1 // Copyright 2014 The Go Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style
3 // license that can be found in the LICENSE file.
11 "runtime/internal/atomic"
12 "runtime/internal/sys"
16 // set using cmd/go/internal/modload.ModInfoProg
19 // Goroutine scheduler
20 // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
22 // The main concepts are:
24 // M - worker thread, or machine.
25 // P - processor, a resource that is required to execute Go code.
26 // M must have an associated P to execute Go code, however it can be
27 // blocked or in a syscall w/o an associated P.
29 // Design doc at https://golang.org/s/go11sched.
31 // Worker thread parking/unparking.
32 // We need to balance between keeping enough running worker threads to utilize
33 // available hardware parallelism and parking excessive running worker threads
34 // to conserve CPU resources and power. This is not simple for two reasons:
35 // (1) scheduler state is intentionally distributed (in particular, per-P work
36 // queues), so it is not possible to compute global predicates on fast paths;
37 // (2) for optimal thread management we would need to know the future (don't park
38 // a worker thread when a new goroutine will be readied in near future).
40 // Three rejected approaches that would work badly:
41 // 1. Centralize all scheduler state (would inhibit scalability).
42 // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
43 // is a spare P, unpark a thread and handoff it the thread and the goroutine.
44 // This would lead to thread state thrashing, as the thread that readied the
45 // goroutine can be out of work the very next moment, we will need to park it.
46 // Also, it would destroy locality of computation as we want to preserve
47 // dependent goroutines on the same thread; and introduce additional latency.
48 // 3. Unpark an additional thread whenever we ready a goroutine and there is an
49 // idle P, but don't do handoff. This would lead to excessive thread parking/
50 // unparking as the additional threads will instantly park without discovering
53 // The current approach:
55 // This approach applies to three primary sources of potential work: readying a
56 // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
57 // additional details.
59 // We unpark an additional thread when we submit work if (this is wakep()):
60 // 1. There is an idle P, and
61 // 2. There are no "spinning" worker threads.
63 // A worker thread is considered spinning if it is out of local work and did
64 // not find work in the global run queue or netpoller; the spinning state is
65 // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
66 // also considered spinning; we don't do goroutine handoff so such threads are
67 // out of work initially. Spinning threads spin on looking for work in per-P
68 // run queues and timer heaps or from the GC before parking. If a spinning
69 // thread finds work it takes itself out of the spinning state and proceeds to
70 // execution. If it does not find work it takes itself out of the spinning
71 // state and then parks.
73 // If there is at least one spinning thread (sched.nmspinning>1), we don't
74 // unpark new threads when submitting work. To compensate for that, if the last
75 // spinning thread finds work and stops spinning, it must unpark a new spinning
76 // thread. This approach smooths out unjustified spikes of thread unparking,
77 // but at the same time guarantees eventual maximal CPU parallelism
80 // The main implementation complication is that we need to be very careful
81 // during spinning->non-spinning thread transition. This transition can race
82 // with submission of new work, and either one part or another needs to unpark
83 // another worker thread. If they both fail to do that, we can end up with
84 // semi-persistent CPU underutilization.
86 // The general pattern for submission is:
87 // 1. Submit work to the local run queue, timer heap, or GC state.
88 // 2. #StoreLoad-style memory barrier.
89 // 3. Check sched.nmspinning.
91 // The general pattern for spinning->non-spinning transition is:
92 // 1. Decrement nmspinning.
93 // 2. #StoreLoad-style memory barrier.
94 // 3. Check all per-P work queues and GC for new work.
96 // Note that all this complexity does not apply to global run queue as we are
97 // not sloppy about thread unparking when submitting to global queue. Also see
98 // comments for nmspinning manipulation.
100 // How these different sources of work behave varies, though it doesn't affect
101 // the synchronization approach:
102 // * Ready goroutine: this is an obvious source of work; the goroutine is
103 // immediately ready and must run on some thread eventually.
104 // * New/modified-earlier timer: The current timer implementation (see time.go)
105 // uses netpoll in a thread with no work available to wait for the soonest
106 // timer. If there is no thread waiting, we want a new spinning thread to go
108 // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
109 // background GC work (note: currently disabled per golang.org/issue/19112).
110 // Also see golang.org/issue/44313, as this should be extended to all GC
120 //go:linkname runtime_inittask runtime..inittask
121 var runtime_inittask initTask
123 //go:linkname main_inittask main..inittask
124 var main_inittask initTask
126 // main_init_done is a signal used by cgocallbackg that initialization
127 // has been completed. It is made before _cgo_notify_runtime_init_done,
128 // so all cgo calls can rely on it existing. When main_init is complete,
129 // it is closed, meaning cgocallbackg can reliably receive from it.
130 var main_init_done chan bool
132 //go:linkname main_main main.main
135 // mainStarted indicates that the main M has started.
138 // runtimeInitTime is the nanotime() at which the runtime started.
139 var runtimeInitTime int64
141 // Value to use for signal mask for newly created M's.
142 var initSigmask sigset
144 // The main goroutine.
148 // Racectx of m0->g0 is used only as the parent of the main goroutine.
149 // It must not be used for anything else.
152 // Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
153 // Using decimal instead of binary GB and MB because
154 // they look nicer in the stack overflow failure message.
155 if goarch.PtrSize == 8 {
156 maxstacksize = 1000000000
158 maxstacksize = 250000000
161 // An upper limit for max stack size. Used to avoid random crashes
162 // after calling SetMaxStack and trying to allocate a stack that is too big,
163 // since stackalloc works with 32-bit sizes.
164 maxstackceiling = 2 * maxstacksize
166 // Allow newproc to start new Ms.
169 if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
171 newm(sysmon, nil, -1)
175 // Lock the main goroutine onto this, the main OS thread,
176 // during initialization. Most programs won't care, but a few
177 // do require certain calls to be made by the main thread.
178 // Those can arrange for main.main to run in the main thread
179 // by calling runtime.LockOSThread during initialization
180 // to preserve the lock.
184 throw("runtime.main not on m0")
187 // Record when the world started.
188 // Must be before doInit for tracing init.
189 runtimeInitTime = nanotime()
190 if runtimeInitTime == 0 {
191 throw("nanotime returning zero")
194 if debug.inittrace != 0 {
195 inittrace.id = getg().goid
196 inittrace.active = true
199 doInit(&runtime_inittask) // Must be before defer.
201 // Defer unlock so that runtime.Goexit during init does the unlock too.
211 main_init_done = make(chan bool)
213 if _cgo_thread_start == nil {
214 throw("_cgo_thread_start missing")
216 if GOOS != "windows" {
217 if _cgo_setenv == nil {
218 throw("_cgo_setenv missing")
220 if _cgo_unsetenv == nil {
221 throw("_cgo_unsetenv missing")
224 if _cgo_notify_runtime_init_done == nil {
225 throw("_cgo_notify_runtime_init_done missing")
227 // Start the template thread in case we enter Go from
228 // a C-created thread and need to create a new thread.
229 startTemplateThread()
230 cgocall(_cgo_notify_runtime_init_done, nil)
233 doInit(&main_inittask)
235 // Disable init tracing after main init done to avoid overhead
236 // of collecting statistics in malloc and newproc
237 inittrace.active = false
239 close(main_init_done)
244 if isarchive || islibrary {
245 // A program compiled with -buildmode=c-archive or c-shared
246 // has a main, but it is not executed.
249 fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
252 runExitHooks(0) // run hooks now, since racefini does not return
256 // Make racy client program work: if panicking on
257 // another goroutine at the same time as main returns,
258 // let the other goroutine finish printing the panic trace.
259 // Once it does, it will exit. See issues 3934 and 20018.
260 if runningPanicDefers.Load() != 0 {
261 // Running deferred functions should not take long.
262 for c := 0; c < 1000; c++ {
263 if runningPanicDefers.Load() == 0 {
269 if panicking.Load() != 0 {
270 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
281 // os_beforeExit is called from os.Exit(0).
283 //go:linkname os_beforeExit os.runtime_beforeExit
284 func os_beforeExit(exitCode int) {
285 runExitHooks(exitCode)
286 if exitCode == 0 && raceenabled {
291 // start forcegc helper goroutine
296 func forcegchelper() {
298 lockInit(&forcegc.lock, lockRankForcegc)
301 if forcegc.idle.Load() {
302 throw("forcegc: phase error")
304 forcegc.idle.Store(true)
305 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
306 // this goroutine is explicitly resumed by sysmon
307 if debug.gctrace > 0 {
310 // Time-triggered, fully concurrent.
311 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
317 // Gosched yields the processor, allowing other goroutines to run. It does not
318 // suspend the current goroutine, so execution resumes automatically.
324 // goschedguarded yields the processor like gosched, but also checks
325 // for forbidden states and opts out of the yield in those cases.
328 func goschedguarded() {
329 mcall(goschedguarded_m)
332 // goschedIfBusy yields the processor like gosched, but only does so if
333 // there are no idle Ps or if we're on the only P and there's nothing in
334 // the run queue. In both cases, there is freely available idle time.
337 func goschedIfBusy() {
339 // Call gosched if gp.preempt is set; we may be in a tight loop that
340 // doesn't otherwise yield.
341 if !gp.preempt && sched.npidle.Load() > 0 {
347 // Puts the current goroutine into a waiting state and calls unlockf on the
350 // If unlockf returns false, the goroutine is resumed.
352 // unlockf must not access this G's stack, as it may be moved between
353 // the call to gopark and the call to unlockf.
355 // Note that because unlockf is called after putting the G into a waiting
356 // state, the G may have already been readied by the time unlockf is called
357 // unless there is external synchronization preventing the G from being
358 // readied. If unlockf returns false, it must guarantee that the G cannot be
359 // externally readied.
361 // Reason explains why the goroutine has been parked. It is displayed in stack
362 // traces and heap dumps. Reasons should be unique and descriptive. Do not
363 // re-use reasons, add new ones.
364 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
365 if reason != waitReasonSleep {
366 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
370 status := readgstatus(gp)
371 if status != _Grunning && status != _Gscanrunning {
372 throw("gopark: bad g status")
375 mp.waitunlockf = unlockf
376 gp.waitreason = reason
377 mp.waittraceev = traceEv
378 mp.waittraceskip = traceskip
380 // can't do anything that might move the G between Ms here.
384 // Puts the current goroutine into a waiting state and unlocks the lock.
385 // The goroutine can be made runnable again by calling goready(gp).
386 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
387 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
390 func goready(gp *g, traceskip int) {
392 ready(gp, traceskip, true)
397 func acquireSudog() *sudog {
398 // Delicate dance: the semaphore implementation calls
399 // acquireSudog, acquireSudog calls new(sudog),
400 // new calls malloc, malloc can call the garbage collector,
401 // and the garbage collector calls the semaphore implementation
403 // Break the cycle by doing acquirem/releasem around new(sudog).
404 // The acquirem/releasem increments m.locks during new(sudog),
405 // which keeps the garbage collector from being invoked.
408 if len(pp.sudogcache) == 0 {
409 lock(&sched.sudoglock)
410 // First, try to grab a batch from central cache.
411 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
412 s := sched.sudogcache
413 sched.sudogcache = s.next
415 pp.sudogcache = append(pp.sudogcache, s)
417 unlock(&sched.sudoglock)
418 // If the central cache is empty, allocate a new one.
419 if len(pp.sudogcache) == 0 {
420 pp.sudogcache = append(pp.sudogcache, new(sudog))
423 n := len(pp.sudogcache)
424 s := pp.sudogcache[n-1]
425 pp.sudogcache[n-1] = nil
426 pp.sudogcache = pp.sudogcache[:n-1]
428 throw("acquireSudog: found s.elem != nil in cache")
435 func releaseSudog(s *sudog) {
437 throw("runtime: sudog with non-nil elem")
440 throw("runtime: sudog with non-false isSelect")
443 throw("runtime: sudog with non-nil next")
446 throw("runtime: sudog with non-nil prev")
448 if s.waitlink != nil {
449 throw("runtime: sudog with non-nil waitlink")
452 throw("runtime: sudog with non-nil c")
456 throw("runtime: releaseSudog with non-nil gp.param")
458 mp := acquirem() // avoid rescheduling to another P
460 if len(pp.sudogcache) == cap(pp.sudogcache) {
461 // Transfer half of local cache to the central cache.
462 var first, last *sudog
463 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
464 n := len(pp.sudogcache)
465 p := pp.sudogcache[n-1]
466 pp.sudogcache[n-1] = nil
467 pp.sudogcache = pp.sudogcache[:n-1]
475 lock(&sched.sudoglock)
476 last.next = sched.sudogcache
477 sched.sudogcache = first
478 unlock(&sched.sudoglock)
480 pp.sudogcache = append(pp.sudogcache, s)
484 // called from assembly.
485 func badmcall(fn func(*g)) {
486 throw("runtime: mcall called on m->g0 stack")
489 func badmcall2(fn func(*g)) {
490 throw("runtime: mcall function returned")
493 func badreflectcall() {
494 panic(plainError("arg size to reflect.call more than 1GB"))
498 //go:nowritebarrierrec
499 func badmorestackg0() {
500 writeErrStr("fatal: morestack on g0\n")
504 //go:nowritebarrierrec
505 func badmorestackgsignal() {
506 writeErrStr("fatal: morestack on gsignal\n")
514 func lockedOSThread() bool {
516 return gp.lockedm != 0 && gp.m.lockedg != 0
520 // allgs contains all Gs ever created (including dead Gs), and thus
523 // Access via the slice is protected by allglock or stop-the-world.
524 // Readers that cannot take the lock may (carefully!) use the atomic
529 // allglen and allgptr are atomic variables that contain len(allgs) and
530 // &allgs[0] respectively. Proper ordering depends on totally-ordered
531 // loads and stores. Writes are protected by allglock.
533 // allgptr is updated before allglen. Readers should read allglen
534 // before allgptr to ensure that allglen is always <= len(allgptr). New
535 // Gs appended during the race can be missed. For a consistent view of
536 // all Gs, allglock must be held.
538 // allgptr copies should always be stored as a concrete type or
539 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
540 // even if it points to a stale array.
545 func allgadd(gp *g) {
546 if readgstatus(gp) == _Gidle {
547 throw("allgadd: bad status Gidle")
551 allgs = append(allgs, gp)
552 if &allgs[0] != allgptr {
553 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
555 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
559 // allGsSnapshot returns a snapshot of the slice of all Gs.
561 // The world must be stopped or allglock must be held.
562 func allGsSnapshot() []*g {
563 assertWorldStoppedOrLockHeld(&allglock)
565 // Because the world is stopped or allglock is held, allgadd
566 // cannot happen concurrently with this. allgs grows
567 // monotonically and existing entries never change, so we can
568 // simply return a copy of the slice header. For added safety,
569 // we trim everything past len because that can still change.
570 return allgs[:len(allgs):len(allgs)]
573 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
574 func atomicAllG() (**g, uintptr) {
575 length := atomic.Loaduintptr(&allglen)
576 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
580 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
581 func atomicAllGIndex(ptr **g, i uintptr) *g {
582 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
585 // forEachG calls fn on every G from allgs.
587 // forEachG takes a lock to exclude concurrent addition of new Gs.
588 func forEachG(fn func(gp *g)) {
590 for _, gp := range allgs {
596 // forEachGRace calls fn on every G from allgs.
598 // forEachGRace avoids locking, but does not exclude addition of new Gs during
599 // execution, which may be missed.
600 func forEachGRace(fn func(gp *g)) {
601 ptr, length := atomicAllG()
602 for i := uintptr(0); i < length; i++ {
603 gp := atomicAllGIndex(ptr, i)
610 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
611 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
615 // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
616 // value of the GODEBUG environment variable.
617 func cpuinit(env string) {
619 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
620 cpu.DebugOptions = true
624 // Support cpu feature variables are used in code generated by the compiler
625 // to guard execution of instructions that can not be assumed to be always supported.
628 x86HasPOPCNT = cpu.X86.HasPOPCNT
629 x86HasSSE41 = cpu.X86.HasSSE41
630 x86HasFMA = cpu.X86.HasFMA
633 armHasVFPv4 = cpu.ARM.HasVFPv4
636 arm64HasATOMICS = cpu.ARM64.HasATOMICS
640 // getGodebugEarly extracts the environment variable GODEBUG from the environment on
641 // Unix-like operating systems and returns it. This function exists to extract GODEBUG
642 // early before much of the runtime is initialized.
643 func getGodebugEarly() string {
644 const prefix = "GODEBUG="
647 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
648 // Similar to goenv_unix but extracts the environment value for
650 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
652 for argv_index(argv, argc+1+n) != nil {
656 for i := int32(0); i < n; i++ {
657 p := argv_index(argv, argc+1+i)
658 s := unsafe.String(p, findnull(p))
660 if hasPrefix(s, prefix) {
661 env = gostring(p)[len(prefix):]
669 // The bootstrap sequence is:
673 // make & queue new G
674 // call runtime·mstart
676 // The new G calls runtime·main.
678 lockInit(&sched.lock, lockRankSched)
679 lockInit(&sched.sysmonlock, lockRankSysmon)
680 lockInit(&sched.deferlock, lockRankDefer)
681 lockInit(&sched.sudoglock, lockRankSudog)
682 lockInit(&deadlock, lockRankDeadlock)
683 lockInit(&paniclk, lockRankPanic)
684 lockInit(&allglock, lockRankAllg)
685 lockInit(&allpLock, lockRankAllp)
686 lockInit(&reflectOffs.lock, lockRankReflectOffs)
687 lockInit(&finlock, lockRankFin)
688 lockInit(&trace.bufLock, lockRankTraceBuf)
689 lockInit(&trace.stringsLock, lockRankTraceStrings)
690 lockInit(&trace.lock, lockRankTrace)
691 lockInit(&cpuprof.lock, lockRankCpuprof)
692 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
693 // Enforce that this lock is always a leaf lock.
694 // All of this lock's critical sections should be
696 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
698 // raceinit must be the first call to race detector.
699 // In particular, it must be done before mallocinit below calls racemapshadow.
702 gp.racectx, raceprocctx0 = raceinit()
705 sched.maxmcount = 10000
707 // The world starts stopped.
