1 // Copyright 2014 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
11 "runtime/internal/atomic"
12 "runtime/internal/sys"
16 // set using cmd/go/internal/modload.ModInfoProg
19 // Goroutine scheduler
20 // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
22 // The main concepts are:
24 // M - worker thread, or machine.
25 // P - processor, a resource that is required to execute Go code.
26 // M must have an associated P to execute Go code, however it can be
27 // blocked or in a syscall w/o an associated P.
29 // Design doc at https://golang.org/s/go11sched.
31 // Worker thread parking/unparking.
32 // We need to balance between keeping enough running worker threads to utilize
33 // available hardware parallelism and parking excessive running worker threads
34 // to conserve CPU resources and power. This is not simple for two reasons:
35 // (1) scheduler state is intentionally distributed (in particular, per-P work
36 // queues), so it is not possible to compute global predicates on fast paths;
37 // (2) for optimal thread management we would need to know the future (don't park
38 // a worker thread when a new goroutine will be readied in near future).
40 // Three rejected approaches that would work badly:
41 // 1. Centralize all scheduler state (would inhibit scalability).
42 // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
43 // is a spare P, unpark a thread and handoff it the thread and the goroutine.
44 // This would lead to thread state thrashing, as the thread that readied the
45 // goroutine can be out of work the very next moment, we will need to park it.
46 // Also, it would destroy locality of computation as we want to preserve
47 // dependent goroutines on the same thread; and introduce additional latency.
48 // 3. Unpark an additional thread whenever we ready a goroutine and there is an
49 // idle P, but don't do handoff. This would lead to excessive thread parking/
50 // unparking as the additional threads will instantly park without discovering
53 // The current approach:
55 // This approach applies to three primary sources of potential work: readying a
56 // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
57 // additional details.
59 // We unpark an additional thread when we submit work if (this is wakep()):
60 // 1. There is an idle P, and
61 // 2. There are no "spinning" worker threads.
63 // A worker thread is considered spinning if it is out of local work and did
64 // not find work in the global run queue or netpoller; the spinning state is
65 // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
66 // also considered spinning; we don't do goroutine handoff so such threads are
67 // out of work initially. Spinning threads spin on looking for work in per-P
68 // run queues and timer heaps or from the GC before parking. If a spinning
69 // thread finds work it takes itself out of the spinning state and proceeds to
70 // execution. If it does not find work it takes itself out of the spinning
71 // state and then parks.
73 // If there is at least one spinning thread (sched.nmspinning>1), we don't
74 // unpark new threads when submitting work. To compensate for that, if the last
75 // spinning thread finds work and stops spinning, it must unpark a new spinning
76 // thread. This approach smooths out unjustified spikes of thread unparking,
77 // but at the same time guarantees eventual maximal CPU parallelism
80 // The main implementation complication is that we need to be very careful
81 // during spinning->non-spinning thread transition. This transition can race
82 // with submission of new work, and either one part or another needs to unpark
83 // another worker thread. If they both fail to do that, we can end up with
84 // semi-persistent CPU underutilization.
86 // The general pattern for submission is:
87 // 1. Submit work to the local run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
120 // This slice records the initializing tasks that need to be
121 // done to start up the runtime. It is built by the linker.
122 var runtime_inittasks []*initTask
124 // main_init_done is a signal used by cgocallbackg that initialization
125 // has been completed. It is made before _cgo_notify_runtime_init_done,
126 // so all cgo calls can rely on it existing. When main_init is complete,
127 // it is closed, meaning cgocallbackg can reliably receive from it.
128 var main_init_done chan bool
130 //go:linkname main_main main.main
133 // mainStarted indicates that the main M has started.
136 // runtimeInitTime is the nanotime() at which the runtime started.
137 var runtimeInitTime int64
139 // Value to use for signal mask for newly created M's.
140 var initSigmask sigset
142 // The main goroutine.
146 // Racectx of m0->g0 is used only as the parent of the main goroutine.
147 // It must not be used for anything else.
150 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
151 // Using decimal instead of binary GB and MB because
152 // they look nicer in the stack overflow failure message.
153 if goarch.PtrSize == 8 {
154 maxstacksize = 1000000000
156 maxstacksize = 250000000
159 // An upper limit for max stack size. Used to avoid random crashes
160 // after calling SetMaxStack and trying to allocate a stack that is too big,
161 // since stackalloc works with 32-bit sizes.
162 maxstackceiling = 2 * maxstacksize
164 // Allow newproc to start new Ms.
167 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
169 newm(sysmon, nil, -1)
173 // Lock the main goroutine onto this, the main OS thread,
174 // during initialization. Most programs won't care, but a few
175 // do require certain calls to be made by the main thread.
176 // Those can arrange for main.main to run in the main thread
177 // by calling runtime.LockOSThread during initialization
178 // to preserve the lock.
182 throw("runtime.main not on m0")
185 // Record when the world started.
186 // Must be before doInit for tracing init.
187 runtimeInitTime = nanotime()
188 if runtimeInitTime == 0 {
189 throw("nanotime returning zero")
192 if debug.inittrace != 0 {
193 inittrace.id = getg().goid
194 inittrace.active = true
197 doInit(runtime_inittasks) // Must be before defer.
199 // Defer unlock so that runtime.Goexit during init does the unlock too.
209 main_init_done = make(chan bool)
211 if _cgo_thread_start == nil {
212 throw("_cgo_thread_start missing")
214 if GOOS != "windows" {
215 if _cgo_setenv == nil {
216 throw("_cgo_setenv missing")
218 if _cgo_unsetenv == nil {
219 throw("_cgo_unsetenv missing")
222 if _cgo_notify_runtime_init_done == nil {
223 throw("_cgo_notify_runtime_init_done missing")
225 // Start the template thread in case we enter Go from
226 // a C-created thread and need to create a new thread.
227 startTemplateThread()
228 cgocall(_cgo_notify_runtime_init_done, nil)
231 // Run the initializing tasks. Depending on build mode this
232 // list can arrive a few different ways, but it will always
233 // contain the init tasks computed by the linker for all the
234 // packages in the program (excluding those added at runtime
235 // by package plugin).
236 for _, m := range activeModules() {
240 // Disable init tracing after main init done to avoid overhead
241 // of collecting statistics in malloc and newproc
242 inittrace.active = false
244 close(main_init_done)
249 if isarchive || islibrary {
250 // A program compiled with -buildmode=c-archive or c-shared
251 // has a main, but it is not executed.
254 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
257 runExitHooks(0) // run hooks now, since racefini does not return
261 // Make racy client program work: if panicking on
262 // another goroutine at the same time as main returns,
263 // let the other goroutine finish printing the panic trace.
264 // Once it does, it will exit. See issues 3934 and 20018.
265 if runningPanicDefers.Load() != 0 {
266 // Running deferred functions should not take long.
267 for c := 0; c < 1000; c++ {
268 if runningPanicDefers.Load() == 0 {
274 if panicking.Load() != 0 {
275 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
286 // os_beforeExit is called from os.Exit(0).
288 //go:linkname os_beforeExit os.runtime_beforeExit
289 func os_beforeExit(exitCode int) {
290 runExitHooks(exitCode)
291 if exitCode == 0 && raceenabled {
296 // start forcegc helper goroutine
301 func forcegchelper() {
303 lockInit(&forcegc.lock, lockRankForcegc)
306 if forcegc.idle.Load() {
307 throw("forcegc: phase error")
309 forcegc.idle.Store(true)
310 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
311 // this goroutine is explicitly resumed by sysmon
312 if debug.gctrace > 0 {
315 // Time-triggered, fully concurrent.
316 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
320 // Gosched yields the processor, allowing other goroutines to run. It does not
321 // suspend the current goroutine, so execution resumes automatically.
329 // goschedguarded yields the processor like gosched, but also checks
330 // for forbidden states and opts out of the yield in those cases.
333 func goschedguarded() {
334 mcall(goschedguarded_m)
337 // goschedIfBusy yields the processor like gosched, but only does so if
338 // there are no idle Ps or if we're on the only P and there's nothing in
339 // the run queue. In both cases, there is freely available idle time.
342 func goschedIfBusy() {
344 // Call gosched if gp.preempt is set; we may be in a tight loop that
345 // doesn't otherwise yield.
346 if !gp.preempt && sched.npidle.Load() > 0 {
352 // Puts the current goroutine into a waiting state and calls unlockf on the
355 // If unlockf returns false, the goroutine is resumed.
357 // unlockf must not access this G's stack, as it may be moved between
358 // the call to gopark and the call to unlockf.
360 // Note that because unlockf is called after putting the G into a waiting
361 // state, the G may have already been readied by the time unlockf is called
362 // unless there is external synchronization preventing the G from being
363 // readied. If unlockf returns false, it must guarantee that the G cannot be
364 // externally readied.
366 // Reason explains why the goroutine has been parked. It is displayed in stack
367 // traces and heap dumps. Reasons should be unique and descriptive. Do not
368 // re-use reasons, add new ones.
369 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
370 if reason != waitReasonSleep {
371 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
375 status := readgstatus(gp)
376 if status != _Grunning && status != _Gscanrunning {
377 throw("gopark: bad g status")
380 mp.waitunlockf = unlockf
381 gp.waitreason = reason
382 mp.waittraceev = traceEv
383 mp.waittraceskip = traceskip
385 // can't do anything that might move the G between Ms here.
389 // Puts the current goroutine into a waiting state and unlocks the lock.
390 // The goroutine can be made runnable again by calling goready(gp).
391 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
392 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
395 func goready(gp *g, traceskip int) {
397 ready(gp, traceskip, true)
402 func acquireSudog() *sudog {
403 // Delicate dance: the semaphore implementation calls
404 // acquireSudog, acquireSudog calls new(sudog),
405 // new calls malloc, malloc can call the garbage collector,
406 // and the garbage collector calls the semaphore implementation
408 // Break the cycle by doing acquirem/releasem around new(sudog).
409 // The acquirem/releasem increments m.locks during new(sudog),
410 // which keeps the garbage collector from being invoked.
413 if len(pp.sudogcache) == 0 {
414 lock(&sched.sudoglock)
415 // First, try to grab a batch from central cache.
416 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
417 s := sched.sudogcache
418 sched.sudogcache = s.next
420 pp.sudogcache = append(pp.sudogcache, s)
422 unlock(&sched.sudoglock)
423 // If the central cache is empty, allocate a new one.
424 if len(pp.sudogcache) == 0 {
425 pp.sudogcache = append(pp.sudogcache, new(sudog))
428 n := len(pp.sudogcache)
429 s := pp.sudogcache[n-1]
430 pp.sudogcache[n-1] = nil
431 pp.sudogcache = pp.sudogcache[:n-1]
433 throw("acquireSudog: found s.elem != nil in cache")
440 func releaseSudog(s *sudog) {
442 throw("runtime: sudog with non-nil elem")
445 throw("runtime: sudog with non-false isSelect")
448 throw("runtime: sudog with non-nil next")
451 throw("runtime: sudog with non-nil prev")
453 if s.waitlink != nil {
454 throw("runtime: sudog with non-nil waitlink")
457 throw("runtime: sudog with non-nil c")
461 throw("runtime: releaseSudog with non-nil gp.param")
463 mp := acquirem() // avoid rescheduling to another P
465 if len(pp.sudogcache) == cap(pp.sudogcache) {
466 // Transfer half of local cache to the central cache.
467 var first, last *sudog
468 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
469 n := len(pp.sudogcache)
470 p := pp.sudogcache[n-1]
471 pp.sudogcache[n-1] = nil
472 pp.sudogcache = pp.sudogcache[:n-1]
480 lock(&sched.sudoglock)
481 last.next = sched.sudogcache
482 sched.sudogcache = first
483 unlock(&sched.sudoglock)
485 pp.sudogcache = append(pp.sudogcache, s)
489 // called from assembly.
490 func badmcall(fn func(*g)) {
491 throw("runtime: mcall called on m->g0 stack")
494 func badmcall2(fn func(*g)) {
495 throw("runtime: mcall function returned")
498 func badreflectcall() {
499 panic(plainError("arg size to reflect.call more than 1GB"))
503 //go:nowritebarrierrec
504 func badmorestackg0() {
505 writeErrStr("fatal: morestack on g0\n")
509 //go:nowritebarrierrec
510 func badmorestackgsignal() {
511 writeErrStr("fatal: morestack on gsignal\n")
519 func lockedOSThread() bool {
521 return gp.lockedm != 0 && gp.m.lockedg != 0
525 // allgs contains all Gs ever created (including dead Gs), and thus
528 // Access via the slice is protected by allglock or stop-the-world.
529 // Readers that cannot take the lock may (carefully!) use the atomic
534 // allglen and allgptr are atomic variables that contain len(allgs) and
535 // &allgs[0] respectively. Proper ordering depends on totally-ordered
536 // loads and stores. Writes are protected by allglock.
538 // allgptr is updated before allglen. Readers should read allglen
539 // before allgptr to ensure that allglen is always <= len(allgptr). New
540 // Gs appended during the race can be missed. For a consistent view of
541 // all Gs, allglock must be held.
543 // allgptr copies should always be stored as a concrete type or
544 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
545 // even if it points to a stale array.
550 func allgadd(gp *g) {
551 if readgstatus(gp) == _Gidle {
552 throw("allgadd: bad status Gidle")
556 allgs = append(allgs, gp)
557 if &allgs[0] != allgptr {
558 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
560 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
564 // allGsSnapshot returns a snapshot of the slice of all Gs.
566 // The world must be stopped or allglock must be held.
567 func allGsSnapshot() []*g {
568 assertWorldStoppedOrLockHeld(&allglock)
570 // Because the world is stopped or allglock is held, allgadd
571 // cannot happen concurrently with this. allgs grows
572 // monotonically and existing entries never change, so we can
573 // simply return a copy of the slice header. For added safety,
574 // we trim everything past len because that can still change.
575 return allgs[:len(allgs):len(allgs)]
578 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
579 func atomicAllG() (**g, uintptr) {
580 length := atomic.Loaduintptr(&allglen)
581 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
585 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
586 func atomicAllGIndex(ptr **g, i uintptr) *g {
587 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
590 // forEachG calls fn on every G from allgs.
592 // forEachG takes a lock to exclude concurrent addition of new Gs.
593 func forEachG(fn func(gp *g)) {
595 for _, gp := range allgs {
601 // forEachGRace calls fn on every G from allgs.
603 // forEachGRace avoids locking, but does not exclude addition of new Gs during
604 // execution, which may be missed.
605 func forEachGRace(fn func(gp *g)) {
606 ptr, length := atomicAllG()
607 for i := uintptr(0); i < length; i++ {
608 gp := atomicAllGIndex(ptr, i)
615 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
616 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
620 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
621 // value of the GODEBUG environment variable.
622 func cpuinit(env string) {
624 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
625 cpu.DebugOptions = true
629 // Support cpu feature variables are used in code generated by the compiler
630 // to guard execution of instructions that can not be assumed to be always supported.
633 x86HasPOPCNT = cpu.X86.HasPOPCNT
634 x86HasSSE41 = cpu.X86.HasSSE41
635 x86HasFMA = cpu.X86.HasFMA
638 armHasVFPv4 = cpu.ARM.HasVFPv4
641 arm64HasATOMICS = cpu.ARM64.HasATOMICS
645 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
646 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
647 // early before much of the runtime is initialized.
648 func getGodebugEarly() string {
649 const prefix = "GODEBUG="
652 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
653 // Similar to goenv_unix but extracts the environment value for
655 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
657 for argv_index(argv, argc+1+n) != nil {
661 for i := int32(0); i < n; i++ {
662 p := argv_index(argv, argc+1+i)
663 s := unsafe.String(p, findnull(p))
665 if hasPrefix(s, prefix) {
666 env = gostring(p)[len(prefix):]
674 // The bootstrap sequence is:
678 // make & queue new G
679 // call runtime·mstart
681 // The new G calls runtime·main.
683 lockInit(&sched.lock, lockRankSched)
684 lockInit(&sched.sysmonlock, lockRankSysmon)
685 lockInit(&sched.deferlock, lockRankDefer)
686 lockInit(&sched.sudoglock, lockRankSudog)
687 lockInit(&deadlock, lockRankDeadlock)
688 lockInit(&paniclk, lockRankPanic)
689 lockInit(&allglock, lockRankAllg)
690 lockInit(&allpLock, lockRankAllp)
691 lockInit(&reflectOffs.lock, lockRankReflectOffs)
692 lockInit(&finlock, lockRankFin)
693 lockInit(&trace.bufLock, lockRankTraceBuf)
694 lockInit(&trace.stringsLock, lockRankTraceStrings)
695 lockInit(&trace.lock, lockRankTrace)
696 lockInit(&cpuprof.lock, lockRankCpuprof)
697 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
698 // Enforce that this lock is always a leaf lock.
699 // All of this lock's critical sections should be
701 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
703 // raceinit must be the first call to race detector.
704 // In particular, it must be done before mallocinit below calls racemapshadow.
707 gp.racectx, raceprocctx0 = raceinit()
710 sched.maxmcount = 10000
712 // The world starts stopped.
718 godebug := getGodebugEarly()
719 initPageTrace(godebug) // must run after mallocinit but before anything allocates
720 cpuinit(godebug) // must run before alginit
721 alginit() // maps, hash, fastrand must not be used before this call
722 fastrandinit() // must run before mcommoninit
723 mcommoninit(gp.m, -1)
724 modulesinit() // provides activeModules
725 typelinksinit() // uses maps, activeModules
726 itabsinit() // uses activeModules
727 stkobjinit() // must run before GC starts
729 sigsave(&gp.m.sigmask)
730 initSigmask = gp.m.sigmask
737 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
738 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
739 // set to true by the linker, it means that nothing is consuming the profile, it is
740 // safe to set MemProfileRate to 0.
741 if disableMemoryProfiling {
746 sched.lastpoll.Store(nanotime())
748 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
751 if procresize(procs) != nil {
752 throw("unknown runnable goroutine during bootstrap")
756 // World is effectively started now, as P's can run.
