1 // Copyright 2014 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
11 "runtime/internal/atomic"
12 "runtime/internal/sys"
16 // set using cmd/go/internal/modload.ModInfoProg
19 // Goroutine scheduler
20 // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
22 // The main concepts are:
24 // M - worker thread, or machine.
25 // P - processor, a resource that is required to execute Go code.
26 // M must have an associated P to execute Go code, however it can be
27 // blocked or in a syscall w/o an associated P.
29 // Design doc at https://golang.org/s/go11sched.
31 // Worker thread parking/unparking.
32 // We need to balance between keeping enough running worker threads to utilize
33 // available hardware parallelism and parking excessive running worker threads
34 // to conserve CPU resources and power. This is not simple for two reasons:
35 // (1) scheduler state is intentionally distributed (in particular, per-P work
36 // queues), so it is not possible to compute global predicates on fast paths;
37 // (2) for optimal thread management we would need to know the future (don't park
38 // a worker thread when a new goroutine will be readied in near future).
40 // Three rejected approaches that would work badly:
41 // 1. Centralize all scheduler state (would inhibit scalability).
42 // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
43 // is a spare P, unpark a thread and handoff it the thread and the goroutine.
44 // This would lead to thread state thrashing, as the thread that readied the
45 // goroutine can be out of work the very next moment, we will need to park it.
46 // Also, it would destroy locality of computation as we want to preserve
47 // dependent goroutines on the same thread; and introduce additional latency.
48 // 3. Unpark an additional thread whenever we ready a goroutine and there is an
49 // idle P, but don't do handoff. This would lead to excessive thread parking/
50 // unparking as the additional threads will instantly park without discovering
53 // The current approach:
55 // This approach applies to three primary sources of potential work: readying a
56 // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
57 // additional details.
59 // We unpark an additional thread when we submit work if (this is wakep()):
60 // 1. There is an idle P, and
61 // 2. There are no "spinning" worker threads.
63 // A worker thread is considered spinning if it is out of local work and did
64 // not find work in the global run queue or netpoller; the spinning state is
65 // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
66 // also considered spinning; we don't do goroutine handoff so such threads are
67 // out of work initially. Spinning threads spin on looking for work in per-P
68 // run queues and timer heaps or from the GC before parking. If a spinning
69 // thread finds work it takes itself out of the spinning state and proceeds to
70 // execution. If it does not find work it takes itself out of the spinning
71 // state and then parks.
73 // If there is at least one spinning thread (sched.nmspinning>1), we don't
74 // unpark new threads when submitting work. To compensate for that, if the last
75 // spinning thread finds work and stops spinning, it must unpark a new spinning
76 // thread. This approach smooths out unjustified spikes of thread unparking,
77 // but at the same time guarantees eventual maximal CPU parallelism
80 // The main implementation complication is that we need to be very careful
81 // during spinning->non-spinning thread transition. This transition can race
82 // with submission of new work, and either one part or another needs to unpark
83 // another worker thread. If they both fail to do that, we can end up with
84 // semi-persistent CPU underutilization.
86 // The general pattern for submission is:
87 // 1. Submit work to the local run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
121 // This slice records the initializing tasks that need to be
122 // done to start up the runtime. It is built by the linker.
123 var runtime_inittasks []*initTask
125 // main_init_done is a signal used by cgocallbackg that initialization
126 // has been completed. It is made before _cgo_notify_runtime_init_done,
127 // so all cgo calls can rely on it existing. When main_init is complete,
128 // it is closed, meaning cgocallbackg can reliably receive from it.
129 var main_init_done chan bool
131 //go:linkname main_main main.main
134 // mainStarted indicates that the main M has started.
137 // runtimeInitTime is the nanotime() at which the runtime started.
138 var runtimeInitTime int64
140 // Value to use for signal mask for newly created M's.
141 var initSigmask sigset
143 // The main goroutine.
147 // Racectx of m0->g0 is used only as the parent of the main goroutine.
148 // It must not be used for anything else.
151 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
152 // Using decimal instead of binary GB and MB because
153 // they look nicer in the stack overflow failure message.
154 if goarch.PtrSize == 8 {
155 maxstacksize = 1000000000
157 maxstacksize = 250000000
160 // An upper limit for max stack size. Used to avoid random crashes
161 // after calling SetMaxStack and trying to allocate a stack that is too big,
162 // since stackalloc works with 32-bit sizes.
163 maxstackceiling = 2 * maxstacksize
165 // Allow newproc to start new Ms.
168 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
170 newm(sysmon, nil, -1)
174 // Lock the main goroutine onto this, the main OS thread,
175 // during initialization. Most programs won't care, but a few
176 // do require certain calls to be made by the main thread.
177 // Those can arrange for main.main to run in the main thread
178 // by calling runtime.LockOSThread during initialization
179 // to preserve the lock.
183 throw("runtime.main not on m0")
186 // Record when the world started.
187 // Must be before doInit for tracing init.
188 runtimeInitTime = nanotime()
189 if runtimeInitTime == 0 {
190 throw("nanotime returning zero")
193 if debug.inittrace != 0 {
194 inittrace.id = getg().goid
195 inittrace.active = true
198 doInit(runtime_inittasks) // Must be before defer.
200 // Defer unlock so that runtime.Goexit during init does the unlock too.
210 main_init_done = make(chan bool)
212 if _cgo_thread_start == nil {
213 throw("_cgo_thread_start missing")
215 if GOOS != "windows" {
216 if _cgo_setenv == nil {
217 throw("_cgo_setenv missing")
219 if _cgo_unsetenv == nil {
220 throw("_cgo_unsetenv missing")
223 if _cgo_notify_runtime_init_done == nil {
224 throw("_cgo_notify_runtime_init_done missing")
226 // Start the template thread in case we enter Go from
227 // a C-created thread and need to create a new thread.
228 startTemplateThread()
229 cgocall(_cgo_notify_runtime_init_done, nil)
232 // Run the initializing tasks. Depending on build mode this
233 // list can arrive a few different ways, but it will always
234 // contain the init tasks computed by the linker for all the
235 // packages in the program (excluding those added at runtime
236 // by package plugin).
237 for _, m := range activeModules() {
241 // Disable init tracing after main init done to avoid overhead
242 // of collecting statistics in malloc and newproc
243 inittrace.active = false
245 close(main_init_done)
250 if isarchive || islibrary {
251 // A program compiled with -buildmode=c-archive or c-shared
252 // has a main, but it is not executed.
255 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
258 runExitHooks(0) // run hooks now, since racefini does not return
262 // Make racy client program work: if panicking on
263 // another goroutine at the same time as main returns,
264 // let the other goroutine finish printing the panic trace.
265 // Once it does, it will exit. See issues 3934 and 20018.
266 if runningPanicDefers.Load() != 0 {
267 // Running deferred functions should not take long.
268 for c := 0; c < 1000; c++ {
269 if runningPanicDefers.Load() == 0 {
275 if panicking.Load() != 0 {
276 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
287 // os_beforeExit is called from os.Exit(0).
289 //go:linkname os_beforeExit os.runtime_beforeExit
290 func os_beforeExit(exitCode int) {
291 runExitHooks(exitCode)
292 if exitCode == 0 && raceenabled {
297 // start forcegc helper goroutine
302 func forcegchelper() {
304 lockInit(&forcegc.lock, lockRankForcegc)
307 if forcegc.idle.Load() {
308 throw("forcegc: phase error")
310 forcegc.idle.Store(true)
311 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
312 // this goroutine is explicitly resumed by sysmon
313 if debug.gctrace > 0 {
316 // Time-triggered, fully concurrent.
317 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
321 // Gosched yields the processor, allowing other goroutines to run. It does not
322 // suspend the current goroutine, so execution resumes automatically.
330 // goschedguarded yields the processor like gosched, but also checks
331 // for forbidden states and opts out of the yield in those cases.
334 func goschedguarded() {
335 mcall(goschedguarded_m)
338 // goschedIfBusy yields the processor like gosched, but only does so if
339 // there are no idle Ps or if we're on the only P and there's nothing in
340 // the run queue. In both cases, there is freely available idle time.
343 func goschedIfBusy() {
345 // Call gosched if gp.preempt is set; we may be in a tight loop that
346 // doesn't otherwise yield.
347 if !gp.preempt && sched.npidle.Load() > 0 {
353 // Puts the current goroutine into a waiting state and calls unlockf on the
356 // If unlockf returns false, the goroutine is resumed.
358 // unlockf must not access this G's stack, as it may be moved between
359 // the call to gopark and the call to unlockf.
361 // Note that because unlockf is called after putting the G into a waiting
362 // state, the G may have already been readied by the time unlockf is called
363 // unless there is external synchronization preventing the G from being
364 // readied. If unlockf returns false, it must guarantee that the G cannot be
365 // externally readied.
367 // Reason explains why the goroutine has been parked. It is displayed in stack
368 // traces and heap dumps. Reasons should be unique and descriptive. Do not
369 // re-use reasons, add new ones.
370 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
371 if reason != waitReasonSleep {
372 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
376 status := readgstatus(gp)
377 if status != _Grunning && status != _Gscanrunning {
378 throw("gopark: bad g status")
381 mp.waitunlockf = unlockf
382 gp.waitreason = reason
383 mp.waittraceev = traceEv
384 mp.waittraceskip = traceskip
386 // can't do anything that might move the G between Ms here.
390 // Puts the current goroutine into a waiting state and unlocks the lock.
391 // The goroutine can be made runnable again by calling goready(gp).
392 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
393 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
396 func goready(gp *g, traceskip int) {
398 ready(gp, traceskip, true)
403 func acquireSudog() *sudog {
404 // Delicate dance: the semaphore implementation calls
405 // acquireSudog, acquireSudog calls new(sudog),
406 // new calls malloc, malloc can call the garbage collector,
407 // and the garbage collector calls the semaphore implementation
409 // Break the cycle by doing acquirem/releasem around new(sudog).
410 // The acquirem/releasem increments m.locks during new(sudog),
411 // which keeps the garbage collector from being invoked.
414 if len(pp.sudogcache) == 0 {
415 lock(&sched.sudoglock)
416 // First, try to grab a batch from central cache.
417 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
418 s := sched.sudogcache
419 sched.sudogcache = s.next
421 pp.sudogcache = append(pp.sudogcache, s)
423 unlock(&sched.sudoglock)
424 // If the central cache is empty, allocate a new one.
425 if len(pp.sudogcache) == 0 {
426 pp.sudogcache = append(pp.sudogcache, new(sudog))
429 n := len(pp.sudogcache)
430 s := pp.sudogcache[n-1]
431 pp.sudogcache[n-1] = nil
432 pp.sudogcache = pp.sudogcache[:n-1]
434 throw("acquireSudog: found s.elem != nil in cache")
441 func releaseSudog(s *sudog) {
443 throw("runtime: sudog with non-nil elem")
446 throw("runtime: sudog with non-false isSelect")
449 throw("runtime: sudog with non-nil next")
452 throw("runtime: sudog with non-nil prev")
454 if s.waitlink != nil {
455 throw("runtime: sudog with non-nil waitlink")
458 throw("runtime: sudog with non-nil c")
462 throw("runtime: releaseSudog with non-nil gp.param")
464 mp := acquirem() // avoid rescheduling to another P
466 if len(pp.sudogcache) == cap(pp.sudogcache) {
467 // Transfer half of local cache to the central cache.
468 var first, last *sudog
469 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
470 n := len(pp.sudogcache)
471 p := pp.sudogcache[n-1]
472 pp.sudogcache[n-1] = nil
473 pp.sudogcache = pp.sudogcache[:n-1]
481 lock(&sched.sudoglock)
482 last.next = sched.sudogcache
483 sched.sudogcache = first
484 unlock(&sched.sudoglock)
486 pp.sudogcache = append(pp.sudogcache, s)
490 // called from assembly.
491 func badmcall(fn func(*g)) {
492 throw("runtime: mcall called on m->g0 stack")
495 func badmcall2(fn func(*g)) {
496 throw("runtime: mcall function returned")
499 func badreflectcall() {
500 panic(plainError("arg size to reflect.call more than 1GB"))
504 //go:nowritebarrierrec
505 func badmorestackg0() {
506 writeErrStr("fatal: morestack on g0\n")
510 //go:nowritebarrierrec
511 func badmorestackgsignal() {
512 writeErrStr("fatal: morestack on gsignal\n")
520 func lockedOSThread() bool {
522 return gp.lockedm != 0 && gp.m.lockedg != 0
526 // allgs contains all Gs ever created (including dead Gs), and thus
529 // Access via the slice is protected by allglock or stop-the-world.
530 // Readers that cannot take the lock may (carefully!) use the atomic
535 // allglen and allgptr are atomic variables that contain len(allgs) and
536 // &allgs[0] respectively. Proper ordering depends on totally-ordered
537 // loads and stores. Writes are protected by allglock.
539 // allgptr is updated before allglen. Readers should read allglen
540 // before allgptr to ensure that allglen is always <= len(allgptr). New
541 // Gs appended during the race can be missed. For a consistent view of
542 // all Gs, allglock must be held.
544 // allgptr copies should always be stored as a concrete type or
545 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
546 // even if it points to a stale array.
551 func allgadd(gp *g) {
552 if readgstatus(gp) == _Gidle {
553 throw("allgadd: bad status Gidle")
557 allgs = append(allgs, gp)
558 if &allgs[0] != allgptr {
559 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
561 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
565 // allGsSnapshot returns a snapshot of the slice of all Gs.
567 // The world must be stopped or allglock must be held.
568 func allGsSnapshot() []*g {
569 assertWorldStoppedOrLockHeld(&allglock)
571 // Because the world is stopped or allglock is held, allgadd
572 // cannot happen concurrently with this. allgs grows
573 // monotonically and existing entries never change, so we can
574 // simply return a copy of the slice header. For added safety,
575 // we trim everything past len because that can still change.
576 return allgs[:len(allgs):len(allgs)]
579 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
580 func atomicAllG() (**g, uintptr) {
581 length := atomic.Loaduintptr(&allglen)
582 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
586 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
587 func atomicAllGIndex(ptr **g, i uintptr) *g {
588 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
591 // forEachG calls fn on every G from allgs.
593 // forEachG takes a lock to exclude concurrent addition of new Gs.
594 func forEachG(fn func(gp *g)) {
596 for _, gp := range allgs {
602 // forEachGRace calls fn on every G from allgs.
604 // forEachGRace avoids locking, but does not exclude addition of new Gs during
605 // execution, which may be missed.
606 func forEachGRace(fn func(gp *g)) {
607 ptr, length := atomicAllG()
608 for i := uintptr(0); i < length; i++ {
609 gp := atomicAllGIndex(ptr, i)
616 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
617 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
621 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
622 // value of the GODEBUG environment variable.
623 func cpuinit(env string) {
625 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
626 cpu.DebugOptions = true
630 // Support cpu feature variables are used in code generated by the compiler
631 // to guard execution of instructions that can not be assumed to be always supported.
634 x86HasPOPCNT = cpu.X86.HasPOPCNT
635 x86HasSSE41 = cpu.X86.HasSSE41
636 x86HasFMA = cpu.X86.HasFMA
639 armHasVFPv4 = cpu.ARM.HasVFPv4
642 arm64HasATOMICS = cpu.ARM64.HasATOMICS
646 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
647 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
648 // early before much of the runtime is initialized.
649 func getGodebugEarly() string {
650 const prefix = "GODEBUG="
653 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
654 // Similar to goenv_unix but extracts the environment value for
656 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
658 for argv_index(argv, argc+1+n) != nil {
662 for i := int32(0); i < n; i++ {
663 p := argv_index(argv, argc+1+i)
664 s := unsafe.String(p, findnull(p))
666 if hasPrefix(s, prefix) {
667 env = gostring(p)[len(prefix):]
675 // The bootstrap sequence is:
679 // make & queue new G
680 // call runtime·mstart
682 // The new G calls runtime·main.
684 lockInit(&sched.lock, lockRankSched)
685 lockInit(&sched.sysmonlock, lockRankSysmon)
686 lockInit(&sched.deferlock, lockRankDefer)
687 lockInit(&sched.sudoglock, lockRankSudog)
688 lockInit(&deadlock, lockRankDeadlock)
689 lockInit(&paniclk, lockRankPanic)
690 lockInit(&allglock, lockRankAllg)
691 lockInit(&allpLock, lockRankAllp)
692 lockInit(&reflectOffs.lock, lockRankReflectOffs)
693 lockInit(&finlock, lockRankFin)
694 lockInit(&cpuprof.lock, lockRankCpuprof)
696 // Enforce that this lock is always a leaf lock.
697 // All of this lock's critical sections should be
699 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
701 // raceinit must be the first call to race detector.
702 // In particular, it must be done before mallocinit below calls racemapshadow.
705 gp.racectx, raceprocctx0 = raceinit()
708 sched.maxmcount = 10000
710 // The world starts stopped.
716 godebug := getGodebugEarly()
717 initPageTrace(godebug) // must run after mallocinit but before anything allocates
718 cpuinit(godebug) // must run before alginit
719 alginit() // maps, hash, fastrand must not be used before this call
720 fastrandinit() // must run before mcommoninit
721 mcommoninit(gp.m, -1)
722 modulesinit() // provides activeModules
723 typelinksinit() // uses maps, activeModules
724 itabsinit() // uses activeModules
725 stkobjinit() // must run before GC starts
727 sigsave(&gp.m.sigmask)
728 initSigmask = gp.m.sigmask
735 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
736 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
737 // set to true by the linker, it means that nothing is consuming the profile, it is
738 // safe to set MemProfileRate to 0.
739 if disableMemoryProfiling {
744 sched.lastpoll.Store(nanotime())
746 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
749 if procresize(procs) != nil {
750 throw("unknown runnable goroutine during bootstrap")
754 // World is effectively started now, as P's can run.
757 if buildVersion == "" {
758 // Condition should never trigger. This code just serves
759 // to ensure runtime·buildVersion is kept in the resulting binary.
760 buildVersion = "unknown"
762 if len(modinfo) == 1 {
763 // Condition should never trigger. This code just serves
764 // to ensure runtime·modinfo is kept in the resulting binary.
769 func dumpgstatus(gp *g) {
771 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
772 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
775 // sched.lock must be held.
