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() {
338 if sched.npidle.Load() > 0 {
344 // Puts the current goroutine into a waiting state and calls unlockf on the
347 // If unlockf returns false, the goroutine is resumed.
349 // unlockf must not access this G's stack, as it may be moved between
350 // the call to gopark and the call to unlockf.
352 // Note that because unlockf is called after putting the G into a waiting
353 // state, the G may have already been readied by the time unlockf is called
354 // unless there is external synchronization preventing the G from being
355 // readied. If unlockf returns false, it must guarantee that the G cannot be
356 // externally readied.
358 // Reason explains why the goroutine has been parked. It is displayed in stack
359 // traces and heap dumps. Reasons should be unique and descriptive. Do not
360 // re-use reasons, add new ones.
361 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
362 if reason != waitReasonSleep {
363 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
367 status := readgstatus(gp)
368 if status != _Grunning && status != _Gscanrunning {
369 throw("gopark: bad g status")
372 mp.waitunlockf = unlockf
373 gp.waitreason = reason
374 mp.waittraceev = traceEv
375 mp.waittraceskip = traceskip
377 // can't do anything that might move the G between Ms here.
381 // Puts the current goroutine into a waiting state and unlocks the lock.
382 // The goroutine can be made runnable again by calling goready(gp).
383 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
384 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
387 func goready(gp *g, traceskip int) {
389 ready(gp, traceskip, true)
394 func acquireSudog() *sudog {
395 // Delicate dance: the semaphore implementation calls
396 // acquireSudog, acquireSudog calls new(sudog),
397 // new calls malloc, malloc can call the garbage collector,
398 // and the garbage collector calls the semaphore implementation
400 // Break the cycle by doing acquirem/releasem around new(sudog).
401 // The acquirem/releasem increments m.locks during new(sudog),
402 // which keeps the garbage collector from being invoked.
405 if len(pp.sudogcache) == 0 {
406 lock(&sched.sudoglock)
407 // First, try to grab a batch from central cache.
408 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
409 s := sched.sudogcache
410 sched.sudogcache = s.next
412 pp.sudogcache = append(pp.sudogcache, s)
414 unlock(&sched.sudoglock)
415 // If the central cache is empty, allocate a new one.
416 if len(pp.sudogcache) == 0 {
417 pp.sudogcache = append(pp.sudogcache, new(sudog))
420 n := len(pp.sudogcache)
421 s := pp.sudogcache[n-1]
422 pp.sudogcache[n-1] = nil
423 pp.sudogcache = pp.sudogcache[:n-1]
425 throw("acquireSudog: found s.elem != nil in cache")
432 func releaseSudog(s *sudog) {
434 throw("runtime: sudog with non-nil elem")
437 throw("runtime: sudog with non-false isSelect")
440 throw("runtime: sudog with non-nil next")
443 throw("runtime: sudog with non-nil prev")
445 if s.waitlink != nil {
446 throw("runtime: sudog with non-nil waitlink")
449 throw("runtime: sudog with non-nil c")
453 throw("runtime: releaseSudog with non-nil gp.param")
455 mp := acquirem() // avoid rescheduling to another P
457 if len(pp.sudogcache) == cap(pp.sudogcache) {
458 // Transfer half of local cache to the central cache.
459 var first, last *sudog
460 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
461 n := len(pp.sudogcache)
462 p := pp.sudogcache[n-1]
463 pp.sudogcache[n-1] = nil
464 pp.sudogcache = pp.sudogcache[:n-1]
472 lock(&sched.sudoglock)
473 last.next = sched.sudogcache
474 sched.sudogcache = first
475 unlock(&sched.sudoglock)
477 pp.sudogcache = append(pp.sudogcache, s)
481 // called from assembly
482 func badmcall(fn func(*g)) {
483 throw("runtime: mcall called on m->g0 stack")
486 func badmcall2(fn func(*g)) {
487 throw("runtime: mcall function returned")
490 func badreflectcall() {
491 panic(plainError("arg size to reflect.call more than 1GB"))
494 var badmorestackg0Msg = "fatal: morestack on g0\n"
497 //go:nowritebarrierrec
498 func badmorestackg0() {
499 sp := stringStructOf(&badmorestackg0Msg)
500 write(2, sp.str, int32(sp.len))
503 var badmorestackgsignalMsg = "fatal: morestack on gsignal\n"
506 //go:nowritebarrierrec
507 func badmorestackgsignal() {
508 sp := stringStructOf(&badmorestackgsignalMsg)
509 write(2, sp.str, int32(sp.len))
517 func lockedOSThread() bool {
519 return gp.lockedm != 0 && gp.m.lockedg != 0
523 // allgs contains all Gs ever created (including dead Gs), and thus
526 // Access via the slice is protected by allglock or stop-the-world.
527 // Readers that cannot take the lock may (carefully!) use the atomic
532 // allglen and allgptr are atomic variables that contain len(allgs) and
533 // &allgs[0] respectively. Proper ordering depends on totally-ordered
534 // loads and stores. Writes are protected by allglock.
536 // allgptr is updated before allglen. Readers should read allglen
537 // before allgptr to ensure that allglen is always <= len(allgptr). New
538 // Gs appended during the race can be missed. For a consistent view of
539 // all Gs, allglock must be held.
541 // allgptr copies should always be stored as a concrete type or
542 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
543 // even if it points to a stale array.
548 func allgadd(gp *g) {
549 if readgstatus(gp) == _Gidle {
550 throw("allgadd: bad status Gidle")
554 allgs = append(allgs, gp)
555 if &allgs[0] != allgptr {
556 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
558 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
562 // allGsSnapshot returns a snapshot of the slice of all Gs.
564 // The world must be stopped or allglock must be held.
565 func allGsSnapshot() []*g {
566 assertWorldStoppedOrLockHeld(&allglock)
568 // Because the world is stopped or allglock is held, allgadd
569 // cannot happen concurrently with this. allgs grows
570 // monotonically and existing entries never change, so we can
571 // simply return a copy of the slice header. For added safety,
572 // we trim everything past len because that can still change.
573 return allgs[:len(allgs):len(allgs)]
576 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
577 func atomicAllG() (**g, uintptr) {
578 length := atomic.Loaduintptr(&allglen)
579 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
583 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
584 func atomicAllGIndex(ptr **g, i uintptr) *g {
585 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
588 // forEachG calls fn on every G from allgs.
590 // forEachG takes a lock to exclude concurrent addition of new Gs.
591 func forEachG(fn func(gp *g)) {
593 for _, gp := range allgs {
599 // forEachGRace calls fn on every G from allgs.
601 // forEachGRace avoids locking, but does not exclude addition of new Gs during
602 // execution, which may be missed.
603 func forEachGRace(fn func(gp *g)) {
604 ptr, length := atomicAllG()
605 for i := uintptr(0); i < length; i++ {
606 gp := atomicAllGIndex(ptr, i)
613 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
614 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
618 // cpuinit extracts the environment variable GODEBUG from the environment on
619 // Unix-like operating systems and calls internal/cpu.Initialize.
621 const prefix = "GODEBUG="
625 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
626 cpu.DebugOptions = true
628 // Similar to goenv_unix but extracts the environment value for
630 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
632 for argv_index(argv, argc+1+n) != nil {
636 for i := int32(0); i < n; i++ {
637 p := argv_index(argv, argc+1+i)
638 s := unsafe.String(p, findnull(p))
640 if hasPrefix(s, prefix) {
641 env = gostring(p)[len(prefix):]
649 // Support cpu feature variables are used in code generated by the compiler
650 // to guard execution of instructions that can not be assumed to be always supported.
653 x86HasPOPCNT = cpu.X86.HasPOPCNT
654 x86HasSSE41 = cpu.X86.HasSSE41
655 x86HasFMA = cpu.X86.HasFMA
658 armHasVFPv4 = cpu.ARM.HasVFPv4
661 arm64HasATOMICS = cpu.ARM64.HasATOMICS
665 // The bootstrap sequence is:
669 // make & queue new G
670 // call runtime·mstart
672 // The new G calls runtime·main.