759 if buildVersion == "" {
760 // Condition should never trigger. This code just serves
761 // to ensure runtime·buildVersion is kept in the resulting binary.
762 buildVersion = "unknown"
764 if len(modinfo) == 1 {
765 // Condition should never trigger. This code just serves
766 // to ensure runtime·modinfo is kept in the resulting binary.
771 func dumpgstatus(gp *g) {
773 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
774 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
777 // sched.lock must be held.
779 assertLockHeld(&sched.lock)
781 if mcount() > sched.maxmcount {
782 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
783 throw("thread exhaustion")
787 // mReserveID returns the next ID to use for a new m. This new m is immediately
788 // considered 'running' by checkdead.
790 // sched.lock must be held.
791 func mReserveID() int64 {
792 assertLockHeld(&sched.lock)
794 if sched.mnext+1 < sched.mnext {
795 throw("runtime: thread ID overflow")
803 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
804 func mcommoninit(mp *m, id int64) {
807 // g0 stack won't make sense for user (and is not necessary unwindable).
809 callers(1, mp.createstack[:])
820 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
821 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
825 // Same behavior as for 1.17.
826 // TODO: Simplify this.
827 if goarch.BigEndian {
828 mp.fastrand = uint64(lo)<<32 | uint64(hi)
830 mp.fastrand = uint64(hi)<<32 | uint64(lo)
834 if mp.gsignal != nil {
835 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
838 // Add to allm so garbage collector doesn't free g->m
839 // when it is just in a register or thread-local storage.
842 // NumCgoCall() iterates over allm w/o schedlock,
843 // so we need to publish it safely.
844 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
847 // Allocate memory to hold a cgo traceback if the cgo call crashes.
848 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
849 mp.cgoCallers = new(cgoCallers)
853 func (mp *m) becomeSpinning() {
855 sched.nmspinning.Add(1)
856 sched.needspinning.Store(0)
859 func (mp *m) incgocallback() bool {
860 return (!mp.incgo && mp.ncgo > 0) || mp.isextra
863 var fastrandseed uintptr
865 func fastrandinit() {
866 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
870 // Mark gp ready to run.
871 func ready(gp *g, traceskip int, next bool) {
873 traceGoUnpark(gp, traceskip)
876 status := readgstatus(gp)
879 mp := acquirem() // disable preemption because it can be holding p in a local var
880 if status&^_Gscan != _Gwaiting {
882 throw("bad g->status in ready")
885 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
886 casgstatus(gp, _Gwaiting, _Grunnable)
887 runqput(mp.p.ptr(), gp, next)
892 // freezeStopWait is a large value that freezetheworld sets
893 // sched.stopwait to in order to request that all Gs permanently stop.
894 const freezeStopWait = 0x7fffffff
896 // freezing is set to non-zero if the runtime is trying to freeze the
898 var freezing atomic.Bool
900 // Similar to stopTheWorld but best-effort and can be called several times.
901 // There is no reverse operation, used during crashing.
902 // This function must not lock any mutexes.
903 func freezetheworld() {
905 // stopwait and preemption requests can be lost
906 // due to races with concurrently executing threads,
907 // so try several times
908 for i := 0; i < 5; i++ {
909 // this should tell the scheduler to not start any new goroutines
910 sched.stopwait = freezeStopWait
911 sched.gcwaiting.Store(true)
912 // this should stop running goroutines
914 break // no running goroutines
924 // All reads and writes of g's status go through readgstatus, casgstatus
925 // castogscanstatus, casfrom_Gscanstatus.
928 func readgstatus(gp *g) uint32 {
929 return gp.atomicstatus.Load()
932 // The Gscanstatuses are acting like locks and this releases them.
933 // If it proves to be a performance hit we should be able to make these
934 // simple atomic stores but for now we are going to throw if
935 // we see an inconsistent state.
936 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
939 // Check that transition is valid.
942 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
944 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
950 if newval == oldval&^_Gscan {
951 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
955 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
957 throw("casfrom_Gscanstatus: gp->status is not in scan state")
959 releaseLockRank(lockRankGscan)
962 // This will return false if the gp is not in the expected status and the cas fails.
963 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
964 func castogscanstatus(gp *g, oldval, newval uint32) bool {
970 if newval == oldval|_Gscan {
971 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
973 acquireLockRank(lockRankGscan)
979 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
980 throw("castogscanstatus")
984 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
985 // various latencies on every transition instead of sampling them.
986 var casgstatusAlwaysTrack = false
988 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
989 // and casfrom_Gscanstatus instead.
990 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
991 // put it in the Gscan state is finished.
994 func casgstatus(gp *g, oldval, newval uint32) {
995 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
997 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
998 throw("casgstatus: bad incoming values")
1002 acquireLockRank(lockRankGscan)
1003 releaseLockRank(lockRankGscan)
1005 // See https://golang.org/cl/21503 for justification of the yield delay.
1006 const yieldDelay = 5 * 1000
1009 // loop if gp->atomicstatus is in a scan state giving
1010 // GC time to finish and change the state to oldval.
1011 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1012 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1013 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1016 nextYield = nanotime() + yieldDelay
1018 if nanotime() < nextYield {
1019 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1024 nextYield = nanotime() + yieldDelay/2
1028 if oldval == _Grunning {
1029 // Track every gTrackingPeriod time a goroutine transitions out of running.
1030 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1039 // Handle various kinds of tracking.
1042 // - Time spent in runnable.
1043 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1046 // We transitioned out of runnable, so measure how much
1047 // time we spent in this state and add it to
1050 gp.runnableTime += now - gp.trackingStamp
1051 gp.trackingStamp = 0
1053 if !gp.waitreason.isMutexWait() {
1054 // Not blocking on a lock.
1057 // Blocking on a lock, measure it. Note that because we're
1058 // sampling, we have to multiply by our sampling period to get
1059 // a more representative estimate of the absolute value.
1060 // gTrackingPeriod also represents an accurate sampling period
1061 // because we can only enter this state from _Grunning.
1063 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1064 gp.trackingStamp = 0
1068 if !gp.waitreason.isMutexWait() {
1069 // Not blocking on a lock.
1072 // Blocking on a lock. Write down the timestamp.
1074 gp.trackingStamp = now
1076 // We just transitioned into runnable, so record what
1077 // time that happened.
1079 gp.trackingStamp = now
1081 // We're transitioning into running, so turn off
1082 // tracking and record how much time we spent in
1085 sched.timeToRun.record(gp.runnableTime)
1090 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1092 // Use this over casgstatus when possible to ensure that a waitreason is set.
1093 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1094 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1095 gp.waitreason = reason
1096 casgstatus(gp, old, _Gwaiting)
1099 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1100 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1101 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1102 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1103 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1106 func casgcopystack(gp *g) uint32 {
1108 oldstatus := readgstatus(gp) &^ _Gscan
1109 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1110 throw("copystack: bad status, not Gwaiting or Grunnable")
1112 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1118 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1120 // TODO(austin): This is the only status operation that both changes
1121 // the status and locks the _Gscan bit. Rethink this.
1122 func casGToPreemptScan(gp *g, old, new uint32) {
1123 if old != _Grunning || new != _Gscan|_Gpreempted {
1124 throw("bad g transition")
1126 acquireLockRank(lockRankGscan)
1127 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1131 // casGFromPreempted attempts to transition gp from _Gpreempted to
1132 // _Gwaiting. If successful, the caller is responsible for
1133 // re-scheduling gp.
1134 func casGFromPreempted(gp *g, old, new uint32) bool {
1135 if old != _Gpreempted || new != _Gwaiting {
1136 throw("bad g transition")
1138 gp.waitreason = waitReasonPreempted
1139 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1142 // stopTheWorld stops all P's from executing goroutines, interrupting
1143 // all goroutines at GC safe points and records reason as the reason
1144 // for the stop. On return, only the current goroutine's P is running.
1145 // stopTheWorld must not be called from a system stack and the caller
1146 // must not hold worldsema. The caller must call startTheWorld when
1147 // other P's should resume execution.
1149 // stopTheWorld is safe for multiple goroutines to call at the
1150 // same time. Each will execute its own stop, and the stops will
1153 // This is also used by routines that do stack dumps. If the system is
1154 // in panic or being exited, this may not reliably stop all
1156 func stopTheWorld(reason string) {
1157 semacquire(&worldsema)
1159 gp.m.preemptoff = reason
1160 systemstack(func() {
1161 // Mark the goroutine which called stopTheWorld preemptible so its
1162 // stack may be scanned.
1163 // This lets a mark worker scan us while we try to stop the world
1164 // since otherwise we could get in a mutual preemption deadlock.
1165 // We must not modify anything on the G stack because a stack shrink
1166 // may occur. A stack shrink is otherwise OK though because in order
1167 // to return from this function (and to leave the system stack) we
1168 // must have preempted all goroutines, including any attempting
1169 // to scan our stack, in which case, any stack shrinking will
1170 // have already completed by the time we exit.
1171 // Don't provide a wait reason because we're still executing.
1172 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1173 stopTheWorldWithSema()
1174 casgstatus(gp, _Gwaiting, _Grunning)
1178 // startTheWorld undoes the effects of stopTheWorld.
1179 func startTheWorld() {
1180 systemstack(func() { startTheWorldWithSema(false) })
1182 // worldsema must be held over startTheWorldWithSema to ensure
1183 // gomaxprocs cannot change while worldsema is held.
1185 // Release worldsema with direct handoff to the next waiter, but
1186 // acquirem so that semrelease1 doesn't try to yield our time.
1188 // Otherwise if e.g. ReadMemStats is being called in a loop,
1189 // it might stomp on other attempts to stop the world, such as
1190 // for starting or ending GC. The operation this blocks is
1191 // so heavy-weight that we should just try to be as fair as
1194 // We don't want to just allow us to get preempted between now
1195 // and releasing the semaphore because then we keep everyone
1196 // (including, for example, GCs) waiting longer.
1199 semrelease1(&worldsema, true, 0)
1203 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1204 // until the GC is not running. It also blocks a GC from starting
1205 // until startTheWorldGC is called.
1206 func stopTheWorldGC(reason string) {
1208 stopTheWorld(reason)
1211 // startTheWorldGC undoes the effects of stopTheWorldGC.
1212 func startTheWorldGC() {
1217 // Holding worldsema grants an M the right to try to stop the world.
1218 var worldsema uint32 = 1
1220 // Holding gcsema grants the M the right to block a GC, and blocks
1221 // until the current GC is done. In particular, it prevents gomaxprocs
1222 // from changing concurrently.
1224 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1225 // being changed/enabled during a GC, remove this.
1226 var gcsema uint32 = 1
1228 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1229 // The caller is responsible for acquiring worldsema and disabling
1230 // preemption first and then should stopTheWorldWithSema on the system
1233 // semacquire(&worldsema, 0)
1234 // m.preemptoff = "reason"
1235 // systemstack(stopTheWorldWithSema)
1237 // When finished, the caller must either call startTheWorld or undo
1238 // these three operations separately:
1240 // m.preemptoff = ""
1241 // systemstack(startTheWorldWithSema)
1242 // semrelease(&worldsema)
1244 // It is allowed to acquire worldsema once and then execute multiple
1245 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1246 // Other P's are able to execute between successive calls to
1247 // startTheWorldWithSema and stopTheWorldWithSema.
1248 // Holding worldsema causes any other goroutines invoking
1249 // stopTheWorld to block.
1250 func stopTheWorldWithSema() {
1253 // If we hold a lock, then we won't be able to stop another M
1254 // that is blocked trying to acquire the lock.
1256 throw("stopTheWorld: holding locks")
1260 sched.stopwait = gomaxprocs
1261 sched.gcwaiting.Store(true)
1264 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1266 // try to retake all P's in Psyscall status
1267 for _, pp := range allp {
1269 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1281 pp, _ := pidleget(now)
1285 pp.status = _Pgcstop
1288 wait := sched.stopwait > 0
1291 // wait for remaining P's to stop voluntarily
1294 // wait for 100us, then try to re-preempt in case of any races
1295 if notetsleep(&sched.stopnote, 100*1000) {
1296 noteclear(&sched.stopnote)
1305 if sched.stopwait != 0 {
1306 bad = "stopTheWorld: not stopped (stopwait != 0)"
1308 for _, pp := range allp {
1309 if pp.status != _Pgcstop {
1310 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1314 if freezing.Load() {
1315 // Some other thread is panicking. This can cause the
1316 // sanity checks above to fail if the panic happens in
1317 // the signal handler on a stopped thread. Either way,
1318 // we should halt this thread.
1329 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1330 assertWorldStopped()
1332 mp := acquirem() // disable preemption because it can be holding p in a local var
1333 if netpollinited() {
1334 list := netpoll(0) // non-blocking
1344 p1 := procresize(procs)
1345 sched.gcwaiting.Store(false)
1346 if sched.sysmonwait.Load() {
1347 sched.sysmonwait.Store(false)
1348 notewakeup(&sched.sysmonnote)
1361 throw("startTheWorld: inconsistent mp->nextp")
1364 notewakeup(&mp.park)
1366 // Start M to run P. Do not start another M below.
1371 // Capture start-the-world time before doing clean-up tasks.
1372 startTime := nanotime()
1377 // Wakeup an additional proc in case we have excessive runnable goroutines
1378 // in local queues or in the global queue. If we don't, the proc will park itself.
1379 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1387 // usesLibcall indicates whether this runtime performs system calls
1389 func usesLibcall() bool {
1391 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1394 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1399 // mStackIsSystemAllocated indicates whether this runtime starts on a
1400 // system-allocated stack.
1401 func mStackIsSystemAllocated() bool {
1403 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1407 case "386", "amd64", "arm", "arm64":
1414 // mstart is the entry-point for new Ms.
1415 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1418 // mstart0 is the Go entry-point for new Ms.
1419 // This must not split the stack because we may not even have stack
1420 // bounds set up yet.
1422 // May run during STW (because it doesn't have a P yet), so write
1423 // barriers are not allowed.
1426 //go:nowritebarrierrec
1430 osStack := gp.stack.lo == 0
1432 // Initialize stack bounds from system stack.
1433 // Cgo may have left stack size in stack.hi.
1434 // minit may update the stack bounds.
1436 // Note: these bounds may not be very accurate.
1437 // We set hi to &size, but there are things above
1438 // it. The 1024 is supposed to compensate this,
1439 // but is somewhat arbitrary.
1442 size = 8192 * sys.StackGuardMultiplier
1444 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1445 gp.stack.lo = gp.stack.hi - size + 1024
1447 // Initialize stack guard so that we can start calling regular
1449 gp.stackguard0 = gp.stack.lo + stackGuard
1450 // This is the g0, so we can also call go:systemstack
1451 // functions, which check stackguard1.
1452 gp.stackguard1 = gp.stackguard0
1455 // Exit this thread.
1456 if mStackIsSystemAllocated() {
1457 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1458 // the stack, but put it in gp.stack before mstart,
1459 // so the logic above hasn't set osStack yet.
1465 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1466 // so that we can set up g0.sched to return to the call of mstart1 above.
1473 throw("bad runtime·mstart")
1476 // Set up m.g0.sched as a label returning to just
1477 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1478 // We're never coming back to mstart1 after we call schedule,
1479 // so other calls can reuse the current frame.
1480 // And goexit0 does a gogo that needs to return from mstart1
1481 // and let mstart0 exit the thread.
1482 gp.sched.g = guintptr(unsafe.Pointer(gp))
1483 gp.sched.pc = getcallerpc()
1484 gp.sched.sp = getcallersp()
1489 // Install signal handlers; after minit so that minit can
1490 // prepare the thread to be able to handle the signals.
1495 if fn := gp.m.mstartfn; fn != nil {
1500 acquirep(gp.m.nextp.ptr())
1506 // mstartm0 implements part of mstart1 that only runs on the m0.
1508 // Write barriers are allowed here because we know the GC can't be
1509 // running yet, so they'll be no-ops.
1511 //go:yeswritebarrierrec
1513 // Create an extra M for callbacks on threads not created by Go.
1514 // An extra M is also needed on Windows for callbacks created by
1515 // syscall.NewCallback. See issue #6751 for details.
1516 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1523 // mPark causes a thread to park itself, returning once woken.
1528 notesleep(&gp.m.park)
1529 noteclear(&gp.m.park)
1532 // mexit tears down and exits the current thread.
1534 // Don't call this directly to exit the thread, since it must run at
1535 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1536 // unwind the stack to the point that exits the thread.
1538 // It is entered with m.p != nil, so write barriers are allowed. It
1539 // will release the P before exiting.
1541 //go:yeswritebarrierrec
1542 func mexit(osStack bool) {
1546 // This is the main thread. Just wedge it.
1548 // On Linux, exiting the main thread puts the process
1549 // into a non-waitable zombie state. On Plan 9,
1550 // exiting the main thread unblocks wait even though
1551 // other threads are still running. On Solaris we can
1552 // neither exitThread nor return from mstart. Other
1553 // bad things probably happen on other platforms.
1555 // We could try to clean up this M more before wedging
1556 // it, but that complicates signal handling.
1557 handoffp(releasep())
1563 throw("locked m0 woke up")
1569 // Free the gsignal stack.
1570 if mp.gsignal != nil {
1571 stackfree(mp.gsignal.stack)
1572 // On some platforms, when calling into VDSO (e.g. nanotime)
1573 // we store our g on the gsignal stack, if there is one.
1574 // Now the stack is freed, unlink it from the m, so we
1575 // won't write to it when calling VDSO code.
1579 // Remove m from allm.
1581 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1587 throw("m not found in allm")
1589 // Delay reaping m until it's done with the stack.
1591 // Put mp on the free list, though it will not be reaped while freeWait
1592 // is freeMWait. mp is no longer reachable via allm, so even if it is
1593 // on an OS stack, we must keep a reference to mp alive so that the GC
1594 // doesn't free mp while we are still using it.