777 assertLockHeld(&sched.lock)
779 // Exclude extra M's, which are used for cgocallback from threads
782 // The purpose of the SetMaxThreads limit is to avoid accidental fork
783 // bomb from something like millions of goroutines blocking on system
784 // calls, causing the runtime to create millions of threads. By
785 // definition, this isn't a problem for threads created in C, so we
786 // exclude them from the limit. See https://go.dev/issue/60004.
787 count := mcount() - int32(extraMInUse.Load()) - int32(extraMLength.Load())
788 if count > sched.maxmcount {
789 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
790 throw("thread exhaustion")
794 // mReserveID returns the next ID to use for a new m. This new m is immediately
795 // considered 'running' by checkdead.
797 // sched.lock must be held.
798 func mReserveID() int64 {
799 assertLockHeld(&sched.lock)
801 if sched.mnext+1 < sched.mnext {
802 throw("runtime: thread ID overflow")
810 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
811 func mcommoninit(mp *m, id int64) {
814 // g0 stack won't make sense for user (and is not necessary unwindable).
816 callers(1, mp.createstack[:])
827 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
828 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
832 // Same behavior as for 1.17.
833 // TODO: Simplify this.
834 if goarch.BigEndian {
835 mp.fastrand = uint64(lo)<<32 | uint64(hi)
837 mp.fastrand = uint64(hi)<<32 | uint64(lo)
841 if mp.gsignal != nil {
842 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
845 // Add to allm so garbage collector doesn't free g->m
846 // when it is just in a register or thread-local storage.
849 // NumCgoCall() iterates over allm w/o schedlock,
850 // so we need to publish it safely.
851 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
854 // Allocate memory to hold a cgo traceback if the cgo call crashes.
855 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
856 mp.cgoCallers = new(cgoCallers)
860 func (mp *m) becomeSpinning() {
862 sched.nmspinning.Add(1)
863 sched.needspinning.Store(0)
866 func (mp *m) hasCgoOnStack() bool {
867 return mp.ncgo > 0 || mp.isextra
870 var fastrandseed uintptr
872 func fastrandinit() {
873 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
877 // Mark gp ready to run.
878 func ready(gp *g, traceskip int, next bool) {
880 traceGoUnpark(gp, traceskip)
883 status := readgstatus(gp)
886 mp := acquirem() // disable preemption because it can be holding p in a local var
887 if status&^_Gscan != _Gwaiting {
889 throw("bad g->status in ready")
892 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
893 casgstatus(gp, _Gwaiting, _Grunnable)
894 runqput(mp.p.ptr(), gp, next)
899 // freezeStopWait is a large value that freezetheworld sets
900 // sched.stopwait to in order to request that all Gs permanently stop.
901 const freezeStopWait = 0x7fffffff
903 // freezing is set to non-zero if the runtime is trying to freeze the
905 var freezing atomic.Bool
907 // Similar to stopTheWorld but best-effort and can be called several times.
908 // There is no reverse operation, used during crashing.
909 // This function must not lock any mutexes.
910 func freezetheworld() {
912 // stopwait and preemption requests can be lost
913 // due to races with concurrently executing threads,
914 // so try several times
915 for i := 0; i < 5; i++ {
916 // this should tell the scheduler to not start any new goroutines
917 sched.stopwait = freezeStopWait
918 sched.gcwaiting.Store(true)
919 // this should stop running goroutines
921 break // no running goroutines
931 // All reads and writes of g's status go through readgstatus, casgstatus
932 // castogscanstatus, casfrom_Gscanstatus.
935 func readgstatus(gp *g) uint32 {
936 return gp.atomicstatus.Load()
939 // The Gscanstatuses are acting like locks and this releases them.
940 // If it proves to be a performance hit we should be able to make these
941 // simple atomic stores but for now we are going to throw if
942 // we see an inconsistent state.
943 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
946 // Check that transition is valid.
949 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
951 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
957 if newval == oldval&^_Gscan {
958 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
962 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
964 throw("casfrom_Gscanstatus: gp->status is not in scan state")
966 releaseLockRank(lockRankGscan)
969 // This will return false if the gp is not in the expected status and the cas fails.
970 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
971 func castogscanstatus(gp *g, oldval, newval uint32) bool {
977 if newval == oldval|_Gscan {
978 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
980 acquireLockRank(lockRankGscan)
986 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
987 throw("castogscanstatus")
991 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
992 // various latencies on every transition instead of sampling them.
993 var casgstatusAlwaysTrack = false
995 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
996 // and casfrom_Gscanstatus instead.
997 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
998 // put it in the Gscan state is finished.
1001 func casgstatus(gp *g, oldval, newval uint32) {
1002 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
1003 systemstack(func() {
1004 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
1005 throw("casgstatus: bad incoming values")
1009 acquireLockRank(lockRankGscan)
1010 releaseLockRank(lockRankGscan)
1012 // See https://golang.org/cl/21503 for justification of the yield delay.
1013 const yieldDelay = 5 * 1000
1016 // loop if gp->atomicstatus is in a scan state giving
1017 // GC time to finish and change the state to oldval.
1018 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1019 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1020 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1023 nextYield = nanotime() + yieldDelay
1025 if nanotime() < nextYield {
1026 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1031 nextYield = nanotime() + yieldDelay/2
1035 if oldval == _Grunning {
1036 // Track every gTrackingPeriod time a goroutine transitions out of running.
1037 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1046 // Handle various kinds of tracking.
1049 // - Time spent in runnable.
1050 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1053 // We transitioned out of runnable, so measure how much
1054 // time we spent in this state and add it to
1057 gp.runnableTime += now - gp.trackingStamp
1058 gp.trackingStamp = 0
1060 if !gp.waitreason.isMutexWait() {
1061 // Not blocking on a lock.
1064 // Blocking on a lock, measure it. Note that because we're
1065 // sampling, we have to multiply by our sampling period to get
1066 // a more representative estimate of the absolute value.
1067 // gTrackingPeriod also represents an accurate sampling period
1068 // because we can only enter this state from _Grunning.
1070 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1071 gp.trackingStamp = 0
1075 if !gp.waitreason.isMutexWait() {
1076 // Not blocking on a lock.
1079 // Blocking on a lock. Write down the timestamp.
1081 gp.trackingStamp = now
1083 // We just transitioned into runnable, so record what
1084 // time that happened.
1086 gp.trackingStamp = now
1088 // We're transitioning into running, so turn off
1089 // tracking and record how much time we spent in
1092 sched.timeToRun.record(gp.runnableTime)
1097 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1099 // Use this over casgstatus when possible to ensure that a waitreason is set.
1100 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1101 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1102 gp.waitreason = reason
1103 casgstatus(gp, old, _Gwaiting)
1106 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1107 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1108 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1109 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1110 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1113 func casgcopystack(gp *g) uint32 {
1115 oldstatus := readgstatus(gp) &^ _Gscan
1116 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1117 throw("copystack: bad status, not Gwaiting or Grunnable")
1119 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1125 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1127 // TODO(austin): This is the only status operation that both changes
1128 // the status and locks the _Gscan bit. Rethink this.
1129 func casGToPreemptScan(gp *g, old, new uint32) {
1130 if old != _Grunning || new != _Gscan|_Gpreempted {
1131 throw("bad g transition")
1133 acquireLockRank(lockRankGscan)
1134 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1138 // casGFromPreempted attempts to transition gp from _Gpreempted to
1139 // _Gwaiting. If successful, the caller is responsible for
1140 // re-scheduling gp.
1141 func casGFromPreempted(gp *g, old, new uint32) bool {
1142 if old != _Gpreempted || new != _Gwaiting {
1143 throw("bad g transition")
1145 gp.waitreason = waitReasonPreempted
1146 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1149 // stopTheWorld stops all P's from executing goroutines, interrupting
1150 // all goroutines at GC safe points and records reason as the reason
1151 // for the stop. On return, only the current goroutine's P is running.
1152 // stopTheWorld must not be called from a system stack and the caller
1153 // must not hold worldsema. The caller must call startTheWorld when
1154 // other P's should resume execution.
1156 // stopTheWorld is safe for multiple goroutines to call at the
1157 // same time. Each will execute its own stop, and the stops will
1160 // This is also used by routines that do stack dumps. If the system is
1161 // in panic or being exited, this may not reliably stop all
1163 func stopTheWorld(reason string) {
1164 semacquire(&worldsema)
1166 gp.m.preemptoff = reason
1167 systemstack(func() {
1168 // Mark the goroutine which called stopTheWorld preemptible so its
1169 // stack may be scanned.
1170 // This lets a mark worker scan us while we try to stop the world
1171 // since otherwise we could get in a mutual preemption deadlock.
1172 // We must not modify anything on the G stack because a stack shrink
1173 // may occur. A stack shrink is otherwise OK though because in order
1174 // to return from this function (and to leave the system stack) we
1175 // must have preempted all goroutines, including any attempting
1176 // to scan our stack, in which case, any stack shrinking will
1177 // have already completed by the time we exit.
1178 // Don't provide a wait reason because we're still executing.
1179 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1180 stopTheWorldWithSema()
1181 casgstatus(gp, _Gwaiting, _Grunning)
1185 // startTheWorld undoes the effects of stopTheWorld.
1186 func startTheWorld() {
1187 systemstack(func() { startTheWorldWithSema(false) })
1189 // worldsema must be held over startTheWorldWithSema to ensure
1190 // gomaxprocs cannot change while worldsema is held.
1192 // Release worldsema with direct handoff to the next waiter, but
1193 // acquirem so that semrelease1 doesn't try to yield our time.
1195 // Otherwise if e.g. ReadMemStats is being called in a loop,
1196 // it might stomp on other attempts to stop the world, such as
1197 // for starting or ending GC. The operation this blocks is
1198 // so heavy-weight that we should just try to be as fair as
1201 // We don't want to just allow us to get preempted between now
1202 // and releasing the semaphore because then we keep everyone
1203 // (including, for example, GCs) waiting longer.
1206 semrelease1(&worldsema, true, 0)
1210 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1211 // until the GC is not running. It also blocks a GC from starting
1212 // until startTheWorldGC is called.
1213 func stopTheWorldGC(reason string) {
1215 stopTheWorld(reason)
1218 // startTheWorldGC undoes the effects of stopTheWorldGC.
1219 func startTheWorldGC() {
1224 // Holding worldsema grants an M the right to try to stop the world.
1225 var worldsema uint32 = 1
1227 // Holding gcsema grants the M the right to block a GC, and blocks
1228 // until the current GC is done. In particular, it prevents gomaxprocs
1229 // from changing concurrently.
1231 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1232 // being changed/enabled during a GC, remove this.
1233 var gcsema uint32 = 1
1235 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1236 // The caller is responsible for acquiring worldsema and disabling
1237 // preemption first and then should stopTheWorldWithSema on the system
1240 // semacquire(&worldsema, 0)
1241 // m.preemptoff = "reason"
1242 // systemstack(stopTheWorldWithSema)
1244 // When finished, the caller must either call startTheWorld or undo
1245 // these three operations separately:
1247 // m.preemptoff = ""
1248 // systemstack(startTheWorldWithSema)
1249 // semrelease(&worldsema)
1251 // It is allowed to acquire worldsema once and then execute multiple
1252 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1253 // Other P's are able to execute between successive calls to
1254 // startTheWorldWithSema and stopTheWorldWithSema.
1255 // Holding worldsema causes any other goroutines invoking
1256 // stopTheWorld to block.
1257 func stopTheWorldWithSema() {
1260 // If we hold a lock, then we won't be able to stop another M
1261 // that is blocked trying to acquire the lock.
1263 throw("stopTheWorld: holding locks")
1267 sched.stopwait = gomaxprocs
1268 sched.gcwaiting.Store(true)
1271 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1273 // try to retake all P's in Psyscall status
1274 for _, pp := range allp {
1276 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1288 pp, _ := pidleget(now)
1292 pp.status = _Pgcstop
1295 wait := sched.stopwait > 0
1298 // wait for remaining P's to stop voluntarily
1301 // wait for 100us, then try to re-preempt in case of any races
1302 if notetsleep(&sched.stopnote, 100*1000) {
1303 noteclear(&sched.stopnote)
1312 if sched.stopwait != 0 {
1313 bad = "stopTheWorld: not stopped (stopwait != 0)"
1315 for _, pp := range allp {
1316 if pp.status != _Pgcstop {
1317 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1321 if freezing.Load() {
1322 // Some other thread is panicking. This can cause the
1323 // sanity checks above to fail if the panic happens in
1324 // the signal handler on a stopped thread. Either way,
1325 // we should halt this thread.
1336 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1337 assertWorldStopped()
1339 mp := acquirem() // disable preemption because it can be holding p in a local var
1340 if netpollinited() {
1341 list := netpoll(0) // non-blocking
1351 p1 := procresize(procs)
1352 sched.gcwaiting.Store(false)
1353 if sched.sysmonwait.Load() {
1354 sched.sysmonwait.Store(false)
1355 notewakeup(&sched.sysmonnote)
1368 throw("startTheWorld: inconsistent mp->nextp")
1371 notewakeup(&mp.park)
1373 // Start M to run P. Do not start another M below.
1378 // Capture start-the-world time before doing clean-up tasks.
1379 startTime := nanotime()
1384 // Wakeup an additional proc in case we have excessive runnable goroutines
1385 // in local queues or in the global queue. If we don't, the proc will park itself.
1386 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1394 // usesLibcall indicates whether this runtime performs system calls
1396 func usesLibcall() bool {
1398 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1401 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1406 // mStackIsSystemAllocated indicates whether this runtime starts on a
1407 // system-allocated stack.
1408 func mStackIsSystemAllocated() bool {
1410 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1414 case "386", "amd64", "arm", "arm64":
1421 // mstart is the entry-point for new Ms.
1422 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1425 // mstart0 is the Go entry-point for new Ms.
1426 // This must not split the stack because we may not even have stack
1427 // bounds set up yet.
1429 // May run during STW (because it doesn't have a P yet), so write
1430 // barriers are not allowed.
1433 //go:nowritebarrierrec
1437 osStack := gp.stack.lo == 0
1439 // Initialize stack bounds from system stack.
1440 // Cgo may have left stack size in stack.hi.
1441 // minit may update the stack bounds.
1443 // Note: these bounds may not be very accurate.
1444 // We set hi to &size, but there are things above
1445 // it. The 1024 is supposed to compensate this,
1446 // but is somewhat arbitrary.
1449 size = 8192 * sys.StackGuardMultiplier
1451 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1452 gp.stack.lo = gp.stack.hi - size + 1024
1454 // Initialize stack guard so that we can start calling regular
1456 gp.stackguard0 = gp.stack.lo + stackGuard
1457 // This is the g0, so we can also call go:systemstack
1458 // functions, which check stackguard1.
1459 gp.stackguard1 = gp.stackguard0
1462 // Exit this thread.
1463 if mStackIsSystemAllocated() {
1464 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1465 // the stack, but put it in gp.stack before mstart,
1466 // so the logic above hasn't set osStack yet.
1472 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1473 // so that we can set up g0.sched to return to the call of mstart1 above.
1480 throw("bad runtime·mstart")
1483 // Set up m.g0.sched as a label returning to just
1484 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1485 // We're never coming back to mstart1 after we call schedule,
1486 // so other calls can reuse the current frame.
1487 // And goexit0 does a gogo that needs to return from mstart1
1488 // and let mstart0 exit the thread.
1489 gp.sched.g = guintptr(unsafe.Pointer(gp))
1490 gp.sched.pc = getcallerpc()
1491 gp.sched.sp = getcallersp()
1496 // Install signal handlers; after minit so that minit can
1497 // prepare the thread to be able to handle the signals.
1502 if fn := gp.m.mstartfn; fn != nil {
1507 acquirep(gp.m.nextp.ptr())
1513 // mstartm0 implements part of mstart1 that only runs on the m0.
1515 // Write barriers are allowed here because we know the GC can't be
1516 // running yet, so they'll be no-ops.
1518 //go:yeswritebarrierrec
1520 // Create an extra M for callbacks on threads not created by Go.
1521 // An extra M is also needed on Windows for callbacks created by
1522 // syscall.NewCallback. See issue #6751 for details.
1523 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1530 // mPark causes a thread to park itself, returning once woken.
1535 notesleep(&gp.m.park)
1536 noteclear(&gp.m.park)
1539 // mexit tears down and exits the current thread.
1541 // Don't call this directly to exit the thread, since it must run at
1542 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1543 // unwind the stack to the point that exits the thread.
1545 // It is entered with m.p != nil, so write barriers are allowed. It
1546 // will release the P before exiting.
1548 //go:yeswritebarrierrec
1549 func mexit(osStack bool) {
1553 // This is the main thread. Just wedge it.
1555 // On Linux, exiting the main thread puts the process
1556 // into a non-waitable zombie state. On Plan 9,
1557 // exiting the main thread unblocks wait even though
1558 // other threads are still running. On Solaris we can
1559 // neither exitThread nor return from mstart. Other
1560 // bad things probably happen on other platforms.
1562 // We could try to clean up this M more before wedging
1563 // it, but that complicates signal handling.
1564 handoffp(releasep())
1570 throw("locked m0 woke up")
1576 // Free the gsignal stack.
1577 if mp.gsignal != nil {
1578 stackfree(mp.gsignal.stack)
1579 // On some platforms, when calling into VDSO (e.g. nanotime)
1580 // we store our g on the gsignal stack, if there is one.
1581 // Now the stack is freed, unlink it from the m, so we
1582 // won't write to it when calling VDSO code.
1586 // Remove m from allm.
1588 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1594 throw("m not found in allm")
1596 // Delay reaping m until it's done with the stack.
1598 // Put mp on the free list, though it will not be reaped while freeWait
1599 // is freeMWait. mp is no longer reachable via allm, so even if it is
1600 // on an OS stack, we must keep a reference to mp alive so that the GC
1601 // doesn't free mp while we are still using it.
1603 // Note that the free list must not be linked through alllink because
1604 // some functions walk allm without locking, so may be using alllink.
1605 mp.freeWait.Store(freeMWait)
1606 mp.freelink = sched.freem
1610 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1613 handoffp(releasep())
1614 // After this point we must not have write barriers.
1616 // Invoke the deadlock detector. This must happen after
1617 // handoffp because it may have started a new M to take our
1624 if GOOS == "darwin" || GOOS == "ios" {
1625 // Make sure pendingPreemptSignals is correct when an M exits.