713 godebug := getGodebugEarly()
714 initPageTrace(godebug) // must run after mallocinit but before anything allocates
715 cpuinit(godebug) // must run before alginit
716 alginit() // maps, hash, fastrand must not be used before this call
717 fastrandinit() // must run before mcommoninit
718 mcommoninit(gp.m, -1)
719 modulesinit() // provides activeModules
720 typelinksinit() // uses maps, activeModules
721 itabsinit() // uses activeModules
722 stkobjinit() // must run before GC starts
724 sigsave(&gp.m.sigmask)
725 initSigmask = gp.m.sigmask
732 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
733 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
734 // set to true by the linker, it means that nothing is consuming the profile, it is
735 // safe to set MemProfileRate to 0.
736 if disableMemoryProfiling {
741 sched.lastpoll.Store(nanotime())
743 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
746 if procresize(procs) != nil {
747 throw("unknown runnable goroutine during bootstrap")
751 // World is effectively started now, as P's can run.
754 // For cgocheck > 1, we turn on the write barrier at all times
755 // and check all pointer writes. We can't do this until after
756 // procresize because the write barrier needs a P.
757 if debug.cgocheck > 1 {
758 writeBarrier.cgo = true
759 writeBarrier.enabled = true
760 for _, pp := range allp {
765 if buildVersion == "" {
766 // Condition should never trigger. This code just serves
767 // to ensure runtime·buildVersion is kept in the resulting binary.
768 buildVersion = "unknown"
770 if len(modinfo) == 1 {
771 // Condition should never trigger. This code just serves
772 // to ensure runtime·modinfo is kept in the resulting binary.
777 func dumpgstatus(gp *g) {
779 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
780 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
783 // sched.lock must be held.
785 assertLockHeld(&sched.lock)
787 if mcount() > sched.maxmcount {
788 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
789 throw("thread exhaustion")
793 // mReserveID returns the next ID to use for a new m. This new m is immediately
794 // considered 'running' by checkdead.
796 // sched.lock must be held.
797 func mReserveID() int64 {
798 assertLockHeld(&sched.lock)
800 if sched.mnext+1 < sched.mnext {
801 throw("runtime: thread ID overflow")
809 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
810 func mcommoninit(mp *m, id int64) {
813 // g0 stack won't make sense for user (and is not necessary unwindable).
815 callers(1, mp.createstack[:])
826 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
827 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
831 // Same behavior as for 1.17.
832 // TODO: Simplify this.
833 if goarch.BigEndian {
834 mp.fastrand = uint64(lo)<<32 | uint64(hi)
836 mp.fastrand = uint64(hi)<<32 | uint64(lo)
840 if mp.gsignal != nil {
841 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
844 // Add to allm so garbage collector doesn't free g->m
845 // when it is just in a register or thread-local storage.
848 // NumCgoCall() iterates over allm w/o schedlock,
849 // so we need to publish it safely.
850 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
853 // Allocate memory to hold a cgo traceback if the cgo call crashes.
854 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
855 mp.cgoCallers = new(cgoCallers)
859 func (mp *m) becomeSpinning() {
861 sched.nmspinning.Add(1)
862 sched.needspinning.Store(0)
865 var fastrandseed uintptr
867 func fastrandinit() {
868 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
872 // Mark gp ready to run.
873 func ready(gp *g, traceskip int, next bool) {
875 traceGoUnpark(gp, traceskip)
878 status := readgstatus(gp)
881 mp := acquirem() // disable preemption because it can be holding p in a local var
882 if status&^_Gscan != _Gwaiting {
884 throw("bad g->status in ready")
887 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
888 casgstatus(gp, _Gwaiting, _Grunnable)
889 runqput(mp.p.ptr(), gp, next)
894 // freezeStopWait is a large value that freezetheworld sets
895 // sched.stopwait to in order to request that all Gs permanently stop.
896 const freezeStopWait = 0x7fffffff
898 // freezing is set to non-zero if the runtime is trying to freeze the
900 var freezing atomic.Bool
902 // Similar to stopTheWorld but best-effort and can be called several times.
903 // There is no reverse operation, used during crashing.
904 // This function must not lock any mutexes.
905 func freezetheworld() {
907 // stopwait and preemption requests can be lost
908 // due to races with concurrently executing threads,
909 // so try several times
910 for i := 0; i < 5; i++ {
911 // this should tell the scheduler to not start any new goroutines
912 sched.stopwait = freezeStopWait
913 sched.gcwaiting.Store(true)
914 // this should stop running goroutines
916 break // no running goroutines
926 // All reads and writes of g's status go through readgstatus, casgstatus
927 // castogscanstatus, casfrom_Gscanstatus.
930 func readgstatus(gp *g) uint32 {
931 return gp.atomicstatus.Load()
934 // The Gscanstatuses are acting like locks and this releases them.
935 // If it proves to be a performance hit we should be able to make these
936 // simple atomic stores but for now we are going to throw if
937 // we see an inconsistent state.
938 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
941 // Check that transition is valid.
944 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
946 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
952 if newval == oldval&^_Gscan {
953 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
957 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
959 throw("casfrom_Gscanstatus: gp->status is not in scan state")
961 releaseLockRank(lockRankGscan)
964 // This will return false if the gp is not in the expected status and the cas fails.
965 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
966 func castogscanstatus(gp *g, oldval, newval uint32) bool {
972 if newval == oldval|_Gscan {
973 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
975 acquireLockRank(lockRankGscan)
981 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
982 throw("castogscanstatus")
986 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
987 // various latencies on every transition instead of sampling them.
988 var casgstatusAlwaysTrack = false
990 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
991 // and casfrom_Gscanstatus instead.
992 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
993 // put it in the Gscan state is finished.
996 func casgstatus(gp *g, oldval, newval uint32) {
997 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
999 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
1000 throw("casgstatus: bad incoming values")
1004 acquireLockRank(lockRankGscan)
1005 releaseLockRank(lockRankGscan)
1007 // See https://golang.org/cl/21503 for justification of the yield delay.
1008 const yieldDelay = 5 * 1000
1011 // loop if gp->atomicstatus is in a scan state giving
1012 // GC time to finish and change the state to oldval.
1013 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1014 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1015 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1018 nextYield = nanotime() + yieldDelay
1020 if nanotime() < nextYield {
1021 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1026 nextYield = nanotime() + yieldDelay/2
1030 if oldval == _Grunning {
1031 // Track every gTrackingPeriod time a goroutine transitions out of running.
1032 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1041 // Handle various kinds of tracking.
1044 // - Time spent in runnable.
1045 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1048 // We transitioned out of runnable, so measure how much
1049 // time we spent in this state and add it to
1052 gp.runnableTime += now - gp.trackingStamp
1053 gp.trackingStamp = 0
1055 if !gp.waitreason.isMutexWait() {
1056 // Not blocking on a lock.
1059 // Blocking on a lock, measure it. Note that because we're
1060 // sampling, we have to multiply by our sampling period to get
1061 // a more representative estimate of the absolute value.
1062 // gTrackingPeriod also represents an accurate sampling period
1063 // because we can only enter this state from _Grunning.
1065 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1066 gp.trackingStamp = 0
1070 if !gp.waitreason.isMutexWait() {
1071 // Not blocking on a lock.
1074 // Blocking on a lock. Write down the timestamp.
1076 gp.trackingStamp = now
1078 // We just transitioned into runnable, so record what
1079 // time that happened.
1081 gp.trackingStamp = now
1083 // We're transitioning into running, so turn off
1084 // tracking and record how much time we spent in
1087 sched.timeToRun.record(gp.runnableTime)
1092 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1094 // Use this over casgstatus when possible to ensure that a waitreason is set.
1095 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1096 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1097 gp.waitreason = reason
1098 casgstatus(gp, old, _Gwaiting)
1101 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1102 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1103 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1104 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1105 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1108 func casgcopystack(gp *g) uint32 {
1110 oldstatus := readgstatus(gp) &^ _Gscan
1111 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1112 throw("copystack: bad status, not Gwaiting or Grunnable")
1114 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1120 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1122 // TODO(austin): This is the only status operation that both changes
1123 // the status and locks the _Gscan bit. Rethink this.
1124 func casGToPreemptScan(gp *g, old, new uint32) {
1125 if old != _Grunning || new != _Gscan|_Gpreempted {
1126 throw("bad g transition")
1128 acquireLockRank(lockRankGscan)
1129 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1133 // casGFromPreempted attempts to transition gp from _Gpreempted to
1134 // _Gwaiting. If successful, the caller is responsible for
1135 // re-scheduling gp.
1136 func casGFromPreempted(gp *g, old, new uint32) bool {
1137 if old != _Gpreempted || new != _Gwaiting {
1138 throw("bad g transition")
1140 gp.waitreason = waitReasonPreempted
1141 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1144 // stopTheWorld stops all P's from executing goroutines, interrupting
1145 // all goroutines at GC safe points and records reason as the reason
1146 // for the stop. On return, only the current goroutine's P is running.
1147 // stopTheWorld must not be called from a system stack and the caller
1148 // must not hold worldsema. The caller must call startTheWorld when
1149 // other P's should resume execution.
1151 // stopTheWorld is safe for multiple goroutines to call at the
1152 // same time. Each will execute its own stop, and the stops will
1155 // This is also used by routines that do stack dumps. If the system is
1156 // in panic or being exited, this may not reliably stop all
1158 func stopTheWorld(reason string) {
1159 semacquire(&worldsema)
1161 gp.m.preemptoff = reason
1162 systemstack(func() {
1163 // Mark the goroutine which called stopTheWorld preemptible so its
1164 // stack may be scanned.
1165 // This lets a mark worker scan us while we try to stop the world
1166 // since otherwise we could get in a mutual preemption deadlock.
1167 // We must not modify anything on the G stack because a stack shrink
1168 // may occur. A stack shrink is otherwise OK though because in order
1169 // to return from this function (and to leave the system stack) we
1170 // must have preempted all goroutines, including any attempting
1171 // to scan our stack, in which case, any stack shrinking will
1172 // have already completed by the time we exit.
1173 // Don't provide a wait reason because we're still executing.
1174 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1175 stopTheWorldWithSema()
1176 casgstatus(gp, _Gwaiting, _Grunning)
1180 // startTheWorld undoes the effects of stopTheWorld.
1181 func startTheWorld() {
1182 systemstack(func() { startTheWorldWithSema(false) })
1184 // worldsema must be held over startTheWorldWithSema to ensure
1185 // gomaxprocs cannot change while worldsema is held.
1187 // Release worldsema with direct handoff to the next waiter, but
1188 // acquirem so that semrelease1 doesn't try to yield our time.
1190 // Otherwise if e.g. ReadMemStats is being called in a loop,
1191 // it might stomp on other attempts to stop the world, such as
1192 // for starting or ending GC. The operation this blocks is
1193 // so heavy-weight that we should just try to be as fair as
1196 // We don't want to just allow us to get preempted between now
1197 // and releasing the semaphore because then we keep everyone
1198 // (including, for example, GCs) waiting longer.
1201 semrelease1(&worldsema, true, 0)
1205 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1206 // until the GC is not running. It also blocks a GC from starting
1207 // until startTheWorldGC is called.
1208 func stopTheWorldGC(reason string) {
1210 stopTheWorld(reason)
1213 // startTheWorldGC undoes the effects of stopTheWorldGC.
1214 func startTheWorldGC() {
1219 // Holding worldsema grants an M the right to try to stop the world.
1220 var worldsema uint32 = 1
1222 // Holding gcsema grants the M the right to block a GC, and blocks
1223 // until the current GC is done. In particular, it prevents gomaxprocs
1224 // from changing concurrently.
1226 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1227 // being changed/enabled during a GC, remove this.
1228 var gcsema uint32 = 1
1230 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1231 // The caller is responsible for acquiring worldsema and disabling
1232 // preemption first and then should stopTheWorldWithSema on the system
1235 // semacquire(&worldsema, 0)
1236 // m.preemptoff = "reason"
1237 // systemstack(stopTheWorldWithSema)
1239 // When finished, the caller must either call startTheWorld or undo
1240 // these three operations separately:
1242 // m.preemptoff = ""
1243 // systemstack(startTheWorldWithSema)
1244 // semrelease(&worldsema)
1246 // It is allowed to acquire worldsema once and then execute multiple
1247 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1248 // Other P's are able to execute between successive calls to
1249 // startTheWorldWithSema and stopTheWorldWithSema.
1250 // Holding worldsema causes any other goroutines invoking
1251 // stopTheWorld to block.
1252 func stopTheWorldWithSema() {
1255 // If we hold a lock, then we won't be able to stop another M
1256 // that is blocked trying to acquire the lock.
1258 throw("stopTheWorld: holding locks")
1262 sched.stopwait = gomaxprocs
1263 sched.gcwaiting.Store(true)
1266 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1268 // try to retake all P's in Psyscall status
1269 for _, pp := range allp {
1271 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1283 pp, _ := pidleget(now)
1287 pp.status = _Pgcstop
1290 wait := sched.stopwait > 0
1293 // wait for remaining P's to stop voluntarily
1296 // wait for 100us, then try to re-preempt in case of any races
1297 if notetsleep(&sched.stopnote, 100*1000) {
1298 noteclear(&sched.stopnote)
1307 if sched.stopwait != 0 {
1308 bad = "stopTheWorld: not stopped (stopwait != 0)"
1310 for _, pp := range allp {
1311 if pp.status != _Pgcstop {
1312 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1316 if freezing.Load() {
1317 // Some other thread is panicking. This can cause the
1318 // sanity checks above to fail if the panic happens in
1319 // the signal handler on a stopped thread. Either way,
1320 // we should halt this thread.
1331 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1332 assertWorldStopped()
1334 mp := acquirem() // disable preemption because it can be holding p in a local var
1335 if netpollinited() {
1336 list := netpoll(0) // non-blocking
1346 p1 := procresize(procs)
1347 sched.gcwaiting.Store(false)
1348 if sched.sysmonwait.Load() {
1349 sched.sysmonwait.Store(false)
1350 notewakeup(&sched.sysmonnote)
1363 throw("startTheWorld: inconsistent mp->nextp")
1366 notewakeup(&mp.park)
1368 // Start M to run P. Do not start another M below.
1373 // Capture start-the-world time before doing clean-up tasks.
1374 startTime := nanotime()
1379 // Wakeup an additional proc in case we have excessive runnable goroutines
1380 // in local queues or in the global queue. If we don't, the proc will park itself.
1381 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1389 // usesLibcall indicates whether this runtime performs system calls
1391 func usesLibcall() bool {
1393 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1396 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1401 // mStackIsSystemAllocated indicates whether this runtime starts on a
1402 // system-allocated stack.
1403 func mStackIsSystemAllocated() bool {
1405 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1409 case "386", "amd64", "arm", "arm64":
1416 // mstart is the entry-point for new Ms.
1417 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1420 // mstart0 is the Go entry-point for new Ms.
1421 // This must not split the stack because we may not even have stack
1422 // bounds set up yet.
1424 // May run during STW (because it doesn't have a P yet), so write
1425 // barriers are not allowed.
1428 //go:nowritebarrierrec
1432 osStack := gp.stack.lo == 0
1434 // Initialize stack bounds from system stack.
1435 // Cgo may have left stack size in stack.hi.
1436 // minit may update the stack bounds.
1438 // Note: these bounds may not be very accurate.
1439 // We set hi to &size, but there are things above
1440 // it. The 1024 is supposed to compensate this,
1441 // but is somewhat arbitrary.
1444 size = 8192 * sys.StackGuardMultiplier
1446 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1447 gp.stack.lo = gp.stack.hi - size + 1024
1449 // Initialize stack guard so that we can start calling regular
1451 gp.stackguard0 = gp.stack.lo + _StackGuard
1452 // This is the g0, so we can also call go:systemstack
1453 // functions, which check stackguard1.
1454 gp.stackguard1 = gp.stackguard0
1457 // Exit this thread.
1458 if mStackIsSystemAllocated() {
1459 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1460 // the stack, but put it in gp.stack before mstart,
1461 // so the logic above hasn't set osStack yet.
1467 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1468 // so that we can set up g0.sched to return to the call of mstart1 above.
1475 throw("bad runtime·mstart")
1478 // Set up m.g0.sched as a label returning to just
1479 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1480 // We're never coming back to mstart1 after we call schedule,
1481 // so other calls can reuse the current frame.
1482 // And goexit0 does a gogo that needs to return from mstart1
1483 // and let mstart0 exit the thread.
1484 gp.sched.g = guintptr(unsafe.Pointer(gp))
1485 gp.sched.pc = getcallerpc()
1486 gp.sched.sp = getcallersp()
1491 // Install signal handlers; after minit so that minit can
1492 // prepare the thread to be able to handle the signals.
1497 if fn := gp.m.mstartfn; fn != nil {
1502 acquirep(gp.m.nextp.ptr())
1508 // mstartm0 implements part of mstart1 that only runs on the m0.
1510 // Write barriers are allowed here because we know the GC can't be
1511 // running yet, so they'll be no-ops.
1513 //go:yeswritebarrierrec
1515 // Create an extra M for callbacks on threads not created by Go.
1516 // An extra M is also needed on Windows for callbacks created by
1517 // syscall.NewCallback. See issue #6751 for details.
1518 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1525 // mPark causes a thread to park itself, returning once woken.
1530 notesleep(&gp.m.park)
1531 noteclear(&gp.m.park)
1534 // mexit tears down and exits the current thread.
1536 // Don't call this directly to exit the thread, since it must run at
1537 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1538 // unwind the stack to the point that exits the thread.
1540 // It is entered with m.p != nil, so write barriers are allowed. It
1541 // will release the P before exiting.
1543 //go:yeswritebarrierrec
1544 func mexit(osStack bool) {
1548 // This is the main thread. Just wedge it.
1550 // On Linux, exiting the main thread puts the process
1551 // into a non-waitable zombie state. On Plan 9,
1552 // exiting the main thread unblocks wait even though
1553 // other threads are still running. On Solaris we can
1554 // neither exitThread nor return from mstart. Other
1555 // bad things probably happen on other platforms.
1557 // We could try to clean up this M more before wedging
1558 // it, but that complicates signal handling.
1559 handoffp(releasep())
1565 throw("locked m0 woke up")
1571 // Free the gsignal stack.
1572 if mp.gsignal != nil {
1573 stackfree(mp.gsignal.stack)
1574 // On some platforms, when calling into VDSO (e.g. nanotime)
1575 // we store our g on the gsignal stack, if there is one.