1596 // Note that the free list must not be linked through alllink because
1597 // some functions walk allm without locking, so may be using alllink.
1598 mp.freeWait.Store(freeMWait)
1599 mp.freelink = sched.freem
1603 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1606 handoffp(releasep())
1607 // After this point we must not have write barriers.
1609 // Invoke the deadlock detector. This must happen after
1610 // handoffp because it may have started a new M to take our
1617 if GOOS == "darwin" || GOOS == "ios" {
1618 // Make sure pendingPreemptSignals is correct when an M exits.
1620 if mp.signalPending.Load() != 0 {
1621 pendingPreemptSignals.Add(-1)
1625 // Destroy all allocated resources. After this is called, we may no
1626 // longer take any locks.
1630 // No more uses of mp, so it is safe to drop the reference.
1631 mp.freeWait.Store(freeMRef)
1633 // Return from mstart and let the system thread
1634 // library free the g0 stack and terminate the thread.
1638 // mstart is the thread's entry point, so there's nothing to
1639 // return to. Exit the thread directly. exitThread will clear
1640 // m.freeWait when it's done with the stack and the m can be
1642 exitThread(&mp.freeWait)
1645 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1646 // If a P is currently executing code, this will bring the P to a GC
1647 // safe point and execute fn on that P. If the P is not executing code
1648 // (it is idle or in a syscall), this will call fn(p) directly while
1649 // preventing the P from exiting its state. This does not ensure that
1650 // fn will run on every CPU executing Go code, but it acts as a global
1651 // memory barrier. GC uses this as a "ragged barrier."
1653 // The caller must hold worldsema.
1656 func forEachP(fn func(*p)) {
1658 pp := getg().m.p.ptr()
1661 if sched.safePointWait != 0 {
1662 throw("forEachP: sched.safePointWait != 0")
1664 sched.safePointWait = gomaxprocs - 1
1665 sched.safePointFn = fn
1667 // Ask all Ps to run the safe point function.
1668 for _, p2 := range allp {
1670 atomic.Store(&p2.runSafePointFn, 1)
1675 // Any P entering _Pidle or _Psyscall from now on will observe
1676 // p.runSafePointFn == 1 and will call runSafePointFn when
1677 // changing its status to _Pidle/_Psyscall.
1679 // Run safe point function for all idle Ps. sched.pidle will
1680 // not change because we hold sched.lock.
1681 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1682 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1684 sched.safePointWait--
1688 wait := sched.safePointWait > 0
1691 // Run fn for the current P.
1694 // Force Ps currently in _Psyscall into _Pidle and hand them
1695 // off to induce safe point function execution.
1696 for _, p2 := range allp {
1698 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1708 // Wait for remaining Ps to run fn.
1711 // Wait for 100us, then try to re-preempt in
1712 // case of any races.
1714 // Requires system stack.
1715 if notetsleep(&sched.safePointNote, 100*1000) {
1716 noteclear(&sched.safePointNote)
1722 if sched.safePointWait != 0 {
1723 throw("forEachP: not done")
1725 for _, p2 := range allp {
1726 if p2.runSafePointFn != 0 {
1727 throw("forEachP: P did not run fn")
1732 sched.safePointFn = nil
1737 // runSafePointFn runs the safe point function, if any, for this P.
1738 // This should be called like
1740 // if getg().m.p.runSafePointFn != 0 {
1744 // runSafePointFn must be checked on any transition in to _Pidle or
1745 // _Psyscall to avoid a race where forEachP sees that the P is running
1746 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1747 // nor the P run the safe-point function.
1748 func runSafePointFn() {
1749 p := getg().m.p.ptr()
1750 // Resolve the race between forEachP running the safe-point
1751 // function on this P's behalf and this P running the
1752 // safe-point function directly.
1753 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1756 sched.safePointFn(p)
1758 sched.safePointWait--
1759 if sched.safePointWait == 0 {
1760 notewakeup(&sched.safePointNote)
1765 // When running with cgo, we call _cgo_thread_start
1766 // to start threads for us so that we can play nicely with
1768 var cgoThreadStart unsafe.Pointer
1770 type cgothreadstart struct {
1776 // Allocate a new m unassociated with any thread.
1777 // Can use p for allocation context if needed.
1778 // fn is recorded as the new m's m.mstartfn.
1779 // id is optional pre-allocated m ID. Omit by passing -1.
1781 // This function is allowed to have write barriers even if the caller
1782 // isn't because it borrows pp.
1784 //go:yeswritebarrierrec
1785 func allocm(pp *p, fn func(), id int64) *m {
1788 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1789 // disable preemption to ensure it is not stolen, which would make the
1790 // caller lose ownership.
1795 acquirep(pp) // temporarily borrow p for mallocs in this function
1798 // Release the free M list. We need to do this somewhere and
1799 // this may free up a stack we can use.
1800 if sched.freem != nil {
1803 for freem := sched.freem; freem != nil; {
1804 wait := freem.freeWait.Load()
1805 if wait == freeMWait {
1806 next := freem.freelink
1807 freem.freelink = newList
1812 // Free the stack if needed. For freeMRef, there is
1813 // nothing to do except drop freem from the sched.freem
1815 if wait == freeMStack {
1816 // stackfree must be on the system stack, but allocm is
1817 // reachable off the system stack transitively from
1819 systemstack(func() {
1820 stackfree(freem.g0.stack)
1823 freem = freem.freelink
1825 sched.freem = newList
1833 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1834 // Windows and Plan 9 will layout sched stack on OS stack.
1835 if iscgo || mStackIsSystemAllocated() {
1838 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1842 if pp == gp.m.p.ptr() {
1847 allocmLock.runlock()
1851 // needm is called when a cgo callback happens on a
1852 // thread without an m (a thread not created by Go).
1853 // In this case, needm is expected to find an m to use
1854 // and return with m, g initialized correctly.
1855 // Since m and g are not set now (likely nil, but see below)
1856 // needm is limited in what routines it can call. In particular
1857 // it can only call nosplit functions (textflag 7) and cannot
1858 // do any scheduling that requires an m.
1860 // In order to avoid needing heavy lifting here, we adopt
1861 // the following strategy: there is a stack of available m's
1862 // that can be stolen. Using compare-and-swap
1863 // to pop from the stack has ABA races, so we simulate
1864 // a lock by doing an exchange (via Casuintptr) to steal the stack
1865 // head and replace the top pointer with MLOCKED (1).
1866 // This serves as a simple spin lock that we can use even
1867 // without an m. The thread that locks the stack in this way
1868 // unlocks the stack by storing a valid stack head pointer.
1870 // In order to make sure that there is always an m structure
1871 // available to be stolen, we maintain the invariant that there
1872 // is always one more than needed. At the beginning of the
1873 // program (if cgo is in use) the list is seeded with a single m.
1874 // If needm finds that it has taken the last m off the list, its job
1875 // is - once it has installed its own m so that it can do things like
1876 // allocate memory - to create a spare m and put it on the list.
1878 // Each of these extra m's also has a g0 and a curg that are
1879 // pressed into service as the scheduling stack and current
1880 // goroutine for the duration of the cgo callback.
1882 // When the callback is done with the m, it calls dropm to
1883 // put the m back on the list.
1887 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1888 // Can happen if C/C++ code calls Go from a global ctor.
1889 // Can also happen on Windows if a global ctor uses a
1890 // callback created by syscall.NewCallback. See issue #6751
1893 // Can not throw, because scheduler is not initialized yet.
1894 writeErrStr("fatal error: cgo callback before cgo call\n")
1898 // Save and block signals before getting an M.
1899 // The signal handler may call needm itself,
1900 // and we must avoid a deadlock. Also, once g is installed,
1901 // any incoming signals will try to execute,
1902 // but we won't have the sigaltstack settings and other data
1903 // set up appropriately until the end of minit, which will
1904 // unblock the signals. This is the same dance as when
1905 // starting a new m to run Go code via newosproc.
1910 // Lock extra list, take head, unlock popped list.
1911 // nilokay=false is safe here because of the invariant above,
1912 // that the extra list always contains or will soon contain
1914 mp := lockextra(false)
1916 // Set needextram when we've just emptied the list,
1917 // so that the eventual call into cgocallbackg will
1918 // allocate a new m for the extra list. We delay the
1919 // allocation until then so that it can be done
1920 // after exitsyscall makes sure it is okay to be
1921 // running at all (that is, there's no garbage collection
1922 // running right now).
1923 mp.needextram = mp.schedlink == 0
1925 unlockextra(mp.schedlink.ptr())
1927 // Store the original signal mask for use by minit.
1928 mp.sigmask = sigmask
1930 // Install TLS on some platforms (previously setg
1931 // would do this if necessary).
1934 // Install g (= m->g0) and set the stack bounds
1935 // to match the current stack. We don't actually know
1936 // how big the stack is, like we don't know how big any
1937 // scheduling stack is, but we assume there's at least 32 kB,
1938 // which is more than enough for us.
1941 gp.stack.hi = getcallersp() + 1024
1942 gp.stack.lo = getcallersp() - 32*1024
1943 gp.stackguard0 = gp.stack.lo + stackGuard
1945 // Initialize this thread to use the m.
1949 // mp.curg is now a real goroutine.
1950 casgstatus(mp.curg, _Gdead, _Gsyscall)
1954 // newextram allocates m's and puts them on the extra list.
1955 // It is called with a working local m, so that it can do things
1956 // like call schedlock and allocate.
1958 c := extraMWaiters.Swap(0)
1960 for i := uint32(0); i < c; i++ {
1964 // Make sure there is at least one extra M.
1965 mp := lockextra(true)
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.
2024 mnext := lockextra(true)
2025 mp.schedlink.set(mnext)
2030 // dropm is called when a cgo callback has called needm but is now
2031 // done with the callback and returning back into the non-Go thread.
2032 // It puts the current m back onto the extra list.
2034 // The main expense here is the call to signalstack to release the
2035 // m's signal stack, and then the call to needm on the next callback
2036 // from this thread. It is tempting to try to save the m for next time,
2037 // which would eliminate both these costs, but there might not be
2038 // a next time: the current thread (which Go does not control) might exit.
2039 // If we saved the m for that thread, there would be an m leak each time
2040 // such a thread exited. Instead, we acquire and release an m on each
2041 // call. These should typically not be scheduling operations, just a few
2042 // atomics, so the cost should be small.
2044 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
2045 // variable using pthread_key_create. Unlike the pthread keys we already use
2046 // on OS X, this dummy key would never be read by Go code. It would exist
2047 // only so that we could register at thread-exit-time destructor.
2048 // That destructor would put the m back onto the extra list.
2049 // This is purely a performance optimization. The current version,
2050 // in which dropm happens on each cgo call, is still correct too.
2051 // We may have to keep the current version on systems with cgo
2052 // but without pthreads, like Windows.
2054 // Clear m and g, and return m to the extra list.
2055 // After the call to setg we can only call nosplit functions
2056 // with no pointer manipulation.
2059 // Return mp.curg to dead state.
2060 casgstatus(mp.curg, _Gsyscall, _Gdead)
2061 mp.curg.preemptStop = false
2064 // Block signals before unminit.
2065 // Unminit unregisters the signal handling stack (but needs g on some systems).
2066 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2067 // It's important not to try to handle a signal between those two steps.
2068 sigmask := mp.sigmask
2072 mnext := lockextra(true)
2074 mp.schedlink.set(mnext)
2078 // Commit the release of mp.
2081 msigrestore(sigmask)
2084 // A helper function for EnsureDropM.
2085 func getm() uintptr {
2086 return uintptr(unsafe.Pointer(getg().m))
2089 var extram atomic.Uintptr
2090 var extraMCount uint32 // Protected by lockextra
2091 var extraMWaiters atomic.Uint32
2093 // lockextra locks the extra list and returns the list head.
2094 // The caller must unlock the list by storing a new list head
2095 // to extram. If nilokay is true, then lockextra will
2096 // return a nil list head if that's what it finds. If nilokay is false,
2097 // lockextra will keep waiting until the list head is no longer nil.
2100 func lockextra(nilokay bool) *m {
2105 old := extram.Load()
2110 if old == 0 && !nilokay {
2112 // Add 1 to the number of threads
2113 // waiting for an M.
2114 // This is cleared by newextram.
2115 extraMWaiters.Add(1)
2121 if extram.CompareAndSwap(old, locked) {
2122 return (*m)(unsafe.Pointer(old))
2130 func unlockextra(mp *m) {
2131 extram.Store(uintptr(unsafe.Pointer(mp)))
2135 // allocmLock is locked for read when creating new Ms in allocm and their
2136 // addition to allm. Thus acquiring this lock for write blocks the
2137 // creation of new Ms.
2140 // execLock serializes exec and clone to avoid bugs or unspecified
2141 // behaviour around exec'ing while creating/destroying threads. See
2146 // These errors are reported (via writeErrStr) by some OS-specific
2147 // versions of newosproc and newosproc0.
2149 failthreadcreate = "runtime: failed to create new OS thread\n"
2150 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2153 // newmHandoff contains a list of m structures that need new OS threads.
2154 // This is used by newm in situations where newm itself can't safely
2155 // start an OS thread.
2156 var newmHandoff struct {
2159 // newm points to a list of M structures that need new OS
2160 // threads. The list is linked through m.schedlink.
2163 // waiting indicates that wake needs to be notified when an m
2164 // is put on the list.
2168 // haveTemplateThread indicates that the templateThread has
2169 // been started. This is not protected by lock. Use cas to set
2171 haveTemplateThread uint32
2174 // Create a new m. It will start off with a call to fn, or else the scheduler.
2175 // fn needs to be static and not a heap allocated closure.
2176 // May run with m.p==nil, so write barriers are not allowed.
2178 // id is optional pre-allocated m ID. Omit by passing -1.
2180 //go:nowritebarrierrec
2181 func newm(fn func(), pp *p, id int64) {
2182 // allocm adds a new M to allm, but they do not start until created by
2183 // the OS in newm1 or the template thread.
2185 // doAllThreadsSyscall requires that every M in allm will eventually
2186 // start and be signal-able, even with a STW.
2188 // Disable preemption here until we start the thread to ensure that
2189 // newm is not preempted between allocm and starting the new thread,
2190 // ensuring that anything added to allm is guaranteed to eventually
2194 mp := allocm(pp, fn, id)
2196 mp.sigmask = initSigmask
2197 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2198 // We're on a locked M or a thread that may have been
2199 // started by C. The kernel state of this thread may
2200 // be strange (the user may have locked it for that
2201 // purpose). We don't want to clone that into another
2202 // thread. Instead, ask a known-good thread to create
2203 // the thread for us.
2205 // This is disabled on Plan 9. See golang.org/issue/22227.
2207 // TODO: This may be unnecessary on Windows, which
2208 // doesn't model thread creation off fork.
2209 lock(&newmHandoff.lock)
2210 if newmHandoff.haveTemplateThread == 0 {
2211 throw("on a locked thread with no template thread")
2213 mp.schedlink = newmHandoff.newm
2214 newmHandoff.newm.set(mp)
2215 if newmHandoff.waiting {
2216 newmHandoff.waiting = false
2217 notewakeup(&newmHandoff.wake)
2219 unlock(&newmHandoff.lock)
2220 // The M has not started yet, but the template thread does not
2221 // participate in STW, so it will always process queued Ms and
2222 // it is safe to releasem.
2232 var ts cgothreadstart
2233 if _cgo_thread_start == nil {
2234 throw("_cgo_thread_start missing")
2237 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2238 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2240 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2243 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2245 execLock.rlock() // Prevent process clone.
2246 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2250 execLock.rlock() // Prevent process clone.
2255 // startTemplateThread starts the template thread if it is not already
2258 // The calling thread must itself be in a known-good state.
2259 func startTemplateThread() {
2260 if GOARCH == "wasm" { // no threads on wasm yet
2264 // Disable preemption to guarantee that the template thread will be
2265 // created before a park once haveTemplateThread is set.
2267 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2271 newm(templateThread, nil, -1)
2275 // templateThread is a thread in a known-good state that exists solely
2276 // to start new threads in known-good states when the calling thread
2277 // may not be in a good state.
2279 // Many programs never need this, so templateThread is started lazily
2280 // when we first enter a state that might lead to running on a thread
2281 // in an unknown state.
2283 // templateThread runs on an M without a P, so it must not have write
2286 //go:nowritebarrierrec
2287 func templateThread() {
2294 lock(&newmHandoff.lock)
2295 for newmHandoff.newm != 0 {
2296 newm := newmHandoff.newm.ptr()
2297 newmHandoff.newm = 0
2298 unlock(&newmHandoff.lock)
2300 next := newm.schedlink.ptr()
2305 lock(&newmHandoff.lock)
2307 newmHandoff.waiting = true
2308 noteclear(&newmHandoff.wake)
2309 unlock(&newmHandoff.lock)
2310 notesleep(&newmHandoff.wake)
2314 // Stops execution of the current m until new work is available.
2315 // Returns with acquired P.
2319 if gp.m.locks != 0 {
2320 throw("stopm holding locks")
2323 throw("stopm holding p")
2326 throw("stopm spinning")
2333 acquirep(gp.m.nextp.ptr())
2338 // startm's caller incremented nmspinning. Set the new M's spinning.
2339 getg().m.spinning = true
2342 // Schedules some M to run the p (creates an M if necessary).
2343 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2344 // May run with m.p==nil, so write barriers are not allowed.
2345 // If spinning is set, the caller has incremented nmspinning and must provide a
2346 // P. startm will set m.spinning in the newly started M.
2348 // Callers passing a non-nil P must call from a non-preemptible context. See
2349 // comment on acquirem below.
2351 // Must not have write barriers because this may be called without a P.
2353 //go:nowritebarrierrec
2354 func startm(pp *p, spinning bool) {
2355 // Disable preemption.
