is enabled so that
5558 // it can print that information at the right time.
5560 // It should also not enter deep sleep if there are any active P's so
5561 // that it can retake P's from syscalls, preempt long running G's, and
5562 // poll the network if all P's are busy for long stretches.
5564 // It should wakeup from deep sleep if any P's become active either due
5565 // to exiting a syscall or waking up due to a timer expiring so that it
5566 // can resume performing those duties. If it wakes from a syscall it
5567 // resets idle and delay as a bet that since it had retaken a P from a
5568 // syscall before, it may need to do it again shortly after the
5569 // application starts work again. It does not reset idle when waking
5570 // from a timer to avoid adding system load to applications that spend
5571 // most of their time sleeping.
5573 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5575 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5576 syscallWake := false
5577 next := timeSleepUntil()
5579 sched.sysmonwait.Store(true)
5581 // Make wake-up period small enough
5582 // for the sampling to be correct.
5583 sleep := forcegcperiod / 2
5584 if next-now < sleep {
5587 shouldRelax := sleep >= osRelaxMinNS
5591 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5596 sched.sysmonwait.Store(false)
5597 noteclear(&sched.sysmonnote)
5607 lock(&sched.sysmonlock)
5608 // Update now in case we blocked on sysmonnote or spent a long time
5609 // blocked on schedlock or sysmonlock above.
5612 // trigger libc interceptors if needed
5613 if *cgo_yield != nil {
5614 asmcgocall(*cgo_yield, nil)
5616 // poll network if not polled for more than 10ms
5617 lastpoll := sched.lastpoll.Load()
5618 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5619 sched.lastpoll.CompareAndSwap(lastpoll, now)
5620 list, delta := netpoll(0) // non-blocking - returns list of goroutines
5622 // Need to decrement number of idle locked M's
5623 // (pretending that one more is running) before injectglist.
5624 // Otherwise it can lead to the following situation:
5625 // injectglist grabs all P's but before it starts M's to run the P's,
5626 // another M returns from syscall, finishes running its G,
5627 // observes that there is no work to do and no other running M's
5628 // and reports deadlock.
5632 netpollAdjustWaiters(delta)
5635 if GOOS == "netbsd" && needSysmonWorkaround {
5636 // netpoll is responsible for waiting for timer
5637 // expiration, so we typically don't have to worry
5638 // about starting an M to service timers. (Note that
5639 // sleep for timeSleepUntil above simply ensures sysmon
5640 // starts running again when that timer expiration may
5641 // cause Go code to run again).
5643 // However, netbsd has a kernel bug that sometimes
5644 // misses netpollBreak wake-ups, which can lead to
5645 // unbounded delays servicing timers. If we detect this
5646 // overrun, then startm to get something to handle the
5649 // See issue 42515 and
5650 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5651 if next := timeSleepUntil(); next < now {
5652 startm(nil, false, false)
5655 if scavenger.sysmonWake.Load() != 0 {
5656 // Kick the scavenger awake if someone requested it.
5659 // retake P's blocked in syscalls
5660 // and preempt long running G's
5661 if retake(now) != 0 {
5666 // check if we need to force a GC
5667 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5669 forcegc.idle.Store(false)
5671 list.push(forcegc.g)
5673 unlock(&forcegc.lock)
5675 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5677 schedtrace(debug.scheddetail > 0)
5679 unlock(&sched.sysmonlock)
5683 type sysmontick struct {
5690 // forcePreemptNS is the time slice given to a G before it is
5692 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5694 func retake(now int64) uint32 {
5696 // Prevent allp slice changes. This lock will be completely
5697 // uncontended unless we're already stopping the world.
5699 // We can't use a range loop over allp because we may
5700 // temporarily drop the allpLock. Hence, we need to re-fetch
5701 // allp each time around the loop.
5702 for i := 0; i < len(allp); i++ {
5705 // This can happen if procresize has grown
5706 // allp but not yet created new Ps.
5709 pd := &pp.sysmontick
5712 if s == _Prunning || s == _Psyscall {
5713 // Preempt G if it's running for too long.
5714 t := int64(pp.schedtick)
5715 if int64(pd.schedtick) != t {
5716 pd.schedtick = uint32(t)
5718 } else if pd.schedwhen+forcePreemptNS <= now {
5720 // In case of syscall, preemptone() doesn't
5721 // work, because there is no M wired to P.