1576 // Now the stack is freed, unlink it from the m, so we
1577 // won't write to it when calling VDSO code.
1581 // Remove m from allm.
1583 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1589 throw("m not found in allm")
1591 // Delay reaping m until it's done with the stack.
1593 // Put mp on the free list, though it will not be reaped while freeWait
1594 // is freeMWait. mp is no longer reachable via allm, so even if it is
1595 // on an OS stack, we must keep a reference to mp alive so that the GC
1596 // doesn't free mp while we are still using it.
1598 // Note that the free list must not be linked through alllink because
1599 // some functions walk allm without locking, so may be using alllink.
1600 mp.freeWait.Store(freeMWait)
1601 mp.freelink = sched.freem
1605 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1608 handoffp(releasep())
1609 // After this point we must not have write barriers.
1611 // Invoke the deadlock detector. This must happen after
1612 // handoffp because it may have started a new M to take our
1619 if GOOS == "darwin" || GOOS == "ios" {
1620 // Make sure pendingPreemptSignals is correct when an M exits.
1622 if mp.signalPending.Load() != 0 {
1623 pendingPreemptSignals.Add(-1)
1627 // Destroy all allocated resources. After this is called, we may no
1628 // longer take any locks.
1632 // No more uses of mp, so it is safe to drop the reference.
1633 mp.freeWait.Store(freeMRef)
1635 // Return from mstart and let the system thread
1636 // library free the g0 stack and terminate the thread.
1640 // mstart is the thread's entry point, so there's nothing to
1641 // return to. Exit the thread directly. exitThread will clear
1642 // m.freeWait when it's done with the stack and the m can be
1644 exitThread(&mp.freeWait)
1647 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1648 // If a P is currently executing code, this will bring the P to a GC
1649 // safe point and execute fn on that P. If the P is not executing code
1650 // (it is idle or in a syscall), this will call fn(p) directly while
1651 // preventing the P from exiting its state. This does not ensure that
1652 // fn will run on every CPU executing Go code, but it acts as a global
1653 // memory barrier. GC uses this as a "ragged barrier."
1655 // The caller must hold worldsema.
1658 func forEachP(fn func(*p)) {
1660 pp := getg().m.p.ptr()
1663 if sched.safePointWait != 0 {
1664 throw("forEachP: sched.safePointWait != 0")
1666 sched.safePointWait = gomaxprocs - 1
1667 sched.safePointFn = fn
1669 // Ask all Ps to run the safe point function.
1670 for _, p2 := range allp {
1672 atomic.Store(&p2.runSafePointFn, 1)
1677 // Any P entering _Pidle or _Psyscall from now on will observe
1678 // p.runSafePointFn == 1 and will call runSafePointFn when
1679 // changing its status to _Pidle/_Psyscall.
1681 // Run safe point function for all idle Ps. sched.pidle will
1682 // not change because we hold sched.lock.
1683 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1684 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1686 sched.safePointWait--
1690 wait := sched.safePointWait > 0
1693 // Run fn for the current P.
1696 // Force Ps currently in _Psyscall into _Pidle and hand them
1697 // off to induce safe point function execution.
1698 for _, p2 := range allp {
1700 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1710 // Wait for remaining Ps to run fn.
1713 // Wait for 100us, then try to re-preempt in
1714 // case of any races.
1716 // Requires system stack.
1717 if notetsleep(&sched.safePointNote, 100*1000) {
1718 noteclear(&sched.safePointNote)
1724 if sched.safePointWait != 0 {
1725 throw("forEachP: not done")
1727 for _, p2 := range allp {
1728 if p2.runSafePointFn != 0 {
1729 throw("forEachP: P did not run fn")
1734 sched.safePointFn = nil
1739 // runSafePointFn runs the safe point function, if any, for this P.
1740 // This should be called like
1742 // if getg().m.p.runSafePointFn != 0 {
1746 // runSafePointFn must be checked on any transition in to _Pidle or
1747 // _Psyscall to avoid a race where forEachP sees that the P is running
1748 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1749 // nor the P run the safe-point function.
1750 func runSafePointFn() {
1751 p := getg().m.p.ptr()
1752 // Resolve the race between forEachP running the safe-point
1753 // function on this P's behalf and this P running the
1754 // safe-point function directly.
1755 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1758 sched.safePointFn(p)
1760 sched.safePointWait--
1761 if sched.safePointWait == 0 {
1762 notewakeup(&sched.safePointNote)
1767 // When running with cgo, we call _cgo_thread_start
1768 // to start threads for us so that we can play nicely with
1770 var cgoThreadStart unsafe.Pointer
1772 type cgothreadstart struct {
1778 // Allocate a new m unassociated with any thread.
1779 // Can use p for allocation context if needed.
1780 // fn is recorded as the new m's m.mstartfn.
1781 // id is optional pre-allocated m ID. Omit by passing -1.
1783 // This function is allowed to have write barriers even if the caller
1784 // isn't because it borrows pp.
1786 //go:yeswritebarrierrec
1787 func allocm(pp *p, fn func(), id int64) *m {
1790 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1791 // disable preemption to ensure it is not stolen, which would make the
1792 // caller lose ownership.
1797 acquirep(pp) // temporarily borrow p for mallocs in this function
1800 // Release the free M list. We need to do this somewhere and
1801 // this may free up a stack we can use.
1802 if sched.freem != nil {
1805 for freem := sched.freem; freem != nil; {
1806 wait := freem.freeWait.Load()
1807 if wait == freeMWait {
1808 next := freem.freelink
1809 freem.freelink = newList
1814 // Free the stack if needed. For freeMRef, there is
1815 // nothing to do except drop freem from the sched.freem
1817 if wait == freeMStack {
1818 // stackfree must be on the system stack, but allocm is
1819 // reachable off the system stack transitively from
1821 systemstack(func() {
1822 stackfree(freem.g0.stack)
1825 freem = freem.freelink
1827 sched.freem = newList
1835 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1836 // Windows and Plan 9 will layout sched stack on OS stack.
1837 if iscgo || mStackIsSystemAllocated() {
1840 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1844 if pp == gp.m.p.ptr() {
1849 allocmLock.runlock()
1853 // needm is called when a cgo callback happens on a
1854 // thread without an m (a thread not created by Go).
1855 // In this case, needm is expected to find an m to use
1856 // and return with m, g initialized correctly.
1857 // Since m and g are not set now (likely nil, but see below)
1858 // needm is limited in what routines it can call. In particular
1859 // it can only call nosplit functions (textflag 7) and cannot
1860 // do any scheduling that requires an m.
1862 // In order to avoid needing heavy lifting here, we adopt
1863 // the following strategy: there is a stack of available m's
1864 // that can be stolen. Using compare-and-swap
1865 // to pop from the stack has ABA races, so we simulate
1866 // a lock by doing an exchange (via Casuintptr) to steal the stack
1867 // head and replace the top pointer with MLOCKED (1).
1868 // This serves as a simple spin lock that we can use even
1869 // without an m. The thread that locks the stack in this way
1870 // unlocks the stack by storing a valid stack head pointer.
1872 // In order to make sure that there is always an m structure
1873 // available to be stolen, we maintain the invariant that there
1874 // is always one more than needed. At the beginning of the
1875 // program (if cgo is in use) the list is seeded with a single m.
1876 // If needm finds that it has taken the last m off the list, its job
1877 // is - once it has installed its own m so that it can do things like
1878 // allocate memory - to create a spare m and put it on the list.
1880 // Each of these extra m's also has a g0 and a curg that are
1881 // pressed into service as the scheduling stack and current
1882 // goroutine for the duration of the cgo callback.
1884 // When the callback is done with the m, it calls dropm to
1885 // put the m back on the list.
1889 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1890 // Can happen if C/C++ code calls Go from a global ctor.
1891 // Can also happen on Windows if a global ctor uses a
1892 // callback created by syscall.NewCallback. See issue #6751
1895 // Can not throw, because scheduler is not initialized yet.
1896 writeErrStr("fatal error: cgo callback before cgo call\n")
1900 // Save and block signals before getting an M.
1901 // The signal handler may call needm itself,
1902 // and we must avoid a deadlock. Also, once g is installed,
1903 // any incoming signals will try to execute,
1904 // but we won't have the sigaltstack settings and other data
1905 // set up appropriately until the end of minit, which will
1906 // unblock the signals. This is the same dance as when
1907 // starting a new m to run Go code via newosproc.
1912 // Lock extra list, take head, unlock popped list.
1913 // nilokay=false is safe here because of the invariant above,
1914 // that the extra list always contains or will soon contain
1916 mp := lockextra(false)
1918 // Set needextram when we've just emptied the list,
1919 // so that the eventual call into cgocallbackg will
1920 // allocate a new m for the extra list. We delay the
1921 // allocation until then so that it can be done
1922 // after exitsyscall makes sure it is okay to be
1923 // running at all (that is, there's no garbage collection
1924 // running right now).
1925 mp.needextram = mp.schedlink == 0
1927 unlockextra(mp.schedlink.ptr())
1929 // Store the original signal mask for use by minit.
1930 mp.sigmask = sigmask
1932 // Install TLS on some platforms (previously setg
1933 // would do this if necessary).
1936 // Install g (= m->g0) and set the stack bounds
1937 // to match the current stack. We don't actually know
1938 // how big the stack is, like we don't know how big any
1939 // scheduling stack is, but we assume there's at least 32 kB,
1940 // which is more than enough for us.
1943 gp.stack.hi = getcallersp() + 1024
1944 gp.stack.lo = getcallersp() - 32*1024
1945 gp.stackguard0 = gp.stack.lo + _StackGuard
1947 // Initialize this thread to use the m.
1951 // mp.curg is now a real goroutine.
1952 casgstatus(mp.curg, _Gdead, _Gsyscall)
1956 // newextram allocates m's and puts them on the extra list.
1957 // It is called with a working local m, so that it can do things
1958 // like call schedlock and allocate.
1960 c := extraMWaiters.Swap(0)
1962 for i := uint32(0); i < c; i++ {
1966 // Make sure there is at least one extra M.
1967 mp := lockextra(true)
1975 // oneNewExtraM allocates an m and puts it on the extra list.
1976 func oneNewExtraM() {
1977 // Create extra goroutine locked to extra m.
1978 // The goroutine is the context in which the cgo callback will run.
1979 // The sched.pc will never be returned to, but setting it to
1980 // goexit makes clear to the traceback routines where
1981 // the goroutine stack ends.
1982 mp := allocm(nil, nil, -1)
1984 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
1985 gp.sched.sp = gp.stack.hi
1986 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
1988 gp.sched.g = guintptr(unsafe.Pointer(gp))
1989 gp.syscallpc = gp.sched.pc
1990 gp.syscallsp = gp.sched.sp
1991 gp.stktopsp = gp.sched.sp
1992 // malg returns status as _Gidle. Change to _Gdead before
1993 // adding to allg where GC can see it. We use _Gdead to hide
1994 // this from tracebacks and stack scans since it isn't a
1995 // "real" goroutine until needm grabs it.
1996 casgstatus(gp, _Gidle, _Gdead)
2003 gp.goid = sched.goidgen.Add(1)
2004 gp.sysblocktraced = true
2006 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
2009 // Trigger two trace events for the locked g in the extra m,
2010 // since the next event of the g will be traceEvGoSysExit in exitsyscall,
2011 // while calling from C thread to Go.
2012 traceGoCreate(gp, 0) // no start pc
2014 traceEvent(traceEvGoInSyscall, -1, gp.goid)
2016 // put on allg for garbage collector
2019 // gp is now on the allg list, but we don't want it to be
2020 // counted by gcount. It would be more "proper" to increment
2021 // sched.ngfree, but that requires locking. Incrementing ngsys
2022 // has the same effect.
2025 // Add m to the extra list.
2026 mnext := lockextra(true)
2027 mp.schedlink.set(mnext)
2032 // dropm is called when a cgo callback has called needm but is now
2033 // done with the callback and returning back into the non-Go thread.
2034 // It puts the current m back onto the extra list.
2036 // The main expense here is the call to signalstack to release the
2037 // m's signal stack, and then the call to needm on the next callback
2038 // from this thread. It is tempting to try to save the m for next time,
2039 // which would eliminate both these costs, but there might not be
2040 // a next time: the current thread (which Go does not control) might exit.
2041 // If we saved the m for that thread, there would be an m leak each time
2042 // such a thread exited. Instead, we acquire and release an m on each
2043 // call. These should typically not be scheduling operations, just a few
2044 // atomics, so the cost should be small.
2046 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
2047 // variable using pthread_key_create. Unlike the pthread keys we already use
2048 // on OS X, this dummy key would never be read by Go code. It would exist
2049 // only so that we could register at thread-exit-time destructor.
2050 // That destructor would put the m back onto the extra list.
2051 // This is purely a performance optimization. The current version,
2052 // in which dropm happens on each cgo call, is still correct too.
2053 // We may have to keep the current version on systems with cgo
2054 // but without pthreads, like Windows.
2056 // Clear m and g, and return m to the extra list.
2057 // After the call to setg we can only call nosplit functions
2058 // with no pointer manipulation.
2061 // Return mp.curg to dead state.
2062 casgstatus(mp.curg, _Gsyscall, _Gdead)
2063 mp.curg.preemptStop = false
2066 // Block signals before unminit.
2067 // Unminit unregisters the signal handling stack (but needs g on some systems).
2068 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2069 // It's important not to try to handle a signal between those two steps.
2070 sigmask := mp.sigmask
2074 mnext := lockextra(true)
2076 mp.schedlink.set(mnext)
2080 // Commit the release of mp.
2083 msigrestore(sigmask)
2086 // A helper function for EnsureDropM.
2087 func getm() uintptr {
2088 return uintptr(unsafe.Pointer(getg().m))
2091 var extram atomic.Uintptr
2092 var extraMCount uint32 // Protected by lockextra
2093 var extraMWaiters atomic.Uint32
2095 // lockextra locks the extra list and returns the list head.
2096 // The caller must unlock the list by storing a new list head
2097 // to extram. If nilokay is true, then lockextra will
2098 // return a nil list head if that's what it finds. If nilokay is false,
2099 // lockextra will keep waiting until the list head is no longer nil.
2102 func lockextra(nilokay bool) *m {
2107 old := extram.Load()
2112 if old == 0 && !nilokay {
2114 // Add 1 to the number of threads
2115 // waiting for an M.
2116 // This is cleared by newextram.
2117 extraMWaiters.Add(1)
2123 if extram.CompareAndSwap(old, locked) {
2124 return (*m)(unsafe.Pointer(old))
2132 func unlockextra(mp *m) {
2133 extram.Store(uintptr(unsafe.Pointer(mp)))
2137 // allocmLock is locked for read when creating new Ms in allocm and their
2138 // addition to allm. Thus acquiring this lock for write blocks the
2139 // creation of new Ms.
2142 // execLock serializes exec and clone to avoid bugs or unspecified
2143 // behaviour around exec'ing while creating/destroying threads. See
2148 // These errors are reported (via writeErrStr) by some OS-specific
2149 // versions of newosproc and newosproc0.
2151 failthreadcreate = "runtime: failed to create new OS thread\n"
2152 failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
2155 // newmHandoff contains a list of m structures that need new OS threads.
2156 // This is used by newm in situations where newm itself can't safely
2157 // start an OS thread.
2158 var newmHandoff struct {
2161 // newm points to a list of M structures that need new OS
2162 // threads. The list is linked through m.schedlink.
2165 // waiting indicates that wake needs to be notified when an m
2166 // is put on the list.
2170 // haveTemplateThread indicates that the templateThread has
2171 // been started. This is not protected by lock. Use cas to set
2173 haveTemplateThread uint32
2176 // Create a new m. It will start off with a call to fn, or else the scheduler.
2177 // fn needs to be static and not a heap allocated closure.
2178 // May run with m.p==nil, so write barriers are not allowed.
2180 // id is optional pre-allocated m ID. Omit by passing -1.
2182 //go:nowritebarrierrec
2183 func newm(fn func(), pp *p, id int64) {
2184 // allocm adds a new M to allm, but they do not start until created by
2185 // the OS in newm1 or the template thread.
2187 // doAllThreadsSyscall requires that every M in allm will eventually
2188 // start and be signal-able, even with a STW.
2190 // Disable preemption here until we start the thread to ensure that
2191 // newm is not preempted between allocm and starting the new thread,
2192 // ensuring that anything added to allm is guaranteed to eventually
2196 mp := allocm(pp, fn, id)
2198 mp.sigmask = initSigmask
2199 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2200 // We're on a locked M or a thread that may have been
2201 // started by C. The kernel state of this thread may
2202 // be strange (the user may have locked it for that
2203 // purpose). We don't want to clone that into another
2204 // thread. Instead, ask a known-good thread to create
2205 // the thread for us.
2207 // This is disabled on Plan 9. See golang.org/issue/22227.
2209 // TODO: This may be unnecessary on Windows, which
2210 // doesn't model thread creation off fork.
2211 lock(&newmHandoff.lock)
2212 if newmHandoff.haveTemplateThread == 0 {
2213 throw("on a locked thread with no template thread")
2215 mp.schedlink = newmHandoff.newm
2216 newmHandoff.newm.set(mp)
2217 if newmHandoff.waiting {
2218 newmHandoff.waiting = false
2219 notewakeup(&newmHandoff.wake)
2221 unlock(&newmHandoff.lock)
2222 // The M has not started yet, but the template thread does not
2223 // participate in STW, so it will always process queued Ms and
2224 // it is safe to releasem.
2234 var ts cgothreadstart
2235 if _cgo_thread_start == nil {
2236 throw("_cgo_thread_start missing")
2239 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2240 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2242 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2245 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2247 execLock.rlock() // Prevent process clone.
2248 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2252 execLock.rlock() // Prevent process clone.
2257 // startTemplateThread starts the template thread if it is not already
2260 // The calling thread must itself be in a known-good state.
2261 func startTemplateThread() {
2262 if GOARCH == "wasm" { // no threads on wasm yet
2266 // Disable preemption to guarantee that the template thread will be
2267 // created before a park once haveTemplateThread is set.
2269 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2273 newm(templateThread, nil, -1)
2277 // templateThread is a thread in a known-good state that exists solely
2278 // to start new threads in known-good states when the calling thread
2279 // may not be in a good state.