1627 if mp.signalPending.Load() != 0 {
1628 pendingPreemptSignals.Add(-1)
1632 // Destroy all allocated resources. After this is called, we may no
1633 // longer take any locks.
1637 // No more uses of mp, so it is safe to drop the reference.
1638 mp.freeWait.Store(freeMRef)
1640 // Return from mstart and let the system thread
1641 // library free the g0 stack and terminate the thread.
1645 // mstart is the thread's entry point, so there's nothing to
1646 // return to. Exit the thread directly. exitThread will clear
1647 // m.freeWait when it's done with the stack and the m can be
1649 exitThread(&mp.freeWait)
1652 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1653 // If a P is currently executing code, this will bring the P to a GC
1654 // safe point and execute fn on that P. If the P is not executing code
1655 // (it is idle or in a syscall), this will call fn(p) directly while
1656 // preventing the P from exiting its state. This does not ensure that
1657 // fn will run on every CPU executing Go code, but it acts as a global
1658 // memory barrier. GC uses this as a "ragged barrier."
1660 // The caller must hold worldsema.
1663 func forEachP(fn func(*p)) {
1665 pp := getg().m.p.ptr()
1668 if sched.safePointWait != 0 {
1669 throw("forEachP: sched.safePointWait != 0")
1671 sched.safePointWait = gomaxprocs - 1
1672 sched.safePointFn = fn
1674 // Ask all Ps to run the safe point function.
1675 for _, p2 := range allp {
1677 atomic.Store(&p2.runSafePointFn, 1)
1682 // Any P entering _Pidle or _Psyscall from now on will observe
1683 // p.runSafePointFn == 1 and will call runSafePointFn when
1684 // changing its status to _Pidle/_Psyscall.
1686 // Run safe point function for all idle Ps. sched.pidle will
1687 // not change because we hold sched.lock.
1688 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1689 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1691 sched.safePointWait--
1695 wait := sched.safePointWait > 0
1698 // Run fn for the current P.
1701 // Force Ps currently in _Psyscall into _Pidle and hand them
1702 // off to induce safe point function execution.
1703 for _, p2 := range allp {
1705 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1715 // Wait for remaining Ps to run fn.
1718 // Wait for 100us, then try to re-preempt in
1719 // case of any races.
1721 // Requires system stack.
1722 if notetsleep(&sched.safePointNote, 100*1000) {
1723 noteclear(&sched.safePointNote)
1729 if sched.safePointWait != 0 {
1730 throw("forEachP: not done")
1732 for _, p2 := range allp {
1733 if p2.runSafePointFn != 0 {
1734 throw("forEachP: P did not run fn")
1739 sched.safePointFn = nil
1744 // runSafePointFn runs the safe point function, if any, for this P.
1745 // This should be called like
1747 // if getg().m.p.runSafePointFn != 0 {
1751 // runSafePointFn must be checked on any transition in to _Pidle or
1752 // _Psyscall to avoid a race where forEachP sees that the P is running
1753 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1754 // nor the P run the safe-point function.
1755 func runSafePointFn() {
1756 p := getg().m.p.ptr()
1757 // Resolve the race between forEachP running the safe-point
1758 // function on this P's behalf and this P running the
1759 // safe-point function directly.
1760 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1763 sched.safePointFn(p)
1765 sched.safePointWait--
1766 if sched.safePointWait == 0 {
1767 notewakeup(&sched.safePointNote)
1772 // When running with cgo, we call _cgo_thread_start
1773 // to start threads for us so that we can play nicely with
1775 var cgoThreadStart unsafe.Pointer
1777 type cgothreadstart struct {
1783 // Allocate a new m unassociated with any thread.
1784 // Can use p for allocation context if needed.
1785 // fn is recorded as the new m's m.mstartfn.
1786 // id is optional pre-allocated m ID. Omit by passing -1.
1788 // This function is allowed to have write barriers even if the caller
1789 // isn't because it borrows pp.
1791 //go:yeswritebarrierrec
1792 func allocm(pp *p, fn func(), id int64) *m {
1795 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1796 // disable preemption to ensure it is not stolen, which would make the
1797 // caller lose ownership.
1802 acquirep(pp) // temporarily borrow p for mallocs in this function
1805 // Release the free M list. We need to do this somewhere and
1806 // this may free up a stack we can use.
1807 if sched.freem != nil {
1810 for freem := sched.freem; freem != nil; {
1811 wait := freem.freeWait.Load()
1812 if wait == freeMWait {
1813 next := freem.freelink
1814 freem.freelink = newList
1819 // Free the stack if needed. For freeMRef, there is
1820 // nothing to do except drop freem from the sched.freem
1822 if wait == freeMStack {
1823 // stackfree must be on the system stack, but allocm is
1824 // reachable off the system stack transitively from
1826 systemstack(func() {
1827 stackfree(freem.g0.stack)
1830 freem = freem.freelink
1832 sched.freem = newList
1840 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1841 // Windows and Plan 9 will layout sched stack on OS stack.
1842 if iscgo || mStackIsSystemAllocated() {
1845 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1849 if pp == gp.m.p.ptr() {
1854 allocmLock.runlock()
1858 // needm is called when a cgo callback happens on a
1859 // thread without an m (a thread not created by Go).
1860 // In this case, needm is expected to find an m to use
1861 // and return with m, g initialized correctly.
1862 // Since m and g are not set now (likely nil, but see below)
1863 // needm is limited in what routines it can call. In particular
1864 // it can only call nosplit functions (textflag 7) and cannot
1865 // do any scheduling that requires an m.
1867 // In order to avoid needing heavy lifting here, we adopt
1868 // the following strategy: there is a stack of available m's
1869 // that can be stolen. Using compare-and-swap
1870 // to pop from the stack has ABA races, so we simulate
1871 // a lock by doing an exchange (via Casuintptr) to steal the stack
1872 // head and replace the top pointer with MLOCKED (1).
1873 // This serves as a simple spin lock that we can use even
1874 // without an m. The thread that locks the stack in this way
1875 // unlocks the stack by storing a valid stack head pointer.
1877 // In order to make sure that there is always an m structure
1878 // available to be stolen, we maintain the invariant that there
1879 // is always one more than needed. At the beginning of the
1880 // program (if cgo is in use) the list is seeded with a single m.
1881 // If needm finds that it has taken the last m off the list, its job
1882 // is - once it has installed its own m so that it can do things like
1883 // allocate memory - to create a spare m and put it on the list.
1885 // Each of these extra m's also has a g0 and a curg that are
1886 // pressed into service as the scheduling stack and current
1887 // goroutine for the duration of the cgo callback.
1889 // When the callback is done with the m, it calls dropm to
1890 // put the m back on the list.
1894 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1895 // Can happen if C/C++ code calls Go from a global ctor.
1896 // Can also happen on Windows if a global ctor uses a
1897 // callback created by syscall.NewCallback. See issue #6751
1900 // Can not throw, because scheduler is not initialized yet.
1901 writeErrStr("fatal error: cgo callback before cgo call\n")
1905 // Save and block signals before getting an M.
1906 // The signal handler may call needm itself,
1907 // and we must avoid a deadlock. Also, once g is installed,
1908 // any incoming signals will try to execute,
1909 // but we won't have the sigaltstack settings and other data
1910 // set up appropriately until the end of minit, which will
1911 // unblock the signals. This is the same dance as when
1912 // starting a new m to run Go code via newosproc.
1917 // nilokay=false is safe here because of the invariant above,
1918 // that the extra list always contains or will soon contain
1920 mp, last := getExtraM(false)
1922 // Set needextram when we've just emptied the list,
1923 // so that the eventual call into cgocallbackg will
1924 // allocate a new m for the extra list. We delay the
1925 // allocation until then so that it can be done
1926 // after exitsyscall makes sure it is okay to be
1927 // running at all (that is, there's no garbage collection
1928 // running right now).
1929 mp.needextram = last
1931 // Store the original signal mask for use by minit.
1932 mp.sigmask = sigmask
1934 // Install TLS on some platforms (previously setg
1935 // would do this if necessary).
1938 // Install g (= m->g0) and set the stack bounds
1939 // to match the current stack. We don't actually know
1940 // how big the stack is, like we don't know how big any
1941 // scheduling stack is, but we assume there's at least 32 kB,
1942 // which is more than enough for us.
1945 gp.stack.hi = getcallersp() + 1024
1946 gp.stack.lo = getcallersp() - 32*1024
1947 gp.stackguard0 = gp.stack.lo + stackGuard
1949 // Initialize this thread to use the m.
1953 // mp.curg is now a real goroutine.
1954 casgstatus(mp.curg, _Gdead, _Gsyscall)
1958 // newextram allocates m's and puts them on the extra list.
1959 // It is called with a working local m, so that it can do things
1960 // like call schedlock and allocate.
1962 c := extraMWaiters.Swap(0)
1964 for i := uint32(0); i < c; i++ {
1967 } else if extraMLength.Load() == 0 {
1968 // Make sure there is at least one extra M.
1973 // oneNewExtraM allocates an m and puts it on the extra list.
1974 func oneNewExtraM() {
1975 // Create extra goroutine locked to extra m.
1976 // The goroutine is the context in which the cgo callback will run.
1977 // The sched.pc will never be returned to, but setting it to
1978 // goexit makes clear to the traceback routines where
1979 // the goroutine stack ends.
1980 mp := allocm(nil, nil, -1)
1982 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
1983 gp.sched.sp = gp.stack.hi
1984 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
1986 gp.sched.g = guintptr(unsafe.Pointer(gp))
1987 gp.syscallpc = gp.sched.pc
1988 gp.syscallsp = gp.sched.sp
1989 gp.stktopsp = gp.sched.sp
1990 // malg returns status as _Gidle. Change to _Gdead before
1991 // adding to allg where GC can see it. We use _Gdead to hide
1992 // this from tracebacks and stack scans since it isn't a
1993 // "real" goroutine until needm grabs it.
1994 casgstatus(gp, _Gidle, _Gdead)
2001 gp.goid = sched.goidgen.Add(1)
2002 gp.sysblocktraced = true
2004 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
2007 // Trigger two trace events for the locked g in the extra m,
2008 // since the next event of the g will be traceEvGoSysExit in exitsyscall,
2009 // while calling from C thread to Go.
2010 traceGoCreate(gp, 0) // no start pc
2012 traceEvent(traceEvGoInSyscall, -1, gp.goid)
2014 // put on allg for garbage collector
2017 // gp is now on the allg list, but we don't want it to be
2018 // counted by gcount. It would be more "proper" to increment
2019 // sched.ngfree, but that requires locking. Incrementing ngsys
2020 // has the same effect.
2023 // Add m to the extra list.
2027 // dropm is called when a cgo callback has called needm but is now
2028 // done with the callback and returning back into the non-Go thread.
2029 // It puts the current m back onto the extra list.
2031 // The main expense here is the call to signalstack to release the
2032 // m's signal stack, and then the call to needm on the next callback
2033 // from this thread. It is tempting to try to save the m for next time,
2034 // which would eliminate both these costs, but there might not be
2035 // a next time: the current thread (which Go does not control) might exit.
2036 // If we saved the m for that thread, there would be an m leak each time
2037 // such a thread exited. Instead, we acquire and release an m on each
2038 // call. These should typically not be scheduling operations, just a few
2039 // atomics, so the cost should be small.
2041 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
2042 // variable using pthread_key_create. Unlike the pthread keys we already use
2043 // on OS X, this dummy key would never be read by Go code. It would exist
2044 // only so that we could register at thread-exit-time destructor.
2045 // That destructor would put the m back onto the extra list.
2046 // This is purely a performance optimization. The current version,
2047 // in which dropm happens on each cgo call, is still correct too.
2048 // We may have to keep the current version on systems with cgo
2049 // but without pthreads, like Windows.
2051 // Clear m and g, and return m to the extra list.
2052 // After the call to setg we can only call nosplit functions
2053 // with no pointer manipulation.
2056 // Return mp.curg to dead state.
2057 casgstatus(mp.curg, _Gsyscall, _Gdead)
2058 mp.curg.preemptStop = false
2061 // Block signals before unminit.
2062 // Unminit unregisters the signal handling stack (but needs g on some systems).
2063 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2064 // It's important not to try to handle a signal between those two steps.
2065 sigmask := mp.sigmask
2073 msigrestore(sigmask)
2076 // A helper function for EnsureDropM.
2077 func getm() uintptr {
2078 return uintptr(unsafe.Pointer(getg().m))
2082 // Locking linked list of extra M's, via mp.schedlink. Must be accessed
2083 // only via lockextra/unlockextra.
2085 // Can't be atomic.Pointer[m] because we use an invalid pointer as a
2086 // "locked" sentinel value. M's on this list remain visible to the GC
2087 // because their mp.curg is on allgs.
2088 extraM atomic.Uintptr
2089 // Number of M's in the extraM list.
2090 extraMLength atomic.Uint32
2091 // Number of waiters in lockextra.
2092 extraMWaiters atomic.Uint32
2094 // Number of extra M's in use by threads.
2095 extraMInUse atomic.Uint32
2098 // lockextra locks the extra list and returns the list head.
2099 // The caller must unlock the list by storing a new list head
2100 // to extram. If nilokay is true, then lockextra will
2101 // return a nil list head if that's what it finds. If nilokay is false,
2102 // lockextra will keep waiting until the list head is no longer nil.
2105 func lockextra(nilokay bool) *m {
2110 old := extraM.Load()
2115 if old == 0 && !nilokay {
2117 // Add 1 to the number of threads
2118 // waiting for an M.
2119 // This is cleared by newextram.
2120 extraMWaiters.Add(1)
2126 if extraM.CompareAndSwap(old, locked) {
2128 return (*m)(unsafe.Pointer(old))
2136 func unlockextra(mp *m, delta int32) {
2137 extraMLength.Add(delta)
2138 extraM.Store(uintptr(unsafe.Pointer(mp)))
2141 // Return an M from the extra M list. Returns last == true if the list becomes
2142 // empty because of this call.
2145 func getExtraM(nilokay bool) (mp *m, last bool) {
2146 mp = lockextra(nilokay)
2151 unlockextra(mp.schedlink.ptr(), -1)
2152 return mp, mp.schedlink.ptr() == nil
2155 // Returns an extra M back to the list. mp must be from getExtraM. Newly
2156 // allocated M's should use addExtraM.
2159 func putExtraM(mp *m) {
2164 // Adds a newly allocated M to the extra M list.
2167 func addExtraM(mp *m) {
2168 mnext := lockextra(true)
2169 mp.schedlink.set(mnext)
2174 // allocmLock is locked for read when creating new Ms in allocm and their
2175 // addition to allm. Thus acquiring this lock for write blocks the
2176 // creation of new Ms.
2179 // execLock serializes exec and clone to avoid bugs or unspecified
2180 // behaviour around exec'ing while creating/destroying threads. See
2185 // These errors are reported (via writeErrStr) by some OS-specific
2186 // versions of newosproc and newosproc0.
2188 failthreadcreate = "runtime: failed to create new OS thread\n"
2189 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2192 // newmHandoff contains a list of m structures that need new OS threads.
2193 // This is used by newm in situations where newm itself can't safely
2194 // start an OS thread.
2195 var newmHandoff struct {
2198 // newm points to a list of M structures that need new OS
2199 // threads. The list is linked through m.schedlink.
2202 // waiting indicates that wake needs to be notified when an m
2203 // is put on the list.
2207 // haveTemplateThread indicates that the templateThread has
2208 // been started. This is not protected by lock. Use cas to set
2210 haveTemplateThread uint32
2213 // Create a new m. It will start off with a call to fn, or else the scheduler.
2214 // fn needs to be static and not a heap allocated closure.
2215 // May run with m.p==nil, so write barriers are not allowed.
2217 // id is optional pre-allocated m ID. Omit by passing -1.
2219 //go:nowritebarrierrec
2220 func newm(fn func(), pp *p, id int64) {
2221 // allocm adds a new M to allm, but they do not start until created by
2222 // the OS in newm1 or the template thread.
2224 // doAllThreadsSyscall requires that every M in allm will eventually
2225 // start and be signal-able, even with a STW.
2227 // Disable preemption here until we start the thread to ensure that
2228 // newm is not preempted between allocm and starting the new thread,
2229 // ensuring that anything added to allm is guaranteed to eventually
2233 mp := allocm(pp, fn, id)
2235 mp.sigmask = initSigmask
2236 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2237 // We're on a locked M or a thread that may have been
2238 // started by C. The kernel state of this thread may
2239 // be strange (the user may have locked it for that
2240 // purpose). We don't want to clone that into another
2241 // thread. Instead, ask a known-good thread to create
2242 // the thread for us.
2244 // This is disabled on Plan 9. See golang.org/issue/22227.
2246 // TODO: This may be unnecessary on Windows, which
2247 // doesn't model thread creation off fork.
2248 lock(&newmHandoff.lock)
2249 if newmHandoff.haveTemplateThread == 0 {
2250 throw("on a locked thread with no template thread")
2252 mp.schedlink = newmHandoff.newm
2253 newmHandoff.newm.set(mp)
2254 if newmHandoff.waiting {
2255 newmHandoff.waiting = false
2256 notewakeup(&newmHandoff.wake)
2258 unlock(&newmHandoff.lock)
2259 // The M has not started yet, but the template thread does not
2260 // participate in STW, so it will always process queued Ms and
2261 // it is safe to releasem.
2271 var ts cgothreadstart
2272 if _cgo_thread_start == nil {
2273 throw("_cgo_thread_start missing")
2276 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2277 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2279 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2282 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2284 execLock.rlock() // Prevent process clone.
2285 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2289 execLock.rlock() // Prevent process clone.
2294 // startTemplateThread starts the template thread if it is not already
2297 // The calling thread must itself be in a known-good state.
2298 func startTemplateThread() {
2299 if GOARCH == "wasm" { // no threads on wasm yet
2303 // Disable preemption to guarantee that the template thread will be
2304 // created before a park once haveTemplateThread is set.
2306 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2310 newm(templateThread, nil, -1)
2314 // templateThread is a thread in a known-good state that exists solely
2315 // to start new threads in known-good states when the calling thread
2316 // may not be in a good state.
2318 // Many programs never need this, so templateThread is started lazily
2319 // when we first enter a state that might lead to running on a thread
2320 // in an unknown state.