674 lockInit(&sched.lock, lockRankSched)
675 lockInit(&sched.sysmonlock, lockRankSysmon)
676 lockInit(&sched.deferlock, lockRankDefer)
677 lockInit(&sched.sudoglock, lockRankSudog)
678 lockInit(&deadlock, lockRankDeadlock)
679 lockInit(&paniclk, lockRankPanic)
680 lockInit(&allglock, lockRankAllg)
681 lockInit(&allpLock, lockRankAllp)
682 lockInit(&reflectOffs.lock, lockRankReflectOffs)
683 lockInit(&finlock, lockRankFin)
684 lockInit(&trace.bufLock, lockRankTraceBuf)
685 lockInit(&trace.stringsLock, lockRankTraceStrings)
686 lockInit(&trace.lock, lockRankTrace)
687 lockInit(&cpuprof.lock, lockRankCpuprof)
688 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
689 // Enforce that this lock is always a leaf lock.
690 // All of this lock's critical sections should be
692 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
694 // raceinit must be the first call to race detector.
695 // In particular, it must be done before mallocinit below calls racemapshadow.
698 gp.racectx, raceprocctx0 = raceinit()
701 sched.maxmcount = 10000
703 // The world starts stopped.
709 cpuinit() // must run before alginit
710 alginit() // maps, hash, fastrand must not be used before this call
711 fastrandinit() // must run before mcommoninit
712 mcommoninit(gp.m, -1)
713 modulesinit() // provides activeModules
714 typelinksinit() // uses maps, activeModules
715 itabsinit() // uses activeModules
716 stkobjinit() // must run before GC starts
718 sigsave(&gp.m.sigmask)
719 initSigmask = gp.m.sigmask
726 // if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
727 // Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
728 // set to true by the linker, it means that nothing is consuming the profile, it is
729 // safe to set MemProfileRate to 0.
730 if disableMemoryProfiling {
735 sched.lastpoll.Store(nanotime())
737 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
740 if procresize(procs) != nil {
741 throw("unknown runnable goroutine during bootstrap")
745 // World is effectively started now, as P's can run.
748 // For cgocheck > 1, we turn on the write barrier at all times
749 // and check all pointer writes. We can't do this until after
750 // procresize because the write barrier needs a P.
751 if debug.cgocheck > 1 {
752 writeBarrier.cgo = true
753 writeBarrier.enabled = true
754 for _, pp := range allp {
759 if buildVersion == "" {
760 // Condition should never trigger. This code just serves
761 // to ensure runtime·buildVersion is kept in the resulting binary.
762 buildVersion = "unknown"
764 if len(modinfo) == 1 {
765 // Condition should never trigger. This code just serves
766 // to ensure runtime·modinfo is kept in the resulting binary.
771 func dumpgstatus(gp *g) {
773 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
774 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
777 // sched.lock must be held.
779 assertLockHeld(&sched.lock)
781 if mcount() > sched.maxmcount {
782 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
783 throw("thread exhaustion")
787 // mReserveID returns the next ID to use for a new m. This new m is immediately
788 // considered 'running' by checkdead.
790 // sched.lock must be held.
791 func mReserveID() int64 {
792 assertLockHeld(&sched.lock)
794 if sched.mnext+1 < sched.mnext {
795 throw("runtime: thread ID overflow")
803 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
804 func mcommoninit(mp *m, id int64) {
807 // g0 stack won't make sense for user (and is not necessary unwindable).
809 callers(1, mp.createstack[:])
820 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
821 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
825 // Same behavior as for 1.17.
826 // TODO: Simplify ths.
827 if goarch.BigEndian {
828 mp.fastrand = uint64(lo)<<32 | uint64(hi)
830 mp.fastrand = uint64(hi)<<32 | uint64(lo)
834 if mp.gsignal != nil {
835 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
838 // Add to allm so garbage collector doesn't free g->m
839 // when it is just in a register or thread-local storage.
842 // NumCgoCall() iterates over allm w/o schedlock,
843 // so we need to publish it safely.
844 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
847 // Allocate memory to hold a cgo traceback if the cgo call crashes.
848 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
849 mp.cgoCallers = new(cgoCallers)
853 func (mp *m) becomeSpinning() {
855 sched.nmspinning.Add(1)
856 sched.needspinning.Store(0)
859 var fastrandseed uintptr
861 func fastrandinit() {
862 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
866 // Mark gp ready to run.
867 func ready(gp *g, traceskip int, next bool) {
869 traceGoUnpark(gp, traceskip)
872 status := readgstatus(gp)
875 mp := acquirem() // disable preemption because it can be holding p in a local var
876 if status&^_Gscan != _Gwaiting {
878 throw("bad g->status in ready")
881 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
882 casgstatus(gp, _Gwaiting, _Grunnable)
883 runqput(mp.p.ptr(), gp, next)
888 // freezeStopWait is a large value that freezetheworld sets
889 // sched.stopwait to in order to request that all Gs permanently stop.
890 const freezeStopWait = 0x7fffffff
892 // freezing is set to non-zero if the runtime is trying to freeze the
894 var freezing atomic.Bool
896 // Similar to stopTheWorld but best-effort and can be called several times.
897 // There is no reverse operation, used during crashing.
898 // This function must not lock any mutexes.
899 func freezetheworld() {
901 // stopwait and preemption requests can be lost
902 // due to races with concurrently executing threads,
903 // so try several times
904 for i := 0; i < 5; i++ {
905 // this should tell the scheduler to not start any new goroutines
906 sched.stopwait = freezeStopWait
907 sched.gcwaiting.Store(true)
908 // this should stop running goroutines
910 break // no running goroutines
920 // All reads and writes of g's status go through readgstatus, casgstatus
921 // castogscanstatus, casfrom_Gscanstatus.
924 func readgstatus(gp *g) uint32 {
925 return gp.atomicstatus.Load()
928 // The Gscanstatuses are acting like locks and this releases them.
929 // If it proves to be a performance hit we should be able to make these
930 // simple atomic stores but for now we are going to throw if
931 // we see an inconsistent state.
932 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
935 // Check that transition is valid.
938 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
940 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
946 if newval == oldval&^_Gscan {
947 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
951 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
953 throw("casfrom_Gscanstatus: gp->status is not in scan state")
955 releaseLockRank(lockRankGscan)
958 // This will return false if the gp is not in the expected status and the cas fails.
959 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
960 func castogscanstatus(gp *g, oldval, newval uint32) bool {
966 if newval == oldval|_Gscan {
967 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
969 acquireLockRank(lockRankGscan)
975 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
976 throw("castogscanstatus")
980 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
981 // various latencies on every transition instead of sampling them.
982 var casgstatusAlwaysTrack = false
984 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
985 // and casfrom_Gscanstatus instead.
986 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
987 // put it in the Gscan state is finished.
990 func casgstatus(gp *g, oldval, newval uint32) {
991 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
993 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
994 throw("casgstatus: bad incoming values")
998 acquireLockRank(lockRankGscan)
999 releaseLockRank(lockRankGscan)
1001 // See https://golang.org/cl/21503 for justification of the yield delay.
1002 const yieldDelay = 5 * 1000
1005 // loop if gp->atomicstatus is in a scan state giving
1006 // GC time to finish and change the state to oldval.
1007 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
1008 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
1009 throw("casgstatus: waiting for Gwaiting but is Grunnable")
1012 nextYield = nanotime() + yieldDelay
1014 if nanotime() < nextYield {
1015 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
1020 nextYield = nanotime() + yieldDelay/2
1024 if oldval == _Grunning {
1025 // Track every gTrackingPeriod time a goroutine transitions out of running.
1026 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1035 // Handle various kinds of tracking.
1038 // - Time spent in runnable.
1039 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1042 // We transitioned out of runnable, so measure how much
1043 // time we spent in this state and add it to
1046 gp.runnableTime += now - gp.trackingStamp
1047 gp.trackingStamp = 0
1049 if !gp.waitreason.isMutexWait() {
1050 // Not blocking on a lock.
1053 // Blocking on a lock, measure it. Note that because we're
1054 // sampling, we have to multiply by our sampling period to get
1055 // a more representative estimate of the absolute value.
1056 // gTrackingPeriod also represents an accurate sampling period
1057 // because we can only enter this state from _Grunning.
1059 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1060 gp.trackingStamp = 0
1064 if !gp.waitreason.isMutexWait() {
1065 // Not blocking on a lock.
1068 // Blocking on a lock. Write down the timestamp.
1070 gp.trackingStamp = now
1072 // We just transitioned into runnable, so record what
1073 // time that happened.