2281 // Many programs never need this, so templateThread is started lazily
2282 // when we first enter a state that might lead to running on a thread
2283 // in an unknown state.
2285 // templateThread runs on an M without a P, so it must not have write
2288 //go:nowritebarrierrec
2289 func templateThread() {
2296 lock(&newmHandoff.lock)
2297 for newmHandoff.newm != 0 {
2298 newm := newmHandoff.newm.ptr()
2299 newmHandoff.newm = 0
2300 unlock(&newmHandoff.lock)
2302 next := newm.schedlink.ptr()
2307 lock(&newmHandoff.lock)
2309 newmHandoff.waiting = true
2310 noteclear(&newmHandoff.wake)
2311 unlock(&newmHandoff.lock)
2312 notesleep(&newmHandoff.wake)
2316 // Stops execution of the current m until new work is available.
2317 // Returns with acquired P.
2321 if gp.m.locks != 0 {
2322 throw("stopm holding locks")
2325 throw("stopm holding p")
2328 throw("stopm spinning")
2335 acquirep(gp.m.nextp.ptr())
2340 // startm's caller incremented nmspinning. Set the new M's spinning.
2341 getg().m.spinning = true
2344 // Schedules some M to run the p (creates an M if necessary).
2345 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2346 // May run with m.p==nil, so write barriers are not allowed.
2347 // If spinning is set, the caller has incremented nmspinning and must provide a
2348 // P. startm will set m.spinning in the newly started M.
2350 // Callers passing a non-nil P must call from a non-preemptible context. See
2351 // comment on acquirem below.
2353 // Must not have write barriers because this may be called without a P.
2355 //go:nowritebarrierrec
2356 func startm(pp *p, spinning bool) {
2357 // Disable preemption.
2359 // Every owned P must have an owner that will eventually stop it in the
2360 // event of a GC stop request. startm takes transient ownership of a P
2361 // (either from argument or pidleget below) and transfers ownership to
2362 // a started M, which will be responsible for performing the stop.
2364 // Preemption must be disabled during this transient ownership,
2365 // otherwise the P this is running on may enter GC stop while still
2366 // holding the transient P, leaving that P in limbo and deadlocking the
2369 // Callers passing a non-nil P must already be in non-preemptible
2370 // context, otherwise such preemption could occur on function entry to
2371 // startm. Callers passing a nil P may be preemptible, so we must
2372 // disable preemption before acquiring a P from pidleget below.
2377 // TODO(prattmic): All remaining calls to this function
2378 // with _p_ == nil could be cleaned up to find a P
2379 // before calling startm.
2380 throw("startm: P required for spinning=true")
2391 // No M is available, we must drop sched.lock and call newm.
2392 // However, we already own a P to assign to the M.
2394 // Once sched.lock is released, another G (e.g., in a syscall),
2395 // could find no idle P while checkdead finds a runnable G but
2396 // no running M's because this new M hasn't started yet, thus
2397 // throwing in an apparent deadlock.
2399 // Avoid this situation by pre-allocating the ID for the new M,
2400 // thus marking it as 'running' before we drop sched.lock. This
2401 // new M will eventually run the scheduler to execute any
2408 // The caller incremented nmspinning, so set m.spinning in the new M.
2412 // Ownership transfer of pp committed by start in newm.
2413 // Preemption is now safe.
2419 throw("startm: m is spinning")
2422 throw("startm: m has p")
2424 if spinning && !runqempty(pp) {
2425 throw("startm: p has runnable gs")
2427 // The caller incremented nmspinning, so set m.spinning in the new M.
2428 nmp.spinning = spinning
2430 notewakeup(&nmp.park)
2431 // Ownership transfer of pp committed by wakeup. Preemption is now
2436 // Hands off P from syscall or locked M.
2437 // Always runs without a P, so write barriers are not allowed.
2439 //go:nowritebarrierrec
2440 func handoffp(pp *p) {
2441 // handoffp must start an M in any situation where
2442 // findrunnable would return a G to run on pp.
2444 // if it has local work, start it straight away
2445 if !runqempty(pp) || sched.runqsize != 0 {
2449 // if there's trace work to do, start it straight away
2450 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2454 // if it has GC work, start it straight away
2455 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2459 // no local work, check that there are no spinning/idle M's,
2460 // otherwise our help is not required
2461 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2462 sched.needspinning.Store(0)
2467 if sched.gcwaiting.Load() {
2468 pp.status = _Pgcstop
2470 if sched.stopwait == 0 {
2471 notewakeup(&sched.stopnote)
2476 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2477 sched.safePointFn(pp)
2478 sched.safePointWait--
2479 if sched.safePointWait == 0 {
2480 notewakeup(&sched.safePointNote)
2483 if sched.runqsize != 0 {
2488 // If this is the last running P and nobody is polling network,
2489 // need to wakeup another M to poll network.
2490 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2496 // The scheduler lock cannot be held when calling wakeNetPoller below
2497 // because wakeNetPoller may call wakep which may call startm.
2498 when := nobarrierWakeTime(pp)
2507 // Tries to add one more P to execute G's.
2508 // Called when a G is made runnable (newproc, ready).
2509 // Must be called with a P.
2511 // Be conservative about spinning threads, only start one if none exist
2513 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2517 // Disable preemption until ownership of pp transfers to the next M in
2518 // startm. Otherwise preemption here would leave pp stuck waiting to
2521 // See preemption comment on acquirem in startm for more details.
2526 pp, _ = pidlegetSpinning(0)
2528 if sched.nmspinning.Add(-1) < 0 {
2529 throw("wakep: negative nmspinning")
2535 // Since we always have a P, the race in the "No M is available"
2536 // comment in startm doesn't apply during the small window between the
2537 // unlock here and lock in startm. A checkdead in between will always
2538 // see at least one running M (ours).
2546 // Stops execution of the current m that is locked to a g until the g is runnable again.
2547 // Returns with acquired P.
2548 func stoplockedm() {
2551 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2552 throw("stoplockedm: inconsistent locking")
2555 // Schedule another M to run this p.
2560 // Wait until another thread schedules lockedg again.
2562 status := readgstatus(gp.m.lockedg.ptr())
2563 if status&^_Gscan != _Grunnable {
2564 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2565 dumpgstatus(gp.m.lockedg.ptr())
2566 throw("stoplockedm: not runnable")
2568 acquirep(gp.m.nextp.ptr())
2572 // Schedules the locked m to run the locked gp.
2573 // May run during STW, so write barriers are not allowed.
2575 //go:nowritebarrierrec
2576 func startlockedm(gp *g) {
2577 mp := gp.lockedm.ptr()
2579 throw("startlockedm: locked to me")
2582 throw("startlockedm: m has p")
2584 // directly handoff current P to the locked m
2588 notewakeup(&mp.park)
2592 // Stops the current m for stopTheWorld.
2593 // Returns when the world is restarted.
2597 if !sched.gcwaiting.Load() {
2598 throw("gcstopm: not waiting for gc")
2601 gp.m.spinning = false
2602 // OK to just drop nmspinning here,
2603 // startTheWorld will unpark threads as necessary.
2604 if sched.nmspinning.Add(-1) < 0 {
2605 throw("gcstopm: negative nmspinning")
2610 pp.status = _Pgcstop
2612 if sched.stopwait == 0 {
2613 notewakeup(&sched.stopnote)
2619 // Schedules gp to run on the current M.
2620 // If inheritTime is true, gp inherits the remaining time in the
2621 // current time slice. Otherwise, it starts a new time slice.
2624 // Write barriers are allowed because this is called immediately after
2625 // acquiring a P in several places.
2627 //go:yeswritebarrierrec
2628 func execute(gp *g, inheritTime bool) {
2631 if goroutineProfile.active {
2632 // Make sure that gp has had its stack written out to the goroutine
2633 // profile, exactly as it was when the goroutine profiler first stopped
2635 tryRecordGoroutineProfile(gp, osyield)
2638 // Assign gp.m before entering _Grunning so running Gs have an
2642 casgstatus(gp, _Grunnable, _Grunning)
2645 gp.stackguard0 = gp.stack.lo + _StackGuard
2647 mp.p.ptr().schedtick++
2650 // Check whether the profiler needs to be turned on or off.
2651 hz := sched.profilehz
2652 if mp.profilehz != hz {
2653 setThreadCPUProfiler(hz)
2657 // GoSysExit has to happen when we have a P, but before GoStart.
2658 // So we emit it here.
2659 if gp.syscallsp != 0 && gp.sysblocktraced {
2660 traceGoSysExit(gp.sysexitticks)
2668 // Finds a runnable goroutine to execute.
2669 // Tries to steal from other P's, get g from local or global queue, poll network.
2670 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2671 // reader) so the caller should try to wake a P.
2672 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2675 // The conditions here and in handoffp must agree: if
2676 // findrunnable would return a G to run, handoffp must start
2681 if sched.gcwaiting.Load() {
2685 if pp.runSafePointFn != 0 {
2689 // now and pollUntil are saved for work stealing later,
2690 // which may steal timers. It's important that between now
2691 // and then, nothing blocks, so these numbers remain mostly
2693 now, pollUntil, _ := checkTimers(pp, 0)
2695 // Try to schedule the trace reader.
2696 if trace.enabled || trace.shutdown {
2699 casgstatus(gp, _Gwaiting, _Grunnable)
2700 traceGoUnpark(gp, 0)
2701 return gp, false, true
2705 // Try to schedule a GC worker.
2706 if gcBlackenEnabled != 0 {
2707 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2709 return gp, false, true
2714 // Check the global runnable queue once in a while to ensure fairness.
2715 // Otherwise two goroutines can completely occupy the local runqueue
2716 // by constantly respawning each other.
2717 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2719 gp := globrunqget(pp, 1)
2722 return gp, false, false
2726 // Wake up the finalizer G.
2727 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2728 if gp := wakefing(); gp != nil {
2732 if *cgo_yield != nil {
2733 asmcgocall(*cgo_yield, nil)
2737 if gp, inheritTime := runqget(pp); gp != nil {
2738 return gp, inheritTime, false
2742 if sched.runqsize != 0 {
2744 gp := globrunqget(pp, 0)
2747 return gp, false, false
2752 // This netpoll is only an optimization before we resort to stealing.
2753 // We can safely skip it if there are no waiters or a thread is blocked
2754 // in netpoll already. If there is any kind of logical race with that
2755 // blocked thread (e.g. it has already returned from netpoll, but does
2756 // not set lastpoll yet), this thread will do blocking netpoll below
2758 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2759 if list := netpoll(0); !list.empty() { // non-blocking
2762 casgstatus(gp, _Gwaiting, _Grunnable)
2764 traceGoUnpark(gp, 0)
2766 return gp, false, false
2770 // Spinning Ms: steal work from other Ps.
2772 // Limit the number of spinning Ms to half the number of busy Ps.
2773 // This is necessary to prevent excessive CPU consumption when
2774 // GOMAXPROCS>>1 but the program parallelism is low.
2775 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2780 gp, inheritTime, tnow, w, newWork := stealWork(now)
2782 // Successfully stole.
2783 return gp, inheritTime, false
2786 // There may be new timer or GC work; restart to
2792 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2793 // Earlier timer to wait for.
2798 // We have nothing to do.
2800 // If we're in the GC mark phase, can safely scan and blacken objects,
2801 // and have work to do, run idle-time marking rather than give up the P.
2802 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2803 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2805 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2807 casgstatus(gp, _Gwaiting, _Grunnable)
2809 traceGoUnpark(gp, 0)
2811 return gp, false, false
2813 gcController.removeIdleMarkWorker()
2817 // If a callback returned and no other goroutine is awake,
2818 // then wake event handler goroutine which pauses execution
2819 // until a callback was triggered.
2820 gp, otherReady := beforeIdle(now, pollUntil)
2822 casgstatus(gp, _Gwaiting, _Grunnable)
2824 traceGoUnpark(gp, 0)
2826 return gp, false, false
2832 // Before we drop our P, make a snapshot of the allp slice,
2833 // which can change underfoot once we no longer block
2834 // safe-points. We don't need to snapshot the contents because
2835 // everything up to cap(allp) is immutable.
2836 allpSnapshot := allp
2837 // Also snapshot masks. Value changes are OK, but we can't allow
2838 // len to change out from under us.
2839 idlepMaskSnapshot := idlepMask
2840 timerpMaskSnapshot := timerpMask
2842 // return P and block
2844 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2848 if sched.runqsize != 0 {
2849 gp := globrunqget(pp, 0)
2851 return gp, false, false
2853 if !mp.spinning && sched.needspinning.Load() == 1 {
2854 // See "Delicate dance" comment below.
2859 if releasep() != pp {
2860 throw("findrunnable: wrong p")
2862 now = pidleput(pp, now)
2865 // Delicate dance: thread transitions from spinning to non-spinning
2866 // state, potentially concurrently with submission of new work. We must
2867 // drop nmspinning first and then check all sources again (with
2868 // #StoreLoad memory barrier in between). If we do it the other way
2869 // around, another thread can submit work after we've checked all
2870 // sources but before we drop nmspinning; as a result nobody will
2871 // unpark a thread to run the work.
2873 // This applies to the following sources of work:
2875 // * Goroutines added to a per-P run queue.
2876 // * New/modified-earlier timers on a per-P timer heap.
2877 // * Idle-priority GC work (barring golang.org/issue/19112).
2879 // If we discover new work below, we need to restore m.spinning as a
2880 // signal for resetspinning to unpark a new worker thread (because
2881 // there can be more than one starving goroutine).
2883 // However, if after discovering new work we also observe no idle Ps
2884 // (either here or in resetspinning), we have a problem. We may be
2885 // racing with a non-spinning M in the block above, having found no
2886 // work and preparing to release its P and park. Allowing that P to go
2887 // idle will result in loss of work conservation (idle P while there is
2888 // runnable work). This could result in complete deadlock in the
2889 // unlikely event that we discover new work (from netpoll) right as we
2890 // are racing with _all_ other Ps going idle.
2892 // We use sched.needspinning to synchronize with non-spinning Ms going
2893 // idle. If needspinning is set when they are about to drop their P,
2894 // they abort the drop and instead become a new spinning M on our
2895 // behalf. If we are not racing and the system is truly fully loaded
2896 // then no spinning threads are required, and the next thread to
2897 // naturally become spinning will clear the flag.
2899 // Also see "Worker thread parking/unparking" comment at the top of the
2901 wasSpinning := mp.spinning
2904 if sched.nmspinning.Add(-1) < 0 {
2905 throw("findrunnable: negative nmspinning")
2908 // Note the for correctness, only the last M transitioning from
2909 // spinning to non-spinning must perform these rechecks to
2910 // ensure no missed work. However, the runtime has some cases
2911 // of transient increments of nmspinning that are decremented
2912 // without going through this path, so we must be conservative
2913 // and perform the check on all spinning Ms.
2915 // See https://go.dev/issue/43997.
2917 // Check all runqueues once again.
2918 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2925 // Check for idle-priority GC work again.
2926 pp, gp := checkIdleGCNoP()
2931 // Run the idle worker.
2932 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2933 casgstatus(gp, _Gwaiting, _Grunnable)
2935 traceGoUnpark(gp, 0)
2937 return gp, false, false
2940 // Finally, check for timer creation or expiry concurrently with
2941 // transitioning from spinning to non-spinning.
2943 // Note that we cannot use checkTimers here because it calls
2944 // adjusttimers which may need to allocate memory, and that isn't
2945 // allowed when we don't have an active P.
2946 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
2949 // Poll network until next timer.
2950 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
2951 sched.pollUntil.Store(pollUntil)
2953 throw("findrunnable: netpoll with p")
2956 throw("findrunnable: netpoll with spinning")
2962 delay = pollUntil - now
2968 // When using fake time, just poll.
2971 list := netpoll(delay) // block until new work is available
2972 sched.pollUntil.Store(0)
2973 sched.lastpoll.Store(now)
2974 if faketime != 0 && list.empty() {
2975 // Using fake time and nothing is ready; stop M.
2976 // When all M's stop, checkdead will call timejump.
2981 pp, _ := pidleget(now)
2990 casgstatus(gp, _Gwaiting, _Grunnable)
2992 traceGoUnpark(gp, 0)
2994 return gp, false, false
3001 } else if pollUntil != 0 && netpollinited() {
3002 pollerPollUntil := sched.pollUntil.Load()
3003 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
3011 // pollWork reports whether there is non-background work this P could
3012 // be doing. This is a fairly lightweight check to be used for
3013 // background work loops, like idle GC. It checks a subset of the
3014 // conditions checked by the actual scheduler.
3015 func pollWork() bool {
3016 if sched.runqsize != 0 {
3019 p := getg().m.p.ptr()
3023 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3024 if list := netpoll(0); !list.empty() {
3032 // stealWork attempts to steal a runnable goroutine or timer from any P.
3034 // If newWork is true, new work may have been readied.
3036 // If now is not 0 it is the current time. stealWork returns the passed time or
3037 // the current time if now was passed as 0.
3038 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3039 pp := getg().m.p.ptr()
3043 const stealTries = 4
3044 for i := 0; i < stealTries; i++ {
3045 stealTimersOrRunNextG := i == stealTries-1
3047 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3048 if sched.gcwaiting.Load() {
3049 // GC work may be available.
3050 return nil, false, now, pollUntil, true
3052 p2 := allp[enum.position()]
3057 // Steal timers from p2. This call to checkTimers is the only place
3058 // where we might hold a lock on a different P's timers. We do this
3059 // once on the last pass before checking runnext because stealing
3060 // from the other P's runnext should be the last resort, so if there
3061 // are timers to steal do that first.
3063 // We only check timers on one of the stealing iterations because
3064 // the time stored in now doesn't change in this loop and checking
3065 // the timers for each P more than once with the same value of now
3066 // is probably a waste of time.
3068 // timerpMask tells us whether the P may have timers at all. If it
3069 // can't, no need to check at all.
3070 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3071 tnow, w, ran := checkTimers(p2, now)
3073 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3077 // Running the timers may have
3078 // made an arbitrary number of G's
3079 // ready and added them to this P's
3080 // local run queue. That invalidates
3081 // the assumption of runqsteal
3082 // that it always has room to add
3083 // stolen G's. So check now if there
3084 // is a local G to run.