2322 // templateThread runs on an M without a P, so it must not have write
2325 //go:nowritebarrierrec
2326 func templateThread() {
2333 lock(&newmHandoff.lock)
2334 for newmHandoff.newm != 0 {
2335 newm := newmHandoff.newm.ptr()
2336 newmHandoff.newm = 0
2337 unlock(&newmHandoff.lock)
2339 next := newm.schedlink.ptr()
2344 lock(&newmHandoff.lock)
2346 newmHandoff.waiting = true
2347 noteclear(&newmHandoff.wake)
2348 unlock(&newmHandoff.lock)
2349 notesleep(&newmHandoff.wake)
2353 // Stops execution of the current m until new work is available.
2354 // Returns with acquired P.
2358 if gp.m.locks != 0 {
2359 throw("stopm holding locks")
2362 throw("stopm holding p")
2365 throw("stopm spinning")
2372 acquirep(gp.m.nextp.ptr())
2377 // startm's caller incremented nmspinning. Set the new M's spinning.
2378 getg().m.spinning = true
2381 // Schedules some M to run the p (creates an M if necessary).
2382 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2383 // May run with m.p==nil, so write barriers are not allowed.
2384 // If spinning is set, the caller has incremented nmspinning and must provide a
2385 // P. startm will set m.spinning in the newly started M.
2387 // Callers passing a non-nil P must call from a non-preemptible context. See
2388 // comment on acquirem below.
2390 // Argument lockheld indicates whether the caller already acquired the
2391 // scheduler lock. Callers holding the lock when making the call must pass
2392 // true. The lock might be temporarily dropped, but will be reacquired before
2395 // Must not have write barriers because this may be called without a P.
2397 //go:nowritebarrierrec
2398 func startm(pp *p, spinning, lockheld bool) {
2399 // Disable preemption.
2401 // Every owned P must have an owner that will eventually stop it in the
2402 // event of a GC stop request. startm takes transient ownership of a P
2403 // (either from argument or pidleget below) and transfers ownership to
2404 // a started M, which will be responsible for performing the stop.
2406 // Preemption must be disabled during this transient ownership,
2407 // otherwise the P this is running on may enter GC stop while still
2408 // holding the transient P, leaving that P in limbo and deadlocking the
2411 // Callers passing a non-nil P must already be in non-preemptible
2412 // context, otherwise such preemption could occur on function entry to
2413 // startm. Callers passing a nil P may be preemptible, so we must
2414 // disable preemption before acquiring a P from pidleget below.
2421 // TODO(prattmic): All remaining calls to this function
2422 // with _p_ == nil could be cleaned up to find a P
2423 // before calling startm.
2424 throw("startm: P required for spinning=true")
2437 // No M is available, we must drop sched.lock and call newm.
2438 // However, we already own a P to assign to the M.
2440 // Once sched.lock is released, another G (e.g., in a syscall),
2441 // could find no idle P while checkdead finds a runnable G but
2442 // no running M's because this new M hasn't started yet, thus
2443 // throwing in an apparent deadlock.
2444 // This apparent deadlock is possible when startm is called
2445 // from sysmon, which doesn't count as a running M.
2447 // Avoid this situation by pre-allocating the ID for the new M,
2448 // thus marking it as 'running' before we drop sched.lock. This
2449 // new M will eventually run the scheduler to execute any
2456 // The caller incremented nmspinning, so set m.spinning in the new M.
2464 // Ownership transfer of pp committed by start in newm.
2465 // Preemption is now safe.
2473 throw("startm: m is spinning")
2476 throw("startm: m has p")
2478 if spinning && !runqempty(pp) {
2479 throw("startm: p has runnable gs")
2481 // The caller incremented nmspinning, so set m.spinning in the new M.
2482 nmp.spinning = spinning
2484 notewakeup(&nmp.park)
2485 // Ownership transfer of pp committed by wakeup. Preemption is now
2490 // Hands off P from syscall or locked M.
2491 // Always runs without a P, so write barriers are not allowed.
2493 //go:nowritebarrierrec
2494 func handoffp(pp *p) {
2495 // handoffp must start an M in any situation where
2496 // findrunnable would return a G to run on pp.
2498 // if it has local work, start it straight away
2499 if !runqempty(pp) || sched.runqsize != 0 {
2500 startm(pp, false, false)
2503 // if there's trace work to do, start it straight away
2504 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2505 startm(pp, false, false)
2508 // if it has GC work, start it straight away
2509 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2510 startm(pp, false, false)
2513 // no local work, check that there are no spinning/idle M's,
2514 // otherwise our help is not required
2515 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2516 sched.needspinning.Store(0)
2517 startm(pp, true, false)
2521 if sched.gcwaiting.Load() {
2522 pp.status = _Pgcstop
2524 if sched.stopwait == 0 {
2525 notewakeup(&sched.stopnote)
2530 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2531 sched.safePointFn(pp)
2532 sched.safePointWait--
2533 if sched.safePointWait == 0 {
2534 notewakeup(&sched.safePointNote)
2537 if sched.runqsize != 0 {
2539 startm(pp, false, false)
2542 // If this is the last running P and nobody is polling network,
2543 // need to wakeup another M to poll network.
2544 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2546 startm(pp, false, false)
2550 // The scheduler lock cannot be held when calling wakeNetPoller below
2551 // because wakeNetPoller may call wakep which may call startm.
2552 when := nobarrierWakeTime(pp)
2561 // Tries to add one more P to execute G's.
2562 // Called when a G is made runnable (newproc, ready).
2563 // Must be called with a P.
2565 // Be conservative about spinning threads, only start one if none exist
2567 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2571 // Disable preemption until ownership of pp transfers to the next M in
2572 // startm. Otherwise preemption here would leave pp stuck waiting to
2575 // See preemption comment on acquirem in startm for more details.
2580 pp, _ = pidlegetSpinning(0)
2582 if sched.nmspinning.Add(-1) < 0 {
2583 throw("wakep: negative nmspinning")
2589 // Since we always have a P, the race in the "No M is available"
2590 // comment in startm doesn't apply during the small window between the
2591 // unlock here and lock in startm. A checkdead in between will always
2592 // see at least one running M (ours).
2595 startm(pp, true, false)
2600 // Stops execution of the current m that is locked to a g until the g is runnable again.
2601 // Returns with acquired P.
2602 func stoplockedm() {
2605 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2606 throw("stoplockedm: inconsistent locking")
2609 // Schedule another M to run this p.
2614 // Wait until another thread schedules lockedg again.
2616 status := readgstatus(gp.m.lockedg.ptr())
2617 if status&^_Gscan != _Grunnable {
2618 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2619 dumpgstatus(gp.m.lockedg.ptr())
2620 throw("stoplockedm: not runnable")
2622 acquirep(gp.m.nextp.ptr())
2626 // Schedules the locked m to run the locked gp.
2627 // May run during STW, so write barriers are not allowed.
2629 //go:nowritebarrierrec
2630 func startlockedm(gp *g) {
2631 mp := gp.lockedm.ptr()
2633 throw("startlockedm: locked to me")
2636 throw("startlockedm: m has p")
2638 // directly handoff current P to the locked m
2642 notewakeup(&mp.park)
2646 // Stops the current m for stopTheWorld.
2647 // Returns when the world is restarted.
2651 if !sched.gcwaiting.Load() {
2652 throw("gcstopm: not waiting for gc")
2655 gp.m.spinning = false
2656 // OK to just drop nmspinning here,
2657 // startTheWorld will unpark threads as necessary.
2658 if sched.nmspinning.Add(-1) < 0 {
2659 throw("gcstopm: negative nmspinning")
2664 pp.status = _Pgcstop
2666 if sched.stopwait == 0 {
2667 notewakeup(&sched.stopnote)
2673 // Schedules gp to run on the current M.
2674 // If inheritTime is true, gp inherits the remaining time in the
2675 // current time slice. Otherwise, it starts a new time slice.
2678 // Write barriers are allowed because this is called immediately after
2679 // acquiring a P in several places.
2681 //go:yeswritebarrierrec
2682 func execute(gp *g, inheritTime bool) {
2685 if goroutineProfile.active {
2686 // Make sure that gp has had its stack written out to the goroutine
2687 // profile, exactly as it was when the goroutine profiler first stopped
2689 tryRecordGoroutineProfile(gp, osyield)
2692 // Assign gp.m before entering _Grunning so running Gs have an
2696 casgstatus(gp, _Grunnable, _Grunning)
2699 gp.stackguard0 = gp.stack.lo + stackGuard
2701 mp.p.ptr().schedtick++
2704 // Check whether the profiler needs to be turned on or off.
2705 hz := sched.profilehz
2706 if mp.profilehz != hz {
2707 setThreadCPUProfiler(hz)
2711 // GoSysExit has to happen when we have a P, but before GoStart.
2712 // So we emit it here.
2713 if gp.syscallsp != 0 && gp.sysblocktraced {
2714 traceGoSysExit(gp.sysexitticks)
2722 // Finds a runnable goroutine to execute.
2723 // Tries to steal from other P's, get g from local or global queue, poll network.
2724 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2725 // reader) so the caller should try to wake a P.
2726 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2729 // The conditions here and in handoffp must agree: if
2730 // findrunnable would return a G to run, handoffp must start
2735 if sched.gcwaiting.Load() {
2739 if pp.runSafePointFn != 0 {
2743 // now and pollUntil are saved for work stealing later,
2744 // which may steal timers. It's important that between now
2745 // and then, nothing blocks, so these numbers remain mostly
2747 now, pollUntil, _ := checkTimers(pp, 0)
2749 // Try to schedule the trace reader.
2750 if trace.enabled || trace.shutdown {
2753 casgstatus(gp, _Gwaiting, _Grunnable)
2754 traceGoUnpark(gp, 0)
2755 return gp, false, true
2759 // Try to schedule a GC worker.
2760 if gcBlackenEnabled != 0 {
2761 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2763 return gp, false, true
2768 // Check the global runnable queue once in a while to ensure fairness.
2769 // Otherwise two goroutines can completely occupy the local runqueue
2770 // by constantly respawning each other.
2771 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2773 gp := globrunqget(pp, 1)
2776 return gp, false, false
2780 // Wake up the finalizer G.
2781 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2782 if gp := wakefing(); gp != nil {
2786 if *cgo_yield != nil {
2787 asmcgocall(*cgo_yield, nil)
2791 if gp, inheritTime := runqget(pp); gp != nil {
2792 return gp, inheritTime, false
2796 if sched.runqsize != 0 {
2798 gp := globrunqget(pp, 0)
2801 return gp, false, false
2806 // This netpoll is only an optimization before we resort to stealing.
2807 // We can safely skip it if there are no waiters or a thread is blocked
2808 // in netpoll already. If there is any kind of logical race with that
2809 // blocked thread (e.g. it has already returned from netpoll, but does
2810 // not set lastpoll yet), this thread will do blocking netpoll below
2812 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2813 if list := netpoll(0); !list.empty() { // non-blocking
2816 casgstatus(gp, _Gwaiting, _Grunnable)
2818 traceGoUnpark(gp, 0)
2820 return gp, false, false
2824 // Spinning Ms: steal work from other Ps.
2826 // Limit the number of spinning Ms to half the number of busy Ps.
2827 // This is necessary to prevent excessive CPU consumption when
2828 // GOMAXPROCS>>1 but the program parallelism is low.
2829 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2834 gp, inheritTime, tnow, w, newWork := stealWork(now)
2836 // Successfully stole.
2837 return gp, inheritTime, false
2840 // There may be new timer or GC work; restart to
2846 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2847 // Earlier timer to wait for.
2852 // We have nothing to do.
2854 // If we're in the GC mark phase, can safely scan and blacken objects,
2855 // and have work to do, run idle-time marking rather than give up the P.
2856 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2857 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2859 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2861 casgstatus(gp, _Gwaiting, _Grunnable)
2863 traceGoUnpark(gp, 0)
2865 return gp, false, false
2867 gcController.removeIdleMarkWorker()
2871 // If a callback returned and no other goroutine is awake,
2872 // then wake event handler goroutine which pauses execution
2873 // until a callback was triggered.
2874 gp, otherReady := beforeIdle(now, pollUntil)
2876 casgstatus(gp, _Gwaiting, _Grunnable)
2878 traceGoUnpark(gp, 0)
2880 return gp, false, false
2886 // Before we drop our P, make a snapshot of the allp slice,
2887 // which can change underfoot once we no longer block
2888 // safe-points. We don't need to snapshot the contents because
2889 // everything up to cap(allp) is immutable.
2890 allpSnapshot := allp
2891 // Also snapshot masks. Value changes are OK, but we can't allow
2892 // len to change out from under us.
2893 idlepMaskSnapshot := idlepMask
2894 timerpMaskSnapshot := timerpMask
2896 // return P and block
2898 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2902 if sched.runqsize != 0 {
2903 gp := globrunqget(pp, 0)
2905 return gp, false, false
2907 if !mp.spinning && sched.needspinning.Load() == 1 {
2908 // See "Delicate dance" comment below.
2913 if releasep() != pp {
2914 throw("findrunnable: wrong p")
2916 now = pidleput(pp, now)
2919 // Delicate dance: thread transitions from spinning to non-spinning
2920 // state, potentially concurrently with submission of new work. We must
2921 // drop nmspinning first and then check all sources again (with
2922 // #StoreLoad memory barrier in between). If we do it the other way
2923 // around, another thread can submit work after we've checked all
2924 // sources but before we drop nmspinning; as a result nobody will
2925 // unpark a thread to run the work.
2927 // This applies to the following sources of work:
2929 // * Goroutines added to a per-P run queue.
2930 // * New/modified-earlier timers on a per-P timer heap.
2931 // * Idle-priority GC work (barring golang.org/issue/19112).
2933 // If we discover new work below, we need to restore m.spinning as a
2934 // signal for resetspinning to unpark a new worker thread (because
2935 // there can be more than one starving goroutine).
2937 // However, if after discovering new work we also observe no idle Ps
2938 // (either here or in resetspinning), we have a problem. We may be
2939 // racing with a non-spinning M in the block above, having found no
2940 // work and preparing to release its P and park. Allowing that P to go
2941 // idle will result in loss of work conservation (idle P while there is
2942 // runnable work). This could result in complete deadlock in the
2943 // unlikely event that we discover new work (from netpoll) right as we
2944 // are racing with _all_ other Ps going idle.
2946 // We use sched.needspinning to synchronize with non-spinning Ms going
2947 // idle. If needspinning is set when they are about to drop their P,
2948 // they abort the drop and instead become a new spinning M on our
2949 // behalf. If we are not racing and the system is truly fully loaded
2950 // then no spinning threads are required, and the next thread to
2951 // naturally become spinning will clear the flag.
2953 // Also see "Worker thread parking/unparking" comment at the top of the
2955 wasSpinning := mp.spinning
2958 if sched.nmspinning.Add(-1) < 0 {
2959 throw("findrunnable: negative nmspinning")
2962 // Note the for correctness, only the last M transitioning from
2963 // spinning to non-spinning must perform these rechecks to
2964 // ensure no missed work. However, the runtime has some cases
2965 // of transient increments of nmspinning that are decremented
2966 // without going through this path, so we must be conservative
2967 // and perform the check on all spinning Ms.
2969 // See https://go.dev/issue/43997.
2971 // Check all runqueues once again.
2972 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2979 // Check for idle-priority GC work again.
2980 pp, gp := checkIdleGCNoP()
2985 // Run the idle worker.
2986 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2987 casgstatus(gp, _Gwaiting, _Grunnable)
2989 traceGoUnpark(gp, 0)
2991 return gp, false, false
2994 // Finally, check for timer creation or expiry concurrently with
2995 // transitioning from spinning to non-spinning.
2997 // Note that we cannot use checkTimers here because it calls
2998 // adjusttimers which may need to allocate memory, and that isn't
2999 // allowed when we don't have an active P.
3000 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
3003 // Poll network until next timer.
3004 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
3005 sched.pollUntil.Store(pollUntil)
3007 throw("findrunnable: netpoll with p")
3010 throw("findrunnable: netpoll with spinning")
3016 delay = pollUntil - now
3022 // When using fake time, just poll.
3025 list := netpoll(delay) // block until new work is available
3026 sched.pollUntil.Store(0)
3027 sched.lastpoll.Store(now)
3028 if faketime != 0 && list.empty() {
3029 // Using fake time and nothing is ready; stop M.
3030 // When all M's stop, checkdead will call timejump.
3035 pp, _ := pidleget(now)
3044 casgstatus(gp, _Gwaiting, _Grunnable)
3046 traceGoUnpark(gp, 0)
3048 return gp, false, false
3055 } else if pollUntil != 0 && netpollinited() {
3056 pollerPollUntil := sched.pollUntil.Load()
3057 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3065 // pollWork reports whether there is non-background work this P could
3066 // be doing. This is a fairly lightweight check to be used for
3067 // background work loops, like idle GC. It checks a subset of the
3068 // conditions checked by the actual scheduler.
3069 func pollWork() bool {
3070 if sched.runqsize != 0 {
3073 p := getg().m.p.ptr()
3077 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3078 if list := netpoll(0); !list.empty() {
3086 // stealWork attempts to steal a runnable goroutine or timer from any P.
3088 // If newWork is true, new work may have been readied.
3090 // If now is not 0 it is the current time. stealWork returns the passed time or
3091 // the current time if now was passed as 0.
3092 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3093 pp := getg().m.p.ptr()
3097 const stealTries = 4
3098 for i := 0; i < stealTries; i++ {
3099 stealTimersOrRunNextG := i == stealTries-1
3101 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3102 if sched.gcwaiting.Load() {
3103 // GC work may be available.
3104 return nil, false, now, pollUntil, true
3106 p2 := allp[enum.position()]
3111 // Steal timers from p2. This call to checkTimers is the only place
3112 // where we might hold a lock on a different P's timers. We do this
3113 // once on the last pass before checking runnext because stealing
3114 // from the other P's runnext should be the last resort, so if there
3115 // are timers to steal do that first.
3117 // We only check timers on one of the stealing iterations because
3118 // the time stored in now doesn't change in this loop and checking
3119 // the timers for each P more than once with the same value of now
3120 // is probably a waste of time.
3122 // timerpMask tells us whether the P may have timers at all. If it
3123 // can't, no need to check at all.
3124 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3125 tnow, w, ran := checkTimers(p2, now)
3127 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3131 // Running the timers may have
3132 // made an arbitrary number of G's
3133 // ready and added them to this P's
3134 // local run queue. That invalidates
3135 // the assumption of runqsteal
3136 // that it always has room to add
3137 // stolen G's. So check now if there
3138 // is a local G to run.