1075 gp.trackingStamp = now
1077 // We're transitioning into running, so turn off
1078 // tracking and record how much time we spent in
1081 sched.timeToRun.record(gp.runnableTime)
1086 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1088 // Use this over casgstatus when possible to ensure that a waitreason is set.
1089 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1090 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1091 gp.waitreason = reason
1092 casgstatus(gp, old, _Gwaiting)
1095 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1096 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1097 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1098 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1099 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1102 func casgcopystack(gp *g) uint32 {
1104 oldstatus := readgstatus(gp) &^ _Gscan
1105 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1106 throw("copystack: bad status, not Gwaiting or Grunnable")
1108 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1114 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1116 // TODO(austin): This is the only status operation that both changes
1117 // the status and locks the _Gscan bit. Rethink this.
1118 func casGToPreemptScan(gp *g, old, new uint32) {
1119 if old != _Grunning || new != _Gscan|_Gpreempted {
1120 throw("bad g transition")
1122 acquireLockRank(lockRankGscan)
1123 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1127 // casGFromPreempted attempts to transition gp from _Gpreempted to
1128 // _Gwaiting. If successful, the caller is responsible for
1129 // re-scheduling gp.
1130 func casGFromPreempted(gp *g, old, new uint32) bool {
1131 if old != _Gpreempted || new != _Gwaiting {
1132 throw("bad g transition")
1134 gp.waitreason = waitReasonPreempted
1135 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1138 // stopTheWorld stops all P's from executing goroutines, interrupting
1139 // all goroutines at GC safe points and records reason as the reason
1140 // for the stop. On return, only the current goroutine's P is running.
1141 // stopTheWorld must not be called from a system stack and the caller
1142 // must not hold worldsema. The caller must call startTheWorld when
1143 // other P's should resume execution.
1145 // stopTheWorld is safe for multiple goroutines to call at the
1146 // same time. Each will execute its own stop, and the stops will
1149 // This is also used by routines that do stack dumps. If the system is
1150 // in panic or being exited, this may not reliably stop all
1152 func stopTheWorld(reason string) {
1153 semacquire(&worldsema)
1155 gp.m.preemptoff = reason
1156 systemstack(func() {
1157 // Mark the goroutine which called stopTheWorld preemptible so its
1158 // stack may be scanned.
1159 // This lets a mark worker scan us while we try to stop the world
1160 // since otherwise we could get in a mutual preemption deadlock.
1161 // We must not modify anything on the G stack because a stack shrink
1162 // may occur. A stack shrink is otherwise OK though because in order
1163 // to return from this function (and to leave the system stack) we
1164 // must have preempted all goroutines, including any attempting
1165 // to scan our stack, in which case, any stack shrinking will
1166 // have already completed by the time we exit.
1167 // Don't provide a wait reason because we're still executing.
1168 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1169 stopTheWorldWithSema()
1170 casgstatus(gp, _Gwaiting, _Grunning)
1174 // startTheWorld undoes the effects of stopTheWorld.
1175 func startTheWorld() {
1176 systemstack(func() { startTheWorldWithSema(false) })
1178 // worldsema must be held over startTheWorldWithSema to ensure
1179 // gomaxprocs cannot change while worldsema is held.
1181 // Release worldsema with direct handoff to the next waiter, but
1182 // acquirem so that semrelease1 doesn't try to yield our time.
1184 // Otherwise if e.g. ReadMemStats is being called in a loop,
1185 // it might stomp on other attempts to stop the world, such as
1186 // for starting or ending GC. The operation this blocks is
1187 // so heavy-weight that we should just try to be as fair as
1190 // We don't want to just allow us to get preempted between now
1191 // and releasing the semaphore because then we keep everyone
1192 // (including, for example, GCs) waiting longer.
1195 semrelease1(&worldsema, true, 0)
1199 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1200 // until the GC is not running. It also blocks a GC from starting
1201 // until startTheWorldGC is called.
1202 func stopTheWorldGC(reason string) {
1204 stopTheWorld(reason)
1207 // startTheWorldGC undoes the effects of stopTheWorldGC.
1208 func startTheWorldGC() {
1213 // Holding worldsema grants an M the right to try to stop the world.
1214 var worldsema uint32 = 1
1216 // Holding gcsema grants the M the right to block a GC, and blocks
1217 // until the current GC is done. In particular, it prevents gomaxprocs
1218 // from changing concurrently.
1220 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1221 // being changed/enabled during a GC, remove this.
1222 var gcsema uint32 = 1
1224 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1225 // The caller is responsible for acquiring worldsema and disabling
1226 // preemption first and then should stopTheWorldWithSema on the system
1229 // semacquire(&worldsema, 0)
1230 // m.preemptoff = "reason"
1231 // systemstack(stopTheWorldWithSema)
1233 // When finished, the caller must either call startTheWorld or undo
1234 // these three operations separately:
1236 // m.preemptoff = ""
1237 // systemstack(startTheWorldWithSema)
1238 // semrelease(&worldsema)
1240 // It is allowed to acquire worldsema once and then execute multiple
1241 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1242 // Other P's are able to execute between successive calls to
1243 // startTheWorldWithSema and stopTheWorldWithSema.
1244 // Holding worldsema causes any other goroutines invoking
1245 // stopTheWorld to block.
1246 func stopTheWorldWithSema() {
1249 // If we hold a lock, then we won't be able to stop another M
1250 // that is blocked trying to acquire the lock.
1252 throw("stopTheWorld: holding locks")
1256 sched.stopwait = gomaxprocs
1257 sched.gcwaiting.Store(true)
1260 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1262 // try to retake all P's in Psyscall status
1263 for _, pp := range allp {
1265 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1277 pp, _ := pidleget(now)
1281 pp.status = _Pgcstop
1284 wait := sched.stopwait > 0
1287 // wait for remaining P's to stop voluntarily
1290 // wait for 100us, then try to re-preempt in case of any races
1291 if notetsleep(&sched.stopnote, 100*1000) {
1292 noteclear(&sched.stopnote)
1301 if sched.stopwait != 0 {
1302 bad = "stopTheWorld: not stopped (stopwait != 0)"
1304 for _, pp := range allp {
1305 if pp.status != _Pgcstop {
1306 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1310 if freezing.Load() {
1311 // Some other thread is panicking. This can cause the
1312 // sanity checks above to fail if the panic happens in
1313 // the signal handler on a stopped thread. Either way,
1314 // we should halt this thread.
1325 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1326 assertWorldStopped()
1328 mp := acquirem() // disable preemption because it can be holding p in a local var
1329 if netpollinited() {
1330 list := netpoll(0) // non-blocking
1340 p1 := procresize(procs)
1341 sched.gcwaiting.Store(false)
1342 if sched.sysmonwait.Load() {
1343 sched.sysmonwait.Store(false)
1344 notewakeup(&sched.sysmonnote)
1357 throw("startTheWorld: inconsistent mp->nextp")
1360 notewakeup(&mp.park)
1362 // Start M to run P. Do not start another M below.
1367 // Capture start-the-world time before doing clean-up tasks.
1368 startTime := nanotime()
1373 // Wakeup an additional proc in case we have excessive runnable goroutines
1374 // in local queues or in the global queue. If we don't, the proc will park itself.
1375 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1383 // usesLibcall indicates whether this runtime performs system calls
1385 func usesLibcall() bool {
1387 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1390 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1395 // mStackIsSystemAllocated indicates whether this runtime starts on a
1396 // system-allocated stack.
1397 func mStackIsSystemAllocated() bool {
1399 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1403 case "386", "amd64", "arm", "arm64":
1410 // mstart is the entry-point for new Ms.
1411 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1414 // mstart0 is the Go entry-point for new Ms.
1415 // This must not split the stack because we may not even have stack
1416 // bounds set up yet.
1418 // May run during STW (because it doesn't have a P yet), so write
1419 // barriers are not allowed.
1422 //go:nowritebarrierrec
1426 osStack := gp.stack.lo == 0
1428 // Initialize stack bounds from system stack.
1429 // Cgo may have left stack size in stack.hi.
1430 // minit may update the stack bounds.
1432 // Note: these bounds may not be very accurate.
1433 // We set hi to &size, but there are things above
1434 // it. The 1024 is supposed to compensate this,
1435 // but is somewhat arbitrary.