3085 if gp, inheritTime := runqget(pp); gp != nil {
3086 return gp, inheritTime, now, pollUntil, ranTimer
3092 // Don't bother to attempt to steal if p2 is idle.
3093 if !idlepMask.read(enum.position()) {
3094 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3095 return gp, false, now, pollUntil, ranTimer
3101 // No goroutines found to steal. Regardless, running a timer may have
3102 // made some goroutine ready that we missed. Indicate the next timer to
3104 return nil, false, now, pollUntil, ranTimer
3107 // Check all Ps for a runnable G to steal.
3109 // On entry we have no P. If a G is available to steal and a P is available,
3110 // the P is returned which the caller should acquire and attempt to steal the
3112 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3113 for id, p2 := range allpSnapshot {
3114 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3116 pp, _ := pidlegetSpinning(0)
3118 // Can't get a P, don't bother checking remaining Ps.
3127 // No work available.
3131 // Check all Ps for a timer expiring sooner than pollUntil.
3133 // Returns updated pollUntil value.
3134 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3135 for id, p2 := range allpSnapshot {
3136 if timerpMaskSnapshot.read(uint32(id)) {
3137 w := nobarrierWakeTime(p2)
3138 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3147 // Check for idle-priority GC, without a P on entry.
3149 // If some GC work, a P, and a worker G are all available, the P and G will be
3150 // returned. The returned P has not been wired yet.
3151 func checkIdleGCNoP() (*p, *g) {
3152 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3153 // must check again after acquiring a P. As an optimization, we also check
3154 // if an idle mark worker is needed at all. This is OK here, because if we
3155 // observe that one isn't needed, at least one is currently running. Even if
3156 // it stops running, its own journey into the scheduler should schedule it
3157 // again, if need be (at which point, this check will pass, if relevant).
3158 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3161 if !gcMarkWorkAvailable(nil) {
3165 // Work is available; we can start an idle GC worker only if there is
3166 // an available P and available worker G.
3168 // We can attempt to acquire these in either order, though both have
3169 // synchronization concerns (see below). Workers are almost always
3170 // available (see comment in findRunnableGCWorker for the one case
3171 // there may be none). Since we're slightly less likely to find a P,
3172 // check for that first.
3174 // Synchronization: note that we must hold sched.lock until we are
3175 // committed to keeping it. Otherwise we cannot put the unnecessary P
3176 // back in sched.pidle without performing the full set of idle
3177 // transition checks.
3179 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3180 // the assumption in gcControllerState.findRunnableGCWorker that an
3181 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3183 pp, now := pidlegetSpinning(0)
3189 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3190 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3196 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3200 gcController.removeIdleMarkWorker()
3206 return pp, node.gp.ptr()
3209 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3210 // going to wake up before the when argument; or it wakes an idle P to service
3211 // timers and the network poller if there isn't one already.
3212 func wakeNetPoller(when int64) {
3213 if sched.lastpoll.Load() == 0 {
3214 // In findrunnable we ensure that when polling the pollUntil
3215 // field is either zero or the time to which the current
3216 // poll is expected to run. This can have a spurious wakeup
3217 // but should never miss a wakeup.
3218 pollerPollUntil := sched.pollUntil.Load()
3219 if pollerPollUntil == 0 || pollerPollUntil > when {
3223 // There are no threads in the network poller, try to get
3224 // one there so it can handle new timers.
3225 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3231 func resetspinning() {
3234 throw("resetspinning: not a spinning m")
3236 gp.m.spinning = false
3237 nmspinning := sched.nmspinning.Add(-1)
3239 throw("findrunnable: negative nmspinning")
3241 // M wakeup policy is deliberately somewhat conservative, so check if we
3242 // need to wakeup another P here. See "Worker thread parking/unparking"
3243 // comment at the top of the file for details.
3247 // injectglist adds each runnable G on the list to some run queue,
3248 // and clears glist. If there is no current P, they are added to the
3249 // global queue, and up to npidle M's are started to run them.
3250 // Otherwise, for each idle P, this adds a G to the global queue
3251 // and starts an M. Any remaining G's are added to the current P's
3253 // This may temporarily acquire sched.lock.
3254 // Can run concurrently with GC.
3255 func injectglist(glist *gList) {
3260 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3261 traceGoUnpark(gp, 0)
3265 // Mark all the goroutines as runnable before we put them
3266 // on the run queues.
3267 head := glist.head.ptr()
3270 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3273 casgstatus(gp, _Gwaiting, _Grunnable)
3276 // Turn the gList into a gQueue.
3282 startIdle := func(n int) {
3283 for i := 0; i < n; i++ {
3284 mp := acquirem() // See comment in startm.
3287 pp, _ := pidlegetSpinning(0)
3300 pp := getg().m.p.ptr()
3303 globrunqputbatch(&q, int32(qsize))
3309 npidle := int(sched.npidle.Load())
3312 for n = 0; n < npidle && !q.empty(); n++ {
3318 globrunqputbatch(&globq, int32(n))
3325 runqputbatch(pp, &q, qsize)
3329 // One round of scheduler: find a runnable goroutine and execute it.
3335 throw("schedule: holding locks")
3338 if mp.lockedg != 0 {
3340 execute(mp.lockedg.ptr(), false) // Never returns.
3343 // We should not schedule away from a g that is executing a cgo call,
3344 // since the cgo call is using the m's g0 stack.
3346 throw("schedule: in cgo")
3353 // Safety check: if we are spinning, the run queue should be empty.
3354 // Check this before calling checkTimers, as that might call
3355 // goready to put a ready goroutine on the local run queue.
3356 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3357 throw("schedule: spinning with local work")
3360 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3362 // This thread is going to run a goroutine and is not spinning anymore,
3363 // so if it was marked as spinning we need to reset it now and potentially
3364 // start a new spinning M.
3369 if sched.disable.user && !schedEnabled(gp) {
3370 // Scheduling of this goroutine is disabled. Put it on
3371 // the list of pending runnable goroutines for when we
3372 // re-enable user scheduling and look again.
3374 if schedEnabled(gp) {
3375 // Something re-enabled scheduling while we
3376 // were acquiring the lock.
3379 sched.disable.runnable.pushBack(gp)
3386 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3387 // wake a P if there is one.
3391 if gp.lockedm != 0 {
3392 // Hands off own p to the locked m,
3393 // then blocks waiting for a new p.
3398 execute(gp, inheritTime)
3401 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3402 // Typically a caller sets gp's status away from Grunning and then
3403 // immediately calls dropg to finish the job. The caller is also responsible
3404 // for arranging that gp will be restarted using ready at an
3405 // appropriate time. After calling dropg and arranging for gp to be
3406 // readied later, the caller can do other work but eventually should
3407 // call schedule to restart the scheduling of goroutines on this m.
3411 setMNoWB(&gp.m.curg.m, nil)
3412 setGNoWB(&gp.m.curg, nil)
3415 // checkTimers runs any timers for the P that are ready.
3416 // If now is not 0 it is the current time.
3417 // It returns the passed time or the current time if now was passed as 0.
3418 // and the time when the next timer should run or 0 if there is no next timer,
3419 // and reports whether it ran any timers.
3420 // If the time when the next timer should run is not 0,
3421 // it is always larger than the returned time.
3422 // We pass now in and out to avoid extra calls of nanotime.
3424 //go:yeswritebarrierrec
3425 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3426 // If it's not yet time for the first timer, or the first adjusted
3427 // timer, then there is nothing to do.
3428 next := pp.timer0When.Load()
3429 nextAdj := pp.timerModifiedEarliest.Load()
3430 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3435 // No timers to run or adjust.
3436 return now, 0, false
3443 // Next timer is not ready to run, but keep going
3444 // if we would clear deleted timers.
3445 // This corresponds to the condition below where
3446 // we decide whether to call clearDeletedTimers.
3447 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3448 return now, next, false
3452 lock(&pp.timersLock)
3454 if len(pp.timers) > 0 {
3455 adjusttimers(pp, now)
3456 for len(pp.timers) > 0 {
3457 // Note that runtimer may temporarily unlock
3459 if tw := runtimer(pp, now); tw != 0 {
3469 // If this is the local P, and there are a lot of deleted timers,
3470 // clear them out. We only do this for the local P to reduce
3471 // lock contention on timersLock.
3472 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3473 clearDeletedTimers(pp)
3476 unlock(&pp.timersLock)
3478 return now, pollUntil, ran
3481 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3482 unlock((*mutex)(lock))
3486 // park continuation on g0.
3487 func park_m(gp *g) {
3491 traceGoPark(mp.waittraceev, mp.waittraceskip)
3494 // N.B. Not using casGToWaiting here because the waitreason is
3495 // set by park_m's caller.
3496 casgstatus(gp, _Grunning, _Gwaiting)
3499 if fn := mp.waitunlockf; fn != nil {
3500 ok := fn(gp, mp.waitlock)
3501 mp.waitunlockf = nil
3505 traceGoUnpark(gp, 2)
3507 casgstatus(gp, _Gwaiting, _Grunnable)
3508 execute(gp, true) // Schedule it back, never returns.
3514 func goschedImpl(gp *g) {
3515 status := readgstatus(gp)
3516 if status&^_Gscan != _Grunning {
3518 throw("bad g status")
3520 casgstatus(gp, _Grunning, _Grunnable)
3529 // Gosched continuation on g0.
3530 func gosched_m(gp *g) {
3537 // goschedguarded is a forbidden-states-avoided version of gosched_m.
3538 func goschedguarded_m(gp *g) {
3540 if !canPreemptM(gp.m) {
3541 gogo(&gp.sched) // never return
3550 func gopreempt_m(gp *g) {
3557 // preemptPark parks gp and puts it in _Gpreempted.
3560 func preemptPark(gp *g) {
3562 traceGoPark(traceEvGoBlock, 0)
3564 status := readgstatus(gp)
3565 if status&^_Gscan != _Grunning {
3567 throw("bad g status")
3570 if gp.asyncSafePoint {
3571 // Double-check that async preemption does not
3572 // happen in SPWRITE assembly functions.
3573 // isAsyncSafePoint must exclude this case.
3574 f := findfunc(gp.sched.pc)
3576 throw("preempt at unknown pc")
3578 if f.flag&funcFlag_SPWRITE != 0 {
3579 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3580 throw("preempt SPWRITE")
3584 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3585 // be in _Grunning when we dropg because then we'd be running
3586 // without an M, but the moment we're in _Gpreempted,
3587 // something could claim this G before we've fully cleaned it
3588 // up. Hence, we set the scan bit to lock down further
3589 // transitions until we can dropg.
3590 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3592 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3596 // goyield is like Gosched, but it:
3597 // - emits a GoPreempt trace event instead of a GoSched trace event
3598 // - puts the current G on the runq of the current P instead of the globrunq
3604 func goyield_m(gp *g) {
3609 casgstatus(gp, _Grunning, _Grunnable)
3611 runqput(pp, gp, false)
3615 // Finishes execution of the current goroutine.
3626 // goexit continuation on g0.
3627 func goexit0(gp *g) {
3631 casgstatus(gp, _Grunning, _Gdead)
3632 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3633 if isSystemGoroutine(gp, false) {
3637 locked := gp.lockedm != 0
3640 gp.preemptStop = false
3641 gp.paniconfault = false
3642 gp._defer = nil // should be true already but just in case.
3643 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3645 gp.waitreason = waitReasonZero
3650 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3651 // Flush assist credit to the global pool. This gives
3652 // better information to pacing if the application is
3653 // rapidly creating an exiting goroutines.
3654 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3655 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3656 gcController.bgScanCredit.Add(scanCredit)
3657 gp.gcAssistBytes = 0
3662 if GOARCH == "wasm" { // no threads yet on wasm
3664 schedule() // never returns
3667 if mp.lockedInt != 0 {
3668 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3669 throw("internal lockOSThread error")
3673 // The goroutine may have locked this thread because
3674 // it put it in an unusual kernel state. Kill it
3675 // rather than returning it to the thread pool.
3677 // Return to mstart, which will release the P and exit
3679 if GOOS != "plan9" { // See golang.org/issue/22227.
3682 // Clear lockedExt on plan9 since we may end up re-using
3690 // save updates getg().sched to refer to pc and sp so that a following
3691 // gogo will restore pc and sp.
3693 // save must not have write barriers because invoking a write barrier
3694 // can clobber getg().sched.
3697 //go:nowritebarrierrec
3698 func save(pc, sp uintptr) {
3701 if gp == gp.m.g0 || gp == gp.m.gsignal {
3702 // m.g0.sched is special and must describe the context
3703 // for exiting the thread. mstart1 writes to it directly.
3704 // m.gsignal.sched should not be used at all.
3705 // This check makes sure save calls do not accidentally
3706 // run in contexts where they'd write to system g's.
3707 throw("save on system g not allowed")
3714 // We need to ensure ctxt is zero, but can't have a write
3715 // barrier here. However, it should always already be zero.
3717 if gp.sched.ctxt != nil {
3722 // The goroutine g is about to enter a system call.
3723 // Record that it's not using the cpu anymore.
3724 // This is called only from the go syscall library and cgocall,
3725 // not from the low-level system calls used by the runtime.
3727 // Entersyscall cannot split the stack: the save must
3728 // make g->sched refer to the caller's stack segment, because
3729 // entersyscall is going to return immediately after.
3731 // Nothing entersyscall calls can split the stack either.
3732 // We cannot safely move the stack during an active call to syscall,
3733 // because we do not know which of the uintptr arguments are
3734 // really pointers (back into the stack).
3735 // In practice, this means that we make the fast path run through
3736 // entersyscall doing no-split things, and the slow path has to use systemstack
3737 // to run bigger things on the system stack.
3739 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3740 // saved SP and PC are restored. This is needed when exitsyscall will be called
3741 // from a function further up in the call stack than the parent, as g->syscallsp
3742 // must always point to a valid stack frame. entersyscall below is the normal
3743 // entry point for syscalls, which obtains the SP and PC from the caller.
3746 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3747 // If the syscall does not block, that is it, we do not emit any other events.
3748 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3749 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3750 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3751 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3752 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3753 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3754 // and we wait for the increment before emitting traceGoSysExit.
3755 // Note that the increment is done even if tracing is not enabled,
3756 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3759 func reentersyscall(pc, sp uintptr) {
3762 // Disable preemption because during this function g is in Gsyscall status,
3763 // but can have inconsistent g->sched, do not let GC observe it.
3766 // Entersyscall must not call any function that might split/grow the stack.
3767 // (See details in comment above.)
3768 // Catch calls that might, by replacing the stack guard with something that
3769 // will trip any stack check and leaving a flag to tell newstack to die.
3770 gp.stackguard0 = stackPreempt
3771 gp.throwsplit = true
3773 // Leave SP around for GC and traceback.
3777 casgstatus(gp, _Grunning, _Gsyscall)
3778 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3779 systemstack(func() {
3780 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3781 throw("entersyscall")
3786 systemstack(traceGoSysCall)
3787 // systemstack itself clobbers g.sched.{pc,sp} and we might
3788 // need them later when the G is genuinely blocked in a
3793 if sched.sysmonwait.Load() {
3794 systemstack(entersyscall_sysmon)
3798 if gp.m.p.ptr().runSafePointFn != 0 {
3799 // runSafePointFn may stack split if run on this stack
3800 systemstack(runSafePointFn)
3804 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3805 gp.sysblocktraced = true
3810 atomic.Store(&pp.status, _Psyscall)
3811 if sched.gcwaiting.Load() {
3812 systemstack(entersyscall_gcwait)
3819 // Standard syscall entry used by the go syscall library and normal cgo calls.
3821 // This is exported via linkname to assembly in the syscall package and x/sys.
3824 //go:linkname entersyscall
3825 func entersyscall() {
3826 reentersyscall(getcallerpc(), getcallersp())
3829 func entersyscall_sysmon() {
3831 if sched.sysmonwait.Load() {
3832 sched.sysmonwait.Store(false)
3833 notewakeup(&sched.sysmonnote)
3838 func entersyscall_gcwait() {
3840 pp := gp.m.oldp.ptr()
3843 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3849 if sched.stopwait--; sched.stopwait == 0 {
3850 notewakeup(&sched.stopnote)
3856 // The same as entersyscall(), but with a hint that the syscall is blocking.
3859 func entersyscallblock() {
3862 gp.m.locks++ // see comment in entersyscall
3863 gp.throwsplit = true
3864 gp.stackguard0 = stackPreempt // see comment in entersyscall
3865 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3866 gp.sysblocktraced = true
3867 gp.m.p.ptr().syscalltick++
3869 // Leave SP around for GC and traceback.
3873 gp.syscallsp = gp.sched.sp
3874 gp.syscallpc = gp.sched.pc
3875 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3879 systemstack(func() {
3880 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3881 throw("entersyscallblock")
3884 casgstatus(gp, _Grunning, _Gsyscall)
3885 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3886 systemstack(func() {
3887 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3888 throw("entersyscallblock")
3892 systemstack(entersyscallblock_handoff)
3894 // Resave for traceback during blocked call.
3895 save(getcallerpc(), getcallersp())
3900 func entersyscallblock_handoff() {
3903 traceGoSysBlock(getg().m.p.ptr())
3905 handoffp(releasep())
3908 // The goroutine g exited its system call.
3909 // Arrange for it to run on a cpu again.
3910 // This is called only from the go syscall library, not
3911 // from the low-level system calls used by the runtime.
3913 // Write barriers are not allowed because our P may have been stolen.
3915 // This is exported via linkname to assembly in the syscall package.
3918 //go:nowritebarrierrec
3919 //go:linkname exitsyscall
3920 func exitsyscall() {
3923 gp.m.locks++ // see comment in entersyscall
3924 if getcallersp() > gp.syscallsp {
3925 throw("exitsyscall: syscall frame is no longer valid")
3929 oldp := gp.m.oldp.ptr()
3931 if exitsyscallfast(oldp) {
3932 // When exitsyscallfast returns success, we have a P so can now use
3934 if goroutineProfile.active {
3935 // Make sure that gp has had its stack written out to the goroutine
3936 // profile, exactly as it was when the goroutine profiler first
3937 // stopped the world.
3938 systemstack(func() {
3939 tryRecordGoroutineProfileWB(gp)
3943 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3944 systemstack(traceGoStart)
3947 // There's a cpu for us, so we can run.