2357 // Every owned P must have an owner that will eventually stop it in the
2358 // event of a GC stop request. startm takes transient ownership of a P
2359 // (either from argument or pidleget below) and transfers ownership to
2360 // a started M, which will be responsible for performing the stop.
2362 // Preemption must be disabled during this transient ownership,
2363 // otherwise the P this is running on may enter GC stop while still
2364 // holding the transient P, leaving that P in limbo and deadlocking the
2367 // Callers passing a non-nil P must already be in non-preemptible
2368 // context, otherwise such preemption could occur on function entry to
2369 // startm. Callers passing a nil P may be preemptible, so we must
2370 // disable preemption before acquiring a P from pidleget below.
2375 // TODO(prattmic): All remaining calls to this function
2376 // with _p_ == nil could be cleaned up to find a P
2377 // before calling startm.
2378 throw("startm: P required for spinning=true")
2389 // No M is available, we must drop sched.lock and call newm.
2390 // However, we already own a P to assign to the M.
2392 // Once sched.lock is released, another G (e.g., in a syscall),
2393 // could find no idle P while checkdead finds a runnable G but
2394 // no running M's because this new M hasn't started yet, thus
2395 // throwing in an apparent deadlock.
2397 // Avoid this situation by pre-allocating the ID for the new M,
2398 // thus marking it as 'running' before we drop sched.lock. This
2399 // new M will eventually run the scheduler to execute any
2406 // The caller incremented nmspinning, so set m.spinning in the new M.
2410 // Ownership transfer of pp committed by start in newm.
2411 // Preemption is now safe.
2417 throw("startm: m is spinning")
2420 throw("startm: m has p")
2422 if spinning && !runqempty(pp) {
2423 throw("startm: p has runnable gs")
2425 // The caller incremented nmspinning, so set m.spinning in the new M.
2426 nmp.spinning = spinning
2428 notewakeup(&nmp.park)
2429 // Ownership transfer of pp committed by wakeup. Preemption is now
2434 // Hands off P from syscall or locked M.
2435 // Always runs without a P, so write barriers are not allowed.
2437 //go:nowritebarrierrec
2438 func handoffp(pp *p) {
2439 // handoffp must start an M in any situation where
2440 // findrunnable would return a G to run on pp.
2442 // if it has local work, start it straight away
2443 if !runqempty(pp) || sched.runqsize != 0 {
2447 // if there's trace work to do, start it straight away
2448 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2452 // if it has GC work, start it straight away
2453 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2457 // no local work, check that there are no spinning/idle M's,
2458 // otherwise our help is not required
2459 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2460 sched.needspinning.Store(0)
2465 if sched.gcwaiting.Load() {
2466 pp.status = _Pgcstop
2468 if sched.stopwait == 0 {
2469 notewakeup(&sched.stopnote)
2474 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2475 sched.safePointFn(pp)
2476 sched.safePointWait--
2477 if sched.safePointWait == 0 {
2478 notewakeup(&sched.safePointNote)
2481 if sched.runqsize != 0 {
2486 // If this is the last running P and nobody is polling network,
2487 // need to wakeup another M to poll network.
2488 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2494 // The scheduler lock cannot be held when calling wakeNetPoller below
2495 // because wakeNetPoller may call wakep which may call startm.
2496 when := nobarrierWakeTime(pp)
2505 // Tries to add one more P to execute G's.
2506 // Called when a G is made runnable (newproc, ready).
2507 // Must be called with a P.
2509 // Be conservative about spinning threads, only start one if none exist
2511 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2515 // Disable preemption until ownership of pp transfers to the next M in
2516 // startm. Otherwise preemption here would leave pp stuck waiting to
2519 // See preemption comment on acquirem in startm for more details.
2524 pp, _ = pidlegetSpinning(0)
2526 if sched.nmspinning.Add(-1) < 0 {
2527 throw("wakep: negative nmspinning")
2533 // Since we always have a P, the race in the "No M is available"
2534 // comment in startm doesn't apply during the small window between the
2535 // unlock here and lock in startm. A checkdead in between will always
2536 // see at least one running M (ours).
2544 // Stops execution of the current m that is locked to a g until the g is runnable again.
2545 // Returns with acquired P.
2546 func stoplockedm() {
2549 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2550 throw("stoplockedm: inconsistent locking")
2553 // Schedule another M to run this p.
2558 // Wait until another thread schedules lockedg again.
2560 status := readgstatus(gp.m.lockedg.ptr())
2561 if status&^_Gscan != _Grunnable {
2562 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2563 dumpgstatus(gp.m.lockedg.ptr())
2564 throw("stoplockedm: not runnable")
2566 acquirep(gp.m.nextp.ptr())
2570 // Schedules the locked m to run the locked gp.
2571 // May run during STW, so write barriers are not allowed.
2573 //go:nowritebarrierrec
2574 func startlockedm(gp *g) {
2575 mp := gp.lockedm.ptr()
2577 throw("startlockedm: locked to me")
2580 throw("startlockedm: m has p")
2582 // directly handoff current P to the locked m
2586 notewakeup(&mp.park)
2590 // Stops the current m for stopTheWorld.
2591 // Returns when the world is restarted.
2595 if !sched.gcwaiting.Load() {
2596 throw("gcstopm: not waiting for gc")
2599 gp.m.spinning = false
2600 // OK to just drop nmspinning here,
2601 // startTheWorld will unpark threads as necessary.
2602 if sched.nmspinning.Add(-1) < 0 {
2603 throw("gcstopm: negative nmspinning")
2608 pp.status = _Pgcstop
2610 if sched.stopwait == 0 {
2611 notewakeup(&sched.stopnote)
2617 // Schedules gp to run on the current M.
2618 // If inheritTime is true, gp inherits the remaining time in the
2619 // current time slice. Otherwise, it starts a new time slice.
2622 // Write barriers are allowed because this is called immediately after
2623 // acquiring a P in several places.
2625 //go:yeswritebarrierrec
2626 func execute(gp *g, inheritTime bool) {
2629 if goroutineProfile.active {
2630 // Make sure that gp has had its stack written out to the goroutine
2631 // profile, exactly as it was when the goroutine profiler first stopped
2633 tryRecordGoroutineProfile(gp, osyield)
2636 // Assign gp.m before entering _Grunning so running Gs have an
2640 casgstatus(gp, _Grunnable, _Grunning)
2643 gp.stackguard0 = gp.stack.lo + stackGuard
2645 mp.p.ptr().schedtick++
2648 // Check whether the profiler needs to be turned on or off.
2649 hz := sched.profilehz
2650 if mp.profilehz != hz {
2651 setThreadCPUProfiler(hz)
2655 // GoSysExit has to happen when we have a P, but before GoStart.
2656 // So we emit it here.
2657 if gp.syscallsp != 0 && gp.sysblocktraced {
2658 traceGoSysExit(gp.sysexitticks)
2666 // Finds a runnable goroutine to execute.
2667 // Tries to steal from other P's, get g from local or global queue, poll network.
2668 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2669 // reader) so the caller should try to wake a P.
2670 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2673 // The conditions here and in handoffp must agree: if
2674 // findrunnable would return a G to run, handoffp must start
2679 if sched.gcwaiting.Load() {
2683 if pp.runSafePointFn != 0 {
2687 // now and pollUntil are saved for work stealing later,
2688 // which may steal timers. It's important that between now
2689 // and then, nothing blocks, so these numbers remain mostly
2691 now, pollUntil, _ := checkTimers(pp, 0)
2693 // Try to schedule the trace reader.
2694 if trace.enabled || trace.shutdown {
2697 casgstatus(gp, _Gwaiting, _Grunnable)
2698 traceGoUnpark(gp, 0)
2699 return gp, false, true
2703 // Try to schedule a GC worker.
2704 if gcBlackenEnabled != 0 {
2705 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2707 return gp, false, true
2712 // Check the global runnable queue once in a while to ensure fairness.
2713 // Otherwise two goroutines can completely occupy the local runqueue
2714 // by constantly respawning each other.
2715 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2717 gp := globrunqget(pp, 1)
2720 return gp, false, false
2724 // Wake up the finalizer G.
2725 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2726 if gp := wakefing(); gp != nil {
2730 if *cgo_yield != nil {
2731 asmcgocall(*cgo_yield, nil)
2735 if gp, inheritTime := runqget(pp); gp != nil {
2736 return gp, inheritTime, false
2740 if sched.runqsize != 0 {
2742 gp := globrunqget(pp, 0)
2745 return gp, false, false
2750 // This netpoll is only an optimization before we resort to stealing.
2751 // We can safely skip it if there are no waiters or a thread is blocked
2752 // in netpoll already. If there is any kind of logical race with that
2753 // blocked thread (e.g. it has already returned from netpoll, but does
2754 // not set lastpoll yet), this thread will do blocking netpoll below
2756 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2757 if list := netpoll(0); !list.empty() { // non-blocking
2760 casgstatus(gp, _Gwaiting, _Grunnable)
2762 traceGoUnpark(gp, 0)
2764 return gp, false, false
2768 // Spinning Ms: steal work from other Ps.
2770 // Limit the number of spinning Ms to half the number of busy Ps.
2771 // This is necessary to prevent excessive CPU consumption when
2772 // GOMAXPROCS>>1 but the program parallelism is low.
2773 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2778 gp, inheritTime, tnow, w, newWork := stealWork(now)
2780 // Successfully stole.
2781 return gp, inheritTime, false
2784 // There may be new timer or GC work; restart to
2790 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2791 // Earlier timer to wait for.
2796 // We have nothing to do.
2798 // If we're in the GC mark phase, can safely scan and blacken objects,
2799 // and have work to do, run idle-time marking rather than give up the P.
2800 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2801 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2803 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2805 casgstatus(gp, _Gwaiting, _Grunnable)
2807 traceGoUnpark(gp, 0)
2809 return gp, false, false
2811 gcController.removeIdleMarkWorker()
2815 // If a callback returned and no other goroutine is awake,
2816 // then wake event handler goroutine which pauses execution
2817 // until a callback was triggered.
2818 gp, otherReady := beforeIdle(now, pollUntil)
2820 casgstatus(gp, _Gwaiting, _Grunnable)
2822 traceGoUnpark(gp, 0)
2824 return gp, false, false
2830 // Before we drop our P, make a snapshot of the allp slice,
2831 // which can change underfoot once we no longer block
2832 // safe-points. We don't need to snapshot the contents because
2833 // everything up to cap(allp) is immutable.
2834 allpSnapshot := allp
2835 // Also snapshot masks. Value changes are OK, but we can't allow
2836 // len to change out from under us.
2837 idlepMaskSnapshot := idlepMask
2838 timerpMaskSnapshot := timerpMask
2840 // return P and block
2842 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2846 if sched.runqsize != 0 {
2847 gp := globrunqget(pp, 0)
2849 return gp, false, false
2851 if !mp.spinning && sched.needspinning.Load() == 1 {
2852 // See "Delicate dance" comment below.
2857 if releasep() != pp {
2858 throw("findrunnable: wrong p")
2860 now = pidleput(pp, now)
2863 // Delicate dance: thread transitions from spinning to non-spinning
2864 // state, potentially concurrently with submission of new work. We must
2865 // drop nmspinning first and then check all sources again (with
2866 // #StoreLoad memory barrier in between). If we do it the other way
2867 // around, another thread can submit work after we've checked all
2868 // sources but before we drop nmspinning; as a result nobody will
2869 // unpark a thread to run the work.
2871 // This applies to the following sources of work:
2873 // * Goroutines added to a per-P run queue.
2874 // * New/modified-earlier timers on a per-P timer heap.
2875 // * Idle-priority GC work (barring golang.org/issue/19112).
2877 // If we discover new work below, we need to restore m.spinning as a
2878 // signal for resetspinning to unpark a new worker thread (because
2879 // there can be more than one starving goroutine).
2881 // However, if after discovering new work we also observe no idle Ps
2882 // (either here or in resetspinning), we have a problem. We may be
2883 // racing with a non-spinning M in the block above, having found no
2884 // work and preparing to release its P and park. Allowing that P to go
2885 // idle will result in loss of work conservation (idle P while there is
2886 // runnable work). This could result in complete deadlock in the
2887 // unlikely event that we discover new work (from netpoll) right as we
2888 // are racing with _all_ other Ps going idle.
2890 // We use sched.needspinning to synchronize with non-spinning Ms going
2891 // idle. If needspinning is set when they are about to drop their P,
2892 // they abort the drop and instead become a new spinning M on our
2893 // behalf. If we are not racing and the system is truly fully loaded
2894 // then no spinning threads are required, and the next thread to
2895 // naturally become spinning will clear the flag.
2897 // Also see "Worker thread parking/unparking" comment at the top of the
2899 wasSpinning := mp.spinning
2902 if sched.nmspinning.Add(-1) < 0 {
2903 throw("findrunnable: negative nmspinning")
2906 // Note the for correctness, only the last M transitioning from
2907 // spinning to non-spinning must perform these rechecks to
2908 // ensure no missed work. However, the runtime has some cases
2909 // of transient increments of nmspinning that are decremented
2910 // without going through this path, so we must be conservative
2911 // and perform the check on all spinning Ms.
2913 // See https://go.dev/issue/43997.
2915 // Check all runqueues once again.
2916 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2923 // Check for idle-priority GC work again.
2924 pp, gp := checkIdleGCNoP()
2929 // Run the idle worker.
2930 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2931 casgstatus(gp, _Gwaiting, _Grunnable)
2933 traceGoUnpark(gp, 0)
2935 return gp, false, false
2938 // Finally, check for timer creation or expiry concurrently with
2939 // transitioning from spinning to non-spinning.
2941 // Note that we cannot use checkTimers here because it calls
2942 // adjusttimers which may need to allocate memory, and that isn't
2943 // allowed when we don't have an active P.
2944 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
2947 // Poll network until next timer.
2948 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
2949 sched.pollUntil.Store(pollUntil)
2951 throw("findrunnable: netpoll with p")
2954 throw("findrunnable: netpoll with spinning")
2960 delay = pollUntil - now
2966 // When using fake time, just poll.
2969 list := netpoll(delay) // block until new work is available
2970 sched.pollUntil.Store(0)
2971 sched.lastpoll.Store(now)
2972 if faketime != 0 && list.empty() {
2973 // Using fake time and nothing is ready; stop M.
2974 // When all M's stop, checkdead will call timejump.
2979 pp, _ := pidleget(now)
2988 casgstatus(gp, _Gwaiting, _Grunnable)
2990 traceGoUnpark(gp, 0)
2992 return gp, false, false
2999 } else if pollUntil != 0 && netpollinited() {
3000 pollerPollUntil := sched.pollUntil.Load()
3001 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3009 // pollWork reports whether there is non-background work this P could
3010 // be doing. This is a fairly lightweight check to be used for
3011 // background work loops, like idle GC. It checks a subset of the
3012 // conditions checked by the actual scheduler.
3013 func pollWork() bool {
3014 if sched.runqsize != 0 {
3017 p := getg().m.p.ptr()
3021 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3022 if list := netpoll(0); !list.empty() {
3030 // stealWork attempts to steal a runnable goroutine or timer from any P.
3032 // If newWork is true, new work may have been readied.
3034 // If now is not 0 it is the current time. stealWork returns the passed time or
3035 // the current time if now was passed as 0.
3036 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3037 pp := getg().m.p.ptr()
3041 const stealTries = 4
3042 for i := 0; i < stealTries; i++ {
3043 stealTimersOrRunNextG := i == stealTries-1
3045 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3046 if sched.gcwaiting.Load() {
3047 // GC work may be available.
3048 return nil, false, now, pollUntil, true
3050 p2 := allp[enum.position()]
3055 // Steal timers from p2. This call to checkTimers is the only place
3056 // where we might hold a lock on a different P's timers. We do this
3057 // once on the last pass before checking runnext because stealing
3058 // from the other P's runnext should be the last resort, so if there
3059 // are timers to steal do that first.
3061 // We only check timers on one of the stealing iterations because
3062 // the time stored in now doesn't change in this loop and checking
3063 // the timers for each P more than once with the same value of now
3064 // is probably a waste of time.
3066 // timerpMask tells us whether the P may have timers at all. If it
3067 // can't, no need to check at all.
3068 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3069 tnow, w, ran := checkTimers(p2, now)
3071 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3075 // Running the timers may have
3076 // made an arbitrary number of G's
3077 // ready and added them to this P's
3078 // local run queue. That invalidates
3079 // the assumption of runqsteal
3080 // that it always has room to add
3081 // stolen G's. So check now if there
3082 // is a local G to run.
3083 if gp, inheritTime := runqget(pp); gp != nil {
3084 return gp, inheritTime, now, pollUntil, ranTimer
3090 // Don't bother to attempt to steal if p2 is idle.
3091 if !idlepMask.read(enum.position()) {
3092 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3093 return gp, false, now, pollUntil, ranTimer
3099 // No goroutines found to steal. Regardless, running a timer may have
3100 // made some goroutine ready that we missed. Indicate the next timer to
3102 return nil, false, now, pollUntil, ranTimer
3105 // Check all Ps for a runnable G to steal.
3107 // On entry we have no P. If a G is available to steal and a P is available,
3108 // the P is returned which the caller should acquire and attempt to steal the
3110 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3111 for id, p2 := range allpSnapshot {
3112 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3114 pp, _ := pidlegetSpinning(0)
3116 // Can't get a P, don't bother checking remaining Ps.
3125 // No work available.
3129 // Check all Ps for a timer expiring sooner than pollUntil.
3131 // Returns updated pollUntil value.
3132 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3133 for id, p2 := range allpSnapshot {
3134 if timerpMaskSnapshot.read(uint32(id)) {
3135 w := nobarrierWakeTime(p2)
3136 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3145 // Check for idle-priority GC, without a P on entry.
3147 // If some GC work, a P, and a worker G are all available, the P and G will be
3148 // returned. The returned P has not been wired yet.