1438 size = 8192 * sys.StackGuardMultiplier
1440 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1441 gp.stack.lo = gp.stack.hi - size + 1024
1443 // Initialize stack guard so that we can start calling regular
1445 gp.stackguard0 = gp.stack.lo + _StackGuard
1446 // This is the g0, so we can also call go:systemstack
1447 // functions, which check stackguard1.
1448 gp.stackguard1 = gp.stackguard0
1451 // Exit this thread.
1452 if mStackIsSystemAllocated() {
1453 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1454 // the stack, but put it in gp.stack before mstart,
1455 // so the logic above hasn't set osStack yet.
1461 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1462 // so that we can set up g0.sched to return to the call of mstart1 above.
1469 throw("bad runtime·mstart")
1472 // Set up m.g0.sched as a label returning to just
1473 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1474 // We're never coming back to mstart1 after we call schedule,
1475 // so other calls can reuse the current frame.
1476 // And goexit0 does a gogo that needs to return from mstart1
1477 // and let mstart0 exit the thread.
1478 gp.sched.g = guintptr(unsafe.Pointer(gp))
1479 gp.sched.pc = getcallerpc()
1480 gp.sched.sp = getcallersp()
1485 // Install signal handlers; after minit so that minit can
1486 // prepare the thread to be able to handle the signals.
1491 if fn := gp.m.mstartfn; fn != nil {
1496 acquirep(gp.m.nextp.ptr())
1502 // mstartm0 implements part of mstart1 that only runs on the m0.
1504 // Write barriers are allowed here because we know the GC can't be
1505 // running yet, so they'll be no-ops.
1507 //go:yeswritebarrierrec
1509 // Create an extra M for callbacks on threads not created by Go.
1510 // An extra M is also needed on Windows for callbacks created by
1511 // syscall.NewCallback. See issue #6751 for details.
1512 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1519 // mPark causes a thread to park itself, returning once woken.
1524 notesleep(&gp.m.park)
1525 noteclear(&gp.m.park)
1528 // mexit tears down and exits the current thread.
1530 // Don't call this directly to exit the thread, since it must run at
1531 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1532 // unwind the stack to the point that exits the thread.
1534 // It is entered with m.p != nil, so write barriers are allowed. It
1535 // will release the P before exiting.
1537 //go:yeswritebarrierrec
1538 func mexit(osStack bool) {
1542 // This is the main thread. Just wedge it.
1544 // On Linux, exiting the main thread puts the process
1545 // into a non-waitable zombie state. On Plan 9,
1546 // exiting the main thread unblocks wait even though
1547 // other threads are still running. On Solaris we can
1548 // neither exitThread nor return from mstart. Other
1549 // bad things probably happen on other platforms.
1551 // We could try to clean up this M more before wedging
1552 // it, but that complicates signal handling.
1553 handoffp(releasep())
1559 throw("locked m0 woke up")
1565 // Free the gsignal stack.
1566 if mp.gsignal != nil {
1567 stackfree(mp.gsignal.stack)
1568 // On some platforms, when calling into VDSO (e.g. nanotime)
1569 // we store our g on the gsignal stack, if there is one.
1570 // Now the stack is freed, unlink it from the m, so we
1571 // won't write to it when calling VDSO code.
1575 // Remove m from allm.
1577 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1583 throw("m not found in allm")
1586 // Delay reaping m until it's done with the stack.
1588 // If this is using an OS stack, the OS will free it
1589 // so there's no need for reaping.
1590 atomic.Store(&mp.freeWait, 1)
1591 // Put m on the free list, though it will not be reaped until
1592 // freeWait is 0. Note that the free list must not be linked
1593 // through alllink because some functions walk allm without
1594 // locking, so may be using alllink.
1595 mp.freelink = sched.freem
1600 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1603 handoffp(releasep())
1604 // After this point we must not have write barriers.
1606 // Invoke the deadlock detector. This must happen after
1607 // handoffp because it may have started a new M to take our
1614 if GOOS == "darwin" || GOOS == "ios" {
1615 // Make sure pendingPreemptSignals is correct when an M exits.
1617 if mp.signalPending.Load() != 0 {
1618 pendingPreemptSignals.Add(-1)
1622 // Destroy all allocated resources. After this is called, we may no
1623 // longer take any locks.
1627 // Return from mstart and let the system thread
1628 // library free the g0 stack and terminate the thread.
1632 // mstart is the thread's entry point, so there's nothing to
1633 // return to. Exit the thread directly. exitThread will clear
1634 // m.freeWait when it's done with the stack and the m can be
1636 exitThread(&mp.freeWait)
1639 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1640 // If a P is currently executing code, this will bring the P to a GC
1641 // safe point and execute fn on that P. If the P is not executing code
1642 // (it is idle or in a syscall), this will call fn(p) directly while
1643 // preventing the P from exiting its state. This does not ensure that
1644 // fn will run on every CPU executing Go code, but it acts as a global
1645 // memory barrier. GC uses this as a "ragged barrier."
1647 // The caller must hold worldsema.
1650 func forEachP(fn func(*p)) {
1652 pp := getg().m.p.ptr()
1655 if sched.safePointWait != 0 {
1656 throw("forEachP: sched.safePointWait != 0")
1658 sched.safePointWait = gomaxprocs - 1
1659 sched.safePointFn = fn
1661 // Ask all Ps to run the safe point function.
1662 for _, p2 := range allp {
1664 atomic.Store(&p2.runSafePointFn, 1)
1669 // Any P entering _Pidle or _Psyscall from now on will observe
1670 // p.runSafePointFn == 1 and will call runSafePointFn when
1671 // changing its status to _Pidle/_Psyscall.
1673 // Run safe point function for all idle Ps. sched.pidle will
1674 // not change because we hold sched.lock.
1675 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1676 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1678 sched.safePointWait--
1682 wait := sched.safePointWait > 0
1685 // Run fn for the current P.
1688 // Force Ps currently in _Psyscall into _Pidle and hand them
1689 // off to induce safe point function execution.
1690 for _, p2 := range allp {
1692 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1702 // Wait for remaining Ps to run fn.
1705 // Wait for 100us, then try to re-preempt in
1706 // case of any races.
1708 // Requires system stack.
1709 if notetsleep(&sched.safePointNote, 100*1000) {
1710 noteclear(&sched.safePointNote)
1716 if sched.safePointWait != 0 {
1717 throw("forEachP: not done")
1719 for _, p2 := range allp {
1720 if p2.runSafePointFn != 0 {
1721 throw("forEachP: P did not run fn")
1726 sched.safePointFn = nil
1731 // runSafePointFn runs the safe point function, if any, for this P.
1732 // This should be called like
1734 // if getg().m.p.runSafePointFn != 0 {
1738 // runSafePointFn must be checked on any transition in to _Pidle or
1739 // _Psyscall to avoid a race where forEachP sees that the P is running
1740 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1741 // nor the P run the safe-point function.
1742 func runSafePointFn() {
1743 p := getg().m.p.ptr()
1744 // Resolve the race between forEachP running the safe-point
1745 // function on this P's behalf and this P running the
1746 // safe-point function directly.
1747 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1750 sched.safePointFn(p)
1752 sched.safePointWait--
1753 if sched.safePointWait == 0 {
1754 notewakeup(&sched.safePointNote)
1759 // When running with cgo, we call _cgo_thread_start
1760 // to start threads for us so that we can play nicely with
1762 var cgoThreadStart unsafe.Pointer
1764 type cgothreadstart struct {
1770 // Allocate a new m unassociated with any thread.
1771 // Can use p for allocation context if needed.
1772 // fn is recorded as the new m's m.mstartfn.
1773 // id is optional pre-allocated m ID. Omit by passing -1.
1775 // This function is allowed to have write barriers even if the caller
1776 // isn't because it borrows pp.
1778 //go:yeswritebarrierrec
1779 func allocm(pp *p, fn func(), id int64) *m {
1782 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1783 // disable preemption to ensure it is not stolen, which would make the
1784 // caller lose ownership.
1789 acquirep(pp) // temporarily borrow p for mallocs in this function
1792 // Release the free M list. We need to do this somewhere and
1793 // this may free up a stack we can use.