3948 gp.m.p.ptr().syscalltick++
3949 // We need to cas the status and scan before resuming...
3950 casgstatus(gp, _Gsyscall, _Grunning)
3952 // Garbage collector isn't running (since we are),
3953 // so okay to clear syscallsp.
3957 // restore the preemption request in case we've cleared it in newstack
3958 gp.stackguard0 = stackPreempt
3960 // otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
3961 gp.stackguard0 = gp.stack.lo + _StackGuard
3963 gp.throwsplit = false
3965 if sched.disable.user && !schedEnabled(gp) {
3966 // Scheduling of this goroutine is disabled.
3975 // Wait till traceGoSysBlock event is emitted.
3976 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3977 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
3980 // We can't trace syscall exit right now because we don't have a P.
3981 // Tracing code can invoke write barriers that cannot run without a P.
3982 // So instead we remember the syscall exit time and emit the event
3983 // in execute when we have a P.
3984 gp.sysexitticks = cputicks()
3989 // Call the scheduler.
3992 // Scheduler returned, so we're allowed to run now.
3993 // Delete the syscallsp information that we left for
3994 // the garbage collector during the system call.
3995 // Must wait until now because until gosched returns
3996 // we don't know for sure that the garbage collector
3999 gp.m.p.ptr().syscalltick++
4000 gp.throwsplit = false
4004 func exitsyscallfast(oldp *p) bool {
4007 // Freezetheworld sets stopwait but does not retake P's.
4008 if sched.stopwait == freezeStopWait {
4012 // Try to re-acquire the last P.
4013 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
4014 // There's a cpu for us, so we can run.
4016 exitsyscallfast_reacquired()
4020 // Try to get any other idle P.
4021 if sched.pidle != 0 {
4023 systemstack(func() {
4024 ok = exitsyscallfast_pidle()
4025 if ok && trace.enabled {
4027 // Wait till traceGoSysBlock event is emitted.
4028 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4029 for oldp.syscalltick == gp.m.syscalltick {
4043 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4044 // has successfully reacquired the P it was running on before the
4048 func exitsyscallfast_reacquired() {
4050 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4052 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4053 // traceGoSysBlock for this syscall was already emitted,
4054 // but here we effectively retake the p from the new syscall running on the same p.
4055 systemstack(func() {
4056 // Denote blocking of the new syscall.
4057 traceGoSysBlock(gp.m.p.ptr())
4058 // Denote completion of the current syscall.
4062 gp.m.p.ptr().syscalltick++
4066 func exitsyscallfast_pidle() bool {
4068 pp, _ := pidleget(0)
4069 if pp != nil && sched.sysmonwait.Load() {
4070 sched.sysmonwait.Store(false)
4071 notewakeup(&sched.sysmonnote)
4081 // exitsyscall slow path on g0.
4082 // Failed to acquire P, enqueue gp as runnable.
4084 // Called via mcall, so gp is the calling g from this M.
4086 //go:nowritebarrierrec
4087 func exitsyscall0(gp *g) {
4088 casgstatus(gp, _Gsyscall, _Grunnable)
4092 if schedEnabled(gp) {
4099 // Below, we stoplockedm if gp is locked. globrunqput releases
4100 // ownership of gp, so we must check if gp is locked prior to
4101 // committing the release by unlocking sched.lock, otherwise we
4102 // could race with another M transitioning gp from unlocked to
4104 locked = gp.lockedm != 0
4105 } else if sched.sysmonwait.Load() {
4106 sched.sysmonwait.Store(false)
4107 notewakeup(&sched.sysmonnote)
4112 execute(gp, false) // Never returns.
4115 // Wait until another thread schedules gp and so m again.
4117 // N.B. lockedm must be this M, as this g was running on this M
4118 // before entersyscall.
4120 execute(gp, false) // Never returns.
4123 schedule() // Never returns.
4126 // Called from syscall package before fork.
4128 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4130 func syscall_runtime_BeforeFork() {
4133 // Block signals during a fork, so that the child does not run
4134 // a signal handler before exec if a signal is sent to the process
4135 // group. See issue #18600.
4137 sigsave(&gp.m.sigmask)
4140 // This function is called before fork in syscall package.
4141 // Code between fork and exec must not allocate memory nor even try to grow stack.
4142 // Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
4143 // runtime_AfterFork will undo this in parent process, but not in child.
4144 gp.stackguard0 = stackFork
4147 // Called from syscall package after fork in parent.
4149 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4151 func syscall_runtime_AfterFork() {
4154 // See the comments in beforefork.
4155 gp.stackguard0 = gp.stack.lo + _StackGuard
4157 msigrestore(gp.m.sigmask)
4162 // inForkedChild is true while manipulating signals in the child process.
4163 // This is used to avoid calling libc functions in case we are using vfork.
4164 var inForkedChild bool
4166 // Called from syscall package after fork in child.
4167 // It resets non-sigignored signals to the default handler, and
4168 // restores the signal mask in preparation for the exec.
4170 // Because this might be called during a vfork, and therefore may be
4171 // temporarily sharing address space with the parent process, this must
4172 // not change any global variables or calling into C code that may do so.
4174 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4176 //go:nowritebarrierrec
4177 func syscall_runtime_AfterForkInChild() {
4178 // It's OK to change the global variable inForkedChild here
4179 // because we are going to change it back. There is no race here,
4180 // because if we are sharing address space with the parent process,
4181 // then the parent process can not be running concurrently.
4182 inForkedChild = true
4184 clearSignalHandlers()
4186 // When we are the child we are the only thread running,
4187 // so we know that nothing else has changed gp.m.sigmask.
4188 msigrestore(getg().m.sigmask)
4190 inForkedChild = false
4193 // pendingPreemptSignals is the number of preemption signals
4194 // that have been sent but not received. This is only used on Darwin.
4196 var pendingPreemptSignals atomic.Int32
4198 // Called from syscall package before Exec.
4200 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4201 func syscall_runtime_BeforeExec() {
4202 // Prevent thread creation during exec.
4205 // On Darwin, wait for all pending preemption signals to
4206 // be received. See issue #41702.
4207 if GOOS == "darwin" || GOOS == "ios" {
4208 for pendingPreemptSignals.Load() > 0 {
4214 // Called from syscall package after Exec.
4216 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4217 func syscall_runtime_AfterExec() {
4221 // Allocate a new g, with a stack big enough for stacksize bytes.
4222 func malg(stacksize int32) *g {
4225 stacksize = round2(_StackSystem + stacksize)
4226 systemstack(func() {
4227 newg.stack = stackalloc(uint32(stacksize))
4229 newg.stackguard0 = newg.stack.lo + _StackGuard
4230 newg.stackguard1 = ^uintptr(0)
4231 // Clear the bottom word of the stack. We record g
4232 // there on gsignal stack during VDSO on ARM and ARM64.
4233 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4238 // Create a new g running fn.
4239 // Put it on the queue of g's waiting to run.
4240 // The compiler turns a go statement into a call to this.
4241 func newproc(fn *funcval) {
4244 systemstack(func() {
4245 newg := newproc1(fn, gp, pc)
4247 pp := getg().m.p.ptr()
4248 runqput(pp, newg, true)
4256 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4257 // address of the go statement that created this. The caller is responsible
4258 // for adding the new g to the scheduler.
4259 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4261 fatal("go of nil func value")
4264 mp := acquirem() // disable preemption because we hold M and P in local vars.
4268 newg = malg(_StackMin)
4269 casgstatus(newg, _Gidle, _Gdead)
4270 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4272 if newg.stack.hi == 0 {
4273 throw("newproc1: newg missing stack")
4276 if readgstatus(newg) != _Gdead {
4277 throw("newproc1: new g is not Gdead")
4280 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4281 totalSize = alignUp(totalSize, sys.StackAlign)
4282 sp := newg.stack.hi - totalSize
4286 *(*uintptr)(unsafe.Pointer(sp)) = 0
4288 spArg += sys.MinFrameSize
4291 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4294 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4295 newg.sched.g = guintptr(unsafe.Pointer(newg))
4296 gostartcallfn(&newg.sched, fn)
4297 newg.gopc = callerpc
4298 newg.ancestors = saveAncestors(callergp)
4299 newg.startpc = fn.fn
4300 if isSystemGoroutine(newg, false) {
4303 // Only user goroutines inherit pprof labels.
4305 newg.labels = mp.curg.labels
4307 if goroutineProfile.active {
4308 // A concurrent goroutine profile is running. It should include
4309 // exactly the set of goroutines that were alive when the goroutine
4310 // profiler first stopped the world. That does not include newg, so
4311 // mark it as not needing a profile before transitioning it from
4313 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4316 // Track initial transition?
4317 newg.trackingSeq = uint8(fastrand())
4318 if newg.trackingSeq%gTrackingPeriod == 0 {
4319 newg.tracking = true
4321 casgstatus(newg, _Gdead, _Grunnable)
4322 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4324 if pp.goidcache == pp.goidcacheend {
4325 // Sched.goidgen is the last allocated id,
4326 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4327 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4328 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4329 pp.goidcache -= _GoidCacheBatch - 1
4330 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4332 newg.goid = pp.goidcache
4335 newg.racectx = racegostart(callerpc)
4336 if newg.labels != nil {
4337 // See note in proflabel.go on labelSync's role in synchronizing
4338 // with the reads in the signal handler.
4339 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4343 traceGoCreate(newg, newg.startpc)
4350 // saveAncestors copies previous ancestors of the given caller g and
4351 // includes info for the current caller into a new set of tracebacks for
4352 // a g being created.
4353 func saveAncestors(callergp *g) *[]ancestorInfo {
4354 // Copy all prior info, except for the root goroutine (goid 0).
4355 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4358 var callerAncestors []ancestorInfo
4359 if callergp.ancestors != nil {
4360 callerAncestors = *callergp.ancestors
4362 n := int32(len(callerAncestors)) + 1
4363 if n > debug.tracebackancestors {
4364 n = debug.tracebackancestors
4366 ancestors := make([]ancestorInfo, n)
4367 copy(ancestors[1:], callerAncestors)
4369 var pcs [_TracebackMaxFrames]uintptr
4370 npcs := gcallers(callergp, 0, pcs[:])
4371 ipcs := make([]uintptr, npcs)
4373 ancestors[0] = ancestorInfo{
4375 goid: callergp.goid,
4376 gopc: callergp.gopc,
4379 ancestorsp := new([]ancestorInfo)
4380 *ancestorsp = ancestors
4384 // Put on gfree list.
4385 // If local list is too long, transfer a batch to the global list.
4386 func gfput(pp *p, gp *g) {
4387 if readgstatus(gp) != _Gdead {
4388 throw("gfput: bad status (not Gdead)")
4391 stksize := gp.stack.hi - gp.stack.lo
4393 if stksize != uintptr(startingStackSize) {
4394 // non-standard stack size - free it.
4403 if pp.gFree.n >= 64 {
4409 for pp.gFree.n >= 32 {
4410 gp := pp.gFree.pop()
4412 if gp.stack.lo == 0 {
4419 lock(&sched.gFree.lock)
4420 sched.gFree.noStack.pushAll(noStackQ)
4421 sched.gFree.stack.pushAll(stackQ)
4422 sched.gFree.n += inc
4423 unlock(&sched.gFree.lock)
4427 // Get from gfree list.
4428 // If local list is empty, grab a batch from global list.
4429 func gfget(pp *p) *g {
4431 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4432 lock(&sched.gFree.lock)
4433 // Move a batch of free Gs to the P.
4434 for pp.gFree.n < 32 {
4435 // Prefer Gs with stacks.
4436 gp := sched.gFree.stack.pop()
4438 gp = sched.gFree.noStack.pop()
4447 unlock(&sched.gFree.lock)
4450 gp := pp.gFree.pop()
4455 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4456 // Deallocate old stack. We kept it in gfput because it was the
4457 // right size when the goroutine was put on the free list, but
4458 // the right size has changed since then.
4459 systemstack(func() {
4466 if gp.stack.lo == 0 {
4467 // Stack was deallocated in gfput or just above. Allocate a new one.
4468 systemstack(func() {
4469 gp.stack = stackalloc(startingStackSize)
4471 gp.stackguard0 = gp.stack.lo + _StackGuard
4474 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4477 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4480 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4486 // Purge all cached G's from gfree list to the global list.
4487 func gfpurge(pp *p) {
4493 for !pp.gFree.empty() {
4494 gp := pp.gFree.pop()
4496 if gp.stack.lo == 0 {
4503 lock(&sched.gFree.lock)
4504 sched.gFree.noStack.pushAll(noStackQ)
4505 sched.gFree.stack.pushAll(stackQ)
4506 sched.gFree.n += inc
4507 unlock(&sched.gFree.lock)
4510 // Breakpoint executes a breakpoint trap.
4515 // dolockOSThread is called by LockOSThread and lockOSThread below
4516 // after they modify m.locked. Do not allow preemption during this call,
4517 // or else the m might be different in this function than in the caller.
4520 func dolockOSThread() {
4521 if GOARCH == "wasm" {
4522 return // no threads on wasm yet
4525 gp.m.lockedg.set(gp)
4526 gp.lockedm.set(gp.m)
4531 // LockOSThread wires the calling goroutine to its current operating system thread.
4532 // The calling goroutine will always execute in that thread,
4533 // and no other goroutine will execute in it,
4534 // until the calling goroutine has made as many calls to
4535 // UnlockOSThread as to LockOSThread.
4536 // If the calling goroutine exits without unlocking the thread,
4537 // the thread will be terminated.
4539 // All init functions are run on the startup thread. Calling LockOSThread
4540 // from an init function will cause the main function to be invoked on
4543 // A goroutine should call LockOSThread before calling OS services or
4544 // non-Go library functions that depend on per-thread state.
4545 func LockOSThread() {
4546 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4547 // If we need to start a new thread from the locked
4548 // thread, we need the template thread. Start it now
4549 // while we're in a known-good state.
4550 startTemplateThread()
4554 if gp.m.lockedExt == 0 {
4556 panic("LockOSThread nesting overflow")
4562 func lockOSThread() {
4563 getg().m.lockedInt++
4567 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4568 // after they update m->locked. Do not allow preemption during this call,
4569 // or else the m might be in different in this function than in the caller.
4572 func dounlockOSThread() {
4573 if GOARCH == "wasm" {
4574 return // no threads on wasm yet
4577 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4586 // UnlockOSThread undoes an earlier call to LockOSThread.
4587 // If this drops the number of active LockOSThread calls on the
4588 // calling goroutine to zero, it unwires the calling goroutine from
4589 // its fixed operating system thread.
4590 // If there are no active LockOSThread calls, this is a no-op.
4592 // Before calling UnlockOSThread, the caller must ensure that the OS
4593 // thread is suitable for running other goroutines. If the caller made
4594 // any permanent changes to the state of the thread that would affect
4595 // other goroutines, it should not call this function and thus leave
4596 // the goroutine locked to the OS thread until the goroutine (and
4597 // hence the thread) exits.
4598 func UnlockOSThread() {
4600 if gp.m.lockedExt == 0 {
4608 func unlockOSThread() {
4610 if gp.m.lockedInt == 0 {
4611 systemstack(badunlockosthread)
4617 func badunlockosthread() {
4618 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4621 func gcount() int32 {
4622 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4623 for _, pp := range allp {
4627 // All these variables can be changed concurrently, so the result can be inconsistent.
4628 // But at least the current goroutine is running.
4635 func mcount() int32 {
4636 return int32(sched.mnext - sched.nmfreed)
4640 signalLock atomic.Uint32
4642 // Must hold signalLock to write. Reads may be lock-free, but
4643 // signalLock should be taken to synchronize with changes.
4647 func _System() { _System() }
4648 func _ExternalCode() { _ExternalCode() }
4649 func _LostExternalCode() { _LostExternalCode() }
4650 func _GC() { _GC() }
4651 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4652 func _VDSO() { _VDSO() }
4654 // Called if we receive a SIGPROF signal.
4655 // Called by the signal handler, may run during STW.
4657 //go:nowritebarrierrec
4658 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4659 if prof.hz.Load() == 0 {
4663 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4664 // We must check this to avoid a deadlock between setcpuprofilerate
4665 // and the call to cpuprof.add, below.
4666 if mp != nil && mp.profilehz == 0 {
4670 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4671 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4672 // the critical section, it creates a deadlock (when writing the sample).
4673 // As a workaround, create a counter of SIGPROFs while in critical section
4674 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4675 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4676 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4677 if f := findfunc(pc); f.valid() {
4678 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4679 cpuprof.lostAtomic++
4683 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4684 // runtime/internal/atomic functions call into kernel
4685 // helpers on arm < 7. See
4686 // runtime/internal/atomic/sys_linux_arm.s.
4687 cpuprof.lostAtomic++
4692 // Profiling runs concurrently with GC, so it must not allocate.
4693 // Set a trap in case the code does allocate.
4694 // Note that on windows, one thread takes profiles of all the
4695 // other threads, so mp is usually not getg().m.
4696 // In fact mp may not even be stopped.
4697 // See golang.org/issue/17165.
4698 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 copy(stk[:], mp.cgoCallers[:cgoOff])
4714 mp.cgoCallers[0] = 0
4717 // Collect Go stack that leads to the cgo call.
4718 n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
4722 } else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4723 // Libcall, i.e. runtime syscall on windows.
4724 // Collect Go stack that leads to the call.
4725 n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[n], len(stk[n:]), nil, nil, 0)
4726 } else if mp != nil && mp.vdsoSP != 0 {
4727 // VDSO call, e.g. nanotime1 on Linux.
4728 // Collect Go stack that leads to the call.
4729 n = gentraceback(mp.vdsoPC, mp.vdsoSP, 0, gp, 0, &stk[n], len(stk[n:]), nil, nil, _TraceJumpStack)
4731 n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4735 // Normal traceback is impossible or has failed.
4736 // Account it against abstract "System" or "GC".
4739 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4740 } else if pc > firstmoduledata.etext {
4741 // "ExternalCode" is better than "etext".
4742 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4745 if mp.preemptoff != "" {
4746 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4748 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4752 if prof.hz.Load() != 0 {
4753 // Note: it can happen on Windows that we interrupted a system thread
4754 // with no g, so gp could nil. The other nil checks are done out of
4755 // caution, but not expected to be nil in practice.