3149 func checkIdleGCNoP() (*p, *g) {
3150 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3151 // must check again after acquiring a P. As an optimization, we also check
3152 // if an idle mark worker is needed at all. This is OK here, because if we
3153 // observe that one isn't needed, at least one is currently running. Even if
3154 // it stops running, its own journey into the scheduler should schedule it
3155 // again, if need be (at which point, this check will pass, if relevant).
3156 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3159 if !gcMarkWorkAvailable(nil) {
3163 // Work is available; we can start an idle GC worker only if there is
3164 // an available P and available worker G.
3166 // We can attempt to acquire these in either order, though both have
3167 // synchronization concerns (see below). Workers are almost always
3168 // available (see comment in findRunnableGCWorker for the one case
3169 // there may be none). Since we're slightly less likely to find a P,
3170 // check for that first.
3172 // Synchronization: note that we must hold sched.lock until we are
3173 // committed to keeping it. Otherwise we cannot put the unnecessary P
3174 // back in sched.pidle without performing the full set of idle
3175 // transition checks.
3177 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3178 // the assumption in gcControllerState.findRunnableGCWorker that an
3179 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3181 pp, now := pidlegetSpinning(0)
3187 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3188 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3194 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3198 gcController.removeIdleMarkWorker()
3204 return pp, node.gp.ptr()
3207 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3208 // going to wake up before the when argument; or it wakes an idle P to service
3209 // timers and the network poller if there isn't one already.
3210 func wakeNetPoller(when int64) {
3211 if sched.lastpoll.Load() == 0 {
3212 // In findrunnable we ensure that when polling the pollUntil
3213 // field is either zero or the time to which the current
3214 // poll is expected to run. This can have a spurious wakeup
3215 // but should never miss a wakeup.
3216 pollerPollUntil := sched.pollUntil.Load()
3217 if pollerPollUntil == 0 || pollerPollUntil > when {
3221 // There are no threads in the network poller, try to get
3222 // one there so it can handle new timers.
3223 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3229 func resetspinning() {
3232 throw("resetspinning: not a spinning m")
3234 gp.m.spinning = false
3235 nmspinning := sched.nmspinning.Add(-1)
3237 throw("findrunnable: negative nmspinning")
3239 // M wakeup policy is deliberately somewhat conservative, so check if we
3240 // need to wakeup another P here. See "Worker thread parking/unparking"
3241 // comment at the top of the file for details.
3245 // injectglist adds each runnable G on the list to some run queue,
3246 // and clears glist. If there is no current P, they are added to the
3247 // global queue, and up to npidle M's are started to run them.
3248 // Otherwise, for each idle P, this adds a G to the global queue
3249 // and starts an M. Any remaining G's are added to the current P's
3251 // This may temporarily acquire sched.lock.
3252 // Can run concurrently with GC.
3253 func injectglist(glist *gList) {
3258 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3259 traceGoUnpark(gp, 0)
3263 // Mark all the goroutines as runnable before we put them
3264 // on the run queues.
3265 head := glist.head.ptr()
3268 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3271 casgstatus(gp, _Gwaiting, _Grunnable)
3274 // Turn the gList into a gQueue.
3280 startIdle := func(n int) {
3281 for i := 0; i < n; i++ {
3282 mp := acquirem() // See comment in startm.
3285 pp, _ := pidlegetSpinning(0)
3298 pp := getg().m.p.ptr()
3301 globrunqputbatch(&q, int32(qsize))
3307 npidle := int(sched.npidle.Load())
3310 for n = 0; n < npidle && !q.empty(); n++ {
3316 globrunqputbatch(&globq, int32(n))
3323 runqputbatch(pp, &q, qsize)
3327 // One round of scheduler: find a runnable goroutine and execute it.
3333 throw("schedule: holding locks")
3336 if mp.lockedg != 0 {
3338 execute(mp.lockedg.ptr(), false) // Never returns.
3341 // We should not schedule away from a g that is executing a cgo call,
3342 // since the cgo call is using the m's g0 stack.
3344 throw("schedule: in cgo")
3351 // Safety check: if we are spinning, the run queue should be empty.
3352 // Check this before calling checkTimers, as that might call
3353 // goready to put a ready goroutine on the local run queue.
3354 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3355 throw("schedule: spinning with local work")
3358 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3360 // This thread is going to run a goroutine and is not spinning anymore,
3361 // so if it was marked as spinning we need to reset it now and potentially
3362 // start a new spinning M.
3367 if sched.disable.user && !schedEnabled(gp) {
3368 // Scheduling of this goroutine is disabled. Put it on
3369 // the list of pending runnable goroutines for when we
3370 // re-enable user scheduling and look again.
3372 if schedEnabled(gp) {
3373 // Something re-enabled scheduling while we
3374 // were acquiring the lock.
3377 sched.disable.runnable.pushBack(gp)
3384 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3385 // wake a P if there is one.
3389 if gp.lockedm != 0 {
3390 // Hands off own p to the locked m,
3391 // then blocks waiting for a new p.
3396 execute(gp, inheritTime)
3399 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3400 // Typically a caller sets gp's status away from Grunning and then
3401 // immediately calls dropg to finish the job. The caller is also responsible
3402 // for arranging that gp will be restarted using ready at an
3403 // appropriate time. After calling dropg and arranging for gp to be
3404 // readied later, the caller can do other work but eventually should
3405 // call schedule to restart the scheduling of goroutines on this m.
3409 setMNoWB(&gp.m.curg.m, nil)
3410 setGNoWB(&gp.m.curg, nil)
3413 // checkTimers runs any timers for the P that are ready.
3414 // If now is not 0 it is the current time.
3415 // It returns the passed time or the current time if now was passed as 0.
3416 // and the time when the next timer should run or 0 if there is no next timer,
3417 // and reports whether it ran any timers.
3418 // If the time when the next timer should run is not 0,
3419 // it is always larger than the returned time.
3420 // We pass now in and out to avoid extra calls of nanotime.
3422 //go:yeswritebarrierrec
3423 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3424 // If it's not yet time for the first timer, or the first adjusted
3425 // timer, then there is nothing to do.
3426 next := pp.timer0When.Load()
3427 nextAdj := pp.timerModifiedEarliest.Load()
3428 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3433 // No timers to run or adjust.
3434 return now, 0, false
3441 // Next timer is not ready to run, but keep going
3442 // if we would clear deleted timers.
3443 // This corresponds to the condition below where
3444 // we decide whether to call clearDeletedTimers.
3445 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3446 return now, next, false
3450 lock(&pp.timersLock)
3452 if len(pp.timers) > 0 {
3453 adjusttimers(pp, now)
3454 for len(pp.timers) > 0 {
3455 // Note that runtimer may temporarily unlock
3457 if tw := runtimer(pp, now); tw != 0 {
3467 // If this is the local P, and there are a lot of deleted timers,
3468 // clear them out. We only do this for the local P to reduce
3469 // lock contention on timersLock.
3470 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3471 clearDeletedTimers(pp)
3474 unlock(&pp.timersLock)
3476 return now, pollUntil, ran
3479 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3480 unlock((*mutex)(lock))
3484 // park continuation on g0.
3485 func park_m(gp *g) {
3489 traceGoPark(mp.waittraceev, mp.waittraceskip)
3492 // N.B. Not using casGToWaiting here because the waitreason is
3493 // set by park_m's caller.
3494 casgstatus(gp, _Grunning, _Gwaiting)
3497 if fn := mp.waitunlockf; fn != nil {
3498 ok := fn(gp, mp.waitlock)
3499 mp.waitunlockf = nil
3503 traceGoUnpark(gp, 2)
3505 casgstatus(gp, _Gwaiting, _Grunnable)
3506 execute(gp, true) // Schedule it back, never returns.
3512 func goschedImpl(gp *g) {
3513 status := readgstatus(gp)
3514 if status&^_Gscan != _Grunning {
3516 throw("bad g status")
3518 casgstatus(gp, _Grunning, _Grunnable)
3527 // Gosched continuation on g0.
3528 func gosched_m(gp *g) {
3535 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3536 func goschedguarded_m(gp *g) {
3538 if !canPreemptM(gp.m) {
3539 gogo(&gp.sched) // never return
3548 func gopreempt_m(gp *g) {
3555 // preemptPark parks gp and puts it in _Gpreempted.
3558 func preemptPark(gp *g) {
3560 traceGoPark(traceEvGoBlock, 0)
3562 status := readgstatus(gp)
3563 if status&^_Gscan != _Grunning {
3565 throw("bad g status")
3568 if gp.asyncSafePoint {
3569 // Double-check that async preemption does not
3570 // happen in SPWRITE assembly functions.
3571 // isAsyncSafePoint must exclude this case.
3572 f := findfunc(gp.sched.pc)
3574 throw("preempt at unknown pc")
3576 if f.flag&funcFlag_SPWRITE != 0 {
3577 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3578 throw("preempt SPWRITE")
3582 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3583 // be in _Grunning when we dropg because then we'd be running
3584 // without an M, but the moment we're in _Gpreempted,
3585 // something could claim this G before we've fully cleaned it
3586 // up. Hence, we set the scan bit to lock down further
3587 // transitions until we can dropg.
3588 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3590 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3594 // goyield is like Gosched, but it:
3595 // - emits a GoPreempt trace event instead of a GoSched trace event
3596 // - puts the current G on the runq of the current P instead of the globrunq
3602 func goyield_m(gp *g) {
3607 casgstatus(gp, _Grunning, _Grunnable)
3609 runqput(pp, gp, false)
3613 // Finishes execution of the current goroutine.
3624 // goexit continuation on g0.
3625 func goexit0(gp *g) {
3629 casgstatus(gp, _Grunning, _Gdead)
3630 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3631 if isSystemGoroutine(gp, false) {
3635 locked := gp.lockedm != 0
3638 gp.preemptStop = false
3639 gp.paniconfault = false
3640 gp._defer = nil // should be true already but just in case.
3641 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3643 gp.waitreason = waitReasonZero
3648 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3649 // Flush assist credit to the global pool. This gives
3650 // better information to pacing if the application is
3651 // rapidly creating an exiting goroutines.
3652 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3653 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3654 gcController.bgScanCredit.Add(scanCredit)
3655 gp.gcAssistBytes = 0
3660 if GOARCH == "wasm" { // no threads yet on wasm
3662 schedule() // never returns
3665 if mp.lockedInt != 0 {
3666 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3667 throw("internal lockOSThread error")
3671 // The goroutine may have locked this thread because
3672 // it put it in an unusual kernel state. Kill it
3673 // rather than returning it to the thread pool.
3675 // Return to mstart, which will release the P and exit
3677 if GOOS != "plan9" { // See golang.org/issue/22227.
3680 // Clear lockedExt on plan9 since we may end up re-using
3688 // save updates getg().sched to refer to pc and sp so that a following
3689 // gogo will restore pc and sp.
3691 // save must not have write barriers because invoking a write barrier
3692 // can clobber getg().sched.
3695 //go:nowritebarrierrec
3696 func save(pc, sp uintptr) {
3699 if gp == gp.m.g0 || gp == gp.m.gsignal {
3700 // m.g0.sched is special and must describe the context
3701 // for exiting the thread. mstart1 writes to it directly.
3702 // m.gsignal.sched should not be used at all.
3703 // This check makes sure save calls do not accidentally
3704 // run in contexts where they'd write to system g's.
3705 throw("save on system g not allowed")
3712 // We need to ensure ctxt is zero, but can't have a write
3713 // barrier here. However, it should always already be zero.
3715 if gp.sched.ctxt != nil {
3720 // The goroutine g is about to enter a system call.
3721 // Record that it's not using the cpu anymore.
3722 // This is called only from the go syscall library and cgocall,
3723 // not from the low-level system calls used by the runtime.
3725 // Entersyscall cannot split the stack: the save must
3726 // make g->sched refer to the caller's stack segment, because
3727 // entersyscall is going to return immediately after.
3729 // Nothing entersyscall calls can split the stack either.
3730 // We cannot safely move the stack during an active call to syscall,
3731 // because we do not know which of the uintptr arguments are
3732 // really pointers (back into the stack).
3733 // In practice, this means that we make the fast path run through
3734 // entersyscall doing no-split things, and the slow path has to use systemstack
3735 // to run bigger things on the system stack.
3737 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3738 // saved SP and PC are restored. This is needed when exitsyscall will be called
3739 // from a function further up in the call stack than the parent, as g->syscallsp
3740 // must always point to a valid stack frame. entersyscall below is the normal
3741 // entry point for syscalls, which obtains the SP and PC from the caller.
3744 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3745 // If the syscall does not block, that is it, we do not emit any other events.
3746 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3747 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3748 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3749 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3750 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3751 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3752 // and we wait for the increment before emitting traceGoSysExit.
3753 // Note that the increment is done even if tracing is not enabled,
3754 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3757 func reentersyscall(pc, sp uintptr) {
3760 // Disable preemption because during this function g is in Gsyscall status,
3761 // but can have inconsistent g->sched, do not let GC observe it.
3764 // Entersyscall must not call any function that might split/grow the stack.
3765 // (See details in comment above.)
3766 // Catch calls that might, by replacing the stack guard with something that
3767 // will trip any stack check and leaving a flag to tell newstack to die.
3768 gp.stackguard0 = stackPreempt
3769 gp.throwsplit = true
3771 // Leave SP around for GC and traceback.
3775 casgstatus(gp, _Grunning, _Gsyscall)
3776 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3777 systemstack(func() {
3778 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3779 throw("entersyscall")
3784 systemstack(traceGoSysCall)
3785 // systemstack itself clobbers g.sched.{pc,sp} and we might
3786 // need them later when the G is genuinely blocked in a
3791 if sched.sysmonwait.Load() {
3792 systemstack(entersyscall_sysmon)
3796 if gp.m.p.ptr().runSafePointFn != 0 {
3797 // runSafePointFn may stack split if run on this stack
3798 systemstack(runSafePointFn)
3802 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3803 gp.sysblocktraced = true
3808 atomic.Store(&pp.status, _Psyscall)
3809 if sched.gcwaiting.Load() {
3810 systemstack(entersyscall_gcwait)
3817 // Standard syscall entry used by the go syscall library and normal cgo calls.
3819 // This is exported via linkname to assembly in the syscall package and x/sys.
3822 //go:linkname entersyscall
3823 func entersyscall() {
3824 reentersyscall(getcallerpc(), getcallersp())
3827 func entersyscall_sysmon() {
3829 if sched.sysmonwait.Load() {
3830 sched.sysmonwait.Store(false)
3831 notewakeup(&sched.sysmonnote)
3836 func entersyscall_gcwait() {
3838 pp := gp.m.oldp.ptr()
3841 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3847 if sched.stopwait--; sched.stopwait == 0 {
3848 notewakeup(&sched.stopnote)
3854 // The same as entersyscall(), but with a hint that the syscall is blocking.
3857 func entersyscallblock() {
3860 gp.m.locks++ // see comment in entersyscall
3861 gp.throwsplit = true
3862 gp.stackguard0 = stackPreempt // see comment in entersyscall
3863 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3864 gp.sysblocktraced = true
3865 gp.m.p.ptr().syscalltick++
3867 // Leave SP around for GC and traceback.
3871 gp.syscallsp = gp.sched.sp
3872 gp.syscallpc = gp.sched.pc
3873 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3877 systemstack(func() {
3878 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3879 throw("entersyscallblock")
3882 casgstatus(gp, _Grunning, _Gsyscall)
3883 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3884 systemstack(func() {
3885 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3886 throw("entersyscallblock")
3890 systemstack(entersyscallblock_handoff)
3892 // Resave for traceback during blocked call.
3893 save(getcallerpc(), getcallersp())
3898 func entersyscallblock_handoff() {
3901 traceGoSysBlock(getg().m.p.ptr())
3903 handoffp(releasep())
3906 // The goroutine g exited its system call.
3907 // Arrange for it to run on a cpu again.
3908 // This is called only from the go syscall library, not
3909 // from the low-level system calls used by the runtime.
3911 // Write barriers are not allowed because our P may have been stolen.
3913 // This is exported via linkname to assembly in the syscall package.
3916 //go:nowritebarrierrec
3917 //go:linkname exitsyscall
3918 func exitsyscall() {
3921 gp.m.locks++ // see comment in entersyscall
3922 if getcallersp() > gp.syscallsp {
3923 throw("exitsyscall: syscall frame is no longer valid")
3927 oldp := gp.m.oldp.ptr()
3929 if exitsyscallfast(oldp) {
3930 // When exitsyscallfast returns success, we have a P so can now use
3932 if goroutineProfile.active {
3933 // Make sure that gp has had its stack written out to the goroutine
3934 // profile, exactly as it was when the goroutine profiler first
3935 // stopped the world.
3936 systemstack(func() {
3937 tryRecordGoroutineProfileWB(gp)
3941 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3942 systemstack(traceGoStart)
3945 // There's a cpu for us, so we can run.
3946 gp.m.p.ptr().syscalltick++
3947 // We need to cas the status and scan before resuming...
3948 casgstatus(gp, _Gsyscall, _Grunning)
3950 // Garbage collector isn't running (since we are),
3951 // so okay to clear syscallsp.
3955 // restore the preemption request in case we've cleared it in newstack
3956 gp.stackguard0 = stackPreempt
3958 // otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
3959 gp.stackguard0 = gp.stack.lo + stackGuard
3961 gp.throwsplit = false
3963 if sched.disable.user && !schedEnabled(gp) {
3964 // Scheduling of this goroutine is disabled.
3973 // Wait till traceGoSysBlock event is emitted.
3974 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3975 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
3978 // We can't trace syscall exit right now because we don't have a P.
3979 // Tracing code can invoke write barriers that cannot run without a P.
3980 // So instead we remember the syscall exit time and emit the event
3981 // in execute when we have a P.
3982 gp.sysexitticks = cputicks()
3987 // Call the scheduler.
3990 // Scheduler returned, so we're allowed to run now.
3991 // Delete the syscallsp information that we left for
3992 // the garbage collector during the system call.