1794 if sched.freem != nil {
1797 for freem := sched.freem; freem != nil; {
1798 if freem.freeWait != 0 {
1799 next := freem.freelink
1800 freem.freelink = newList
1805 // stackfree must be on the system stack, but allocm is
1806 // reachable off the system stack transitively from
1808 systemstack(func() {
1809 stackfree(freem.g0.stack)
1811 freem = freem.freelink
1813 sched.freem = newList
1821 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1822 // Windows and Plan 9 will layout sched stack on OS stack.
1823 if iscgo || mStackIsSystemAllocated() {
1826 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1830 if pp == gp.m.p.ptr() {
1835 allocmLock.runlock()
1839 // needm is called when a cgo callback happens on a
1840 // thread without an m (a thread not created by Go).
1841 // In this case, needm is expected to find an m to use
1842 // and return with m, g initialized correctly.
1843 // Since m and g are not set now (likely nil, but see below)
1844 // needm is limited in what routines it can call. In particular
1845 // it can only call nosplit functions (textflag 7) and cannot
1846 // do any scheduling that requires an m.
1848 // In order to avoid needing heavy lifting here, we adopt
1849 // the following strategy: there is a stack of available m's
1850 // that can be stolen. Using compare-and-swap
1851 // to pop from the stack has ABA races, so we simulate
1852 // a lock by doing an exchange (via Casuintptr) to steal the stack
1853 // head and replace the top pointer with MLOCKED (1).
1854 // This serves as a simple spin lock that we can use even
1855 // without an m. The thread that locks the stack in this way
1856 // unlocks the stack by storing a valid stack head pointer.
1858 // In order to make sure that there is always an m structure
1859 // available to be stolen, we maintain the invariant that there
1860 // is always one more than needed. At the beginning of the
1861 // program (if cgo is in use) the list is seeded with a single m.
1862 // If needm finds that it has taken the last m off the list, its job
1863 // is - once it has installed its own m so that it can do things like
1864 // allocate memory - to create a spare m and put it on the list.
1866 // Each of these extra m's also has a g0 and a curg that are
1867 // pressed into service as the scheduling stack and current
1868 // goroutine for the duration of the cgo callback.
1870 // When the callback is done with the m, it calls dropm to
1871 // put the m back on the list.
1875 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1876 // Can happen if C/C++ code calls Go from a global ctor.
1877 // Can also happen on Windows if a global ctor uses a
1878 // callback created by syscall.NewCallback. See issue #6751
1881 // Can not throw, because scheduler is not initialized yet.
1882 write(2, unsafe.Pointer(&earlycgocallback[0]), int32(len(earlycgocallback)))
1886 // Save and block signals before getting an M.
1887 // The signal handler may call needm itself,
1888 // and we must avoid a deadlock. Also, once g is installed,
1889 // any incoming signals will try to execute,
1890 // but we won't have the sigaltstack settings and other data
1891 // set up appropriately until the end of minit, which will
1892 // unblock the signals. This is the same dance as when
1893 // starting a new m to run Go code via newosproc.
1898 // Lock extra list, take head, unlock popped list.
1899 // nilokay=false is safe here because of the invariant above,
1900 // that the extra list always contains or will soon contain
1902 mp := lockextra(false)
1904 // Set needextram when we've just emptied the list,
1905 // so that the eventual call into cgocallbackg will
1906 // allocate a new m for the extra list. We delay the
1907 // allocation until then so that it can be done
1908 // after exitsyscall makes sure it is okay to be
1909 // running at all (that is, there's no garbage collection
1910 // running right now).
1911 mp.needextram = mp.schedlink == 0
1913 unlockextra(mp.schedlink.ptr())
1915 // Store the original signal mask for use by minit.
1916 mp.sigmask = sigmask
1918 // Install TLS on some platforms (previously setg
1919 // would do this if necessary).
1922 // Install g (= m->g0) and set the stack bounds
1923 // to match the current stack. We don't actually know
1924 // how big the stack is, like we don't know how big any
1925 // scheduling stack is, but we assume there's at least 32 kB,
1926 // which is more than enough for us.
1929 gp.stack.hi = getcallersp() + 1024
1930 gp.stack.lo = getcallersp() - 32*1024
1931 gp.stackguard0 = gp.stack.lo + _StackGuard
1933 // Initialize this thread to use the m.
1937 // mp.curg is now a real goroutine.
1938 casgstatus(mp.curg, _Gdead, _Gsyscall)
1942 var earlycgocallback = []byte("fatal error: cgo callback before cgo call\n")
1944 // newextram allocates m's and puts them on the extra list.
1945 // It is called with a working local m, so that it can do things
1946 // like call schedlock and allocate.
1948 c := extraMWaiters.Swap(0)
1950 for i := uint32(0); i < c; i++ {
1954 // Make sure there is at least one extra M.
1955 mp := lockextra(true)
1963 // oneNewExtraM allocates an m and puts it on the extra list.
1964 func oneNewExtraM() {
1965 // Create extra goroutine locked to extra m.
1966 // The goroutine is the context in which the cgo callback will run.
1967 // The sched.pc will never be returned to, but setting it to
1968 // goexit makes clear to the traceback routines where
1969 // the goroutine stack ends.
1970 mp := allocm(nil, nil, -1)
1972 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
1973 gp.sched.sp = gp.stack.hi
1974 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
1976 gp.sched.g = guintptr(unsafe.Pointer(gp))
1977 gp.syscallpc = gp.sched.pc
1978 gp.syscallsp = gp.sched.sp
1979 gp.stktopsp = gp.sched.sp
1980 // malg returns status as _Gidle. Change to _Gdead before
1981 // adding to allg where GC can see it. We use _Gdead to hide
1982 // this from tracebacks and stack scans since it isn't a
1983 // "real" goroutine until needm grabs it.
1984 casgstatus(gp, _Gidle, _Gdead)
1991 gp.goid = sched.goidgen.Add(1)
1992 gp.sysblocktraced = true
1994 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
1997 // Trigger two trace events for the locked g in the extra m,
1998 // since the next event of the g will be traceEvGoSysExit in exitsyscall,
1999 // while calling from C thread to Go.
2000 traceGoCreate(gp, 0) // no start pc
2002 traceEvent(traceEvGoInSyscall, -1, gp.goid)
2004 // put on allg for garbage collector
2007 // gp is now on the allg list, but we don't want it to be
2008 // counted by gcount. It would be more "proper" to increment
2009 // sched.ngfree, but that requires locking. Incrementing ngsys
2010 // has the same effect.
2013 // Add m to the extra list.
2014 mnext := lockextra(true)
2015 mp.schedlink.set(mnext)
2020 // dropm is called when a cgo callback has called needm but is now
2021 // done with the callback and returning back into the non-Go thread.
2022 // It puts the current m back onto the extra list.
2024 // The main expense here is the call to signalstack to release the
2025 // m's signal stack, and then the call to needm on the next callback
2026 // from this thread. It is tempting to try to save the m for next time,
2027 // which would eliminate both these costs, but there might not be
2028 // a next time: the current thread (which Go does not control) might exit.
2029 // If we saved the m for that thread, there would be an m leak each time
2030 // such a thread exited. Instead, we acquire and release an m on each
2031 // call. These should typically not be scheduling operations, just a few
2032 // atomics, so the cost should be small.
2034 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
2035 // variable using pthread_key_create. Unlike the pthread keys we already use
2036 // on OS X, this dummy key would never be read by Go code. It would exist
2037 // only so that we could register at thread-exit-time destructor.
2038 // That destructor would put the m back onto the extra list.
2039 // This is purely a performance optimization. The current version,
2040 // in which dropm happens on each cgo call, is still correct too.
2041 // We may have to keep the current version on systems with cgo
2042 // but without pthreads, like Windows.
2044 // Clear m and g, and return m to the extra list.
2045 // After the call to setg we can only call nosplit functions
2046 // with no pointer manipulation.
2049 // Return mp.curg to dead state.
2050 casgstatus(mp.curg, _Gsyscall, _Gdead)
2051 mp.curg.preemptStop = false
2054 // Block signals before unminit.
2055 // Unminit unregisters the signal handling stack (but needs g on some systems).
2056 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2057 // It's important not to try to handle a signal between those two steps.
2058 sigmask := mp.sigmask
2062 mnext := lockextra(true)
2064 mp.schedlink.set(mnext)
2068 // Commit the release of mp.