4756 var tagPtr *unsafe.Pointer
4757 if gp != nil && gp.m != nil && gp.m.curg != nil {
4758 tagPtr = &gp.m.curg.labels
4760 cpuprof.add(tagPtr, stk[:n])
4764 if gp != nil && gp.m != nil {
4765 if gp.m.curg != nil {
4770 traceCPUSample(gprof, pp, stk[:n])
4772 getg().m.mallocing--
4775 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4776 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4777 func setcpuprofilerate(hz int32) {
4778 // Force sane arguments.
4783 // Disable preemption, otherwise we can be rescheduled to another thread
4784 // that has profiling enabled.
4788 // Stop profiler on this thread so that it is safe to lock prof.
4789 // if a profiling signal came in while we had prof locked,
4790 // it would deadlock.
4791 setThreadCPUProfiler(0)
4793 for !prof.signalLock.CompareAndSwap(0, 1) {
4796 if prof.hz.Load() != hz {
4797 setProcessCPUProfiler(hz)
4800 prof.signalLock.Store(0)
4803 sched.profilehz = hz
4807 setThreadCPUProfiler(hz)
4813 // init initializes pp, which may be a freshly allocated p or a
4814 // previously destroyed p, and transitions it to status _Pgcstop.
4815 func (pp *p) init(id int32) {
4817 pp.status = _Pgcstop
4818 pp.sudogcache = pp.sudogbuf[:0]
4819 pp.deferpool = pp.deferpoolbuf[:0]
4821 if pp.mcache == nil {
4824 throw("missing mcache?")
4826 // Use the bootstrap mcache0. Only one P will get
4827 // mcache0: the one with ID 0.
4830 pp.mcache = allocmcache()
4833 if raceenabled && pp.raceprocctx == 0 {
4835 pp.raceprocctx = raceprocctx0
4836 raceprocctx0 = 0 // bootstrap
4838 pp.raceprocctx = raceproccreate()
4841 lockInit(&pp.timersLock, lockRankTimers)
4843 // This P may get timers when it starts running. Set the mask here
4844 // since the P may not go through pidleget (notably P 0 on startup).
4846 // Similarly, we may not go through pidleget before this P starts
4847 // running if it is P 0 on startup.
4851 // destroy releases all of the resources associated with pp and
4852 // transitions it to status _Pdead.
4854 // sched.lock must be held and the world must be stopped.
4855 func (pp *p) destroy() {
4856 assertLockHeld(&sched.lock)
4857 assertWorldStopped()
4859 // Move all runnable goroutines to the global queue
4860 for pp.runqhead != pp.runqtail {
4861 // Pop from tail of local queue
4863 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4864 // Push onto head of global queue
4867 if pp.runnext != 0 {
4868 globrunqputhead(pp.runnext.ptr())
4871 if len(pp.timers) > 0 {
4872 plocal := getg().m.p.ptr()
4873 // The world is stopped, but we acquire timersLock to
4874 // protect against sysmon calling timeSleepUntil.
4875 // This is the only case where we hold the timersLock of
4876 // more than one P, so there are no deadlock concerns.
4877 lock(&plocal.timersLock)
4878 lock(&pp.timersLock)
4879 moveTimers(plocal, pp.timers)
4881 pp.numTimers.Store(0)
4882 pp.deletedTimers.Store(0)
4883 pp.timer0When.Store(0)
4884 unlock(&pp.timersLock)
4885 unlock(&plocal.timersLock)
4887 // Flush p's write barrier buffer.
4888 if gcphase != _GCoff {
4892 for i := range pp.sudogbuf {
4893 pp.sudogbuf[i] = nil
4895 pp.sudogcache = pp.sudogbuf[:0]
4896 for j := range pp.deferpoolbuf {
4897 pp.deferpoolbuf[j] = nil
4899 pp.deferpool = pp.deferpoolbuf[:0]
4900 systemstack(func() {
4901 for i := 0; i < pp.mspancache.len; i++ {
4902 // Safe to call since the world is stopped.
4903 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4905 pp.mspancache.len = 0
4907 pp.pcache.flush(&mheap_.pages)
4908 unlock(&mheap_.lock)
4910 freemcache(pp.mcache)
4915 if pp.timerRaceCtx != 0 {
4916 // The race detector code uses a callback to fetch
4917 // the proc context, so arrange for that callback
4918 // to see the right thing.
4919 // This hack only works because we are the only
4925 racectxend(pp.timerRaceCtx)
4930 raceprocdestroy(pp.raceprocctx)
4937 // Change number of processors.
4939 // sched.lock must be held, and the world must be stopped.
4941 // gcworkbufs must not be being modified by either the GC or the write barrier
4942 // code, so the GC must not be running if the number of Ps actually changes.
4944 // Returns list of Ps with local work, they need to be scheduled by the caller.
4945 func procresize(nprocs int32) *p {
4946 assertLockHeld(&sched.lock)
4947 assertWorldStopped()
4950 if old < 0 || nprocs <= 0 {
4951 throw("procresize: invalid arg")
4954 traceGomaxprocs(nprocs)
4957 // update statistics
4959 if sched.procresizetime != 0 {
4960 sched.totaltime += int64(old) * (now - sched.procresizetime)
4962 sched.procresizetime = now
4964 maskWords := (nprocs + 31) / 32
4966 // Grow allp if necessary.
4967 if nprocs > int32(len(allp)) {
4968 // Synchronize with retake, which could be running
4969 // concurrently since it doesn't run on a P.
4971 if nprocs <= int32(cap(allp)) {
4972 allp = allp[:nprocs]
4974 nallp := make([]*p, nprocs)
4975 // Copy everything up to allp's cap so we
4976 // never lose old allocated Ps.
4977 copy(nallp, allp[:cap(allp)])
4981 if maskWords <= int32(cap(idlepMask)) {
4982 idlepMask = idlepMask[:maskWords]
4983 timerpMask = timerpMask[:maskWords]
4985 nidlepMask := make([]uint32, maskWords)
4986 // No need to copy beyond len, old Ps are irrelevant.
4987 copy(nidlepMask, idlepMask)
4988 idlepMask = nidlepMask
4990 ntimerpMask := make([]uint32, maskWords)
4991 copy(ntimerpMask, timerpMask)
4992 timerpMask = ntimerpMask
4997 // initialize new P's
4998 for i := old; i < nprocs; i++ {
5004 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
5008 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
5009 // continue to use the current P
5010 gp.m.p.ptr().status = _Prunning
5011 gp.m.p.ptr().mcache.prepareForSweep()
5013 // release the current P and acquire allp[0].
5015 // We must do this before destroying our current P
5016 // because p.destroy itself has write barriers, so we
5017 // need to do that from a valid P.
5020 // Pretend that we were descheduled
5021 // and then scheduled again to keep
5024 traceProcStop(gp.m.p.ptr())
5038 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5041 // release resources from unused P's
5042 for i := nprocs; i < old; i++ {
5045 // can't free P itself because it can be referenced by an M in syscall
5049 if int32(len(allp)) != nprocs {
5051 allp = allp[:nprocs]
5052 idlepMask = idlepMask[:maskWords]
5053 timerpMask = timerpMask[:maskWords]
5058 for i := nprocs - 1; i >= 0; i-- {
5060 if gp.m.p.ptr() == pp {
5068 pp.link.set(runnablePs)
5072 stealOrder.reset(uint32(nprocs))
5073 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5074 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5076 // Notify the limiter that the amount of procs has changed.
5077 gcCPULimiter.resetCapacity(now, nprocs)
5082 // Associate p and the current m.
5084 // This function is allowed to have write barriers even if the caller
5085 // isn't because it immediately acquires pp.
5087 //go:yeswritebarrierrec
5088 func acquirep(pp *p) {
5089 // Do the part that isn't allowed to have write barriers.
5092 // Have p; write barriers now allowed.
5094 // Perform deferred mcache flush before this P can allocate
5095 // from a potentially stale mcache.
5096 pp.mcache.prepareForSweep()
5103 // wirep is the first step of acquirep, which actually associates the
5104 // current M to pp. This is broken out so we can disallow write
5105 // barriers for this part, since we don't yet have a P.
5107 //go:nowritebarrierrec
5113 throw("wirep: already in go")
5115 if pp.m != 0 || pp.status != _Pidle {
5120 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5121 throw("wirep: invalid p state")
5125 pp.status = _Prunning
5128 // Disassociate p and the current m.
5129 func releasep() *p {
5133 throw("releasep: invalid arg")
5136 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5137 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5138 throw("releasep: invalid p state")
5141 traceProcStop(gp.m.p.ptr())
5149 func incidlelocked(v int32) {
5151 sched.nmidlelocked += v
5158 // Check for deadlock situation.
5159 // The check is based on number of running M's, if 0 -> deadlock.
5160 // sched.lock must be held.
5162 assertLockHeld(&sched.lock)
5164 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5165 // there are no running goroutines. The calling program is
5166 // assumed to be running.
5167 if islibrary || isarchive {
5171 // If we are dying because of a signal caught on an already idle thread,
5172 // freezetheworld will cause all running threads to block.
5173 // And runtime will essentially enter into deadlock state,
5174 // except that there is a thread that will call exit soon.
5175 if panicking.Load() > 0 {
5179 // If we are not running under cgo, but we have an extra M then account
5180 // for it. (It is possible to have an extra M on Windows without cgo to
5181 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5184 if !iscgo && cgoHasExtraM {
5185 mp := lockextra(true)
5186 haveExtraM := extraMCount > 0
5193 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5198 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5199 throw("checkdead: inconsistent counts")
5203 forEachG(func(gp *g) {
5204 if isSystemGoroutine(gp, false) {
5207 s := readgstatus(gp)
5208 switch s &^ _Gscan {
5215 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5216 throw("checkdead: runnable g")
5219 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5220 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5221 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5224 // Maybe jump time forward for playground.
5226 if when := timeSleepUntil(); when < maxWhen {
5229 // Start an M to steal the timer.
5230 pp, _ := pidleget(faketime)
5232 // There should always be a free P since
5233 // nothing is running.
5234 throw("checkdead: no p for timer")
5238 // There should always be a free M since
5239 // nothing is running.
5240 throw("checkdead: no m for timer")
5242 // M must be spinning to steal. We set this to be
5243 // explicit, but since this is the only M it would
5244 // become spinning on its own anyways.
5245 sched.nmspinning.Add(1)
5248 notewakeup(&mp.park)
5253 // There are no goroutines running, so we can look at the P's.
5254 for _, pp := range allp {
5255 if len(pp.timers) > 0 {
5260 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5261 fatal("all goroutines are asleep - deadlock!")
5264 // forcegcperiod is the maximum time in nanoseconds between garbage
5265 // collections. If we go this long without a garbage collection, one
5266 // is forced to run.
5268 // This is a variable for testing purposes. It normally doesn't change.
5269 var forcegcperiod int64 = 2 * 60 * 1e9
5271 // needSysmonWorkaround is true if the workaround for
5272 // golang.org/issue/42515 is needed on NetBSD.
5273 var needSysmonWorkaround bool = false
5275 // Always runs without a P, so write barriers are not allowed.
5277 //go:nowritebarrierrec
5284 lasttrace := int64(0)
5285 idle := 0 // how many cycles in succession we had not wokeup somebody
5289 if idle == 0 { // start with 20us sleep...
5291 } else if idle > 50 { // start doubling the sleep after 1ms...
5294 if delay > 10*1000 { // up to 10ms
5299 // sysmon should not enter deep sleep if schedtrace is enabled so that
5300 // it can print that information at the right time.
5302 // It should also not enter deep sleep if there are any active P's so
5303 // that it can retake P's from syscalls, preempt long running G's, and
5304 // poll the network if all P's are busy for long stretches.
5306 // It should wakeup from deep sleep if any P's become active either due
5307 // to exiting a syscall or waking up due to a timer expiring so that it
5308 // can resume performing those duties. If it wakes from a syscall it
5309 // resets idle and delay as a bet that since it had retaken a P from a
5310 // syscall before, it may need to do it again shortly after the
5311 // application starts work again. It does not reset idle when waking
5312 // from a timer to avoid adding system load to applications that spend
5313 // most of their time sleeping.
5315 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5317 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5318 syscallWake := false
5319 next := timeSleepUntil()
5321 sched.sysmonwait.Store(true)
5323 // Make wake-up period small enough
5324 // for the sampling to be correct.
5325 sleep := forcegcperiod / 2
5326 if next-now < sleep {
5329 shouldRelax := sleep >= osRelaxMinNS
5333 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5338 sched.sysmonwait.Store(false)
5339 noteclear(&sched.sysmonnote)
5349 lock(&sched.sysmonlock)
5350 // Update now in case we blocked on sysmonnote or spent a long time
5351 // blocked on schedlock or sysmonlock above.
5354 // trigger libc interceptors if needed
5355 if *cgo_yield != nil {
5356 asmcgocall(*cgo_yield, nil)
5358 // poll network if not polled for more than 10ms
5359 lastpoll := sched.lastpoll.Load()
5360 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5361 sched.lastpoll.CompareAndSwap(lastpoll, now)
5362 list := netpoll(0) // non-blocking - returns list of goroutines
5364 // Need to decrement number of idle locked M's
5365 // (pretending that one more is running) before injectglist.
5366 // Otherwise it can lead to the following situation:
5367 // injectglist grabs all P's but before it starts M's to run the P's,
5368 // another M returns from syscall, finishes running its G,
5369 // observes that there is no work to do and no other running M's
5370 // and reports deadlock.
5376 if GOOS == "netbsd" && needSysmonWorkaround {
5377 // netpoll is responsible for waiting for timer
5378 // expiration, so we typically don't have to worry
5379 // about starting an M to service timers. (Note that
5380 // sleep for timeSleepUntil above simply ensures sysmon
5381 // starts running again when that timer expiration may
5382 // cause Go code to run again).
5384 // However, netbsd has a kernel bug that sometimes
5385 // misses netpollBreak wake-ups, which can lead to
5386 // unbounded delays servicing timers. If we detect this
5387 // overrun, then startm to get something to handle the
5390 // See issue 42515 and
5391 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5392 if next := timeSleepUntil(); next < now {
5396 if scavenger.sysmonWake.Load() != 0 {
5397 // Kick the scavenger awake if someone requested it.
5400 // retake P's blocked in syscalls
5401 // and preempt long running G's
5402 if retake(now) != 0 {
5407 // check if we need to force a GC
5408 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5410 forcegc.idle.Store(false)
5412 list.push(forcegc.g)
5414 unlock(&forcegc.lock)
5416 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5418 schedtrace(debug.scheddetail > 0)
5420 unlock(&sched.sysmonlock)
5424 type sysmontick struct {
5431 // forcePreemptNS is the time slice given to a G before it is
5433 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5435 func retake(now int64) uint32 {
5437 // Prevent allp slice changes. This lock will be completely
5438 // uncontended unless we're already stopping the world.
5440 // We can't use a range loop over allp because we may
5441 // temporarily drop the allpLock. Hence, we need to re-fetch
5442 // allp each time around the loop.
5443 for i := 0; i < len(allp); i++ {
5446 // This can happen if procresize has grown
5447 // allp but not yet created new Ps.
5450 pd := &pp.sysmontick
5453 if s == _Prunning || s == _Psyscall {
5454 // Preempt G if it's running for too long.
5455 t := int64(pp.schedtick)
5456 if int64(pd.schedtick) != t {
5457 pd.schedtick = uint32(t)
5459 } else if pd.schedwhen+forcePreemptNS <= now {
5461 // In case of syscall, preemptone() doesn't
5462 // work, because there is no M wired to P.
5467 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5468 t := int64(pp.syscalltick)
5469 if !sysretake && int64(pd.syscalltick) != t {
5470 pd.syscalltick = uint32(t)
5471 pd.syscallwhen = now
5474 // On the one hand we don't want to retake Ps if there is no other work to do,
5475 // but on the other hand we want to retake them eventually
5476 // because they can prevent the sysmon thread from deep sleep.
5477 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5480 // Drop allpLock so we can take sched.lock.
5482 // Need to decrement number of idle locked M's
5483 // (pretending that one more is running) before the CAS.
5484 // Otherwise the M from which we retake can exit the syscall,
5485 // increment nmidle and report deadlock.
5487 if atomic.Cas(&pp.status, s, _Pidle) {
5504 // Tell all goroutines that they have been preempted and they should stop.
5505 // This function is purely best-effort. It can fail to inform a goroutine if a
5506 // processor just started running it.
5507 // No locks need to be held.
5508 // Returns true if preemption request was issued to at least one goroutine.
5509 func preemptall() bool {
5511 for _, pp := range allp {
5512 if pp.status != _Prunning {
5522 // Tell the goroutine running on processor P to stop.
5523 // This function is purely best-effort. It can incorrectly fail to inform the
5524 // goroutine. It can inform the wrong goroutine. Even if it informs the
5525 // correct goroutine, that goroutine might ignore the request if it is
5526 // simultaneously executing newstack.
5527 // No lock needs to be held.
5528 // Returns true if preemption request was issued.
5529 // The actual preemption will happen at some point in the future
5530 // and will be indicated by the gp->status no longer being
5532 func preemptone(pp *p) bool {
5534 if mp == nil || mp == getg().m {
5538 if gp == nil || gp == mp.g0 {
5544 // Every call in a goroutine checks for stack overflow by
5545 // comparing the current stack pointer to gp->stackguard0.
5546 // Setting gp->stackguard0 to StackPreempt folds
5547 // preemption into the normal stack overflow check.
5548 gp.stackguard0 = stackPreempt
5550 // Request an async preemption of this P.
5551 if preemptMSupported && debug.asyncpreemptoff == 0 {
5561 func schedtrace(detailed bool) {
5568 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)
5570 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5572 // We must be careful while reading data from P's, M's and G's.
5573 // Even if we hold schedlock, most data can be changed concurrently.
5574 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5575 for i, pp := range allp {
5577 h := atomic.Load(&pp.runqhead)
5578 t := atomic.Load(&pp.runqtail)
5580 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5586 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5588 // In non-detailed mode format lengths of per-P run queues as:
5589 // [len1 len2 len3 len4]
5595 if i == len(allp)-1 {
5606 for mp := allm; mp != nil; mp = mp.alllink {
5608 print(" M", mp.id, ": p=")
5620 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5621 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5629 forEachG(func(gp *g) {
5630 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5637 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5647 // schedEnableUser enables or disables the scheduling of user
5650 // This does not stop already running user goroutines, so the caller
5651 // should first stop the world when disabling user goroutines.