3993 // Must wait until now because until gosched returns
3994 // we don't know for sure that the garbage collector
3997 gp.m.p.ptr().syscalltick++
3998 gp.throwsplit = false
4002 func exitsyscallfast(oldp *p) bool {
4005 // Freezetheworld sets stopwait but does not retake P's.
4006 if sched.stopwait == freezeStopWait {
4010 // Try to re-acquire the last P.
4011 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4012 // There's a cpu for us, so we can run.
4014 exitsyscallfast_reacquired()
4018 // Try to get any other idle P.
4019 if sched.pidle != 0 {
4021 systemstack(func() {
4022 ok = exitsyscallfast_pidle()
4023 if ok && trace.enabled {
4025 // Wait till traceGoSysBlock event is emitted.
4026 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4027 for oldp.syscalltick == gp.m.syscalltick {
4041 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4042 // has successfully reacquired the P it was running on before the
4046 func exitsyscallfast_reacquired() {
4048 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4050 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4051 // traceGoSysBlock for this syscall was already emitted,
4052 // but here we effectively retake the p from the new syscall running on the same p.
4053 systemstack(func() {
4054 // Denote blocking of the new syscall.
4055 traceGoSysBlock(gp.m.p.ptr())
4056 // Denote completion of the current syscall.
4060 gp.m.p.ptr().syscalltick++
4064 func exitsyscallfast_pidle() bool {
4066 pp, _ := pidleget(0)
4067 if pp != nil && sched.sysmonwait.Load() {
4068 sched.sysmonwait.Store(false)
4069 notewakeup(&sched.sysmonnote)
4079 // exitsyscall slow path on g0.
4080 // Failed to acquire P, enqueue gp as runnable.
4082 // Called via mcall, so gp is the calling g from this M.
4084 //go:nowritebarrierrec
4085 func exitsyscall0(gp *g) {
4086 casgstatus(gp, _Gsyscall, _Grunnable)
4090 if schedEnabled(gp) {
4097 // Below, we stoplockedm if gp is locked. globrunqput releases
4098 // ownership of gp, so we must check if gp is locked prior to
4099 // committing the release by unlocking sched.lock, otherwise we
4100 // could race with another M transitioning gp from unlocked to
4102 locked = gp.lockedm != 0
4103 } else if sched.sysmonwait.Load() {
4104 sched.sysmonwait.Store(false)
4105 notewakeup(&sched.sysmonnote)
4110 execute(gp, false) // Never returns.
4113 // Wait until another thread schedules gp and so m again.
4115 // N.B. lockedm must be this M, as this g was running on this M
4116 // before entersyscall.
4118 execute(gp, false) // Never returns.
4121 schedule() // Never returns.
4124 // Called from syscall package before fork.
4126 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4128 func syscall_runtime_BeforeFork() {
4131 // Block signals during a fork, so that the child does not run
4132 // a signal handler before exec if a signal is sent to the process
4133 // group. See issue #18600.
4135 sigsave(&gp.m.sigmask)
4138 // This function is called before fork in syscall package.
4139 // Code between fork and exec must not allocate memory nor even try to grow stack.
4140 // Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
4141 // runtime_AfterFork will undo this in parent process, but not in child.
4142 gp.stackguard0 = stackFork
4145 // Called from syscall package after fork in parent.
4147 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4149 func syscall_runtime_AfterFork() {
4152 // See the comments in beforefork.
4153 gp.stackguard0 = gp.stack.lo + stackGuard
4155 msigrestore(gp.m.sigmask)
4160 // inForkedChild is true while manipulating signals in the child process.
4161 // This is used to avoid calling libc functions in case we are using vfork.
4162 var inForkedChild bool
4164 // Called from syscall package after fork in child.
4165 // It resets non-sigignored signals to the default handler, and
4166 // restores the signal mask in preparation for the exec.
4168 // Because this might be called during a vfork, and therefore may be
4169 // temporarily sharing address space with the parent process, this must
4170 // not change any global variables or calling into C code that may do so.
4172 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4174 //go:nowritebarrierrec
4175 func syscall_runtime_AfterForkInChild() {
4176 // It's OK to change the global variable inForkedChild here
4177 // because we are going to change it back. There is no race here,
4178 // because if we are sharing address space with the parent process,
4179 // then the parent process can not be running concurrently.
4180 inForkedChild = true
4182 clearSignalHandlers()
4184 // When we are the child we are the only thread running,
4185 // so we know that nothing else has changed gp.m.sigmask.
4186 msigrestore(getg().m.sigmask)
4188 inForkedChild = false
4191 // pendingPreemptSignals is the number of preemption signals
4192 // that have been sent but not received. This is only used on Darwin.
4194 var pendingPreemptSignals atomic.Int32
4196 // Called from syscall package before Exec.
4198 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4199 func syscall_runtime_BeforeExec() {
4200 // Prevent thread creation during exec.
4203 // On Darwin, wait for all pending preemption signals to
4204 // be received. See issue #41702.
4205 if GOOS == "darwin" || GOOS == "ios" {
4206 for pendingPreemptSignals.Load() > 0 {
4212 // Called from syscall package after Exec.
4214 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4215 func syscall_runtime_AfterExec() {
4219 // Allocate a new g, with a stack big enough for stacksize bytes.
4220 func malg(stacksize int32) *g {
4223 stacksize = round2(stackSystem + stacksize)
4224 systemstack(func() {
4225 newg.stack = stackalloc(uint32(stacksize))
4227 newg.stackguard0 = newg.stack.lo + stackGuard
4228 newg.stackguard1 = ^uintptr(0)
4229 // Clear the bottom word of the stack. We record g
4230 // there on gsignal stack during VDSO on ARM and ARM64.
4231 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4236 // Create a new g running fn.
4237 // Put it on the queue of g's waiting to run.
4238 // The compiler turns a go statement into a call to this.
4239 func newproc(fn *funcval) {
4242 systemstack(func() {
4243 newg := newproc1(fn, gp, pc)
4245 pp := getg().m.p.ptr()
4246 runqput(pp, newg, true)
4254 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4255 // address of the go statement that created this. The caller is responsible
4256 // for adding the new g to the scheduler.
4257 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4259 fatal("go of nil func value")
4262 mp := acquirem() // disable preemption because we hold M and P in local vars.
4266 newg = malg(stackMin)
4267 casgstatus(newg, _Gidle, _Gdead)
4268 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4270 if newg.stack.hi == 0 {
4271 throw("newproc1: newg missing stack")
4274 if readgstatus(newg) != _Gdead {
4275 throw("newproc1: new g is not Gdead")
4278 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4279 totalSize = alignUp(totalSize, sys.StackAlign)
4280 sp := newg.stack.hi - totalSize
4284 *(*uintptr)(unsafe.Pointer(sp)) = 0
4286 spArg += sys.MinFrameSize
4289 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4292 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4293 newg.sched.g = guintptr(unsafe.Pointer(newg))
4294 gostartcallfn(&newg.sched, fn)
4295 newg.parentGoid = callergp.goid
4296 newg.gopc = callerpc
4297 newg.ancestors = saveAncestors(callergp)
4298 newg.startpc = fn.fn
4299 if isSystemGoroutine(newg, false) {
4302 // Only user goroutines inherit pprof labels.
4304 newg.labels = mp.curg.labels
4306 if goroutineProfile.active {
4307 // A concurrent goroutine profile is running. It should include
4308 // exactly the set of goroutines that were alive when the goroutine
4309 // profiler first stopped the world. That does not include newg, so
4310 // mark it as not needing a profile before transitioning it from
4312 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4315 // Track initial transition?
4316 newg.trackingSeq = uint8(fastrand())
4317 if newg.trackingSeq%gTrackingPeriod == 0 {
4318 newg.tracking = true
4320 casgstatus(newg, _Gdead, _Grunnable)
4321 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4323 if pp.goidcache == pp.goidcacheend {
4324 // Sched.goidgen is the last allocated id,
4325 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4326 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4327 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4328 pp.goidcache -= _GoidCacheBatch - 1
4329 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4331 newg.goid = pp.goidcache
4334 newg.racectx = racegostart(callerpc)
4335 if newg.labels != nil {
4336 // See note in proflabel.go on labelSync's role in synchronizing
4337 // with the reads in the signal handler.
4338 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4342 traceGoCreate(newg, newg.startpc)
4349 // saveAncestors copies previous ancestors of the given caller g and
4350 // includes info for the current caller into a new set of tracebacks for
4351 // a g being created.
4352 func saveAncestors(callergp *g) *[]ancestorInfo {
4353 // Copy all prior info, except for the root goroutine (goid 0).
4354 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4357 var callerAncestors []ancestorInfo
4358 if callergp.ancestors != nil {
4359 callerAncestors = *callergp.ancestors
4361 n := int32(len(callerAncestors)) + 1
4362 if n > debug.tracebackancestors {
4363 n = debug.tracebackancestors
4365 ancestors := make([]ancestorInfo, n)
4366 copy(ancestors[1:], callerAncestors)
4368 var pcs [tracebackInnerFrames]uintptr
4369 npcs := gcallers(callergp, 0, pcs[:])
4370 ipcs := make([]uintptr, npcs)
4372 ancestors[0] = ancestorInfo{
4374 goid: callergp.goid,
4375 gopc: callergp.gopc,
4378 ancestorsp := new([]ancestorInfo)
4379 *ancestorsp = ancestors
4383 // Put on gfree list.
4384 // If local list is too long, transfer a batch to the global list.
4385 func gfput(pp *p, gp *g) {
4386 if readgstatus(gp) != _Gdead {
4387 throw("gfput: bad status (not Gdead)")
4390 stksize := gp.stack.hi - gp.stack.lo
4392 if stksize != uintptr(startingStackSize) {
4393 // non-standard stack size - free it.
4402 if pp.gFree.n >= 64 {
4408 for pp.gFree.n >= 32 {
4409 gp := pp.gFree.pop()
4411 if gp.stack.lo == 0 {
4418 lock(&sched.gFree.lock)
4419 sched.gFree.noStack.pushAll(noStackQ)
4420 sched.gFree.stack.pushAll(stackQ)
4421 sched.gFree.n += inc
4422 unlock(&sched.gFree.lock)
4426 // Get from gfree list.
4427 // If local list is empty, grab a batch from global list.
4428 func gfget(pp *p) *g {
4430 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4431 lock(&sched.gFree.lock)
4432 // Move a batch of free Gs to the P.
4433 for pp.gFree.n < 32 {
4434 // Prefer Gs with stacks.
4435 gp := sched.gFree.stack.pop()
4437 gp = sched.gFree.noStack.pop()
4446 unlock(&sched.gFree.lock)
4449 gp := pp.gFree.pop()
4454 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4455 // Deallocate old stack. We kept it in gfput because it was the
4456 // right size when the goroutine was put on the free list, but
4457 // the right size has changed since then.
4458 systemstack(func() {
4465 if gp.stack.lo == 0 {
4466 // Stack was deallocated in gfput or just above. Allocate a new one.
4467 systemstack(func() {
4468 gp.stack = stackalloc(startingStackSize)
4470 gp.stackguard0 = gp.stack.lo + stackGuard
4473 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4476 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4479 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4485 // Purge all cached G's from gfree list to the global list.
4486 func gfpurge(pp *p) {
4492 for !pp.gFree.empty() {
4493 gp := pp.gFree.pop()
4495 if gp.stack.lo == 0 {
4502 lock(&sched.gFree.lock)
4503 sched.gFree.noStack.pushAll(noStackQ)
4504 sched.gFree.stack.pushAll(stackQ)
4505 sched.gFree.n += inc
4506 unlock(&sched.gFree.lock)
4509 // Breakpoint executes a breakpoint trap.
4514 // dolockOSThread is called by LockOSThread and lockOSThread below
4515 // after they modify m.locked. Do not allow preemption during this call,
4516 // or else the m might be different in this function than in the caller.
4519 func dolockOSThread() {
4520 if GOARCH == "wasm" {
4521 return // no threads on wasm yet
4524 gp.m.lockedg.set(gp)
4525 gp.lockedm.set(gp.m)
4528 // LockOSThread wires the calling goroutine to its current operating system thread.
4529 // The calling goroutine will always execute in that thread,
4530 // and no other goroutine will execute in it,
4531 // until the calling goroutine has made as many calls to
4532 // UnlockOSThread as to LockOSThread.
4533 // If the calling goroutine exits without unlocking the thread,
4534 // the thread will be terminated.
4536 // All init functions are run on the startup thread. Calling LockOSThread
4537 // from an init function will cause the main function to be invoked on
4540 // A goroutine should call LockOSThread before calling OS services or
4541 // non-Go library functions that depend on per-thread state.
4544 func LockOSThread() {
4545 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4546 // If we need to start a new thread from the locked
4547 // thread, we need the template thread. Start it now
4548 // while we're in a known-good state.
4549 startTemplateThread()
4553 if gp.m.lockedExt == 0 {
4555 panic("LockOSThread nesting overflow")
4561 func lockOSThread() {
4562 getg().m.lockedInt++
4566 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4567 // after they update m->locked. Do not allow preemption during this call,
4568 // or else the m might be in different in this function than in the caller.
4571 func dounlockOSThread() {
4572 if GOARCH == "wasm" {
4573 return // no threads on wasm yet
4576 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4583 // UnlockOSThread undoes an earlier call to LockOSThread.
4584 // If this drops the number of active LockOSThread calls on the
4585 // calling goroutine to zero, it unwires the calling goroutine from
4586 // its fixed operating system thread.
4587 // If there are no active LockOSThread calls, this is a no-op.
4589 // Before calling UnlockOSThread, the caller must ensure that the OS
4590 // thread is suitable for running other goroutines. If the caller made
4591 // any permanent changes to the state of the thread that would affect
4592 // other goroutines, it should not call this function and thus leave
4593 // the goroutine locked to the OS thread until the goroutine (and
4594 // hence the thread) exits.
4597 func UnlockOSThread() {
4599 if gp.m.lockedExt == 0 {
4607 func unlockOSThread() {
4609 if gp.m.lockedInt == 0 {
4610 systemstack(badunlockosthread)
4616 func badunlockosthread() {
4617 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4620 func gcount() int32 {
4621 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4622 for _, pp := range allp {
4626 // All these variables can be changed concurrently, so the result can be inconsistent.
4627 // But at least the current goroutine is running.
4634 func mcount() int32 {
4635 return int32(sched.mnext - sched.nmfreed)
4639 signalLock atomic.Uint32
4641 // Must hold signalLock to write. Reads may be lock-free, but
4642 // signalLock should be taken to synchronize with changes.
4646 func _System() { _System() }
4647 func _ExternalCode() { _ExternalCode() }
4648 func _LostExternalCode() { _LostExternalCode() }
4649 func _GC() { _GC() }
4650 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4651 func _VDSO() { _VDSO() }
4653 // Called if we receive a SIGPROF signal.
4654 // Called by the signal handler, may run during STW.
4656 //go:nowritebarrierrec
4657 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4658 if prof.hz.Load() == 0 {
4662 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4663 // We must check this to avoid a deadlock between setcpuprofilerate
4664 // and the call to cpuprof.add, below.
4665 if mp != nil && mp.profilehz == 0 {
4669 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4670 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4671 // the critical section, it creates a deadlock (when writing the sample).
4672 // As a workaround, create a counter of SIGPROFs while in critical section
4673 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4674 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4675 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4676 if f := findfunc(pc); f.valid() {
4677 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4678 cpuprof.lostAtomic++
4682 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4683 // runtime/internal/atomic functions call into kernel
4684 // helpers on arm < 7. See
4685 // runtime/internal/atomic/sys_linux_arm.s.
4686 cpuprof.lostAtomic++
4691 // Profiling runs concurrently with GC, so it must not allocate.
4692 // Set a trap in case the code does allocate.
4693 // Note that on windows, one thread takes profiles of all the
4694 // other threads, so mp is usually not getg().m.
4695 // In fact mp may not even be stopped.
4696 // See golang.org/issue/17165.
4697 getg().m.mallocing++
4700 var stk [maxCPUProfStack]uintptr
4702 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4704 // Check cgoCallersUse to make sure that we are not
4705 // interrupting other code that is fiddling with
4706 // cgoCallers. We are running in a signal handler
4707 // with all signals blocked, so we don't have to worry
4708 // about any other code interrupting us.
4709 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4710 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4713 n += copy(stk[:], mp.cgoCallers[:cgoOff])
4714 mp.cgoCallers[0] = 0
4717 // Collect Go stack that leads to the cgo call.
4718 u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
4719 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4720 // Libcall, i.e. runtime syscall on windows.
4721 // Collect Go stack that leads to the call.
4722 u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
4723 } else if mp != nil && mp.vdsoSP != 0 {
4724 // VDSO call, e.g. nanotime1 on Linux.
4725 // Collect Go stack that leads to the call.
4726 u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
4728 u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
4730 n += tracebackPCs(&u, 0, stk[n:])
4733 // Normal traceback is impossible or has failed.
4734 // Account it against abstract "System" or "GC".
4737 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4738 } else if pc > firstmoduledata.etext {
4739 // "ExternalCode" is better than "etext".
4740 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4743 if mp.preemptoff != "" {
4744 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4746 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4750 if prof.hz.Load() != 0 {
4751 // Note: it can happen on Windows that we interrupted a system thread
4752 // with no g, so gp could nil. The other nil checks are done out of
4753 // caution, but not expected to be nil in practice.
4754 var tagPtr *unsafe.Pointer
4755 if gp != nil && gp.m != nil && gp.m.curg != nil {
4756 tagPtr = &gp.m.curg.labels
4758 cpuprof.add(tagPtr, stk[:n])
4762 if gp != nil && gp.m != nil {
4763 if gp.m.curg != nil {
4768 traceCPUSample(gprof, pp, stk[:n])
4770 getg().m.mallocing--
4773 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4774 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4775 func setcpuprofilerate(hz int32) {
4776 // Force sane arguments.
4781 // Disable preemption, otherwise we can be rescheduled to another thread
4782 // that has profiling enabled.