2071 msigrestore(sigmask)
2074 // A helper function for EnsureDropM.
2075 func getm() uintptr {
2076 return uintptr(unsafe.Pointer(getg().m))
2079 var extram atomic.Uintptr
2080 var extraMCount uint32 // Protected by lockextra
2081 var extraMWaiters atomic.Uint32
2083 // lockextra locks the extra list and returns the list head.
2084 // The caller must unlock the list by storing a new list head
2085 // to extram. If nilokay is true, then lockextra will
2086 // return a nil list head if that's what it finds. If nilokay is false,
2087 // lockextra will keep waiting until the list head is no longer nil.
2090 func lockextra(nilokay bool) *m {
2095 old := extram.Load()
2100 if old == 0 && !nilokay {
2102 // Add 1 to the number of threads
2103 // waiting for an M.
2104 // This is cleared by newextram.
2105 extraMWaiters.Add(1)
2111 if extram.CompareAndSwap(old, locked) {
2112 return (*m)(unsafe.Pointer(old))
2120 func unlockextra(mp *m) {
2121 extram.Store(uintptr(unsafe.Pointer(mp)))
2125 // allocmLock is locked for read when creating new Ms in allocm and their
2126 // addition to allm. Thus acquiring this lock for write blocks the
2127 // creation of new Ms.
2130 // execLock serializes exec and clone to avoid bugs or unspecified
2131 // behaviour around exec'ing while creating/destroying threads. See
2136 // newmHandoff contains a list of m structures that need new OS threads.
2137 // This is used by newm in situations where newm itself can't safely
2138 // start an OS thread.
2139 var newmHandoff struct {
2142 // newm points to a list of M structures that need new OS
2143 // threads. The list is linked through m.schedlink.
2146 // waiting indicates that wake needs to be notified when an m
2147 // is put on the list.
2151 // haveTemplateThread indicates that the templateThread has
2152 // been started. This is not protected by lock. Use cas to set
2154 haveTemplateThread uint32
2157 // Create a new m. It will start off with a call to fn, or else the scheduler.
2158 // fn needs to be static and not a heap allocated closure.
2159 // May run with m.p==nil, so write barriers are not allowed.
2161 // id is optional pre-allocated m ID. Omit by passing -1.
2163 //go:nowritebarrierrec
2164 func newm(fn func(), pp *p, id int64) {
2165 // allocm adds a new M to allm, but they do not start until created by
2166 // the OS in newm1 or the template thread.
2168 // doAllThreadsSyscall requires that every M in allm will eventually
2169 // start and be signal-able, even with a STW.
2171 // Disable preemption here until we start the thread to ensure that
2172 // newm is not preempted between allocm and starting the new thread,
2173 // ensuring that anything added to allm is guaranteed to eventually
2177 mp := allocm(pp, fn, id)
2179 mp.sigmask = initSigmask
2180 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2181 // We're on a locked M or a thread that may have been
2182 // started by C. The kernel state of this thread may
2183 // be strange (the user may have locked it for that
2184 // purpose). We don't want to clone that into another
2185 // thread. Instead, ask a known-good thread to create
2186 // the thread for us.
2188 // This is disabled on Plan 9. See golang.org/issue/22227.
2190 // TODO: This may be unnecessary on Windows, which
2191 // doesn't model thread creation off fork.
2192 lock(&newmHandoff.lock)
2193 if newmHandoff.haveTemplateThread == 0 {
2194 throw("on a locked thread with no template thread")
2196 mp.schedlink = newmHandoff.newm
2197 newmHandoff.newm.set(mp)
2198 if newmHandoff.waiting {
2199 newmHandoff.waiting = false
2200 notewakeup(&newmHandoff.wake)
2202 unlock(&newmHandoff.lock)
2203 // The M has not started yet, but the template thread does not
2204 // participate in STW, so it will always process queued Ms and
2205 // it is safe to releasem.
2215 var ts cgothreadstart
2216 if _cgo_thread_start == nil {
2217 throw("_cgo_thread_start missing")
2220 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2221 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2223 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2226 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2228 execLock.rlock() // Prevent process clone.
2229 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2233 execLock.rlock() // Prevent process clone.
2238 // startTemplateThread starts the template thread if it is not already
2241 // The calling thread must itself be in a known-good state.
2242 func startTemplateThread() {
2243 if GOARCH == "wasm" { // no threads on wasm yet
2247 // Disable preemption to guarantee that the template thread will be
2248 // created before a park once haveTemplateThread is set.
2250 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2254 newm(templateThread, nil, -1)
2258 // templateThread is a thread in a known-good state that exists solely
2259 // to start new threads in known-good states when the calling thread
2260 // may not be in a good state.
2262 // Many programs never need this, so templateThread is started lazily
2263 // when we first enter a state that might lead to running on a thread
2264 // in an unknown state.
2266 // templateThread runs on an M without a P, so it must not have write
2269 //go:nowritebarrierrec
2270 func templateThread() {
2277 lock(&newmHandoff.lock)
2278 for newmHandoff.newm != 0 {
2279 newm := newmHandoff.newm.ptr()
2280 newmHandoff.newm = 0
2281 unlock(&newmHandoff.lock)
2283 next := newm.schedlink.ptr()
2288 lock(&newmHandoff.lock)
2290 newmHandoff.waiting = true
2291 noteclear(&newmHandoff.wake)
2292 unlock(&newmHandoff.lock)
2293 notesleep(&newmHandoff.wake)
2297 // Stops execution of the current m until new work is available.
2298 // Returns with acquired P.
2302 if gp.m.locks != 0 {
2303 throw("stopm holding locks")
2306 throw("stopm holding p")
2309 throw("stopm spinning")
2316 acquirep(gp.m.nextp.ptr())
2321 // startm's caller incremented nmspinning. Set the new M's spinning.
2322 getg().m.spinning = true
2325 // Schedules some M to run the p (creates an M if necessary).
2326 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2327 // May run with m.p==nil, so write barriers are not allowed.
2328 // If spinning is set, the caller has incremented nmspinning and must provide a
2329 // P. startm will set m.spinning in the newly started M.
2331 // Callers passing a non-nil P must call from a non-preemptible context. See
2332 // comment on acquirem below.
2334 // Must not have write barriers because this may be called without a P.
2336 //go:nowritebarrierrec
2337 func startm(pp *p, spinning bool) {
2338 // Disable preemption.
2340 // Every owned P must have an owner that will eventually stop it in the
2341 // event of a GC stop request. startm takes transient ownership of a P
2342 // (either from argument or pidleget below) and transfers ownership to
2343 // a started M, which will be responsible for performing the stop.
2345 // Preemption must be disabled during this transient ownership,
2346 // otherwise the P this is running on may enter GC stop while still
2347 // holding the transient P, leaving that P in limbo and deadlocking the
2350 // Callers passing a non-nil P must already be in non-preemptible
2351 // context, otherwise such preemption could occur on function entry to
2352 // startm. Callers passing a nil P may be preemptible, so we must
2353 // disable preemption before acquiring a P from pidleget below.
2358 // TODO(prattmic): All remaining calls to this function
2359 // with _p_ == nil could be cleaned up to find a P
2360 // before calling startm.
2361 throw("startm: P required for spinning=true")
2372 // No M is available, we must drop sched.lock and call newm.
2373 // However, we already own a P to assign to the M.
2375 // Once sched.lock is released, another G (e.g., in a syscall),
2376 // could find no idle P while checkdead finds a runnable G but
2377 // no running M's because this new M hasn't started yet, thus
2378 // throwing in an apparent deadlock.
2380 // Avoid this situation by pre-allocating the ID for the new M,
2381 // thus marking it as 'running' before we drop sched.lock. This
2382 // new M will eventually run the scheduler to execute any
2389 // The caller incremented nmspinning, so set m.spinning in the new M.
2393 // Ownership transfer of pp committed by start in newm.
2394 // Preemption is now safe.
2400 throw("startm: m is spinning")
2403 throw("startm: m has p")
2405 if spinning && !runqempty(pp) {
2406 throw("startm: p has runnable gs")
2408 // The caller incremented nmspinning, so set m.spinning in the new M.
2409 nmp.spinning = spinning
2411 notewakeup(&nmp.park)
2412 // Ownership transfer of pp committed by wakeup. Preemption is now
2417 // Hands off P from syscall or locked M.