5726 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5727 t := int64(pp.syscalltick)
5728 if !sysretake && int64(pd.syscalltick) != t {
5729 pd.syscalltick = uint32(t)
5730 pd.syscallwhen = now
5733 // On the one hand we don't want to retake Ps if there is no other work to do,
5734 // but on the other hand we want to retake them eventually
5735 // because they can prevent the sysmon thread from deep sleep.
5736 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5739 // Drop allpLock so we can take sched.lock.
5741 // Need to decrement number of idle locked M's
5742 // (pretending that one more is running) before the CAS.
5743 // Otherwise the M from which we retake can exit the syscall,
5744 // increment nmidle and report deadlock.
5746 if atomic.Cas(&pp.status, s, _Pidle) {
5763 // Tell all goroutines that they have been preempted and they should stop.
5764 // This function is purely best-effort. It can fail to inform a goroutine if a
5765 // processor just started running it.
5766 // No locks need to be held.
5767 // Returns true if preemption request was issued to at least one goroutine.
5768 func preemptall() bool {
5770 for _, pp := range allp {
5771 if pp.status != _Prunning {
5781 // Tell the goroutine running on processor P to stop.
5782 // This function is purely best-effort. It can incorrectly fail to inform the
5783 // goroutine. It can inform the wrong goroutine. Even if it informs the
5784 // correct goroutine, that goroutine might ignore the request if it is
5785 // simultaneously executing newstack.
5786 // No lock needs to be held.
5787 // Returns true if preemption request was issued.
5788 // The actual preemption will happen at some point in the future
5789 // and will be indicated by the gp->status no longer being
5791 func preemptone(pp *p) bool {
5793 if mp == nil || mp == getg().m {
5797 if gp == nil || gp == mp.g0 {
5803 // Every call in a goroutine checks for stack overflow by
5804 // comparing the current stack pointer to gp->stackguard0.
5805 // Setting gp->stackguard0 to StackPreempt folds
5806 // preemption into the normal stack overflow check.
5807 gp.stackguard0 = stackPreempt
5809 // Request an async preemption of this P.
5810 if preemptMSupported && debug.asyncpreemptoff == 0 {
5820 func schedtrace(detailed bool) {
5827 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)
5829 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5831 // We must be careful while reading data from P's, M's and G's.
5832 // Even if we hold schedlock, most data can be changed concurrently.
5833 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5834 for i, pp := range allp {
5836 h := atomic.Load(&pp.runqhead)
5837 t := atomic.Load(&pp.runqtail)
5839 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5845 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5847 // In non-detailed mode format lengths of per-P run queues as:
5848 // [len1 len2 len3 len4]
5854 if i == len(allp)-1 {
5865 for mp := allm; mp != nil; mp = mp.alllink {
5867 print(" M", mp.id, ": p=")
5879 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5880 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5888 forEachG(func(gp *g) {
5889 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5896 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5906 // schedEnableUser enables or disables the scheduling of user
5909 // This does not stop already running user goroutines, so the caller
5910 // should first stop the world when disabling user goroutines.
5911 func schedEnableUser(enable bool) {
5913 if sched.disable.user == !enable {
5917 sched.disable.user = !enable
5919 n := sched.disable.n
5921 globrunqputbatch(&sched.disable.runnable, n)
5923 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5924 startm(nil, false, false)
5931 // schedEnabled reports whether gp should be scheduled. It returns
5932 // false is scheduling of gp is disabled.
5934 // sched.lock must be held.
5935 func schedEnabled(gp *g) bool {
5936 assertLockHeld(&sched.lock)
5938 if sched.disable.user {
5939 return isSystemGoroutine(gp, true)
5944 // Put mp on midle list.
5945 // sched.lock must be held.
5946 // May run during STW, so write barriers are not allowed.
5948 //go:nowritebarrierrec
5950 assertLockHeld(&sched.lock)
5952 mp.schedlink = sched.midle
5958 // Try to get an m from midle list.
5959 // sched.lock must be held.
5960 // May run during STW, so write barriers are not allowed.
5962 //go:nowritebarrierrec
5964 assertLockHeld(&sched.lock)
5966 mp := sched.midle.ptr()
5968 sched.midle = mp.schedlink
5974 // Put gp on the global runnable queue.
5975 // sched.lock must be held.