4786 // Stop profiler on this thread so that it is safe to lock prof.
4787 // if a profiling signal came in while we had prof locked,
4788 // it would deadlock.
4789 setThreadCPUProfiler(0)
4791 for !prof.signalLock.CompareAndSwap(0, 1) {
4794 if prof.hz.Load() != hz {
4795 setProcessCPUProfiler(hz)
4798 prof.signalLock.Store(0)
4801 sched.profilehz = hz
4805 setThreadCPUProfiler(hz)
4811 // init initializes pp, which may be a freshly allocated p or a
4812 // previously destroyed p, and transitions it to status _Pgcstop.
4813 func (pp *p) init(id int32) {
4815 pp.status = _Pgcstop
4816 pp.sudogcache = pp.sudogbuf[:0]
4817 pp.deferpool = pp.deferpoolbuf[:0]
4819 if pp.mcache == nil {
4822 throw("missing mcache?")
4824 // Use the bootstrap mcache0. Only one P will get
4825 // mcache0: the one with ID 0.
4828 pp.mcache = allocmcache()
4831 if raceenabled && pp.raceprocctx == 0 {
4833 pp.raceprocctx = raceprocctx0
4834 raceprocctx0 = 0 // bootstrap
4836 pp.raceprocctx = raceproccreate()
4839 lockInit(&pp.timersLock, lockRankTimers)
4841 // This P may get timers when it starts running. Set the mask here
4842 // since the P may not go through pidleget (notably P 0 on startup).
4844 // Similarly, we may not go through pidleget before this P starts
4845 // running if it is P 0 on startup.
4849 // destroy releases all of the resources associated with pp and
4850 // transitions it to status _Pdead.
4852 // sched.lock must be held and the world must be stopped.
4853 func (pp *p) destroy() {
4854 assertLockHeld(&sched.lock)
4855 assertWorldStopped()
4857 // Move all runnable goroutines to the global queue
4858 for pp.runqhead != pp.runqtail {
4859 // Pop from tail of local queue
4861 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4862 // Push onto head of global queue
4865 if pp.runnext != 0 {
4866 globrunqputhead(pp.runnext.ptr())
4869 if len(pp.timers) > 0 {
4870 plocal := getg().m.p.ptr()
4871 // The world is stopped, but we acquire timersLock to
4872 // protect against sysmon calling timeSleepUntil.
4873 // This is the only case where we hold the timersLock of
4874 // more than one P, so there are no deadlock concerns.
4875 lock(&plocal.timersLock)
4876 lock(&pp.timersLock)
4877 moveTimers(plocal, pp.timers)
4879 pp.numTimers.Store(0)
4880 pp.deletedTimers.Store(0)
4881 pp.timer0When.Store(0)
4882 unlock(&pp.timersLock)
4883 unlock(&plocal.timersLock)
4885 // Flush p's write barrier buffer.
4886 if gcphase != _GCoff {
4890 for i := range pp.sudogbuf {
4891 pp.sudogbuf[i] = nil
4893 pp.sudogcache = pp.sudogbuf[:0]
4894 for j := range pp.deferpoolbuf {
4895 pp.deferpoolbuf[j] = nil
4897 pp.deferpool = pp.deferpoolbuf[:0]
4898 systemstack(func() {
4899 for i := 0; i < pp.mspancache.len; i++ {
4900 // Safe to call since the world is stopped.
4901 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4903 pp.mspancache.len = 0
4905 pp.pcache.flush(&mheap_.pages)
4906 unlock(&mheap_.lock)
4908 freemcache(pp.mcache)
4913 if pp.timerRaceCtx != 0 {
4914 // The race detector code uses a callback to fetch
4915 // the proc context, so arrange for that callback
4916 // to see the right thing.
4917 // This hack only works because we are the only
4923 racectxend(pp.timerRaceCtx)
4928 raceprocdestroy(pp.raceprocctx)
4935 // Change number of processors.
4937 // sched.lock must be held, and the world must be stopped.
4939 // gcworkbufs must not be being modified by either the GC or the write barrier
4940 // code, so the GC must not be running if the number of Ps actually changes.
4942 // Returns list of Ps with local work, they need to be scheduled by the caller.
4943 func procresize(nprocs int32) *p {
4944 assertLockHeld(&sched.lock)
4945 assertWorldStopped()
4948 if old < 0 || nprocs <= 0 {
4949 throw("procresize: invalid arg")
4952 traceGomaxprocs(nprocs)
4955 // update statistics
4957 if sched.procresizetime != 0 {
4958 sched.totaltime += int64(old) * (now - sched.procresizetime)
4960 sched.procresizetime = now
4962 maskWords := (nprocs + 31) / 32
4964 // Grow allp if necessary.
4965 if nprocs > int32(len(allp)) {
4966 // Synchronize with retake, which could be running
4967 // concurrently since it doesn't run on a P.
4969 if nprocs <= int32(cap(allp)) {
4970 allp = allp[:nprocs]
4972 nallp := make([]*p, nprocs)
4973 // Copy everything up to allp's cap so we
4974 // never lose old allocated Ps.
4975 copy(nallp, allp[:cap(allp)])
4979 if maskWords <= int32(cap(idlepMask)) {
4980 idlepMask = idlepMask[:maskWords]
4981 timerpMask = timerpMask[:maskWords]
4983 nidlepMask := make([]uint32, maskWords)
4984 // No need to copy beyond len, old Ps are irrelevant.
4985 copy(nidlepMask, idlepMask)
4986 idlepMask = nidlepMask
4988 ntimerpMask := make([]uint32, maskWords)
4989 copy(ntimerpMask, timerpMask)
4990 timerpMask = ntimerpMask
4995 // initialize new P's
4996 for i := old; i < nprocs; i++ {
5002 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5006 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5007 // continue to use the current P
5008 gp.m.p.ptr().status = _Prunning
5009 gp.m.p.ptr().mcache.prepareForSweep()
5011 // release the current P and acquire allp[0].
5013 // We must do this before destroying our current P
5014 // because p.destroy itself has write barriers, so we
5015 // need to do that from a valid P.
5018 // Pretend that we were descheduled
5019 // and then scheduled again to keep
5022 traceProcStop(gp.m.p.ptr())
5036 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5039 // release resources from unused P's
5040 for i := nprocs; i < old; i++ {
5043 // can't free P itself because it can be referenced by an M in syscall
5047 if int32(len(allp)) != nprocs {
5049 allp = allp[:nprocs]
5050 idlepMask = idlepMask[:maskWords]
5051 timerpMask = timerpMask[:maskWords]
5056 for i := nprocs - 1; i >= 0; i-- {
5058 if gp.m.p.ptr() == pp {
5066 pp.link.set(runnablePs)
5070 stealOrder.reset(uint32(nprocs))
5071 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5072 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5074 // Notify the limiter that the amount of procs has changed.
5075 gcCPULimiter.resetCapacity(now, nprocs)
5080 // Associate p and the current m.
5082 // This function is allowed to have write barriers even if the caller
5083 // isn't because it immediately acquires pp.
5085 //go:yeswritebarrierrec
5086 func acquirep(pp *p) {
5087 // Do the part that isn't allowed to have write barriers.
5090 // Have p; write barriers now allowed.
5092 // Perform deferred mcache flush before this P can allocate
5093 // from a potentially stale mcache.
5094 pp.mcache.prepareForSweep()
5101 // wirep is the first step of acquirep, which actually associates the
5102 // current M to pp. This is broken out so we can disallow write
5103 // barriers for this part, since we don't yet have a P.
5105 //go:nowritebarrierrec
5111 throw("wirep: already in go")
5113 if pp.m != 0 || pp.status != _Pidle {
5118 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5119 throw("wirep: invalid p state")
5123 pp.status = _Prunning
5126 // Disassociate p and the current m.
5127 func releasep() *p {
5131 throw("releasep: invalid arg")
5134 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5135 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5136 throw("releasep: invalid p state")
5139 traceProcStop(gp.m.p.ptr())
5147 func incidlelocked(v int32) {
5149 sched.nmidlelocked += v
5156 // Check for deadlock situation.
5157 // The check is based on number of running M's, if 0 -> deadlock.
5158 // sched.lock must be held.
5160 assertLockHeld(&sched.lock)
5162 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5163 // there are no running goroutines. The calling program is
5164 // assumed to be running.
5165 if islibrary || isarchive {
5169 // If we are dying because of a signal caught on an already idle thread,
5170 // freezetheworld will cause all running threads to block.
5171 // And runtime will essentially enter into deadlock state,
5172 // except that there is a thread that will call exit soon.
5173 if panicking.Load() > 0 {
5177 // If we are not running under cgo, but we have an extra M then account
5178 // for it. (It is possible to have an extra M on Windows without cgo to
5179 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5182 if !iscgo && cgoHasExtraM {
5183 mp := lockextra(true)
5184 haveExtraM := extraMCount > 0
5191 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5196 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5197 throw("checkdead: inconsistent counts")
5201 forEachG(func(gp *g) {
5202 if isSystemGoroutine(gp, false) {
5205 s := readgstatus(gp)
5206 switch s &^ _Gscan {
5213 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5214 throw("checkdead: runnable g")
5217 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5218 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5219 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5222 // Maybe jump time forward for playground.
5224 if when := timeSleepUntil(); when < maxWhen {
5227 // Start an M to steal the timer.
5228 pp, _ := pidleget(faketime)
5230 // There should always be a free P since
5231 // nothing is running.
5232 throw("checkdead: no p for timer")
5236 // There should always be a free M since
5237 // nothing is running.
5238 throw("checkdead: no m for timer")
5240 // M must be spinning to steal. We set this to be
5241 // explicit, but since this is the only M it would
5242 // become spinning on its own anyways.
5243 sched.nmspinning.Add(1)
5246 notewakeup(&mp.park)
5251 // There are no goroutines running, so we can look at the P's.
5252 for _, pp := range allp {
5253 if len(pp.timers) > 0 {
5258 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5259 fatal("all goroutines are asleep - deadlock!")
5262 // forcegcperiod is the maximum time in nanoseconds between garbage
5263 // collections. If we go this long without a garbage collection, one
5264 // is forced to run.
5266 // This is a variable for testing purposes. It normally doesn't change.
5267 var forcegcperiod int64 = 2 * 60 * 1e9
5269 // needSysmonWorkaround is true if the workaround for
5270 // golang.org/issue/42515 is needed on NetBSD.
5271 var needSysmonWorkaround bool = false
5273 // Always runs without a P, so write barriers are not allowed.
5275 //go:nowritebarrierrec
5282 lasttrace := int64(0)
5283 idle := 0 // how many cycles in succession we had not wokeup somebody
5287 if idle == 0 { // start with 20us sleep...
5289 } else if idle > 50 { // start doubling the sleep after 1ms...
5292 if delay > 10*1000 { // up to 10ms
5297 // sysmon should not enter deep sleep if schedtrace is enabled so that
5298 // it can print that information at the right time.
5300 // It should also not enter deep sleep if there are any active P's so
5301 // that it can retake P's from syscalls, preempt long running G's, and
5302 // poll the network if all P's are busy for long stretches.
5304 // It should wakeup from deep sleep if any P's become active either due
5305 // to exiting a syscall or waking up due to a timer expiring so that it
5306 // can resume performing those duties. If it wakes from a syscall it
5307 // resets idle and delay as a bet that since it had retaken a P from a
5308 // syscall before, it may need to do it again shortly after the
5309 // application starts work again. It does not reset idle when waking
5310 // from a timer to avoid adding system load to applications that spend
5311 // most of their time sleeping.
5313 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5315 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5316 syscallWake := false
5317 next := timeSleepUntil()
5319 sched.sysmonwait.Store(true)
5321 // Make wake-up period small enough
5322 // for the sampling to be correct.
5323 sleep := forcegcperiod / 2
5324 if next-now < sleep {
5327 shouldRelax := sleep >= osRelaxMinNS
5331 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5336 sched.sysmonwait.Store(false)
5337 noteclear(&sched.sysmonnote)
5347 lock(&sched.sysmonlock)
5348 // Update now in case we blocked on sysmonnote or spent a long time
5349 // blocked on schedlock or sysmonlock above.
5352 // trigger libc interceptors if needed
5353 if *cgo_yield != nil {
5354 asmcgocall(*cgo_yield, nil)
5356 // poll network if not polled for more than 10ms
5357 lastpoll := sched.lastpoll.Load()
5358 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5359 sched.lastpoll.CompareAndSwap(lastpoll, now)
5360 list := netpoll(0) // non-blocking - returns list of goroutines
5362 // Need to decrement number of idle locked M's
5363 // (pretending that one more is running) before injectglist.
5364 // Otherwise it can lead to the following situation:
5365 // injectglist grabs all P's but before it starts M's to run the P's,
5366 // another M returns from syscall, finishes running its G,
5367 // observes that there is no work to do and no other running M's
5368 // and reports deadlock.
5374 if GOOS == "netbsd" && needSysmonWorkaround {
5375 // netpoll is responsible for waiting for timer
5376 // expiration, so we typically don't have to worry
5377 // about starting an M to service timers. (Note that
5378 // sleep for timeSleepUntil above simply ensures sysmon
5379 // starts running again when that timer expiration may
5380 // cause Go code to run again).
5382 // However, netbsd has a kernel bug that sometimes
5383 // misses netpollBreak wake-ups, which can lead to
5384 // unbounded delays servicing timers. If we detect this
5385 // overrun, then startm to get something to handle the
5388 // See issue 42515 and
5389 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5390 if next := timeSleepUntil(); next < now {
5394 if scavenger.sysmonWake.Load() != 0 {
5395 // Kick the scavenger awake if someone requested it.
5398 // retake P's blocked in syscalls
5399 // and preempt long running G's
5400 if retake(now) != 0 {
5405 // check if we need to force a GC
5406 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5408 forcegc.idle.Store(false)
5410 list.push(forcegc.g)
5412 unlock(&forcegc.lock)
5414 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5416 schedtrace(debug.scheddetail > 0)
5418 unlock(&sched.sysmonlock)
5422 type sysmontick struct {
5429 // forcePreemptNS is the time slice given to a G before it is
5431 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5433 func retake(now int64) uint32 {
5435 // Prevent allp slice changes. This lock will be completely
5436 // uncontended unless we're already stopping the world.
5438 // We can't use a range loop over allp because we may
5439 // temporarily drop the allpLock. Hence, we need to re-fetch
5440 // allp each time around the loop.
5441 for i := 0; i < len(allp); i++ {
5444 // This can happen if procresize has grown
5445 // allp but not yet created new Ps.
5448 pd := &pp.sysmontick
5451 if s == _Prunning || s == _Psyscall {
5452 // Preempt G if it's running for too long.
5453 t := int64(pp.schedtick)
5454 if int64(pd.schedtick) != t {
5455 pd.schedtick = uint32(t)
5457 } else if pd.schedwhen+forcePreemptNS <= now {
5459 // In case of syscall, preemptone() doesn't
5460 // work, because there is no M wired to P.
5465 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5466 t := int64(pp.syscalltick)
5467 if !sysretake && int64(pd.syscalltick) != t {
5468 pd.syscalltick = uint32(t)
5469 pd.syscallwhen = now
5472 // On the one hand we don't want to retake Ps if there is no other work to do,
5473 // but on the other hand we want to retake them eventually
5474 // because they can prevent the sysmon thread from deep sleep.
5475 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5478 // Drop allpLock so we can take sched.lock.
5480 // Need to decrement number of idle locked M's
5481 // (pretending that one more is running) before the CAS.
5482 // Otherwise the M from which we retake can exit the syscall,
5483 // increment nmidle and report deadlock.
5485 if atomic.Cas(&pp.status, s, _Pidle) {
5502 // Tell all goroutines that they have been preempted and they should stop.
5503 // This function is purely best-effort. It can fail to inform a goroutine if a
5504 // processor just started running it.
5505 // No locks need to be held.
5506 // Returns true if preemption request was issued to at least one goroutine.
5507 func preemptall() bool {
5509 for _, pp := range allp {
5510 if pp.status != _Prunning {
5520 // Tell the goroutine running on processor P to stop.
5521 // This function is purely best-effort. It can incorrectly fail to inform the
5522 // goroutine. It can inform the wrong goroutine. Even if it informs the
5523 // correct goroutine, that goroutine might ignore the request if it is
5524 // simultaneously executing newstack.
5525 // No lock needs to be held.
5526 // Returns true if preemption request was issued.
5527 // The actual preemption will happen at some point in the future
5528 // and will be indicated by the gp->status no longer being
5530 func preemptone(pp *p) bool {
5532 if mp == nil || mp == getg().m {
5536 if gp == nil || gp == mp.g0 {
5542 // Every call in a goroutine checks for stack overflow by
5543 // comparing the current stack pointer to gp->stackguard0.
5544 // Setting gp->stackguard0 to StackPreempt folds
5545 // preemption into the normal stack overflow check.
5546 gp.stackguard0 = stackPreempt
5548 // Request an async preemption of this P.
5549 if preemptMSupported && debug.asyncpreemptoff == 0 {
5559 func schedtrace(detailed bool) {
5566 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)
5568 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5570 // We must be careful while reading data from P's, M's and G's.
5571 // Even if we hold schedlock, most data can be changed concurrently.
5572 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5573 for i, pp := range allp {
5575 h := atomic.Load(&pp.runqhead)
5576 t := atomic.Load(&pp.runqtail)
5578 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5584 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5586 // In non-detailed mode format lengths of per-P run queues as:
5587 // [len1 len2 len3 len4]
5593 if i == len(allp)-1 {
5604 for mp := allm; mp != nil; mp = mp.alllink {
5606 print(" M", mp.id, ": p=")
5618 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5619 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5627 forEachG(func(gp *g) {
5628 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5635 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5645 // schedEnableUser enables or disables the scheduling of user
5648 // This does not stop already running user goroutines, so the caller
5649 // should first stop the world when disabling user goroutines.