2418 // Always runs without a P, so write barriers are not allowed.
2420 //go:nowritebarrierrec
2421 func handoffp(pp *p) {
2422 // handoffp must start an M in any situation where
2423 // findrunnable would return a G to run on pp.
2425 // if it has local work, start it straight away
2426 if !runqempty(pp) || sched.runqsize != 0 {
2430 // if there's trace work to do, start it straight away
2431 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2435 // if it has GC work, start it straight away
2436 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2440 // no local work, check that there are no spinning/idle M's,
2441 // otherwise our help is not required
2442 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2443 sched.needspinning.Store(0)
2448 if sched.gcwaiting.Load() {
2449 pp.status = _Pgcstop
2451 if sched.stopwait == 0 {
2452 notewakeup(&sched.stopnote)
2457 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2458 sched.safePointFn(pp)
2459 sched.safePointWait--
2460 if sched.safePointWait == 0 {
2461 notewakeup(&sched.safePointNote)
2464 if sched.runqsize != 0 {
2469 // If this is the last running P and nobody is polling network,
2470 // need to wakeup another M to poll network.
2471 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2477 // The scheduler lock cannot be held when calling wakeNetPoller below
2478 // because wakeNetPoller may call wakep which may call startm.
2479 when := nobarrierWakeTime(pp)
2488 // Tries to add one more P to execute G's.
2489 // Called when a G is made runnable (newproc, ready).
2490 // Must be called with a P.
2492 // Be conservative about spinning threads, only start one if none exist
2494 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2498 // Disable preemption until ownership of pp transfers to the next M in
2499 // startm. Otherwise preemption here would leave pp stuck waiting to
2502 // See preemption comment on acquirem in startm for more details.
2507 pp, _ = pidlegetSpinning(0)
2509 if sched.nmspinning.Add(-1) < 0 {
2510 throw("wakep: negative nmspinning")
2516 // Since we always have a P, the race in the "No M is available"
2517 // comment in startm doesn't apply during the small window between the
2518 // unlock here and lock in startm. A checkdead in between will always
2519 // see at least one running M (ours).
2527 // Stops execution of the current m that is locked to a g until the g is runnable again.
2528 // Returns with acquired P.
2529 func stoplockedm() {
2532 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2533 throw("stoplockedm: inconsistent locking")
2536 // Schedule another M to run this p.
2541 // Wait until another thread schedules lockedg again.
2543 status := readgstatus(gp.m.lockedg.ptr())
2544 if status&^_Gscan != _Grunnable {
2545 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2546 dumpgstatus(gp.m.lockedg.ptr())
2547 throw("stoplockedm: not runnable")
2549 acquirep(gp.m.nextp.ptr())
2553 // Schedules the locked m to run the locked gp.
2554 // May run during STW, so write barriers are not allowed.
2556 //go:nowritebarrierrec
2557 func startlockedm(gp *g) {
2558 mp := gp.lockedm.ptr()
2560 throw("startlockedm: locked to me")
2563 throw("startlockedm: m has p")
2565 // directly handoff current P to the locked m
2569 notewakeup(&mp.park)
2573 // Stops the current m for stopTheWorld.
2574 // Returns when the world is restarted.
2578 if !sched.gcwaiting.Load() {
2579 throw("gcstopm: not waiting for gc")
2582 gp.m.spinning = false
2583 // OK to just drop nmspinning here,
2584 // startTheWorld will unpark threads as necessary.
2585 if sched.nmspinning.Add(-1) < 0 {
2586 throw("gcstopm: negative nmspinning")
2591 pp.status = _Pgcstop
2593 if sched.stopwait == 0 {
2594 notewakeup(&sched.stopnote)
2600 // Schedules gp to run on the current M.
2601 // If inheritTime is true, gp inherits the remaining time in the
2602 // current time slice. Otherwise, it starts a new time slice.
2605 // Write barriers are allowed because this is called immediately after
2606 // acquiring a P in several places.
2608 //go:yeswritebarrierrec
2609 func execute(gp *g, inheritTime bool) {
2612 if goroutineProfile.active {
2613 // Make sure that gp has had its stack written out to the goroutine
2614 // profile, exactly as it was when the goroutine profiler first stopped
2616 tryRecordGoroutineProfile(gp, osyield)
2619 // Assign gp.m before entering _Grunning so running Gs have an
2623 casgstatus(gp, _Grunnable, _Grunning)
2626 gp.stackguard0 = gp.stack.lo + _StackGuard
2628 mp.p.ptr().schedtick++
2631 // Check whether the profiler needs to be turned on or off.
2632 hz := sched.profilehz
2633 if mp.profilehz != hz {
2634 setThreadCPUProfiler(hz)
2638 // GoSysExit has to happen when we have a P, but before GoStart.
2639 // So we emit it here.
2640 if gp.syscallsp != 0 && gp.sysblocktraced {
2641 traceGoSysExit(gp.sysexitticks)
2649 // Finds a runnable goroutine to execute.
2650 // Tries to steal from other P's, get g from local or global queue, poll network.
2651 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2652 // reader) so the caller should try to wake a P.
2653 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2656 // The conditions here and in handoffp must agree: if
2657 // findrunnable would return a G to run, handoffp must start
2662 if sched.gcwaiting.Load() {
2666 if pp.runSafePointFn != 0 {
2670 // now and pollUntil are saved for work stealing later,
2671 // which may steal timers. It's important that between now
2672 // and then, nothing blocks, so these numbers remain mostly
2674 now, pollUntil, _ := checkTimers(pp, 0)
2676 // Try to schedule the trace reader.
2677 if trace.enabled || trace.shutdown {
2680 casgstatus(gp, _Gwaiting, _Grunnable)
2681 traceGoUnpark(gp, 0)
2682 return gp, false, true
2686 // Try to schedule a GC worker.
2687 if gcBlackenEnabled != 0 {
2688 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2690 return gp, false, true
2695 // Check the global runnable queue once in a while to ensure fairness.
2696 // Otherwise two goroutines can completely occupy the local runqueue
2697 // by constantly respawning each other.
2698 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2700 gp := globrunqget(pp, 1)
2703 return gp, false, false
2707 // Wake up the finalizer G.
2708 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2709 if gp := wakefing(); gp != nil {
2713 if *cgo_yield != nil {
2714 asmcgocall(*cgo_yield, nil)
2718 if gp, inheritTime := runqget(pp); gp != nil {
2719 return gp, inheritTime, false
2723 if sched.runqsize != 0 {
2725 gp := globrunqget(pp, 0)
2728 return gp, false, false
2733 // This netpoll is only an optimization before we resort to stealing.
2734 // We can safely skip it if there are no waiters or a thread is blocked
2735 // in netpoll already. If there is any kind of logical race with that
2736 // blocked thread (e.g. it has already returned from netpoll, but does
2737 // not set lastpoll yet), this thread will do blocking netpoll below
2739 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2740 if list := netpoll(0); !list.empty() { // non-blocking
2743 casgstatus(gp, _Gwaiting, _Grunnable)
2745 traceGoUnpark(gp, 0)
2747 return gp, false, false
2751 // Spinning Ms: steal work from other Ps.
2753 // Limit the number of spinning Ms to half the number of busy Ps.
2754 // This is necessary to prevent excessive CPU consumption when
2755 // GOMAXPROCS>>1 but the program parallelism is low.
2756 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2761 gp, inheritTime, tnow, w, newWork := stealWork(now)
2763 // Successfully stole.
2764 return gp, inheritTime, false
2767 // There may be new timer or GC work; restart to
2773 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2774 // Earlier timer to wait for.
2779 // We have nothing to do.
2781 // If we're in the GC mark phase, can safely scan and blacken objects,
2782 // and have work to do, run idle-time marking rather than give up the P.
2783 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2784 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2786 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2788 casgstatus(gp, _Gwaiting, _Grunnable)
2790 traceGoUnpark(gp, 0)
2792 return gp, false, false
2794 gcController.removeIdleMarkWorker()
2798 // If a callback returned and no other goroutine is awake,
2799 // then wake event handler goroutine which pauses execution
2800 // until a callback was triggered.