5976 // May run during STW, so write barriers are not allowed.
5978 //go:nowritebarrierrec
5979 func globrunqput(gp *g) {
5980 assertLockHeld(&sched.lock)
5982 sched.runq.pushBack(gp)
5986 // Put gp at the head of the global runnable queue.
5987 // sched.lock must be held.
5988 // May run during STW, so write barriers are not allowed.
5990 //go:nowritebarrierrec
5991 func globrunqputhead(gp *g) {
5992 assertLockHeld(&sched.lock)
5998 // Put a batch of runnable goroutines on the global runnable queue.
5999 // This clears *batch.
6000 // sched.lock must be held.
6001 // May run during STW, so write barriers are not allowed.
6003 //go:nowritebarrierrec
6004 func globrunqputbatch(batch *gQueue, n int32) {
6005 assertLockHeld(&sched.lock)
6007 sched.runq.pushBackAll(*batch)
6012 // Try get a batch of G's from the global runnable queue.
6013 // sched.lock must be held.
6014 func globrunqget(pp *p, max int32) *g {
6015 assertLockHeld(&sched.lock)
6017 if sched.runqsize == 0 {
6021 n := sched.runqsize/gomaxprocs + 1
6022 if n > sched.runqsize {
6025 if max > 0 && n > max {
6028 if n > int32(len(pp.runq))/2 {
6029 n = int32(len(pp.runq)) / 2
6034 gp := sched.runq.pop()
6037 gp1 := sched.runq.pop()
6038 runqput(pp, gp1, false)
6043 // pMask is an atomic bitstring with one bit per P.
6046 // read returns true if P id's bit is set.
6047 func (p pMask) read(id uint32) bool {
6049 mask := uint32(1) << (id % 32)
6050 return (atomic.Load(&p[word]) & mask) != 0
6053 // set sets P id's bit.
6054 func (p pMask) set(id int32) {
6056 mask := uint32(1) << (id % 32)
6057 atomic.Or(&p[word], mask)
6060 // clear clears P id's bit.
6061 func (p pMask) clear(id int32) {
6063 mask := uint32(1) << (id % 32)
6064 atomic.And(&p[word], ^mask)
6067 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
6069 // Ideally, the timer mask would be kept immediately consistent on any timer
6070 // operations. Unfortunately, updating a shared global data structure in the
6071 // timer hot path adds too much overhead in applications frequently switching
6072 // between no timers and some timers.
6074 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
6075 // running P (returned by pidleget) may add a timer at any time, so its mask
6076 // must be set. An idle P (passed to pidleput) cannot add new timers while
6077 // idle, so if it has no timers at that time, its mask may be cleared.
6079 // Thus, we get the following effects on timer-stealing in findrunnable:
6081 // - Idle Ps with no timers when they go idle are never checked in findrunnable
6082 // (for work- or timer-stealing; this is the ideal case).
6083 // - Running Ps must always be checked.
6084 // - Idle Ps whose timers are stolen must continue to be checked until they run
6085 // again, even after timer expiration.
6087 // When the P starts running again, the mask should be set, as a timer may be
6088 // added at any time.
6090 // TODO(prattmic): Additional targeted updates may improve the above cases.
6091 // e.g., updating the mask when stealing a timer.
6092 func updateTimerPMask(pp *p) {
6093 if pp.numTimers.Load() > 0 {
6097 // Looks like there are no timers, however another P may transiently
6098 // decrement numTimers when handling a timerModified timer in
6099 // checkTimers. We must take timersLock to serialize with these changes.
6100 lock(&pp.timersLock)
6101 if pp.numTimers.Load() == 0 {
6102 timerpMask.clear(pp.id)
6104 unlock(&pp.timersLock)
6107 // pidleput puts p on the _Pidle list. now must be a relatively recent call
6108 // to nanotime or zero. Returns now or the current time if now was zero.
6110 // This releases ownership of p. Once sched.lock is released it is no longer
6113 // sched.lock must be held.
6115 // May run during STW, so write barriers are not allowed.
6117 //go:nowritebarrierrec
6118 func pidleput(pp *p, now int64) int64 {
6119 assertLockHeld(&sched.lock)
6122 throw("pidleput: P has non-empty run queue")
6127 updateTimerPMask(pp) // clear if there are no timers.