3139 if gp, inheritTime := runqget(pp); gp != nil {
3140 return gp, inheritTime, now, pollUntil, ranTimer
3146 // Don't bother to attempt to steal if p2 is idle.
3147 if !idlepMask.read(enum.position()) {
3148 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3149 return gp, false, now, pollUntil, ranTimer
3155 // No goroutines found to steal. Regardless, running a timer may have
3156 // made some goroutine ready that we missed. Indicate the next timer to
3158 return nil, false, now, pollUntil, ranTimer
3161 // Check all Ps for a runnable G to steal.
3163 // On entry we have no P. If a G is available to steal and a P is available,
3164 // the P is returned which the caller should acquire and attempt to steal the
3166 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3167 for id, p2 := range allpSnapshot {
3168 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3170 pp, _ := pidlegetSpinning(0)
3172 // Can't get a P, don't bother checking remaining Ps.
3181 // No work available.
3185 // Check all Ps for a timer expiring sooner than pollUntil.
3187 // Returns updated pollUntil value.
3188 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3189 for id, p2 := range allpSnapshot {
3190 if timerpMaskSnapshot.read(uint32(id)) {
3191 w := nobarrierWakeTime(p2)
3192 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3201 // Check for idle-priority GC, without a P on entry.
3203 // If some GC work, a P, and a worker G are all available, the P and G will be
3204 // returned. The returned P has not been wired yet.
3205 func checkIdleGCNoP() (*p, *g) {
3206 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3207 // must check again after acquiring a P. As an optimization, we also check
3208 // if an idle mark worker is needed at all. This is OK here, because if we
3209 // observe that one isn't needed, at least one is currently running. Even if
3210 // it stops running, its own journey into the scheduler should schedule it
3211 // again, if need be (at which point, this check will pass, if relevant).
3212 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3215 if !gcMarkWorkAvailable(nil) {
3219 // Work is available; we can start an idle GC worker only if there is
3220 // an available P and available worker G.
3222 // We can attempt to acquire these in either order, though both have
3223 // synchronization concerns (see below). Workers are almost always
3224 // available (see comment in findRunnableGCWorker for the one case
3225 // there may be none). Since we're slightly less likely to find a P,
3226 // check for that first.
3228 // Synchronization: note that we must hold sched.lock until we are
3229 // committed to keeping it. Otherwise we cannot put the unnecessary P
3230 // back in sched.pidle without performing the full set of idle
3231 // transition checks.
3233 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3234 // the assumption in gcControllerState.findRunnableGCWorker that an
3235 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3237 pp, now := pidlegetSpinning(0)
3243 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3244 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3250 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3254 gcController.removeIdleMarkWorker()
3260 return pp, node.gp.ptr()
3263 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3264 // going to wake up before the when argument; or it wakes an idle P to service
3265 // timers and the network poller if there isn't one already.
3266 func wakeNetPoller(when int64) {
3267 if sched.lastpoll.Load() == 0 {
3268 // In findrunnable we ensure that when polling the pollUntil
3269 // field is either zero or the time to which the current
3270 // poll is expected to run. This can have a spurious wakeup
3271 // but should never miss a wakeup.
3272 pollerPollUntil := sched.pollUntil.Load()
3273 if pollerPollUntil == 0 || pollerPollUntil > when {
3277 // There are no threads in the network poller, try to get
3278 // one there so it can handle new timers.
3279 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3285 func resetspinning() {
3288 throw("resetspinning: not a spinning m")
3290 gp.m.spinning = false
3291 nmspinning := sched.nmspinning.Add(-1)
3293 throw("findrunnable: negative nmspinning")
3295 // M wakeup policy is deliberately somewhat conservative, so check if we
3296 // need to wakeup another P here. See "Worker thread parking/unparking"
3297 // comment at the top of the file for details.
3301 // injectglist adds each runnable G on the list to some run queue,
3302 // and clears glist. If there is no current P, they are added to the
3303 // global queue, and up to npidle M's are started to run them.
3304 // Otherwise, for each idle P, this adds a G to the global queue
3305 // and starts an M. Any remaining G's are added to the current P's
3307 // This may temporarily acquire sched.lock.
3308 // Can run concurrently with GC.
3309 func injectglist(glist *gList) {
3314 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3315 traceGoUnpark(gp, 0)
3319 // Mark all the goroutines as runnable before we put them
3320 // on the run queues.
3321 head := glist.head.ptr()
3324 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3327 casgstatus(gp, _Gwaiting, _Grunnable)
3330 // Turn the gList into a gQueue.
3336 startIdle := func(n int) {
3337 for i := 0; i < n; i++ {
3338 mp := acquirem() // See comment in startm.
3341 pp, _ := pidlegetSpinning(0)
3348 startm(pp, false, true)
3354 pp := getg().m.p.ptr()
3357 globrunqputbatch(&q, int32(qsize))
3363 npidle := int(sched.npidle.Load())
3366 for n = 0; n < npidle && !q.empty(); n++ {
3372 globrunqputbatch(&globq, int32(n))
3379 runqputbatch(pp, &q, qsize)
3383 // One round of scheduler: find a runnable goroutine and execute it.
3389 throw("schedule: holding locks")
3392 if mp.lockedg != 0 {
3394 execute(mp.lockedg.ptr(), false) // Never returns.
3397 // We should not schedule away from a g that is executing a cgo call,
3398 // since the cgo call is using the m's g0 stack.
3400 throw("schedule: in cgo")
3407 // Safety check: if we are spinning, the run queue should be empty.
3408 // Check this before calling checkTimers, as that might call
3409 // goready to put a ready goroutine on the local run queue.
3410 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3411 throw("schedule: spinning with local work")
3414 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3416 // This thread is going to run a goroutine and is not spinning anymore,
3417 // so if it was marked as spinning we need to reset it now and potentially
3418 // start a new spinning M.
3423 if sched.disable.user && !schedEnabled(gp) {
3424 // Scheduling of this goroutine is disabled. Put it on
3425 // the list of pending runnable goroutines for when we
3426 // re-enable user scheduling and look again.
3428 if schedEnabled(gp) {
3429 // Something re-enabled scheduling while we
3430 // were acquiring the lock.
3433 sched.disable.runnable.pushBack(gp)
3440 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3441 // wake a P if there is one.
3445 if gp.lockedm != 0 {
3446 // Hands off own p to the locked m,
3447 // then blocks waiting for a new p.
3452 execute(gp, inheritTime)
3455 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3456 // Typically a caller sets gp's status away from Grunning and then
3457 // immediately calls dropg to finish the job. The caller is also responsible
3458 // for arranging that gp will be restarted using ready at an
3459 // appropriate time. After calling dropg and arranging for gp to be
3460 // readied later, the caller can do other work but eventually should
3461 // call schedule to restart the scheduling of goroutines on this m.
3465 setMNoWB(&gp.m.curg.m, nil)
3466 setGNoWB(&gp.m.curg, nil)
3469 // checkTimers runs any timers for the P that are ready.
3470 // If now is not 0 it is the current time.
3471 // It returns the passed time or the current time if now was passed as 0.
3472 // and the time when the next timer should run or 0 if there is no next timer,
3473 // and reports whether it ran any timers.
3474 // If the time when the next timer should run is not 0,
3475 // it is always larger than the returned time.
3476 // We pass now in and out to avoid extra calls of nanotime.
3478 //go:yeswritebarrierrec
3479 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3480 // If it's not yet time for the first timer, or the first adjusted
3481 // timer, then there is nothing to do.
3482 next := pp.timer0When.Load()
3483 nextAdj := pp.timerModifiedEarliest.Load()
3484 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3489 // No timers to run or adjust.
3490 return now, 0, false
3497 // Next timer is not ready to run, but keep going
3498 // if we would clear deleted timers.
3499 // This corresponds to the condition below where
3500 // we decide whether to call clearDeletedTimers.
3501 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3502 return now, next, false
3506 lock(&pp.timersLock)
3508 if len(pp.timers) > 0 {
3509 adjusttimers(pp, now)
3510 for len(pp.timers) > 0 {
3511 // Note that runtimer may temporarily unlock
3513 if tw := runtimer(pp, now); tw != 0 {
3523 // If this is the local P, and there are a lot of deleted timers,
3524 // clear them out. We only do this for the local P to reduce
3525 // lock contention on timersLock.
3526 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3527 clearDeletedTimers(pp)
3530 unlock(&pp.timersLock)
3532 return now, pollUntil, ran
3535 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3536 unlock((*mutex)(lock))
3540 // park continuation on g0.
3541 func park_m(gp *g) {
3545 traceGoPark(mp.waittraceev, mp.waittraceskip)
3548 // N.B. Not using casGToWaiting here because the waitreason is
3549 // set by park_m's caller.
3550 casgstatus(gp, _Grunning, _Gwaiting)
3553 if fn := mp.waitunlockf; fn != nil {
3554 ok := fn(gp, mp.waitlock)
3555 mp.waitunlockf = nil
3559 traceGoUnpark(gp, 2)
3561 casgstatus(gp, _Gwaiting, _Grunnable)
3562 execute(gp, true) // Schedule it back, never returns.
3568 func goschedImpl(gp *g) {
3569 status := readgstatus(gp)
3570 if status&^_Gscan != _Grunning {
3572 throw("bad g status")
3574 casgstatus(gp, _Grunning, _Grunnable)
3583 // Gosched continuation on g0.
3584 func gosched_m(gp *g) {
3591 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3592 func goschedguarded_m(gp *g) {
3594 if !canPreemptM(gp.m) {
3595 gogo(&gp.sched) // never return
3604 func gopreempt_m(gp *g) {
3611 // preemptPark parks gp and puts it in _Gpreempted.
3614 func preemptPark(gp *g) {
3616 traceGoPark(traceEvGoBlock, 0)
3618 status := readgstatus(gp)
3619 if status&^_Gscan != _Grunning {
3621 throw("bad g status")
3624 if gp.asyncSafePoint {
3625 // Double-check that async preemption does not
3626 // happen in SPWRITE assembly functions.
3627 // isAsyncSafePoint must exclude this case.
3628 f := findfunc(gp.sched.pc)
3630 throw("preempt at unknown pc")
3632 if f.flag&abi.FuncFlagSPWrite != 0 {
3633 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3634 throw("preempt SPWRITE")
3638 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3639 // be in _Grunning when we dropg because then we'd be running
3640 // without an M, but the moment we're in _Gpreempted,
3641 // something could claim this G before we've fully cleaned it
3642 // up. Hence, we set the scan bit to lock down further
3643 // transitions until we can dropg.
3644 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3646 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3650 // goyield is like Gosched, but it:
3651 // - emits a GoPreempt trace event instead of a GoSched trace event
3652 // - puts the current G on the runq of the current P instead of the globrunq
3658 func goyield_m(gp *g) {
3663 casgstatus(gp, _Grunning, _Grunnable)
3665 runqput(pp, gp, false)
3669 // Finishes execution of the current goroutine.
3680 // goexit continuation on g0.
3681 func goexit0(gp *g) {
3685 casgstatus(gp, _Grunning, _Gdead)
3686 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3687 if isSystemGoroutine(gp, false) {
3691 locked := gp.lockedm != 0
3694 gp.preemptStop = false
3695 gp.paniconfault = false
3696 gp._defer = nil // should be true already but just in case.
3697 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3699 gp.waitreason = waitReasonZero
3704 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3705 // Flush assist credit to the global pool. This gives
3706 // better information to pacing if the application is
3707 // rapidly creating an exiting goroutines.
3708 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3709 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3710 gcController.bgScanCredit.Add(scanCredit)
3711 gp.gcAssistBytes = 0
3716 if GOARCH == "wasm" { // no threads yet on wasm
3718 schedule() // never returns
3721 if mp.lockedInt != 0 {
3722 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3723 throw("internal lockOSThread error")
3727 // The goroutine may have locked this thread because
3728 // it put it in an unusual kernel state. Kill it
3729 // rather than returning it to the thread pool.
3731 // Return to mstart, which will release the P and exit
3733 if GOOS != "plan9" { // See golang.org/issue/22227.
3736 // Clear lockedExt on plan9 since we may end up re-using
3744 // save updates getg().sched to refer to pc and sp so that a following
3745 // gogo will restore pc and sp.
3747 // save must not have write barriers because invoking a write barrier
3748 // can clobber getg().sched.
3751 //go:nowritebarrierrec
3752 func save(pc, sp uintptr) {
3755 if gp == gp.m.g0 || gp == gp.m.gsignal {
3756 // m.g0.sched is special and must describe the context
3757 // for exiting the thread. mstart1 writes to it directly.
3758 // m.gsignal.sched should not be used at all.
3759 // This check makes sure save calls do not accidentally
3760 // run in contexts where they'd write to system g's.
3761 throw("save on system g not allowed")
3768 // We need to ensure ctxt is zero, but can't have a write
3769 // barrier here. However, it should always already be zero.
3771 if gp.sched.ctxt != nil {
3776 // The goroutine g is about to enter a system call.
3777 // Record that it's not using the cpu anymore.
3778 // This is called only from the go syscall library and cgocall,
3779 // not from the low-level system calls used by the runtime.
3781 // Entersyscall cannot split the stack: the save must
3782 // make g->sched refer to the caller's stack segment, because
3783 // entersyscall is going to return immediately after.
3785 // Nothing entersyscall calls can split the stack either.
3786 // We cannot safely move the stack during an active call to syscall,
3787 // because we do not know which of the uintptr arguments are
3788 // really pointers (back into the stack).
3789 // In practice, this means that we make the fast path run through
3790 // entersyscall doing no-split things, and the slow path has to use systemstack
3791 // to run bigger things on the system stack.
3793 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3794 // saved SP and PC are restored. This is needed when exitsyscall will be called
3795 // from a function further up in the call stack than the parent, as g->syscallsp
3796 // must always point to a valid stack frame. entersyscall below is the normal
3797 // entry point for syscalls, which obtains the SP and PC from the caller.
3800 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3801 // If the syscall does not block, that is it, we do not emit any other events.
3802 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3803 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3804 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3805 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3806 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3807 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3808 // and we wait for the increment before emitting traceGoSysExit.
3809 // Note that the increment is done even if tracing is not enabled,
3810 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3813 func reentersyscall(pc, sp uintptr) {
3816 // Disable preemption because during this function g is in Gsyscall status,
3817 // but can have inconsistent g->sched, do not let GC observe it.
3820 // Entersyscall must not call any function that might split/grow the stack.
3821 // (See details in comment above.)
3822 // Catch calls that might, by replacing the stack guard with something that
3823 // will trip any stack check and leaving a flag to tell newstack to die.
3824 gp.stackguard0 = stackPreempt
3825 gp.throwsplit = true
3827 // Leave SP around for GC and traceback.
3831 casgstatus(gp, _Grunning, _Gsyscall)
3832 if staticLockRanking {
3833 // When doing static lock ranking casgstatus can call
3834 // systemstack which clobbers g.sched.
3837 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3838 systemstack(func() {
3839 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3840 throw("entersyscall")
3845 systemstack(traceGoSysCall)
3846 // systemstack itself clobbers g.sched.{pc,sp} and we might
3847 // need them later when the G is genuinely blocked in a
3852 if sched.sysmonwait.Load() {
3853 systemstack(entersyscall_sysmon)
3857 if gp.m.p.ptr().runSafePointFn != 0 {
3858 // runSafePointFn may stack split if run on this stack
3859 systemstack(runSafePointFn)
3863 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3864 gp.sysblocktraced = true
3869 atomic.Store(&pp.status, _Psyscall)
3870 if sched.gcwaiting.Load() {
3871 systemstack(entersyscall_gcwait)
3878 // Standard syscall entry used by the go syscall library and normal cgo calls.
3880 // This is exported via linkname to assembly in the syscall package and x/sys.
3883 //go:linkname entersyscall
3884 func entersyscall() {
3885 reentersyscall(getcallerpc(), getcallersp())
3888 func entersyscall_sysmon() {
3890 if sched.sysmonwait.Load() {
3891 sched.sysmonwait.Store(false)
3892 notewakeup(&sched.sysmonnote)
3897 func entersyscall_gcwait() {
3899 pp := gp.m.oldp.ptr()
3902 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3908 if sched.stopwait--; sched.stopwait == 0 {
3909 notewakeup(&sched.stopnote)
3915 // The same as entersyscall(), but with a hint that the syscall is blocking.
3918 func entersyscallblock() {
3921 gp.m.locks++ // see comment in entersyscall
3922 gp.throwsplit = true
3923 gp.stackguard0 = stackPreempt // see comment in entersyscall
3924 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3925 gp.sysblocktraced = true
3926 gp.m.p.ptr().syscalltick++
3928 // Leave SP around for GC and traceback.
3932 gp.syscallsp = gp.sched.sp
3933 gp.syscallpc = gp.sched.pc
3934 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3938 systemstack(func() {
3939 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3940 throw("entersyscallblock")
3943 casgstatus(gp, _Grunning, _Gsyscall)
3944 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3945 systemstack(func() {
3946 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3947 throw("entersyscallblock")
3951 systemstack(entersyscallblock_handoff)
3953 // Resave for traceback during blocked call.
3954 save(getcallerpc(), getcallersp())
3959 func entersyscallblock_handoff() {
3962 traceGoSysBlock(getg().m.p.ptr())
3964 handoffp(releasep())
3967 // The goroutine g exited its system call.
3968 // Arrange for it to run on a cpu again.
3969 // This is called only from the go syscall library, not
3970 // from the low-level system calls used by the runtime.
3972 // Write barriers are not allowed because our P may have been stolen.
3974 // This is exported via linkname to assembly in the syscall package.
3977 //go:nowritebarrierrec
3978 //go:linkname exitsyscall
3979 func exitsyscall() {
3982 gp.m.locks++ // see comment in entersyscall
3983 if getcallersp() > gp.syscallsp {
3984 throw("exitsyscall: syscall frame is no longer valid")
3988 oldp := gp.m.oldp.ptr()
3990 if exitsyscallfast(oldp) {
3991 // When exitsyscallfast returns success, we have a P so can now use
3993 if goroutineProfile.active {
3994 // Make sure that gp has had its stack written out to the goroutine
3995 // profile, exactly as it was when the goroutine profiler first
3996 // stopped the world.