5652 func schedEnableUser(enable bool) {
5654 if sched.disable.user == !enable {
5658 sched.disable.user = !enable
5660 n := sched.disable.n
5662 globrunqputbatch(&sched.disable.runnable, n)
5664 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5672 // schedEnabled reports whether gp should be scheduled. It returns
5673 // false is scheduling of gp is disabled.
5675 // sched.lock must be held.
5676 func schedEnabled(gp *g) bool {
5677 assertLockHeld(&sched.lock)
5679 if sched.disable.user {
5680 return isSystemGoroutine(gp, true)
5685 // Put mp on midle list.
5686 // sched.lock must be held.
5687 // May run during STW, so write barriers are not allowed.
5689 //go:nowritebarrierrec
5691 assertLockHeld(&sched.lock)
5693 mp.schedlink = sched.midle
5699 // Try to get an m from midle list.
5700 // sched.lock must be held.
5701 // May run during STW, so write barriers are not allowed.
5703 //go:nowritebarrierrec
5705 assertLockHeld(&sched.lock)
5707 mp := sched.midle.ptr()
5709 sched.midle = mp.schedlink
5715 // Put gp on the global runnable queue.
5716 // sched.lock must be held.
5717 // May run during STW, so write barriers are not allowed.
5719 //go:nowritebarrierrec
5720 func globrunqput(gp *g) {
5721 assertLockHeld(&sched.lock)
5723 sched.runq.pushBack(gp)
5727 // Put gp at the head of the global runnable queue.
5728 // sched.lock must be held.
5729 // May run during STW, so write barriers are not allowed.
5731 //go:nowritebarrierrec
5732 func globrunqputhead(gp *g) {
5733 assertLockHeld(&sched.lock)
5739 // Put a batch of runnable goroutines on the global runnable queue.
5740 // This clears *batch.
5741 // sched.lock must be held.
5742 // May run during STW, so write barriers are not allowed.
5744 //go:nowritebarrierrec
5745 func globrunqputbatch(batch *gQueue, n int32) {
5746 assertLockHeld(&sched.lock)
5748 sched.runq.pushBackAll(*batch)
5753 // Try get a batch of G's from the global runnable queue.
5754 // sched.lock must be held.
5755 func globrunqget(pp *p, max int32) *g {
5756 assertLockHeld(&sched.lock)
5758 if sched.runqsize == 0 {
5762 n := sched.runqsize/gomaxprocs + 1
5763 if n > sched.runqsize {
5766 if max > 0 && n > max {
5769 if n > int32(len(pp.runq))/2 {
5770 n = int32(len(pp.runq)) / 2
5775 gp := sched.runq.pop()
5778 gp1 := sched.runq.pop()
5779 runqput(pp, gp1, false)
5784 // pMask is an atomic bitstring with one bit per P.
5787 // read returns true if P id's bit is set.
5788 func (p pMask) read(id uint32) bool {
5790 mask := uint32(1) << (id % 32)
5791 return (atomic.Load(&p[word]) & mask) != 0
5794 // set sets P id's bit.
5795 func (p pMask) set(id int32) {
5797 mask := uint32(1) << (id % 32)
5798 atomic.Or(&p[word], mask)
5801 // clear clears P id's bit.
5802 func (p pMask) clear(id int32) {
5804 mask := uint32(1) << (id % 32)
5805 atomic.And(&p[word], ^mask)
5808 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5810 // Ideally, the timer mask would be kept immediately consistent on any timer
5811 // operations. Unfortunately, updating a shared global data structure in the
5812 // timer hot path adds too much overhead in applications frequently switching
5813 // between no timers and some timers.
5815 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5816 // running P (returned by pidleget) may add a timer at any time, so its mask
5817 // must be set. An idle P (passed to pidleput) cannot add new timers while
5818 // idle, so if it has no timers at that time, its mask may be cleared.
5820 // Thus, we get the following effects on timer-stealing in findrunnable:
5822 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5823 // (for work- or timer-stealing; this is the ideal case).
5824 // - Running Ps must always be checked.
5825 // - Idle Ps whose timers are stolen must continue to be checked until they run
5826 // again, even after timer expiration.
5828 // When the P starts running again, the mask should be set, as a timer may be
5829 // added at any time.
5831 // TODO(prattmic): Additional targeted updates may improve the above cases.
5832 // e.g., updating the mask when stealing a timer.
5833 func updateTimerPMask(pp *p) {
5834 if pp.numTimers.Load() > 0 {
5838 // Looks like there are no timers, however another P may transiently
5839 // decrement numTimers when handling a timerModified timer in
5840 // checkTimers. We must take timersLock to serialize with these changes.
5841 lock(&pp.timersLock)
5842 if pp.numTimers.Load() == 0 {
5843 timerpMask.clear(pp.id)
5845 unlock(&pp.timersLock)
5848 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5849 // to nanotime or zero. Returns now or the current time if now was zero.
5851 // This releases ownership of p. Once sched.lock is released it is no longer
5854 // sched.lock must be held.
5856 // May run during STW, so write barriers are not allowed.
5858 //go:nowritebarrierrec
5859 func pidleput(pp *p, now int64) int64 {
5860 assertLockHeld(&sched.lock)
5863 throw("pidleput: P has non-empty run queue")
5868 updateTimerPMask(pp) // clear if there are no timers.
5869 idlepMask.set(pp.id)
5870 pp.link = sched.pidle
5873 if !pp.limiterEvent.start(limiterEventIdle, now) {
5874 throw("must be able to track idle limiter event")
5879 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5881 // sched.lock must be held.
5883 // May run during STW, so write barriers are not allowed.
5885 //go:nowritebarrierrec
5886 func pidleget(now int64) (*p, int64) {
5887 assertLockHeld(&sched.lock)
5889 pp := sched.pidle.ptr()
5891 // Timer may get added at any time now.
5895 timerpMask.set(pp.id)
5896 idlepMask.clear(pp.id)
5897 sched.pidle = pp.link
5898 sched.npidle.Add(-1)
5899 pp.limiterEvent.stop(limiterEventIdle, now)
5904 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5905 // This is called by spinning Ms (or callers than need a spinning M) that have
5906 // found work. If no P is available, this must synchronized with non-spinning
5907 // Ms that may be preparing to drop their P without discovering this work.
5909 // sched.lock must be held.
5911 // May run during STW, so write barriers are not allowed.
5913 //go:nowritebarrierrec
5914 func pidlegetSpinning(now int64) (*p, int64) {
5915 assertLockHeld(&sched.lock)
5917 pp, now := pidleget(now)
5919 // See "Delicate dance" comment in findrunnable. We found work
5920 // that we cannot take, we must synchronize with non-spinning
5921 // Ms that may be preparing to drop their P.
5922 sched.needspinning.Store(1)
5929 // runqempty reports whether pp has no Gs on its local run queue.
5930 // It never returns true spuriously.
5931 func runqempty(pp *p) bool {
5932 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5933 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5934 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5935 // does not mean the queue is empty.
5937 head := atomic.Load(&pp.runqhead)
5938 tail := atomic.Load(&pp.runqtail)
5939 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5940 if tail == atomic.Load(&pp.runqtail) {
5941 return head == tail && runnext == 0
5946 // To shake out latent assumptions about scheduling order,
5947 // we introduce some randomness into scheduling decisions
5948 // when running with the race detector.
5949 // The need for this was made obvious by changing the
5950 // (deterministic) scheduling order in Go 1.5 and breaking
5951 // many poorly-written tests.
5952 // With the randomness here, as long as the tests pass
5953 // consistently with -race, they shouldn't have latent scheduling
5955 const randomizeScheduler = raceenabled
5957 // runqput tries to put g on the local runnable queue.
5958 // If next is false, runqput adds g to the tail of the runnable queue.
5959 // If next is true, runqput puts g in the pp.runnext slot.
5960 // If the run queue is full, runnext puts g on the global queue.
5961 // Executed only by the owner P.
5962 func runqput(pp *p, gp *g, next bool) {
5963 if randomizeScheduler && next && fastrandn(2) == 0 {
5969 oldnext := pp.runnext
5970 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
5976 // Kick the old runnext out to the regular run queue.
5981 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
5983 if t-h < uint32(len(pp.runq)) {
5984 pp.runq[t%uint32(len(pp.runq))].set(gp)
5985 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
5988 if runqputslow(pp, gp, h, t) {
5991 // the queue is not full, now the put above must succeed
5995 // Put g and a batch of work from local runnable queue on global queue.
5996 // Executed only by the owner P.
5997 func runqputslow(pp *p, gp *g, h, t uint32) bool {
5998 var batch [len(pp.runq)/2 + 1]*g
6000 // First, grab a batch from local queue.
6003 if n != uint32(len(pp.runq)/2) {
6004 throw("runqputslow: queue is not full")
6006 for i := uint32(0); i < n; i++ {
6007 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6009 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6014 if randomizeScheduler {
6015 for i := uint32(1); i <= n; i++ {
6016 j := fastrandn(i + 1)
6017 batch[i], batch[j] = batch[j], batch[i]
6021 // Link the goroutines.
6022 for i := uint32(0); i < n; i++ {
6023 batch[i].schedlink.set(batch[i+1])
6026 q.head.set(batch[0])
6027 q.tail.set(batch[n])
6029 // Now put the batch on global queue.
6031 globrunqputbatch(&q, int32(n+1))
6036 // runqputbatch tries to put all the G's on q on the local runnable queue.
6037 // If the queue is full, they are put on the global queue; in that case
6038 // this will temporarily acquire the scheduler lock.
6039 // Executed only by the owner P.
6040 func runqputbatch(pp *p, q *gQueue, qsize int) {
6041 h := atomic.LoadAcq(&pp.runqhead)
6044 for !q.empty() && t-h < uint32(len(pp.runq)) {
6046 pp.runq[t%uint32(len(pp.runq))].set(gp)
6052 if randomizeScheduler {
6053 off := func(o uint32) uint32 {
6054 return (pp.runqtail + o) % uint32(len(pp.runq))
6056 for i := uint32(1); i < n; i++ {
6057 j := fastrandn(i + 1)
6058 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6062 atomic.StoreRel(&pp.runqtail, t)
6065 globrunqputbatch(q, int32(qsize))
6070 // Get g from local runnable queue.
6071 // If inheritTime is true, gp should inherit the remaining time in the
6072 // current time slice. Otherwise, it should start a new time slice.
6073 // Executed only by the owner P.
6074 func runqget(pp *p) (gp *g, inheritTime bool) {
6075 // If there's a runnext, it's the next G to run.
6077 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6078 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6079 // Hence, there's no need to retry this CAS if it fails.
6080 if next != 0 && pp.runnext.cas(next, 0) {
6081 return next.ptr(), true
6085 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6090 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6091 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6097 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6098 // Executed only by the owner P.
6099 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6100 oldNext := pp.runnext
6101 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6102 drainQ.pushBack(oldNext.ptr())
6107 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6113 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6117 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6121 // We've inverted the order in which it gets G's from the local P's runnable queue
6122 // and then advances the head pointer because we don't want to mess up the statuses of G's
6123 // while runqdrain() and runqsteal() are running in parallel.
6124 // Thus we should advance the head pointer before draining the local P into a gQueue,
6125 // so that we can update any gp.schedlink only after we take the full ownership of G,
6126 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6127 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6128 for i := uint32(0); i < qn; i++ {
6129 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6136 // Grabs a batch of goroutines from pp's runnable queue into batch.
6137 // Batch is a ring buffer starting at batchHead.
6138 // Returns number of grabbed goroutines.
6139 // Can be executed by any P.
6140 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6142 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6143 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6148 // Try to steal from pp.runnext.
6149 if next := pp.runnext; next != 0 {
6150 if pp.status == _Prunning {
6151 // Sleep to ensure that pp isn't about to run the g
6152 // we are about to steal.
6153 // The important use case here is when the g running
6154 // on pp ready()s another g and then almost
6155 // immediately blocks. Instead of stealing runnext
6156 // in this window, back off to give pp a chance to
6157 // schedule runnext. This will avoid thrashing gs
6158 // between different Ps.
6159 // A sync chan send/recv takes ~50ns as of time of
6160 // writing, so 3us gives ~50x overshoot.
6161 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6164 // On some platforms system timer granularity is
6165 // 1-15ms, which is way too much for this
6166 // optimization. So just yield.
6170 if !pp.runnext.cas(next, 0) {
6173 batch[batchHead%uint32(len(batch))] = next
6179 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6182 for i := uint32(0); i < n; i++ {
6183 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6184 batch[(batchHead+i)%uint32(len(batch))] = g
6186 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6192 // Steal half of elements from local runnable queue of p2
6193 // and put onto local runnable queue of p.
6194 // Returns one of the stolen elements (or nil if failed).
6195 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6197 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6202 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6206 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6207 if t-h+n >= uint32(len(pp.runq)) {
6208 throw("runqsteal: runq overflow")
6210 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6214 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6215 // be on one gQueue or gList at a time.
6216 type gQueue struct {
6221 // empty reports whether q is empty.
6222 func (q *gQueue) empty() bool {
6226 // push adds gp to the head of q.
6227 func (q *gQueue) push(gp *g) {
6228 gp.schedlink = q.head
6235 // pushBack adds gp to the tail of q.
6236 func (q *gQueue) pushBack(gp *g) {
6239 q.tail.ptr().schedlink.set(gp)
6246 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6248 func (q *gQueue) pushBackAll(q2 gQueue) {
6252 q2.tail.ptr().schedlink = 0
6254 q.tail.ptr().schedlink = q2.head
6261 // pop removes and returns the head of queue q. It returns nil if
6263 func (q *gQueue) pop() *g {
6266 q.head = gp.schedlink
6274 // popList takes all Gs in q and returns them as a gList.
6275 func (q *gQueue) popList() gList {
6276 stack := gList{q.head}
6281 // A gList is a list of Gs linked through g.schedlink. A G can only be
6282 // on one gQueue or gList at a time.
6287 // empty reports whether l is empty.
6288 func (l *gList) empty() bool {
6292 // push adds gp to the head of l.
6293 func (l *gList) push(gp *g) {
6294 gp.schedlink = l.head
6298 // pushAll prepends all Gs in q to l.
6299 func (l *gList) pushAll(q gQueue) {
6301 q.tail.ptr().schedlink = l.head
6306 // pop removes and returns the head of l. If l is empty, it returns nil.
6307 func (l *gList) pop() *g {
6310 l.head = gp.schedlink
6315 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6316 func setMaxThreads(in int) (out int) {
6318 out = int(sched.maxmcount)
6319 if in > 0x7fffffff { // MaxInt32
6320 sched.maxmcount = 0x7fffffff
6322 sched.maxmcount = int32(in)
6330 func procPin() int {
6335 return int(mp.p.ptr().id)
6344 //go:linkname sync_runtime_procPin sync.runtime_procPin
6346 func sync_runtime_procPin() int {
6350 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6352 func sync_runtime_procUnpin() {
6356 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6358 func sync_atomic_runtime_procPin() int {
6362 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6364 func sync_atomic_runtime_procUnpin() {
6368 // Active spinning for sync.Mutex.
6370 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6372 func sync_runtime_canSpin(i int) bool {
6373 // sync.Mutex is cooperative, so we are conservative with spinning.
6374 // Spin only few times and only if running on a multicore machine and
6375 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6376 // As opposed to runtime mutex we don't do passive spinning here,
6377 // because there can be work on global runq or on other Ps.
6378 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6381 if p := getg().m.p.ptr(); !runqempty(p) {
6387 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6389 func sync_runtime_doSpin() {
6390 procyield(active_spin_cnt)
6393 var stealOrder randomOrder
6395 // randomOrder/randomEnum are helper types for randomized work stealing.
6396 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6397 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6398 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6399 type randomOrder struct {
6404 type randomEnum struct {
6411 func (ord *randomOrder) reset(count uint32) {
6413 ord.coprimes = ord.coprimes[:0]
6414 for i := uint32(1); i <= count; i++ {
6415 if gcd(i, count) == 1 {
6416 ord.coprimes = append(ord.coprimes, i)
6421 func (ord *randomOrder) start(i uint32) randomEnum {
6425 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6429 func (enum *randomEnum) done() bool {
6430 return enum.i == enum.count
6433 func (enum *randomEnum) next() {
6435 enum.pos = (enum.pos + enum.inc) % enum.count
6438 func (enum *randomEnum) position() uint32 {
6442 func gcd(a, b uint32) uint32 {
6449 // An initTask represents the set of initializations that need to be done for a package.
6450 // Keep in sync with ../../test/initempty.go:initTask
6451 type initTask struct {
6452 // TODO: pack the first 3 fields more tightly?
6453 state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
6456 // followed by ndeps instances of an *initTask, one per package depended on
6457 // followed by nfns pcs, one per init function to run
6460 // inittrace stores statistics for init functions which are
6461 // updated by malloc and newproc when active is true.
6462 var inittrace tracestat
6464 type tracestat struct {
6465 active bool // init tracing activation status
6466 id uint64 // init goroutine id
6467 allocs uint64 // heap allocations
6468 bytes uint64 // heap allocated bytes
6471 func doInit(t *initTask) {
6473 case 2: // fully initialized
6475 case 1: // initialization in progress
6476 throw("recursive call during initialization - linker skew")
6477 default: // not initialized yet
6478 t.state = 1 // initialization in progress
6480 for i := uintptr(0); i < t.ndeps; i++ {
6481 p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
6482 t2 := *(**initTask)(p)
6487 t.state = 2 // initialization done
6496 if inittrace.active {
6498 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6502 firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
6503 for i := uintptr(0); i < t.nfns; i++ {
6504 p := add(firstFunc, i*goarch.PtrSize)
6505 f := *(*func())(unsafe.Pointer(&p))
6509 if inittrace.active {
6511 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6514 f := *(*func())(unsafe.Pointer(&firstFunc))
6515 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6518 print("init ", pkg, " @")
6519 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6520 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6521 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6522 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6526 t.state = 2 // initialization done