6128 idlepMask.set(pp.id)
6129 pp.link = sched.pidle
6132 if !pp.limiterEvent.start(limiterEventIdle, now) {
6133 throw("must be able to track idle limiter event")
6138 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
6140 // sched.lock must be held.
6142 // May run during STW, so write barriers are not allowed.
6144 //go:nowritebarrierrec
6145 func pidleget(now int64) (*p, int64) {
6146 assertLockHeld(&sched.lock)
6148 pp := sched.pidle.ptr()
6150 // Timer may get added at any time now.
6154 timerpMask.set(pp.id)
6155 idlepMask.clear(pp.id)
6156 sched.pidle = pp.link
6157 sched.npidle.Add(-1)
6158 pp.limiterEvent.stop(limiterEventIdle, now)
6163 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
6164 // This is called by spinning Ms (or callers than need a spinning M) that have
6165 // found work. If no P is available, this must synchronized with non-spinning
6166 // Ms that may be preparing to drop their P without discovering this work.
6168 // sched.lock must be held.
6170 // May run during STW, so write barriers are not allowed.
6172 //go:nowritebarrierrec
6173 func pidlegetSpinning(now int64) (*p, int64) {
6174 assertLockHeld(&sched.lock)
6176 pp, now := pidleget(now)
6178 // See "Delicate dance" comment in findrunnable. We found work
6179 // that we cannot take, we must synchronize with non-spinning
6180 // Ms that may be preparing to drop their P.
6181 sched.needspinning.Store(1)
6188 // runqempty reports whether pp has no Gs on its local run queue.
6189 // It never returns true spuriously.
6190 func runqempty(pp *p) bool {
6191 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
6192 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
6193 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
6194 // does not mean the queue is empty.
6196 head := atomic.Load(&pp.runqhead)
6197 tail := atomic.Load(&pp.runqtail)
6198 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
6199 if tail == atomic.Load(&pp.runqtail) {
6200 return head == tail && runnext == 0
6205 // To shake out latent assumptions about scheduling order,
6206 // we introduce some randomness into scheduling decisions
6207 // when running with the race detector.
6208 // The need for this was made obvious by changing the
6209 // (deterministic) scheduling order in Go 1.5 and breaking
6210 // many poorly-written tests.
6211 // With the randomness here, as long as the tests pass
6212 // consistently with -race, they shouldn't have latent scheduling
6214 const randomizeScheduler = raceenabled
6216 // runqput tries to put g on the local runnable queue.
6217 // If next is false, runqput adds g to the tail of the runnable queue.
6218 // If next is true, runqput puts g in the pp.runnext slot.
6219 // If the run queue is full, runnext puts g on the global queue.
6220 // Executed only by the owner P.
6221 func runqput(pp *p, gp *g, next bool) {
6222 if randomizeScheduler && next && fastrandn(2) == 0 {
6228 oldnext := pp.runnext
6229 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
6235 // Kick the old runnext out to the regular run queue.
6240 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6242 if t-h < uint32(len(pp.runq)) {
6243 pp.runq[t%uint32(len(pp.runq))].set(gp)
6244 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
6247 if runqputslow(pp, gp, h, t) {
6250 // the queue is not full, now the put above must succeed
6254 // Put g and a batch of work from local runnable queue on global queue.
6255 // Executed only by the owner P.
6256 func runqputslow(pp *p, gp *g, h, t uint32) bool {
6257 var batch [len(pp.runq)/2 + 1]*g
6259 // First, grab a batch from local queue.
6262 if n != uint32(len(pp.runq)/2) {
6263 throw("runqputslow: queue is not full")
6265 for i := uint32(0); i < n; i++ {
6266 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6268 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6273 if randomizeScheduler {
6274 for i := uint32(1); i <= n; i++ {
6275 j := fastrandn(i + 1)
6276 batch[i], batch[j] = batch[j], batch[i]
6280 // Link the goroutines.
6281 for i := uint32(0); i < n; i++ {
6282 batch[i].schedlink.set(batch[i+1])
6285 q.head.set(batch[0])
6286 q.tail.set(batch[n])
6288 // Now put the batch on global queue.
6290 globrunqputbatch(&q, int32(n+1))
6295 // runqputbatch tries to put all the G's on q on the local runnable queue.
6296 // If the queue is full, they are put on the global queue; in that case
6297 // this will temporarily acquire the scheduler lock.
6298 // Executed only by the owner P.