3997 systemstack(func() {
3998 tryRecordGoroutineProfileWB(gp)
4002 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4003 systemstack(traceGoStart)
4006 // There's a cpu for us, so we can run.
4007 gp.m.p.ptr().syscalltick++
4008 // We need to cas the status and scan before resuming...
4009 casgstatus(gp, _Gsyscall, _Grunning)
4011 // Garbage collector isn't running (since we are),
4012 // so okay to clear syscallsp.
4016 // restore the preemption request in case we've cleared it in newstack
4017 gp.stackguard0 = stackPreempt
4019 // otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
4020 gp.stackguard0 = gp.stack.lo + stackGuard
4022 gp.throwsplit = false
4024 if sched.disable.user && !schedEnabled(gp) {
4025 // Scheduling of this goroutine is disabled.
4034 // Wait till traceGoSysBlock event is emitted.
4035 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4036 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
4039 // We can't trace syscall exit right now because we don't have a P.
4040 // Tracing code can invoke write barriers that cannot run without a P.
4041 // So instead we remember the syscall exit time and emit the event
4042 // in execute when we have a P.
4043 gp.sysexitticks = cputicks()
4048 // Call the scheduler.
4051 // Scheduler returned, so we're allowed to run now.
4052 // Delete the syscallsp information that we left for
4053 // the garbage collector during the system call.
4054 // Must wait until now because until gosched returns
4055 // we don't know for sure that the garbage collector
4058 gp.m.p.ptr().syscalltick++
4059 gp.throwsplit = false
4063 func exitsyscallfast(oldp *p) bool {
4066 // Freezetheworld sets stopwait but does not retake P's.
4067 if sched.stopwait == freezeStopWait {
4071 // Try to re-acquire the last P.
4072 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4073 // There's a cpu for us, so we can run.
4075 exitsyscallfast_reacquired()
4079 // Try to get any other idle P.
4080 if sched.pidle != 0 {
4082 systemstack(func() {
4083 ok = exitsyscallfast_pidle()
4084 if ok && trace.enabled {
4086 // Wait till traceGoSysBlock event is emitted.
4087 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4088 for oldp.syscalltick == gp.m.syscalltick {
4102 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4103 // has successfully reacquired the P it was running on before the
4107 func exitsyscallfast_reacquired() {
4109 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4111 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4112 // traceGoSysBlock for this syscall was already emitted,
4113 // but here we effectively retake the p from the new syscall running on the same p.
4114 systemstack(func() {
4115 // Denote blocking of the new syscall.
4116 traceGoSysBlock(gp.m.p.ptr())
4117 // Denote completion of the current syscall.
4121 gp.m.p.ptr().syscalltick++
4125 func exitsyscallfast_pidle() bool {
4127 pp, _ := pidleget(0)
4128 if pp != nil && sched.sysmonwait.Load() {
4129 sched.sysmonwait.Store(false)
4130 notewakeup(&sched.sysmonnote)
4140 // exitsyscall slow path on g0.
4141 // Failed to acquire P, enqueue gp as runnable.
4143 // Called via mcall, so gp is the calling g from this M.
4145 //go:nowritebarrierrec
4146 func exitsyscall0(gp *g) {
4147 casgstatus(gp, _Gsyscall, _Grunnable)
4151 if schedEnabled(gp) {
4158 // Below, we stoplockedm if gp is locked. globrunqput releases
4159 // ownership of gp, so we must check if gp is locked prior to
4160 // committing the release by unlocking sched.lock, otherwise we
4161 // could race with another M transitioning gp from unlocked to
4163 locked = gp.lockedm != 0
4164 } else if sched.sysmonwait.Load() {
4165 sched.sysmonwait.Store(false)
4166 notewakeup(&sched.sysmonnote)
4171 execute(gp, false) // Never returns.
4174 // Wait until another thread schedules gp and so m again.
4176 // N.B. lockedm must be this M, as this g was running on this M
4177 // before entersyscall.
4179 execute(gp, false) // Never returns.
4182 schedule() // Never returns.
4185 // Called from syscall package before fork.
4187 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4189 func syscall_runtime_BeforeFork() {
4192 // Block signals during a fork, so that the child does not run
4193 // a signal handler before exec if a signal is sent to the process
4194 // group. See issue #18600.
4196 sigsave(&gp.m.sigmask)
4199 // This function is called before fork in syscall package.
4200 // Code between fork and exec must not allocate memory nor even try to grow stack.
4201 // Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
4202 // runtime_AfterFork will undo this in parent process, but not in child.
4203 gp.stackguard0 = stackFork
4206 // Called from syscall package after fork in parent.
4208 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4210 func syscall_runtime_AfterFork() {
4213 // See the comments in beforefork.
4214 gp.stackguard0 = gp.stack.lo + stackGuard
4216 msigrestore(gp.m.sigmask)
4221 // inForkedChild is true while manipulating signals in the child process.
4222 // This is used to avoid calling libc functions in case we are using vfork.
4223 var inForkedChild bool
4225 // Called from syscall package after fork in child.
4226 // It resets non-sigignored signals to the default handler, and
4227 // restores the signal mask in preparation for the exec.
4229 // Because this might be called during a vfork, and therefore may be
4230 // temporarily sharing address space with the parent process, this must
4231 // not change any global variables or calling into C code that may do so.
4233 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4235 //go:nowritebarrierrec
4236 func syscall_runtime_AfterForkInChild() {
4237 // It's OK to change the global variable inForkedChild here
4238 // because we are going to change it back. There is no race here,
4239 // because if we are sharing address space with the parent process,
4240 // then the parent process can not be running concurrently.
4241 inForkedChild = true
4243 clearSignalHandlers()
4245 // When we are the child we are the only thread running,
4246 // so we know that nothing else has changed gp.m.sigmask.
4247 msigrestore(getg().m.sigmask)
4249 inForkedChild = false
4252 // pendingPreemptSignals is the number of preemption signals
4253 // that have been sent but not received. This is only used on Darwin.
4255 var pendingPreemptSignals atomic.Int32
4257 // Called from syscall package before Exec.
4259 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4260 func syscall_runtime_BeforeExec() {
4261 // Prevent thread creation during exec.
4264 // On Darwin, wait for all pending preemption signals to
4265 // be received. See issue #41702.
4266 if GOOS == "darwin" || GOOS == "ios" {
4267 for pendingPreemptSignals.Load() > 0 {
4273 // Called from syscall package after Exec.
4275 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4276 func syscall_runtime_AfterExec() {
4280 // Allocate a new g, with a stack big enough for stacksize bytes.
4281 func malg(stacksize int32) *g {
4284 stacksize = round2(stackSystem + stacksize)
4285 systemstack(func() {
4286 newg.stack = stackalloc(uint32(stacksize))
4288 newg.stackguard0 = newg.stack.lo + stackGuard
4289 newg.stackguard1 = ^uintptr(0)
4290 // Clear the bottom word of the stack. We record g
4291 // there on gsignal stack during VDSO on ARM and ARM64.
4292 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4297 // Create a new g running fn.
4298 // Put it on the queue of g's waiting to run.
4299 // The compiler turns a go statement into a call to this.
4300 func newproc(fn *funcval) {
4303 systemstack(func() {
4304 newg := newproc1(fn, gp, pc)
4306 pp := getg().m.p.ptr()
4307 runqput(pp, newg, true)
4315 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4316 // address of the go statement that created this. The caller is responsible
4317 // for adding the new g to the scheduler.
4318 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4320 fatal("go of nil func value")
4323 mp := acquirem() // disable preemption because we hold M and P in local vars.
4327 newg = malg(stackMin)
4328 casgstatus(newg, _Gidle, _Gdead)
4329 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4331 if newg.stack.hi == 0 {
4332 throw("newproc1: newg missing stack")
4335 if readgstatus(newg) != _Gdead {
4336 throw("newproc1: new g is not Gdead")
4339 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4340 totalSize = alignUp(totalSize, sys.StackAlign)
4341 sp := newg.stack.hi - totalSize
4345 *(*uintptr)(unsafe.Pointer(sp)) = 0
4347 spArg += sys.MinFrameSize
4350 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4353 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4354 newg.sched.g = guintptr(unsafe.Pointer(newg))
4355 gostartcallfn(&newg.sched, fn)
4356 newg.parentGoid = callergp.goid
4357 newg.gopc = callerpc
4358 newg.ancestors = saveAncestors(callergp)
4359 newg.startpc = fn.fn
4360 if isSystemGoroutine(newg, false) {
4363 // Only user goroutines inherit pprof labels.
4365 newg.labels = mp.curg.labels
4367 if goroutineProfile.active {
4368 // A concurrent goroutine profile is running. It should include
4369 // exactly the set of goroutines that were alive when the goroutine
4370 // profiler first stopped the world. That does not include newg, so
4371 // mark it as not needing a profile before transitioning it from
4373 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4376 // Track initial transition?
4377 newg.trackingSeq = uint8(fastrand())
4378 if newg.trackingSeq%gTrackingPeriod == 0 {
4379 newg.tracking = true
4381 casgstatus(newg, _Gdead, _Grunnable)
4382 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4384 if pp.goidcache == pp.goidcacheend {
4385 // Sched.goidgen is the last allocated id,
4386 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4387 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4388 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4389 pp.goidcache -= _GoidCacheBatch - 1
4390 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4392 newg.goid = pp.goidcache
4395 newg.racectx = racegostart(callerpc)
4396 if newg.labels != nil {
4397 // See note in proflabel.go on labelSync's role in synchronizing
4398 // with the reads in the signal handler.
4399 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4403 traceGoCreate(newg, newg.startpc)
4410 // saveAncestors copies previous ancestors of the given caller g and
4411 // includes info for the current caller into a new set of tracebacks for
4412 // a g being created.
4413 func saveAncestors(callergp *g) *[]ancestorInfo {
4414 // Copy all prior info, except for the root goroutine (goid 0).
4415 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4418 var callerAncestors []ancestorInfo
4419 if callergp.ancestors != nil {
4420 callerAncestors = *callergp.ancestors
4422 n := int32(len(callerAncestors)) + 1
4423 if n > debug.tracebackancestors {
4424 n = debug.tracebackancestors
4426 ancestors := make([]ancestorInfo, n)
4427 copy(ancestors[1:], callerAncestors)
4429 var pcs [tracebackInnerFrames]uintptr
4430 npcs := gcallers(callergp, 0, pcs[:])
4431 ipcs := make([]uintptr, npcs)
4433 ancestors[0] = ancestorInfo{
4435 goid: callergp.goid,
4436 gopc: callergp.gopc,
4439 ancestorsp := new([]ancestorInfo)
4440 *ancestorsp = ancestors
4444 // Put on gfree list.
4445 // If local list is too long, transfer a batch to the global list.
4446 func gfput(pp *p, gp *g) {
4447 if readgstatus(gp) != _Gdead {
4448 throw("gfput: bad status (not Gdead)")
4451 stksize := gp.stack.hi - gp.stack.lo
4453 if stksize != uintptr(startingStackSize) {
4454 // non-standard stack size - free it.
4463 if pp.gFree.n >= 64 {
4469 for pp.gFree.n >= 32 {
4470 gp := pp.gFree.pop()
4472 if gp.stack.lo == 0 {
4479 lock(&sched.gFree.lock)
4480 sched.gFree.noStack.pushAll(noStackQ)
4481 sched.gFree.stack.pushAll(stackQ)
4482 sched.gFree.n += inc
4483 unlock(&sched.gFree.lock)
4487 // Get from gfree list.
4488 // If local list is empty, grab a batch from global list.
4489 func gfget(pp *p) *g {
4491 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4492 lock(&sched.gFree.lock)
4493 // Move a batch of free Gs to the P.
4494 for pp.gFree.n < 32 {
4495 // Prefer Gs with stacks.
4496 gp := sched.gFree.stack.pop()
4498 gp = sched.gFree.noStack.pop()
4507 unlock(&sched.gFree.lock)
4510 gp := pp.gFree.pop()
4515 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4516 // Deallocate old stack. We kept it in gfput because it was the
4517 // right size when the goroutine was put on the free list, but
4518 // the right size has changed since then.
4519 systemstack(func() {
4526 if gp.stack.lo == 0 {
4527 // Stack was deallocated in gfput or just above. Allocate a new one.
4528 systemstack(func() {
4529 gp.stack = stackalloc(startingStackSize)
4531 gp.stackguard0 = gp.stack.lo + stackGuard
4534 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4537 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4540 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4546 // Purge all cached G's from gfree list to the global list.
4547 func gfpurge(pp *p) {
4553 for !pp.gFree.empty() {
4554 gp := pp.gFree.pop()
4556 if gp.stack.lo == 0 {
4563 lock(&sched.gFree.lock)
4564 sched.gFree.noStack.pushAll(noStackQ)
4565 sched.gFree.stack.pushAll(stackQ)
4566 sched.gFree.n += inc
4567 unlock(&sched.gFree.lock)
4570 // Breakpoint executes a breakpoint trap.
4575 // dolockOSThread is called by LockOSThread and lockOSThread below
4576 // after they modify m.locked. Do not allow preemption during this call,
4577 // or else the m might be different in this function than in the caller.
4580 func dolockOSThread() {
4581 if GOARCH == "wasm" {
4582 return // no threads on wasm yet
4585 gp.m.lockedg.set(gp)
4586 gp.lockedm.set(gp.m)
4589 // LockOSThread wires the calling goroutine to its current operating system thread.
4590 // The calling goroutine will always execute in that thread,
4591 // and no other goroutine will execute in it,
4592 // until the calling goroutine has made as many calls to
4593 // UnlockOSThread as to LockOSThread.
4594 // If the calling goroutine exits without unlocking the thread,
4595 // the thread will be terminated.
4597 // All init functions are run on the startup thread. Calling LockOSThread
4598 // from an init function will cause the main function to be invoked on
4601 // A goroutine should call LockOSThread before calling OS services or
4602 // non-Go library functions that depend on per-thread state.
4605 func LockOSThread() {
4606 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4607 // If we need to start a new thread from the locked
4608 // thread, we need the template thread. Start it now
4609 // while we're in a known-good state.
4610 startTemplateThread()
4614 if gp.m.lockedExt == 0 {
4616 panic("LockOSThread nesting overflow")
4622 func lockOSThread() {
4623 getg().m.lockedInt++
4627 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4628 // after they update m->locked. Do not allow preemption during this call,
4629 // or else the m might be in different in this function than in the caller.
4632 func dounlockOSThread() {
4633 if GOARCH == "wasm" {
4634 return // no threads on wasm yet
4637 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4644 // UnlockOSThread undoes an earlier call to LockOSThread.
4645 // If this drops the number of active LockOSThread calls on the
4646 // calling goroutine to zero, it unwires the calling goroutine from
4647 // its fixed operating system thread.
4648 // If there are no active LockOSThread calls, this is a no-op.
4650 // Before calling UnlockOSThread, the caller must ensure that the OS
4651 // thread is suitable for running other goroutines. If the caller made
4652 // any permanent changes to the state of the thread that would affect
4653 // other goroutines, it should not call this function and thus leave
4654 // the goroutine locked to the OS thread until the goroutine (and
4655 // hence the thread) exits.
4658 func UnlockOSThread() {
4660 if gp.m.lockedExt == 0 {
4668 func unlockOSThread() {
4670 if gp.m.lockedInt == 0 {
4671 systemstack(badunlockosthread)
4677 func badunlockosthread() {
4678 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4681 func gcount() int32 {
4682 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4683 for _, pp := range allp {
4687 // All these variables can be changed concurrently, so the result can be inconsistent.
4688 // But at least the current goroutine is running.
4695 func mcount() int32 {
4696 return int32(sched.mnext - sched.nmfreed)
4700 signalLock atomic.Uint32
4702 // Must hold signalLock to write. Reads may be lock-free, but
4703 // signalLock should be taken to synchronize with changes.
4707 func _System() { _System() }
4708 func _ExternalCode() { _ExternalCode() }
4709 func _LostExternalCode() { _LostExternalCode() }
4710 func _GC() { _GC() }
4711 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4712 func _VDSO() { _VDSO() }
4714 // Called if we receive a SIGPROF signal.
4715 // Called by the signal handler, may run during STW.
4717 //go:nowritebarrierrec
4718 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4719 if prof.hz.Load() == 0 {
4723 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4724 // We must check this to avoid a deadlock between setcpuprofilerate
4725 // and the call to cpuprof.add, below.
4726 if mp != nil && mp.profilehz == 0 {
4730 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4731 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4732 // the critical section, it creates a deadlock (when writing the sample).
4733 // As a workaround, create a counter of SIGPROFs while in critical section
4734 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4735 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4736 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4737 if f := findfunc(pc); f.valid() {
4738 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4739 cpuprof.lostAtomic++
4743 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4744 // runtime/internal/atomic functions call into kernel
4745 // helpers on arm < 7. See
4746 // runtime/internal/atomic/sys_linux_arm.s.
4747 cpuprof.lostAtomic++
4752 // Profiling runs concurrently with GC, so it must not allocate.
4753 // Set a trap in case the code does allocate.
4754 // Note that on windows, one thread takes profiles of all the
4755 // other threads, so mp is usually not getg().m.
4756 // In fact mp may not even be stopped.
4757 // See golang.org/issue/17165.
4758 getg().m.mallocing++
4761 var stk [maxCPUProfStack]uintptr
4763 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4765 // Check cgoCallersUse to make sure that we are not
4766 // interrupting other code that is fiddling with
4767 // cgoCallers. We are running in a signal handler
4768 // with all signals blocked, so we don't have to worry
4769 // about any other code interrupting us.
4770 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4771 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4774 n += copy(stk[:], mp.cgoCallers[:cgoOff])
4775 mp.cgoCallers[0] = 0
4778 // Collect Go stack that leads to the cgo call.
4779 u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
4780 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4781 // Libcall, i.e. runtime syscall on windows.
4782 // Collect Go stack that leads to the call.
4783 u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
4784 } else if mp != nil && mp.vdsoSP != 0 {
4785 // VDSO call, e.g. nanotime1 on Linux.
4786 // Collect Go stack that leads to the call.
4787 u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
4789 u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
4791 n += tracebackPCs(&u, 0, stk[n:])
4794 // Normal traceback is impossible or has failed.
4795 // Account it against abstract "System" or "GC".