5650 func schedEnableUser(enable bool) {
5652 if sched.disable.user == !enable {
5656 sched.disable.user = !enable
5658 n := sched.disable.n
5660 globrunqputbatch(&sched.disable.runnable, n)
5662 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5670 // schedEnabled reports whether gp should be scheduled. It returns
5671 // false is scheduling of gp is disabled.
5673 // sched.lock must be held.
5674 func schedEnabled(gp *g) bool {
5675 assertLockHeld(&sched.lock)
5677 if sched.disable.user {
5678 return isSystemGoroutine(gp, true)
5683 // Put mp on midle list.
5684 // sched.lock must be held.
5685 // May run during STW, so write barriers are not allowed.
5687 //go:nowritebarrierrec
5689 assertLockHeld(&sched.lock)
5691 mp.schedlink = sched.midle
5697 // Try to get an m from midle list.
5698 // sched.lock must be held.
5699 // May run during STW, so write barriers are not allowed.
5701 //go:nowritebarrierrec
5703 assertLockHeld(&sched.lock)
5705 mp := sched.midle.ptr()
5707 sched.midle = mp.schedlink
5713 // Put gp on the global runnable queue.
5714 // sched.lock must be held.
5715 // May run during STW, so write barriers are not allowed.
5717 //go:nowritebarrierrec
5718 func globrunqput(gp *g) {
5719 assertLockHeld(&sched.lock)
5721 sched.runq.pushBack(gp)
5725 // Put gp at the head of the global runnable queue.
5726 // sched.lock must be held.
5727 // May run during STW, so write barriers are not allowed.
5729 //go:nowritebarrierrec
5730 func globrunqputhead(gp *g) {
5731 assertLockHeld(&sched.lock)
5737 // Put a batch of runnable goroutines on the global runnable queue.
5738 // This clears *batch.
5739 // sched.lock must be held.
5740 // May run during STW, so write barriers are not allowed.
5742 //go:nowritebarrierrec
5743 func globrunqputbatch(batch *gQueue, n int32) {
5744 assertLockHeld(&sched.lock)
5746 sched.runq.pushBackAll(*batch)
5751 // Try get a batch of G's from the global runnable queue.
5752 // sched.lock must be held.
5753 func globrunqget(pp *p, max int32) *g {
5754 assertLockHeld(&sched.lock)
5756 if sched.runqsize == 0 {
5760 n := sched.runqsize/gomaxprocs + 1
5761 if n > sched.runqsize {
5764 if max > 0 && n > max {
5767 if n > int32(len(pp.runq))/2 {
5768 n = int32(len(pp.runq)) / 2
5773 gp := sched.runq.pop()
5776 gp1 := sched.runq.pop()
5777 runqput(pp, gp1, false)
5782 // pMask is an atomic bitstring with one bit per P.
5785 // read returns true if P id's bit is set.
5786 func (p pMask) read(id uint32) bool {
5788 mask := uint32(1) << (id % 32)
5789 return (atomic.Load(&p[word]) & mask) != 0
5792 // set sets P id's bit.
5793 func (p pMask) set(id int32) {
5795 mask := uint32(1) << (id % 32)
5796 atomic.Or(&p[word], mask)
5799 // clear clears P id's bit.
5800 func (p pMask) clear(id int32) {
5802 mask := uint32(1) << (id % 32)
5803 atomic.And(&p[word], ^mask)
5806 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5808 // Ideally, the timer mask would be kept immediately consistent on any timer
5809 // operations. Unfortunately, updating a shared global data structure in the
5810 // timer hot path adds too much overhead in applications frequently switching
5811 // between no timers and some timers.
5813 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5814 // running P (returned by pidleget) may add a timer at any time, so its mask
5815 // must be set. An idle P (passed to pidleput) cannot add new timers while
5816 // idle, so if it has no timers at that time, its mask may be cleared.
5818 // Thus, we get the following effects on timer-stealing in findrunnable:
5820 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5821 // (for work- or timer-stealing; this is the ideal case).
5822 // - Running Ps must always be checked.
5823 // - Idle Ps whose timers are stolen must continue to be checked until they run
5824 // again, even after timer expiration.
5826 // When the P starts running again, the mask should be set, as a timer may be
5827 // added at any time.
5829 // TODO(prattmic): Additional targeted updates may improve the above cases.
5830 // e.g., updating the mask when stealing a timer.
5831 func updateTimerPMask(pp *p) {
5832 if pp.numTimers.Load() > 0 {
5836 // Looks like there are no timers, however another P may transiently
5837 // decrement numTimers when handling a timerModified timer in
5838 // checkTimers. We must take timersLock to serialize with these changes.
5839 lock(&pp.timersLock)
5840 if pp.numTimers.Load() == 0 {
5841 timerpMask.clear(pp.id)
5843 unlock(&pp.timersLock)
5846 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5847 // to nanotime or zero. Returns now or the current time if now was zero.
5849 // This releases ownership of p. Once sched.lock is released it is no longer
5852 // sched.lock must be held.
5854 // May run during STW, so write barriers are not allowed.
5856 //go:nowritebarrierrec
5857 func pidleput(pp *p, now int64) int64 {
5858 assertLockHeld(&sched.lock)
5861 throw("pidleput: P has non-empty run queue")
5866 updateTimerPMask(pp) // clear if there are no timers.
5867 idlepMask.set(pp.id)
5868 pp.link = sched.pidle
5871 if !pp.limiterEvent.start(limiterEventIdle, now) {
5872 throw("must be able to track idle limiter event")
5877 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5879 // sched.lock must be held.
5881 // May run during STW, so write barriers are not allowed.
5883 //go:nowritebarrierrec
5884 func pidleget(now int64) (*p, int64) {
5885 assertLockHeld(&sched.lock)
5887 pp := sched.pidle.ptr()
5889 // Timer may get added at any time now.
5893 timerpMask.set(pp.id)
5894 idlepMask.clear(pp.id)
5895 sched.pidle = pp.link
5896 sched.npidle.Add(-1)
5897 pp.limiterEvent.stop(limiterEventIdle, now)
5902 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5903 // This is called by spinning Ms (or callers than need a spinning M) that have
5904 // found work. If no P is available, this must synchronized with non-spinning
5905 // Ms that may be preparing to drop their P without discovering this work.
5907 // sched.lock must be held.
5909 // May run during STW, so write barriers are not allowed.
5911 //go:nowritebarrierrec
5912 func pidlegetSpinning(now int64) (*p, int64) {
5913 assertLockHeld(&sched.lock)
5915 pp, now := pidleget(now)
5917 // See "Delicate dance" comment in findrunnable. We found work
5918 // that we cannot take, we must synchronize with non-spinning
5919 // Ms that may be preparing to drop their P.
5920 sched.needspinning.Store(1)
5927 // runqempty reports whether pp has no Gs on its local run queue.
5928 // It never returns true spuriously.
5929 func runqempty(pp *p) bool {
5930 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5931 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5932 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5933 // does not mean the queue is empty.
5935 head := atomic.Load(&pp.runqhead)
5936 tail := atomic.Load(&pp.runqtail)
5937 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5938 if tail == atomic.Load(&pp.runqtail) {
5939 return head == tail && runnext == 0
5944 // To shake out latent assumptions about scheduling order,
5945 // we introduce some randomness into scheduling decisions
5946 // when running with the race detector.
5947 // The need for this was made obvious by changing the
5948 // (deterministic) scheduling order in Go 1.5 and breaking
5949 // many poorly-written tests.
5950 // With the randomness here, as long as the tests pass
5951 // consistently with -race, they shouldn't have latent scheduling
5953 const randomizeScheduler = raceenabled
5955 // runqput tries to put g on the local runnable queue.
5956 // If next is false, runqput adds g to the tail of the runnable queue.
5957 // If next is true, runqput puts g in the pp.runnext slot.
5958 // If the run queue is full, runnext puts g on the global queue.
5959 // Executed only by the owner P.
5960 func runqput(pp *p, gp *g, next bool) {
5961 if randomizeScheduler && next && fastrandn(2) == 0 {
5967 oldnext := pp.runnext
5968 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
5974 // Kick the old runnext out to the regular run queue.
5979 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
5981 if t-h < uint32(len(pp.runq)) {
5982 pp.runq[t%uint32(len(pp.runq))].set(gp)
5983 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
5986 if runqputslow(pp, gp, h, t) {
5989 // the queue is not full, now the put above must succeed
5993 // Put g and a batch of work from local runnable queue on global queue.
5994 // Executed only by the owner P.
5995 func runqputslow(pp *p, gp *g, h, t uint32) bool {
5996 var batch [len(pp.runq)/2 + 1]*g
5998 // First, grab a batch from local queue.
6001 if n != uint32(len(pp.runq)/2) {
6002 throw("runqputslow: queue is not full")
6004 for i := uint32(0); i < n; i++ {
6005 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6007 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6012 if randomizeScheduler {
6013 for i := uint32(1); i <= n; i++ {
6014 j := fastrandn(i + 1)
6015 batch[i], batch[j] = batch[j], batch[i]
6019 // Link the goroutines.
6020 for i := uint32(0); i < n; i++ {
6021 batch[i].schedlink.set(batch[i+1])
6024 q.head.set(batch[0])
6025 q.tail.set(batch[n])
6027 // Now put the batch on global queue.
6029 globrunqputbatch(&q, int32(n+1))
6034 // runqputbatch tries to put all the G's on q on the local runnable queue.
6035 // If the queue is full, they are put on the global queue; in that case
6036 // this will temporarily acquire the scheduler lock.
6037 // Executed only by the owner P.
6038 func runqputbatch(pp *p, q *gQueue, qsize int) {
6039 h := atomic.LoadAcq(&pp.runqhead)
6042 for !q.empty() && t-h < uint32(len(pp.runq)) {
6044 pp.runq[t%uint32(len(pp.runq))].set(gp)
6050 if randomizeScheduler {
6051 off := func(o uint32) uint32 {
6052 return (pp.runqtail + o) % uint32(len(pp.runq))
6054 for i := uint32(1); i < n; i++ {
6055 j := fastrandn(i + 1)
6056 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6060 atomic.StoreRel(&pp.runqtail, t)
6063 globrunqputbatch(q, int32(qsize))
6068 // Get g from local runnable queue.
6069 // If inheritTime is true, gp should inherit the remaining time in the
6070 // current time slice. Otherwise, it should start a new time slice.
6071 // Executed only by the owner P.
6072 func runqget(pp *p) (gp *g, inheritTime bool) {
6073 // If there's a runnext, it's the next G to run.
6075 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6076 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6077 // Hence, there's no need to retry this CAS if it fails.
6078 if next != 0 && pp.runnext.cas(next, 0) {
6079 return next.ptr(), true
6083 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6088 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6089 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6095 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6096 // Executed only by the owner P.
6097 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6098 oldNext := pp.runnext
6099 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6100 drainQ.pushBack(oldNext.ptr())
6105 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6111 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6115 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6119 // We've inverted the order in which it gets G's from the local P's runnable queue
6120 // and then advances the head pointer because we don't want to mess up the statuses of G's
6121 // while runqdrain() and runqsteal() are running in parallel.
6122 // Thus we should advance the head pointer before draining the local P into a gQueue,
6123 // so that we can update any gp.schedlink only after we take the full ownership of G,
6124 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6125 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6126 for i := uint32(0); i < qn; i++ {
6127 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6134 // Grabs a batch of goroutines from pp's runnable queue into batch.
6135 // Batch is a ring buffer starting at batchHead.
6136 // Returns number of grabbed goroutines.
6137 // Can be executed by any P.
6138 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6140 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6141 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6146 // Try to steal from pp.runnext.
6147 if next := pp.runnext; next != 0 {
6148 if pp.status == _Prunning {
6149 // Sleep to ensure that pp isn't about to run the g
6150 // we are about to steal.
6151 // The important use case here is when the g running
6152 // on pp ready()s another g and then almost
6153 // immediately blocks. Instead of stealing runnext
6154 // in this window, back off to give pp a chance to
6155 // schedule runnext. This will avoid thrashing gs
6156 // between different Ps.
6157 // A sync chan send/recv takes ~50ns as of time of
6158 // writing, so 3us gives ~50x overshoot.
6159 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6162 // On some platforms system timer granularity is
6163 // 1-15ms, which is way too much for this
6164 // optimization. So just yield.
6168 if !pp.runnext.cas(next, 0) {
6171 batch[batchHead%uint32(len(batch))] = next
6177 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6180 for i := uint32(0); i < n; i++ {
6181 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6182 batch[(batchHead+i)%uint32(len(batch))] = g
6184 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6190 // Steal half of elements from local runnable queue of p2
6191 // and put onto local runnable queue of p.
6192 // Returns one of the stolen elements (or nil if failed).
6193 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6195 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6200 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6204 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6205 if t-h+n >= uint32(len(pp.runq)) {
6206 throw("runqsteal: runq overflow")
6208 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6212 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6213 // be on one gQueue or gList at a time.
6214 type gQueue struct {
6219 // empty reports whether q is empty.
6220 func (q *gQueue) empty() bool {
6224 // push adds gp to the head of q.
6225 func (q *gQueue) push(gp *g) {
6226 gp.schedlink = q.head
6233 // pushBack adds gp to the tail of q.
6234 func (q *gQueue) pushBack(gp *g) {
6237 q.tail.ptr().schedlink.set(gp)
6244 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6246 func (q *gQueue) pushBackAll(q2 gQueue) {
6250 q2.tail.ptr().schedlink = 0
6252 q.tail.ptr().schedlink = q2.head
6259 // pop removes and returns the head of queue q. It returns nil if
6261 func (q *gQueue) pop() *g {
6264 q.head = gp.schedlink
6272 // popList takes all Gs in q and returns them as a gList.
6273 func (q *gQueue) popList() gList {
6274 stack := gList{q.head}
6279 // A gList is a list of Gs linked through g.schedlink. A G can only be
6280 // on one gQueue or gList at a time.
6285 // empty reports whether l is empty.
6286 func (l *gList) empty() bool {
6290 // push adds gp to the head of l.
6291 func (l *gList) push(gp *g) {
6292 gp.schedlink = l.head
6296 // pushAll prepends all Gs in q to l.
6297 func (l *gList) pushAll(q gQueue) {
6299 q.tail.ptr().schedlink = l.head
6304 // pop removes and returns the head of l. If l is empty, it returns nil.
6305 func (l *gList) pop() *g {
6308 l.head = gp.schedlink
6313 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6314 func setMaxThreads(in int) (out int) {
6316 out = int(sched.maxmcount)
6317 if in > 0x7fffffff { // MaxInt32
6318 sched.maxmcount = 0x7fffffff
6320 sched.maxmcount = int32(in)
6328 func procPin() int {
6333 return int(mp.p.ptr().id)
6342 //go:linkname sync_runtime_procPin sync.runtime_procPin
6344 func sync_runtime_procPin() int {
6348 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6350 func sync_runtime_procUnpin() {
6354 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6356 func sync_atomic_runtime_procPin() int {
6360 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6362 func sync_atomic_runtime_procUnpin() {
6366 // Active spinning for sync.Mutex.
6368 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6370 func sync_runtime_canSpin(i int) bool {
6371 // sync.Mutex is cooperative, so we are conservative with spinning.
6372 // Spin only few times and only if running on a multicore machine and
6373 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6374 // As opposed to runtime mutex we don't do passive spinning here,
6375 // because there can be work on global runq or on other Ps.
6376 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6379 if p := getg().m.p.ptr(); !runqempty(p) {
6385 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6387 func sync_runtime_doSpin() {
6388 procyield(active_spin_cnt)
6391 var stealOrder randomOrder
6393 // randomOrder/randomEnum are helper types for randomized work stealing.
6394 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6395 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6396 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6397 type randomOrder struct {
6402 type randomEnum struct {
6409 func (ord *randomOrder) reset(count uint32) {
6411 ord.coprimes = ord.coprimes[:0]
6412 for i := uint32(1); i <= count; i++ {
6413 if gcd(i, count) == 1 {
6414 ord.coprimes = append(ord.coprimes, i)
6419 func (ord *randomOrder) start(i uint32) randomEnum {
6423 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6427 func (enum *randomEnum) done() bool {
6428 return enum.i == enum.count
6431 func (enum *randomEnum) next() {
6433 enum.pos = (enum.pos + enum.inc) % enum.count
6436 func (enum *randomEnum) position() uint32 {
6440 func gcd(a, b uint32) uint32 {
6447 // An initTask represents the set of initializations that need to be done for a package.
6448 // Keep in sync with ../../test/noinit.go:initTask
6449 type initTask struct {
6450 state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
6452 // followed by nfns pcs, uintptr sized, one per init function to run
6455 // inittrace stores statistics for init functions which are
6456 // updated by malloc and newproc when active is true.
6457 var inittrace tracestat
6459 type tracestat struct {
6460 active bool // init tracing activation status
6461 id uint64 // init goroutine id
6462 allocs uint64 // heap allocations
6463 bytes uint64 // heap allocated bytes
6466 func doInit(ts []*initTask) {
6467 for _, t := range ts {
6472 func doInit1(t *initTask) {
6474 case 2: // fully initialized
6476 case 1: // initialization in progress
6477 throw("recursive call during initialization - linker skew")
6478 default: // not initialized yet
6479 t.state = 1 // initialization in progress
6486 if inittrace.active {
6488 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6493 // We should have pruned all of these in the linker.
6494 throw("inittask with no functions")
6497 firstFunc := add(unsafe.Pointer(t), 8)
6498 for i := uint32(0); i < t.nfns; i++ {
6499 p := add(firstFunc, uintptr(i)*goarch.PtrSize)
6500 f := *(*func())(unsafe.Pointer(&p))
6504 if inittrace.active {
6506 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6509 f := *(*func())(unsafe.Pointer(&firstFunc))
6510 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6513 print("init ", pkg, " @")
6514 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6515 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6516 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6517 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6521 t.state = 2 // initialization done