6299 func runqputbatch(pp *p, q *gQueue, qsize int) {
6300 h := atomic.LoadAcq(&pp.runqhead)
6303 for !q.empty() && t-h < uint32(len(pp.runq)) {
6305 pp.runq[t%uint32(len(pp.runq))].set(gp)
6311 if randomizeScheduler {
6312 off := func(o uint32) uint32 {
6313 return (pp.runqtail + o) % uint32(len(pp.runq))
6315 for i := uint32(1); i < n; i++ {
6316 j := fastrandn(i + 1)
6317 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6321 atomic.StoreRel(&pp.runqtail, t)
6324 globrunqputbatch(q, int32(qsize))
6329 // Get g from local runnable queue.
6330 // If inheritTime is true, gp should inherit the remaining time in the
6331 // current time slice. Otherwise, it should start a new time slice.
6332 // Executed only by the owner P.
6333 func runqget(pp *p) (gp *g, inheritTime bool) {
6334 // If there's a runnext, it's the next G to run.
6336 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6337 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6338 // Hence, there's no need to retry this CAS if it fails.
6339 if next != 0 && pp.runnext.cas(next, 0) {
6340 return next.ptr(), true
6344 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6349 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6350 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6356 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6357 // Executed only by the owner P.
6358 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6359 oldNext := pp.runnext
6360 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6361 drainQ.pushBack(oldNext.ptr())
6366 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6372 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6376 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6380 // We've inverted the order in which it gets G's from the local P's runnable queue
6381 // and then advances the head pointer because we don't want to mess up the statuses of G's
6382 // while runqdrain() and runqsteal() are running in parallel.
6383 // Thus we should advance the head pointer before draining the local P into a gQueue,
6384 // so that we can update any gp.schedlink only after we take the full ownership of G,
6385 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6386 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6387 for i := uint32(0); i < qn; i++ {
6388 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6395 // Grabs a batch of goroutines from pp's runnable queue into batch.
6396 // Batch is a ring buffer starting at batchHead.
6397 // Returns number of grabbed goroutines.
6398 // Can be executed by any P.
6399 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6401 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6402 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6407 // Try to steal from pp.runnext.
6408 if next := pp.runnext; next != 0 {
6409 if pp.status == _Prunning {
6410 // Sleep to ensure that pp isn't about to run the g
6411 // we are about to steal.
6412 // The important use case here is when the g running
6413 // on pp ready()s another g and then almost
6414 // immediately blocks. Instead of stealing runnext
6415 // in this window, back off to give pp a chance to
6416 // schedule runnext. This will avoid thrashing gs
6417 // between different Ps.
6418 // A sync chan send/recv takes ~50ns as of time of
6419 // writing, so 3us gives ~50x overshoot.
6420 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6423 // On some platforms system timer granularity is
6424 // 1-15ms, which is way too much for this
6425 // optimization. So just yield.
6429 if !pp.runnext.cas(next, 0) {
6432 batch[batchHead%uint32(len(batch))] = next
6438 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6441 for i := uint32(0); i < n; i++ {
6442 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6443 batch[(batchHead+i)%uint32(len(batch))] = g
6445 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6451 // Steal half of elements from local runnable queue of p2
6452 // and put onto local runnable queue of p.
6453 // Returns one of the stolen elements (or nil if failed).
6454 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6456 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6461 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6465 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6466 if t-h+n >= uint32(len(pp.runq)) {
6467 throw("runqsteal: runq overflow")
6469 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6473 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6474 // be on one gQueue or gList at a time.
6475 type gQueue struct {
6480 // empty reports whether q is empty.
6481 func (q *gQueue) empty() bool {
6485 // push adds gp to the head of q.
6486 func (q *gQueue) push(gp *g) {
6487 gp.schedlink = q.head
6494 // pushBack adds gp to the tail of q.
6495 func (q *gQueue) pushBack(gp *g) {
6498 q.tail.ptr().schedlink.set(gp)
6505 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6507 func (q *gQueue) pushBackAll(q2 gQueue) {
6511 q2.tail.ptr().schedlink = 0
6513 q.tail.ptr().schedlink = q2.head
6520 // pop removes and returns the head of queue q. It returns nil if
6522 func (q *gQueue) pop() *g {
6525 q.head = gp.schedlink
6533 // popList takes all Gs in q and returns them as a gList.