4798 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4799 } else if pc > firstmoduledata.etext {
4800 // "ExternalCode" is better than "etext".
4801 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4804 if mp.preemptoff != "" {
4805 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4807 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4811 if prof.hz.Load() != 0 {
4812 // Note: it can happen on Windows that we interrupted a system thread
4813 // with no g, so gp could nil. The other nil checks are done out of
4814 // caution, but not expected to be nil in practice.
4815 var tagPtr *unsafe.Pointer
4816 if gp != nil && gp.m != nil && gp.m.curg != nil {
4817 tagPtr = &gp.m.curg.labels
4819 cpuprof.add(tagPtr, stk[:n])
4823 if gp != nil && gp.m != nil {
4824 if gp.m.curg != nil {
4829 traceCPUSample(gprof, pp, stk[:n])
4831 getg().m.mallocing--
4834 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4835 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4836 func setcpuprofilerate(hz int32) {
4837 // Force sane arguments.
4842 // Disable preemption, otherwise we can be rescheduled to another thread
4843 // that has profiling enabled.
4847 // Stop profiler on this thread so that it is safe to lock prof.
4848 // if a profiling signal came in while we had prof locked,
4849 // it would deadlock.
4850 setThreadCPUProfiler(0)
4852 for !prof.signalLock.CompareAndSwap(0, 1) {
4855 if prof.hz.Load() != hz {
4856 setProcessCPUProfiler(hz)
4859 prof.signalLock.Store(0)
4862 sched.profilehz = hz
4866 setThreadCPUProfiler(hz)
4872 // init initializes pp, which may be a freshly allocated p or a
4873 // previously destroyed p, and transitions it to status _Pgcstop.
4874 func (pp *p) init(id int32) {
4876 pp.status = _Pgcstop
4877 pp.sudogcache = pp.sudogbuf[:0]
4878 pp.deferpool = pp.deferpoolbuf[:0]
4880 if pp.mcache == nil {
4883 throw("missing mcache?")
4885 // Use the bootstrap mcache0. Only one P will get
4886 // mcache0: the one with ID 0.
4889 pp.mcache = allocmcache()
4892 if raceenabled && pp.raceprocctx == 0 {
4894 pp.raceprocctx = raceprocctx0
4895 raceprocctx0 = 0 // bootstrap
4897 pp.raceprocctx = raceproccreate()
4900 lockInit(&pp.timersLock, lockRankTimers)
4902 // This P may get timers when it starts running. Set the mask here
4903 // since the P may not go through pidleget (notably P 0 on startup).
4905 // Similarly, we may not go through pidleget before this P starts
4906 // running if it is P 0 on startup.
4910 // destroy releases all of the resources associated with pp and
4911 // transitions it to status _Pdead.
4913 // sched.lock must be held and the world must be stopped.
4914 func (pp *p) destroy() {
4915 assertLockHeld(&sched.lock)
4916 assertWorldStopped()
4918 // Move all runnable goroutines to the global queue
4919 for pp.runqhead != pp.runqtail {
4920 // Pop from tail of local queue
4922 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4923 // Push onto head of global queue
4926 if pp.runnext != 0 {
4927 globrunqputhead(pp.runnext.ptr())
4930 if len(pp.timers) > 0 {
4931 plocal := getg().m.p.ptr()
4932 // The world is stopped, but we acquire timersLock to
4933 // protect against sysmon calling timeSleepUntil.
4934 // This is the only case where we hold the timersLock of
4935 // more than one P, so there are no deadlock concerns.
4936 lock(&plocal.timersLock)
4937 lock(&pp.timersLock)
4938 moveTimers(plocal, pp.timers)
4940 pp.numTimers.Store(0)
4941 pp.deletedTimers.Store(0)
4942 pp.timer0When.Store(0)
4943 unlock(&pp.timersLock)
4944 unlock(&plocal.timersLock)
4946 // Flush p's write barrier buffer.
4947 if gcphase != _GCoff {
4951 for i := range pp.sudogbuf {
4952 pp.sudogbuf[i] = nil
4954 pp.sudogcache = pp.sudogbuf[:0]
4955 for j := range pp.deferpoolbuf {
4956 pp.deferpoolbuf[j] = nil
4958 pp.deferpool = pp.deferpoolbuf[:0]
4959 systemstack(func() {
4960 for i := 0; i < pp.mspancache.len; i++ {
4961 // Safe to call since the world is stopped.
4962 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4964 pp.mspancache.len = 0
4966 pp.pcache.flush(&mheap_.pages)
4967 unlock(&mheap_.lock)
4969 freemcache(pp.mcache)
4974 if pp.timerRaceCtx != 0 {
4975 // The race detector code uses a callback to fetch
4976 // the proc context, so arrange for that callback
4977 // to see the right thing.
4978 // This hack only works because we are the only
4984 racectxend(pp.timerRaceCtx)
4989 raceprocdestroy(pp.raceprocctx)
4996 // Change number of processors.
4998 // sched.lock must be held, and the world must be stopped.
5000 // gcworkbufs must not be being modified by either the GC or the write barrier
5001 // code, so the GC must not be running if the number of Ps actually changes.
5003 // Returns list of Ps with local work, they need to be scheduled by the caller.
5004 func procresize(nprocs int32) *p {
5005 assertLockHeld(&sched.lock)
5006 assertWorldStopped()
5009 if old < 0 || nprocs <= 0 {
5010 throw("procresize: invalid arg")
5013 traceGomaxprocs(nprocs)
5016 // update statistics
5018 if sched.procresizetime != 0 {
5019 sched.totaltime += int64(old) * (now - sched.procresizetime)
5021 sched.procresizetime = now
5023 maskWords := (nprocs + 31) / 32
5025 // Grow allp if necessary.
5026 if nprocs > int32(len(allp)) {
5027 // Synchronize with retake, which could be running
5028 // concurrently since it doesn't run on a P.
5030 if nprocs <= int32(cap(allp)) {
5031 allp = allp[:nprocs]
5033 nallp := make([]*p, nprocs)
5034 // Copy everything up to allp's cap so we
5035 // never lose old allocated Ps.
5036 copy(nallp, allp[:cap(allp)])
5040 if maskWords <= int32(cap(idlepMask)) {
5041 idlepMask = idlepMask[:maskWords]
5042 timerpMask = timerpMask[:maskWords]
5044 nidlepMask := make([]uint32, maskWords)
5045 // No need to copy beyond len, old Ps are irrelevant.
5046 copy(nidlepMask, idlepMask)
5047 idlepMask = nidlepMask
5049 ntimerpMask := make([]uint32, maskWords)
5050 copy(ntimerpMask, timerpMask)
5051 timerpMask = ntimerpMask
5056 // initialize new P's
5057 for i := old; i < nprocs; i++ {
5063 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5067 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5068 // continue to use the current P
5069 gp.m.p.ptr().status = _Prunning
5070 gp.m.p.ptr().mcache.prepareForSweep()
5072 // release the current P and acquire allp[0].
5074 // We must do this before destroying our current P
5075 // because p.destroy itself has write barriers, so we
5076 // need to do that from a valid P.
5079 // Pretend that we were descheduled
5080 // and then scheduled again to keep
5083 traceProcStop(gp.m.p.ptr())
5097 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5100 // release resources from unused P's
5101 for i := nprocs; i < old; i++ {
5104 // can't free P itself because it can be referenced by an M in syscall
5108 if int32(len(allp)) != nprocs {
5110 allp = allp[:nprocs]
5111 idlepMask = idlepMask[:maskWords]
5112 timerpMask = timerpMask[:maskWords]
5117 for i := nprocs - 1; i >= 0; i-- {
5119 if gp.m.p.ptr() == pp {
5127 pp.link.set(runnablePs)
5131 stealOrder.reset(uint32(nprocs))
5132 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5133 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5135 // Notify the limiter that the amount of procs has changed.
5136 gcCPULimiter.resetCapacity(now, nprocs)
5141 // Associate p and the current m.
5143 // This function is allowed to have write barriers even if the caller
5144 // isn't because it immediately acquires pp.
5146 //go:yeswritebarrierrec
5147 func acquirep(pp *p) {
5148 // Do the part that isn't allowed to have write barriers.
5151 // Have p; write barriers now allowed.
5153 // Perform deferred mcache flush before this P can allocate
5154 // from a potentially stale mcache.
5155 pp.mcache.prepareForSweep()
5162 // wirep is the first step of acquirep, which actually associates the
5163 // current M to pp. This is broken out so we can disallow write
5164 // barriers for this part, since we don't yet have a P.
5166 //go:nowritebarrierrec
5172 throw("wirep: already in go")
5174 if pp.m != 0 || pp.status != _Pidle {
5179 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5180 throw("wirep: invalid p state")
5184 pp.status = _Prunning
5187 // Disassociate p and the current m.
5188 func releasep() *p {
5192 throw("releasep: invalid arg")
5195 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5196 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5197 throw("releasep: invalid p state")
5200 traceProcStop(gp.m.p.ptr())
5208 func incidlelocked(v int32) {
5210 sched.nmidlelocked += v
5217 // Check for deadlock situation.
5218 // The check is based on number of running M's, if 0 -> deadlock.
5219 // sched.lock must be held.
5221 assertLockHeld(&sched.lock)
5223 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5224 // there are no running goroutines. The calling program is
5225 // assumed to be running.
5226 if islibrary || isarchive {
5230 // If we are dying because of a signal caught on an already idle thread,
5231 // freezetheworld will cause all running threads to block.
5232 // And runtime will essentially enter into deadlock state,
5233 // except that there is a thread that will call exit soon.
5234 if panicking.Load() > 0 {
5238 // If we are not running under cgo, but we have an extra M then account
5239 // for it. (It is possible to have an extra M on Windows without cgo to
5240 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5243 if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
5247 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5252 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5254 throw("checkdead: inconsistent counts")
5258 forEachG(func(gp *g) {
5259 if isSystemGoroutine(gp, false) {
5262 s := readgstatus(gp)
5263 switch s &^ _Gscan {
5270 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5272 throw("checkdead: runnable g")
5275 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5276 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5277 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5280 // Maybe jump time forward for playground.
5282 if when := timeSleepUntil(); when < maxWhen {
5285 // Start an M to steal the timer.
5286 pp, _ := pidleget(faketime)
5288 // There should always be a free P since
5289 // nothing is running.
5291 throw("checkdead: no p for timer")
5295 // There should always be a free M since
5296 // nothing is running.
5298 throw("checkdead: no m for timer")
5300 // M must be spinning to steal. We set this to be
5301 // explicit, but since this is the only M it would
5302 // become spinning on its own anyways.
5303 sched.nmspinning.Add(1)
5306 notewakeup(&mp.park)
5311 // There are no goroutines running, so we can look at the P's.
5312 for _, pp := range allp {
5313 if len(pp.timers) > 0 {
5318 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5319 fatal("all goroutines are asleep - deadlock!")
5322 // forcegcperiod is the maximum time in nanoseconds between garbage
5323 // collections. If we go this long without a garbage collection, one
5324 // is forced to run.
5326 // This is a variable for testing purposes. It normally doesn't change.
5327 var forcegcperiod int64 = 2 * 60 * 1e9
5329 // needSysmonWorkaround is true if the workaround for
5330 // golang.org/issue/42515 is needed on NetBSD.
5331 var needSysmonWorkaround bool = false
5333 // Always runs without a P, so write barriers are not allowed.
5335 //go:nowritebarrierrec
5342 lasttrace := int64(0)
5343 idle := 0 // how many cycles in succession we had not wokeup somebody
5347 if idle == 0 { // start with 20us sleep...
5349 } else if idle > 50 { // start doubling the sleep after 1ms...
5352 if delay > 10*1000 { // up to 10ms
5357 // sysmon should not enter deep sleep if schedtrace is enabled so that
5358 // it can print that information at the right time.
5360 // It should also not enter deep sleep if there are any active P's so
5361 // that it can retake P's from syscalls, preempt long running G's, and
5362 // poll the network if all P's are busy for long stretches.
5364 // It should wakeup from deep sleep if any P's become active either due
5365 // to exiting a syscall or waking up due to a timer expiring so that it
5366 // can resume performing those duties. If it wakes from a syscall it
5367 // resets idle and delay as a bet that since it had retaken a P from a
5368 // syscall before, it may need to do it again shortly after the
5369 // application starts work again. It does not reset idle when waking
5370 // from a timer to avoid adding system load to applications that spend
5371 // most of their time sleeping.
5373 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5375 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5376 syscallWake := false
5377 next := timeSleepUntil()
5379 sched.sysmonwait.Store(true)
5381 // Make wake-up period small enough
5382 // for the sampling to be correct.
5383 sleep := forcegcperiod / 2
5384 if next-now < sleep {
5387 shouldRelax := sleep >= osRelaxMinNS
5391 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5396 sched.sysmonwait.Store(false)
5397 noteclear(&sched.sysmonnote)
5407 lock(&sched.sysmonlock)
5408 // Update now in case we blocked on sysmonnote or spent a long time
5409 // blocked on schedlock or sysmonlock above.
5412 // trigger libc interceptors if needed
5413 if *cgo_yield != nil {
5414 asmcgocall(*cgo_yield, nil)
5416 // poll network if not polled for more than 10ms
5417 lastpoll := sched.lastpoll.Load()
5418 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5419 sched.lastpoll.CompareAndSwap(lastpoll, now)
5420 list := netpoll(0) // non-blocking - returns list of goroutines
5422 // Need to decrement number of idle locked M's
5423 // (pretending that one more is running) before injectglist.
5424 // Otherwise it can lead to the following situation:
5425 // injectglist grabs all P's but before it starts M's to run the P's,
5426 // another M returns from syscall, finishes running its G,
5427 // observes that there is no work to do and no other running M's
5428 // and reports deadlock.
5434 if GOOS == "netbsd" && needSysmonWorkaround {
5435 // netpoll is responsible for waiting for timer
5436 // expiration, so we typically don't have to worry
5437 // about starting an M to service timers. (Note that
5438 // sleep for timeSleepUntil above simply ensures sysmon
5439 // starts running again when that timer expiration may
5440 // cause Go code to run again).
5442 // However, netbsd has a kernel bug that sometimes
5443 // misses netpollBreak wake-ups, which can lead to
5444 // unbounded delays servicing timers. If we detect this
5445 // overrun, then startm to get something to handle the
5448 // See issue 42515 and
5449 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5450 if next := timeSleepUntil(); next < now {
5451 startm(nil, false, false)
5454 if scavenger.sysmonWake.Load() != 0 {
5455 // Kick the scavenger awake if someone requested it.
5458 // retake P's blocked in syscalls
5459 // and preempt long running G's
5460 if retake(now) != 0 {
5465 // check if we need to force a GC
5466 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5468 forcegc.idle.Store(false)
5470 list.push(forcegc.g)
5472 unlock(&forcegc.lock)
5474 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5476 schedtrace(debug.scheddetail > 0)
5478 unlock(&sched.sysmonlock)
5482 type sysmontick struct {
5489 // forcePreemptNS is the time slice given to a G before it is
5491 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5493 func retake(now int64) uint32 {
5495 // Prevent allp slice changes. This lock will be completely
5496 // uncontended unless we're already stopping the world.
5498 // We can't use a range loop over allp because we may
5499 // temporarily drop the allpLock. Hence, we need to re-fetch
5500 // allp each time around the loop.
5501 for i := 0; i < len(allp); i++ {
5504 // This can happen if procresize has grown
5505 // allp but not yet created new Ps.
5508 pd := &pp.sysmontick
5511 if s == _Prunning || s == _Psyscall {
5512 // Preempt G if it's running for too long.
5513 t := int64(pp.schedtick)
5514 if int64(pd.schedtick) != t {
5515 pd.schedtick = uint32(t)
5517 } else if pd.schedwhen+forcePreemptNS <= now {
5519 // In case of syscall, preemptone() doesn't
5520 // work, because there is no M wired to P.
5525 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5526 t := int64(pp.syscalltick)
5527 if !sysretake && int64(pd.syscalltick) != t {
5528 pd.syscalltick = uint32(t)
5529 pd.syscallwhen = now
5532 // On the one hand we don't want to retake Ps if there is no other work to do,
5533 // but on the other hand we want to retake them eventually
5534 // because they can prevent the sysmon thread from deep sleep.
5535 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5538 // Drop allpLock so we can take sched.lock.
5540 // Need to decrement number of idle locked M's
5541 // (pretending that one more is running) before the CAS.
5542 // Otherwise the M from which we retake can exit the syscall,
5543 // increment nmidle and report deadlock.
5545 if atomic.Cas(&pp.status, s, _Pidle) {
5562 // Tell all goroutines that they have been preempted and they should stop.
5563 // This function is purely best-effort. It can fail to inform a goroutine if a
5564 // processor just started running it.
5565 // No locks need to be held.
5566 // Returns true if preemption request was issued to at least one goroutine.
5567 func preemptall() bool {
5569 for _, pp := range allp {
5570 if pp.status != _Prunning {
5580 // Tell the goroutine running on processor P to stop.
5581 // This function is purely best-effort. It can incorrectly fail to inform the
5582 // goroutine. It can inform the wrong goroutine. Even if it informs the
5583 // correct goroutine, that goroutine might ignore the request if it is
5584 // simultaneously executing newstack.
5585 // No lock needs to be held.
5586 // Returns true if preemption request was issued.
5587 // The actual preemption will happen at some point in the future
5588 // and will be indicated by the gp->status no longer being
5590 func preemptone(pp *p) bool {
5592 if mp == nil || mp == getg().m {
5596 if gp == nil || gp == mp.g0 {
5602 // Every call in a goroutine checks for stack overflow by
5603 // comparing the current stack pointer to gp->stackguard0.
5604 // Setting gp->stackguard0 to StackPreempt folds
5605 // preemption into the normal stack overflow check.
5606 gp.stackguard0 = stackPreempt
5608 // Request an async preemption of this P.
5609 if preemptMSupported && debug.asyncpreemptoff == 0 {
5619 func schedtrace(detailed bool) {
5626 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)
5628 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5630 // We must be careful while reading data from P's, M's and G's.
5631 // Even if we hold schedlock, most data can be changed concurrently.
5632 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5633 for i, pp := range allp {
5635 h := atomic.Load(&pp.runqhead)
5636 t := atomic.Load(&pp.runqtail)
5638 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5644 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5646 // In non-detailed mode format lengths of per-P run queues as:
5647 // [len1 len2 len3 len4]
5653 if i == len(allp)-1 {
5664 for mp := allm; mp != nil; mp = mp.alllink {
5666 print(" M", mp.id, ": p=")
5678 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5679 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5687 forEachG(func(gp *g) {
5688 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5695 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5705 // schedEnableUser enables or disables the scheduling of user
5708 // This does not stop already running user goroutines, so the caller
5709 // should first stop the world when disabling user goroutines.
5710 func schedEnableUser(enable bool) {
5712 if sched.disable.user == !enable {
5716 sched.disable.user = !enable
5718 n := sched.disable.n
5720 globrunqputbatch(&sched.disable.runnable, n)
5722 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5723 startm(nil, false, false)
5730 // schedEnabled reports whether gp should be scheduled. It returns
5731 // false is scheduling of gp is disabled.
5733 // sched.lock must be held.
5734 func schedEnabled(gp *g) bool {
5735 assertLockHeld(&sched.lock)
5737 if sched.disable.user {
5738 return isSystemGoroutine(gp, true)
5743 // Put mp on midle list.
5744 // sched.lock must be held.
5745 // May run during STW, so write barriers are not allowed.
5747 //go:nowritebarrierrec
5749 assertLockHeld(&sched.lock)
5751 mp.schedlink = sched.midle
5757 // Try to get an m from midle list.
5758 // sched.lock must be held.
5759 // May run during STW, so write barriers are not allowed.
5761 //go:nowritebarrierrec
5763 assertLockHeld(&sched.lock)
5765 mp := sched.midle.ptr()
5767 sched.midle = mp.schedlink
5773 // Put gp on the global runnable queue.
5774 // sched.lock must be held.
5775 // May run during STW, so write barriers are not allowed.
5777 //go:nowritebarrierrec
5778 func globrunqput(gp *g) {
5779 assertLockHeld(&sched.lock)
5781 sched.runq.pushBack(gp)
5785 // Put gp at the head of the global runnable queue.
5786 // sched.lock must be held.
5787 // May run during STW, so write barriers are not allowed.
5789 //go:nowritebarrierrec
5790 func globrunqputhead(gp *g) {
5791 assertLockHeld(&sched.lock)
5797 // Put a batch of runnable goroutines on the global runnable queue.
5798 // This clears *batch.
5799 // sched.lock must be held.
5800 // May run during STW, so write barriers are not allowed.
5802 //go:nowritebarrierrec
5803 func globrunqputbatch(batch *gQueue, n int32) {
5804 assertLockHeld(&sched.lock)
5806 sched.runq.pushBackAll(*batch)
5811 // Try get a batch of G's from the global runnable queue.
5812 // sched.lock must be held.
5813 func globrunqget(pp *p, max int32) *g {
5814 assertLockHeld(&sched.lock)
5816 if sched.runqsize == 0 {
5820 n := sched.runqsize/gomaxprocs + 1
5821 if n > sched.runqsize {
5824 if max > 0 && n > max {
5827 if n > int32(len(pp.runq))/2 {
5828 n = int32(len(pp.runq)) / 2
5833 gp := sched.runq.pop()
5836 gp1 := sched.runq.pop()
5837 runqput(pp, gp1, false)
5842 // pMask is an atomic bitstring with one bit per P.
5845 // read returns true if P id's bit is set.
5846 func (p pMask) read(id uint32) bool {
5848 mask := uint32(1) << (id % 32)
5849 return (atomic.Load(&p[word]) & mask) != 0
5852 // set sets P id's bit.
5853 func (p pMask) set(id int32) {
5855 mask := uint32(1) << (id % 32)
5856 atomic.Or(&p[word], mask)
5859 // clear clears P id's bit.
5860 func (p pMask) clear(id int32) {
5862 mask := uint32(1) << (id % 32)
5863 atomic.And(&p[word], ^mask)
5866 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5868 // Ideally, the timer mask would be kept immediately consistent on any timer
5869 // operations. Unfortunately, updating a shared global data structure in the
5870 // timer hot path adds too much overhead in applications frequently switching
5871 // between no timers and some timers.
5873 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5874 // running P (returned by pidleget) may add a timer at any time, so its mask
5875 // must be set. An idle P (passed to pidleput) cannot add new timers while
5876 // idle, so if it has no timers at that time, its mask may be cleared.
5878 // Thus, we get the following effects on timer-stealing in findrunnable:
5880 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5881 // (for work- or timer-stealing; this is the ideal case).
5882 // - Running Ps must always be checked.
5883 // - Idle Ps whose timers are stolen must continue to be checked until they run
5884 // again, even after timer expiration.
5886 // When the P starts running again, the mask should be set, as a timer may be
5887 // added at any time.
5889 // TODO(prattmic): Additional targeted updates may improve the above cases.
5890 // e.g., updating the mask when stealing a timer.
5891 func updateTimerPMask(pp *p) {
5892 if pp.numTimers.Load() > 0 {
5896 // Looks like there are no timers, however another P may transiently
5897 // decrement numTimers when handling a timerModified timer in
5898 // checkTimers. We must take timersLock to serialize with these changes.
5899 lock(&pp.timersLock)
5900 if pp.numTimers.Load() == 0 {
5901 timerpMask.clear(pp.id)
5903 unlock(&pp.timersLock)
5906 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5907 // to nanotime or zero. Returns now or the current time if now was zero.
5909 // This releases ownership of p. Once sched.lock is released it is no longer
5912 // sched.lock must be held.
5914 // May run during STW, so write barriers are not allowed.
5916 //go:nowritebarrierrec
5917 func pidleput(pp *p, now int64) int64 {
5918 assertLockHeld(&sched.lock)
5921 throw("pidleput: P has non-empty run queue")
5926 updateTimerPMask(pp) // clear if there are no timers.
5927 idlepMask.set(pp.id)
5928 pp.link = sched.pidle
5931 if !pp.limiterEvent.start(limiterEventIdle, now) {
5932 throw("must be able to track idle limiter event")
5937 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5939 // sched.lock must be held.
5941 // May run during STW, so write barriers are not allowed.
5943 //go:nowritebarrierrec
5944 func pidleget(now int64) (*p, int64) {
5945 assertLockHeld(&sched.lock)
5947 pp := sched.pidle.ptr()
5949 // Timer may get added at any time now.
5953 timerpMask.set(pp.id)
5954 idlepMask.clear(pp.id)
5955 sched.pidle = pp.link
5956 sched.npidle.Add(-1)
5957 pp.limiterEvent.stop(limiterEventIdle, now)
5962 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5963 // This is called by spinning Ms (or callers than need a spinning M) that have
5964 // found work. If no P is available, this must synchronized with non-spinning
5965 // Ms that may be preparing to drop their P without discovering this work.
5967 // sched.lock must be held.
5969 // May run during STW, so write barriers are not allowed.
5971 //go:nowritebarrierrec
5972 func pidlegetSpinning(now int64) (*p, int64) {
5973 assertLockHeld(&sched.lock)
5975 pp, now := pidleget(now)
5977 // See "Delicate dance" comment in findrunnable. We found work
5978 // that we cannot take, we must synchronize with non-spinning
5979 // Ms that may be preparing to drop their P.
5980 sched.needspinning.Store(1)
5987 // runqempty reports whether pp has no Gs on its local run queue.
5988 // It never returns true spuriously.
5989 func runqempty(pp *p) bool {
5990 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5991 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5992 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5993 // does not mean the queue is empty.
5995 head := atomic.Load(&pp.runqhead)
5996 tail := atomic.Load(&pp.runqtail)
5997 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5998 if tail == atomic.Load(&pp.runqtail) {
5999 return head == tail && runnext == 0
6004 // To shake out latent assumptions about scheduling order,
6005 // we introduce some randomness into scheduling decisions
6006 // when running with the race detector.
6007 // The need for this was made obvious by changing the
6008 // (deterministic) scheduling order in Go 1.5 and breaking
6009 // many poorly-written tests.
6010 // With the randomness here, as long as the tests pass
6011 // consistently with -race, they shouldn't have latent scheduling
6013 const randomizeScheduler = raceenabled
6015 // runqput tries to put g on the local runnable queue.
6016 // If next is false, runqput adds g to the tail of the runnable queue.
6017 // If next is true, runqput puts g in the pp.runnext slot.
6018 // If the run queue is full, runnext puts g on the global queue.
6019 // Executed only by the owner P.
6020 func runqput(pp *p, gp *g, next bool) {
6021 if randomizeScheduler && next && fastrandn(2) == 0 {
6027 oldnext := pp.runnext
6028 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
6034 // Kick the old runnext out to the regular run queue.
6039 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6041 if t-h < uint32(len(pp.runq)) {
6042 pp.runq[t%uint32(len(pp.runq))].set(gp)
6043 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
6046 if runqputslow(pp, gp, h, t) {
6049 // the queue is not full, now the put above must succeed
6053 // Put g and a batch of work from local runnable queue on global queue.
6054 // Executed only by the owner P.
6055 func runqputslow(pp *p, gp *g, h, t uint32) bool {
6056 var batch [len(pp.runq)/2 + 1]*g
6058 // First, grab a batch from local queue.
6061 if n != uint32(len(pp.runq)/2) {
6062 throw("runqputslow: queue is not full")
6064 for i := uint32(0); i < n; i++ {
6065 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6067 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6072 if randomizeScheduler {
6073 for i := uint32(1); i <= n; i++ {
6074 j := fastrandn(i + 1)
6075 batch[i], batch[j] = batch[j], batch[i]
6079 // Link the goroutines.
6080 for i := uint32(0); i < n; i++ {
6081 batch[i].schedlink.set(batch[i+1])
6084 q.head.set(batch[0])
6085 q.tail.set(batch[n])
6087 // Now put the batch on global queue.
6089 globrunqputbatch(&q, int32(n+1))
6094 // runqputbatch tries to put all the G's on q on the local runnable queue.
6095 // If the queue is full, they are put on the global queue; in that case
6096 // this will temporarily acquire the scheduler lock.
6097 // Executed only by the owner P.
6098 func runqputbatch(pp *p, q *gQueue, qsize int) {
6099 h := atomic.LoadAcq(&pp.runqhead)
6102 for !q.empty() && t-h < uint32(len(pp.runq)) {
6104 pp.runq[t%uint32(len(pp.runq))].set(gp)
6110 if randomizeScheduler {
6111 off := func(o uint32) uint32 {
6112 return (pp.runqtail + o) % uint32(len(pp.runq))
6114 for i := uint32(1); i < n; i++ {
6115 j := fastrandn(i + 1)
6116 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6120 atomic.StoreRel(&pp.runqtail, t)
6123 globrunqputbatch(q, int32(qsize))
6128 // Get g from local runnable queue.
6129 // If inheritTime is true, gp should inherit the remaining time in the
6130 // current time slice. Otherwise, it should start a new time slice.
6131 // Executed only by the owner P.
6132 func runqget(pp *p) (gp *g, inheritTime bool) {
6133 // If there's a runnext, it's the next G to run.
6135 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6136 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6137 // Hence, there's no need to retry this CAS if it fails.
6138 if next != 0 && pp.runnext.cas(next, 0) {
6139 return next.ptr(), true
6143 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6148 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6149 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6155 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6156 // Executed only by the owner P.
6157 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6158 oldNext := pp.runnext
6159 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6160 drainQ.pushBack(oldNext.ptr())
6165 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6171 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6175 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6179 // We've inverted the order in which it gets G's from the local P's runnable queue
6180 // and then advances the head pointer because we don't want to mess up the statuses of G's
6181 // while runqdrain() and runqsteal() are running in parallel.
6182 // Thus we should advance the head pointer before draining the local P into a gQueue,
6183 // so that we can update any gp.schedlink only after we take the full ownership of G,
6184 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6185 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6186 for i := uint32(0); i < qn; i++ {
6187 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6194 // Grabs a batch of goroutines from pp's runnable queue into batch.
6195 // Batch is a ring buffer starting at batchHead.
6196 // Returns number of grabbed goroutines.
6197 // Can be executed by any P.
6198 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6200 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6201 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6206 // Try to steal from pp.runnext.
6207 if next := pp.runnext; next != 0 {
6208 if pp.status == _Prunning {
6209 // Sleep to ensure that pp isn't about to run the g
6210 // we are about to steal.
6211 // The important use case here is when the g running
6212 // on pp ready()s another g and then almost
6213 // immediately blocks. Instead of stealing runnext
6214 // in this window, back off to give pp a chance to
6215 // schedule runnext. This will avoid thrashing gs
6216 // between different Ps.
6217 // A sync chan send/recv takes ~50ns as of time of
6218 // writing, so 3us gives ~50x overshoot.
6219 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6222 // On some platforms system timer granularity is
6223 // 1-15ms, which is way too much for this
6224 // optimization. So just yield.
6228 if !pp.runnext.cas(next, 0) {
6231 batch[batchHead%uint32(len(batch))] = next
6237 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6240 for i := uint32(0); i < n; i++ {
6241 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6242 batch[(batchHead+i)%uint32(len(batch))] = g
6244 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6250 // Steal half of elements from local runnable queue of p2
6251 // and put onto local runnable queue of p.
6252 // Returns one of the stolen elements (or nil if failed).
6253 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6255 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6260 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6264 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6265 if t-h+n >= uint32(len(pp.runq)) {
6266 throw("runqsteal: runq overflow")
6268 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6272 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6273 // be on one gQueue or gList at a time.
6274 type gQueue struct {
6279 // empty reports whether q is empty.
6280 func (q *gQueue) empty() bool {
6284 // push adds gp to the head of q.
6285 func (q *gQueue) push(gp *g) {
6286 gp.schedlink = q.head
6293 // pushBack adds gp to the tail of q.
6294 func (q *gQueue) pushBack(gp *g) {
6297 q.tail.ptr().schedlink.set(gp)
6304 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6306 func (q *gQueue) pushBackAll(q2 gQueue) {
6310 q2.tail.ptr().schedlink = 0
6312 q.tail.ptr().schedlink = q2.head
6319 // pop removes and returns the head of queue q. It returns nil if
6321 func (q *gQueue) pop() *g {
6324 q.head = gp.schedlink
6332 // popList takes all Gs in q and returns them as a gList.
6333 func (q *gQueue) popList() gList {
6334 stack := gList{q.head}
6339 // A gList is a list of Gs linked through g.schedlink. A G can only be
6340 // on one gQueue or gList at a time.
6345 // empty reports whether l is empty.
6346 func (l *gList) empty() bool {
6350 // push adds gp to the head of l.
6351 func (l *gList) push(gp *g) {
6352 gp.schedlink = l.head
6356 // pushAll prepends all Gs in q to l.
6357 func (l *gList) pushAll(q gQueue) {
6359 q.tail.ptr().schedlink = l.head
6364 // pop removes and returns the head of l. If l is empty, it returns nil.
6365 func (l *gList) pop() *g {
6368 l.head = gp.schedlink
6373 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6374 func setMaxThreads(in int) (out int) {
6376 out = int(sched.maxmcount)
6377 if in > 0x7fffffff { // MaxInt32
6378 sched.maxmcount = 0x7fffffff
6380 sched.maxmcount = int32(in)
6388 func procPin() int {
6393 return int(mp.p.ptr().id)
6402 //go:linkname sync_runtime_procPin sync.runtime_procPin
6404 func sync_runtime_procPin() int {
6408 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6410 func sync_runtime_procUnpin() {
6414 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6416 func sync_atomic_runtime_procPin() int {
6420 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6422 func sync_atomic_runtime_procUnpin() {
6426 // Active spinning for sync.Mutex.
6428 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6430 func sync_runtime_canSpin(i int) bool {
6431 // sync.Mutex is cooperative, so we are conservative with spinning.
6432 // Spin only few times and only if running on a multicore machine and
6433 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6434 // As opposed to runtime mutex we don't do passive spinning here,
6435 // because there can be work on global runq or on other Ps.
6436 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6439 if p := getg().m.p.ptr(); !runqempty(p) {
6445 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6447 func sync_runtime_doSpin() {
6448 procyield(active_spin_cnt)
6451 var stealOrder randomOrder
6453 // randomOrder/randomEnum are helper types for randomized work stealing.
6454 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6455 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6456 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6457 type randomOrder struct {
6462 type randomEnum struct {
6469 func (ord *randomOrder) reset(count uint32) {
6471 ord.coprimes = ord.coprimes[:0]
6472 for i := uint32(1); i <= count; i++ {
6473 if gcd(i, count) == 1 {
6474 ord.coprimes = append(ord.coprimes, i)
6479 func (ord *randomOrder) start(i uint32) randomEnum {
6483 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6487 func (enum *randomEnum) done() bool {
6488 return enum.i == enum.count
6491 func (enum *randomEnum) next() {
6493 enum.pos = (enum.pos + enum.inc) % enum.count
6496 func (enum *randomEnum) position() uint32 {
6500 func gcd(a, b uint32) uint32 {
6507 // An initTask represents the set of initializations that need to be done for a package.
6508 // Keep in sync with ../../test/noinit.go:initTask
6509 type initTask struct {
6510 state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
6512 // followed by nfns pcs, uintptr sized, one per init function to run
6515 // inittrace stores statistics for init functions which are
6516 // updated by malloc and newproc when active is true.
6517 var inittrace tracestat
6519 type tracestat struct {
6520 active bool // init tracing activation status
6521 id uint64 // init goroutine id
6522 allocs uint64 // heap allocations
6523 bytes uint64 // heap allocated bytes
6526 func doInit(ts []*initTask) {
6527 for _, t := range ts {
6532 func doInit1(t *initTask) {
6534 case 2: // fully initialized
6536 case 1: // initialization in progress
6537 throw("recursive call during initialization - linker skew")
6538 default: // not initialized yet
6539 t.state = 1 // initialization in progress
6546 if inittrace.active {
6548 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6553 // We should have pruned all of these in the linker.
6554 throw("inittask with no functions")
6557 firstFunc := add(unsafe.Pointer(t), 8)
6558 for i := uint32(0); i < t.nfns; i++ {
6559 p := add(firstFunc, uintptr(i)*goarch.PtrSize)
6560 f := *(*func())(unsafe.Pointer(&p))
6564 if inittrace.active {
6566 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6569 f := *(*func())(unsafe.Pointer(&firstFunc))
6570 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6573 print("init ", pkg, " @")
6574 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6575 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6576 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6577 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6581 t.state = 2 // initialization done