6534 func (q *gQueue) popList() gList {
6535 stack := gList{q.head}
6540 // A gList is a list of Gs linked through g.schedlink. A G can only be
6541 // on one gQueue or gList at a time.
6546 // empty reports whether l is empty.
6547 func (l *gList) empty() bool {
6551 // push adds gp to the head of l.
6552 func (l *gList) push(gp *g) {
6553 gp.schedlink = l.head
6557 // pushAll prepends all Gs in q to l.
6558 func (l *gList) pushAll(q gQueue) {
6560 q.tail.ptr().schedlink = l.head
6565 // pop removes and returns the head of l. If l is empty, it returns nil.
6566 func (l *gList) pop() *g {
6569 l.head = gp.schedlink
6574 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6575 func setMaxThreads(in int) (out int) {
6577 out = int(sched.maxmcount)
6578 if in > 0x7fffffff { // MaxInt32
6579 sched.maxmcount = 0x7fffffff
6581 sched.maxmcount = int32(in)
6589 func procPin() int {
6594 return int(mp.p.ptr().id)
6603 //go:linkname sync_runtime_procPin sync.runtime_procPin
6605 func sync_runtime_procPin() int {
6609 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6611 func sync_runtime_procUnpin() {
6615 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6617 func sync_atomic_runtime_procPin() int {
6621 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6623 func sync_atomic_runtime_procUnpin() {
6627 // Active spinning for sync.Mutex.
6629 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6631 func sync_runtime_canSpin(i int) bool {
6632 // sync.Mutex is cooperative, so we are conservative with spinning.
6633 // Spin only few times and only if running on a multicore machine and
6634 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6635 // As opposed to runtime mutex we don't do passive spinning here,
6636 // because there can be work on global runq or on other Ps.
6637 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6640 if p := getg().m.p.ptr(); !runqempty(p) {
6646 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6648 func sync_runtime_doSpin() {
6649 procyield(active_spin_cnt)
6652 var stealOrder randomOrder
6654 // randomOrder/randomEnum are helper types for randomized work stealing.
6655 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6656 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6657 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6658 type randomOrder struct {
6663 type randomEnum struct {
6670 func (ord *randomOrder) reset(count uint32) {
6672 ord.coprimes = ord.coprimes[:0]
6673 for i := uint32(1); i <= count; i++ {
6674 if gcd(i, count) == 1 {
6675 ord.coprimes = append(ord.coprimes, i)
6680 func (ord *randomOrder) start(i uint32) randomEnum {
6684 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6688 func (enum *randomEnum) done() bool {
6689 return enum.i == enum.count
6692 func (enum *randomEnum) next() {
6694 enum.pos = (enum.pos + enum.inc) % enum.count
6697 func (enum *randomEnum) position() uint32 {
6701 func gcd(a, b uint32) uint32 {
6708 // An initTask represents the set of initializations that need to be done for a package.
6709 // Keep in sync with ../../test/noinit.go:initTask
6710 type initTask struct {
6711 state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
6713 // followed by nfns pcs, uintptr sized, one per init function to run
6716 // inittrace stores statistics for init functions which are
6717 // updated by malloc and newproc when active is true.
6718 var inittrace tracestat
6720 type tracestat struct {
6721 active bool // init tracing activation status
6722 id uint64 // init goroutine id
6723 allocs uint64 // heap allocations
6724 bytes uint64 // heap allocated bytes
6727 func doInit(ts []*initTask) {
6728 for _, t := range ts {
6733 func doInit1(t *initTask) {
6735 case 2: // fully initialized
6737 case 1: // initialization in progress
6738 throw("recursive call during initialization - linker skew")
6739 default: // not initialized yet
6740 t.state = 1 // initialization in progress
6747 if inittrace.active {
6749 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6754 // We should have pruned all of these in the linker.
6755 throw("inittask with no functions")
6758 firstFunc := add(unsafe.Pointer(t), 8)
6759 for i := uint32(0); i < t.nfns; i++ {
6760 p := add(firstFunc, uintptr(i)*goarch.PtrSize)
6761 f := *(*func())(unsafe.Pointer(&p))
6765 if inittrace.active {
6767 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6770 f := *(*func())(unsafe.Pointer(&firstFunc))
6771 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6774 print("init ", pkg, " @")
6775 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6776 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6777 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6778 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6782 t.state = 2 // initialization done