2801 gp, otherReady := beforeIdle(now, pollUntil)
2803 casgstatus(gp, _Gwaiting, _Grunnable)
2805 traceGoUnpark(gp, 0)
2807 return gp, false, false
2813 // Before we drop our P, make a snapshot of the allp slice,
2814 // which can change underfoot once we no longer block
2815 // safe-points. We don't need to snapshot the contents because
2816 // everything up to cap(allp) is immutable.
2817 allpSnapshot := allp
2818 // Also snapshot masks. Value changes are OK, but we can't allow
2819 // len to change out from under us.
2820 idlepMaskSnapshot := idlepMask
2821 timerpMaskSnapshot := timerpMask
2823 // return P and block
2825 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2829 if sched.runqsize != 0 {
2830 gp := globrunqget(pp, 0)
2832 return gp, false, false
2834 if !mp.spinning && sched.needspinning.Load() == 1 {
2835 // See "Delicate dance" comment below.
2840 if releasep() != pp {
2841 throw("findrunnable: wrong p")
2843 now = pidleput(pp, now)
2846 // Delicate dance: thread transitions from spinning to non-spinning
2847 // state, potentially concurrently with submission of new work. We must
2848 // drop nmspinning first and then check all sources again (with
2849 // #StoreLoad memory barrier in between). If we do it the other way
2850 // around, another thread can submit work after we've checked all
2851 // sources but before we drop nmspinning; as a result nobody will
2852 // unpark a thread to run the work.
2854 // This applies to the following sources of work:
2856 // * Goroutines added to a per-P run queue.
2857 // * New/modified-earlier timers on a per-P timer heap.
2858 // * Idle-priority GC work (barring golang.org/issue/19112).
2860 // If we discover new work below, we need to restore m.spinning as a
2861 // signal for resetspinning to unpark a new worker thread (because
2862 // there can be more than one starving goroutine).
2864 // However, if after discovering new work we also observe no idle Ps
2865 // (either here or in resetspinning), we have a problem. We may be
2866 // racing with a non-spinning M in the block above, having found no
2867 // work and preparing to release its P and park. Allowing that P to go
2868 // idle will result in loss of work conservation (idle P while there is
2869 // runnable work). This could result in complete deadlock in the
2870 // unlikely event that we discover new work (from netpoll) right as we
2871 // are racing with _all_ other Ps going idle.
2873 // We use sched.needspinning to synchronize with non-spinning Ms going
2874 // idle. If needspinning is set when they are about to drop their P,
2875 // they abort the drop and instead become a new spinning M on our
2876 // behalf. If we are not racing and the system is truly fully loaded
2877 // then no spinning threads are required, and the next thread to
2878 // naturally become spinning will clear the flag.
2880 // Also see "Worker thread parking/unparking" comment at the top of the
2882 wasSpinning := mp.spinning
2885 if sched.nmspinning.Add(-1) < 0 {
2886 throw("findrunnable: negative nmspinning")
2889 // Note the for correctness, only the last M transitioning from
2890 // spinning to non-spinning must perform these rechecks to
2891 // ensure no missed work. However, the runtime has some cases
2892 // of transient increments of nmspinning that are decremented
2893 // without going through this path, so we must be conservative
2894 // and perform the check on all spinning Ms.
2896 // See https://go.dev/issue/43997.
2898 // Check all runqueues once again.
2899 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2906 // Check for idle-priority GC work again.
2907 pp, gp := checkIdleGCNoP()
2912 // Run the idle worker.
2913 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2914 casgstatus(gp, _Gwaiting, _Grunnable)
2916 traceGoUnpark(gp, 0)
2918 return gp, false, false
2921 // Finally, check for timer creation or expiry concurrently with
2922 // transitioning from spinning to non-spinning.
2924 // Note that we cannot use checkTimers here because it calls
2925 // adjusttimers which may need to allocate memory, and that isn't
2926 // allowed when we don't have an active P.
2927 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
2930 // Poll network until next timer.
2931 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
2932 sched.pollUntil.Store(pollUntil)
2934 throw("findrunnable: netpoll with p")
2937 throw("findrunnable: netpoll with spinning")
2943 delay = pollUntil - now
2949 // When using fake time, just poll.
2952 list := netpoll(delay) // block until new work is available
2953 sched.pollUntil.Store(0)
2954 sched.lastpoll.Store(now)
2955 if faketime != 0 && list.empty() {
2956 // Using fake time and nothing is ready; stop M.
2957 // When all M's stop, checkdead will call timejump.
2962 pp, _ := pidleget(now)
2971 casgstatus(gp, _Gwaiting, _Grunnable)
2973 traceGoUnpark(gp, 0)
2975 return gp, false, false
2982 } else if pollUntil != 0 && netpollinited() {
2983 pollerPollUntil := sched.pollUntil.Load()
2984 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
2992 // pollWork reports whether there is non-background work this P could
2993 // be doing. This is a fairly lightweight check to be used for
2994 // background work loops, like idle GC. It checks a subset of the
2995 // conditions checked by the actual scheduler.
2996 func pollWork() bool {
2997 if sched.runqsize != 0 {
3000 p := getg().m.p.ptr()
3004 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
3005 if list := netpoll(0); !list.empty() {
3013 // stealWork attempts to steal a runnable goroutine or timer from any P.
3015 // If newWork is true, new work may have been readied.
3017 // If now is not 0 it is the current time. stealWork returns the passed time or
3018 // the current time if now was passed as 0.
3019 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
3020 pp := getg().m.p.ptr()
3024 const stealTries = 4
3025 for i := 0; i < stealTries; i++ {
3026 stealTimersOrRunNextG := i == stealTries-1
3028 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
3029 if sched.gcwaiting.Load() {
3030 // GC work may be available.
3031 return nil, false, now, pollUntil, true
3033 p2 := allp[enum.position()]
3038 // Steal timers from p2. This call to checkTimers is the only place
3039 // where we might hold a lock on a different P's timers. We do this
3040 // once on the last pass before checking runnext because stealing
3041 // from the other P's runnext should be the last resort, so if there
3042 // are timers to steal do that first.
3044 // We only check timers on one of the stealing iterations because
3045 // the time stored in now doesn't change in this loop and checking
3046 // the timers for each P more than once with the same value of now
3047 // is probably a waste of time.
3049 // timerpMask tells us whether the P may have timers at all. If it
3050 // can't, no need to check at all.
3051 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3052 tnow, w, ran := checkTimers(p2, now)
3054 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3058 // Running the timers may have
3059 // made an arbitrary number of G's
3060 // ready and added them to this P's
3061 // local run queue. That invalidates
3062 // the assumption of runqsteal
3063 // that it always has room to add
3064 // stolen G's. So check now if there
3065 // is a local G to run.
3066 if gp, inheritTime := runqget(pp); gp != nil {
3067 return gp, inheritTime, now, pollUntil, ranTimer
3073 // Don't bother to attempt to steal if p2 is idle.
3074 if !idlepMask.read(enum.position()) {
3075 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3076 return gp, false, now, pollUntil, ranTimer
3082 // No goroutines found to steal. Regardless, running a timer may have
3083 // made some goroutine ready that we missed. Indicate the next timer to
3085 return nil, false, now, pollUntil, ranTimer
3088 // Check all Ps for a runnable G to steal.
3090 // On entry we have no P. If a G is available to steal and a P is available,
3091 // the P is returned which the caller should acquire and attempt to steal the
3093 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3094 for id, p2 := range allpSnapshot {
3095 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3097 pp, _ := pidlegetSpinning(0)
3099 // Can't get a P, don't bother checking remaining Ps.
3108 // No work available.
3112 // Check all Ps for a timer expiring sooner than pollUntil.
3114 // Returns updated pollUntil value.
3115 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3116 for id, p2 := range allpSnapshot {
3117 if timerpMaskSnapshot.read(uint32(id)) {
3118 w := nobarrierWakeTime(p2)
3119 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3128 // Check for idle-priority GC, without a P on entry.
3130 // If some GC work, a P, and a worker G are all available, the P and G will be
3131 // returned. The returned P has not been wired yet.
3132 func checkIdleGCNoP() (*p, *g) {
3133 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3134 // must check again after acquiring a P. As an optimization, we also check
3135 // if an idle mark worker is needed at all. This is OK here, because if we
3136 // observe that one isn't needed, at least one is currently running. Even if
3137 // it stops running, its own journey into the scheduler should schedule it
3138 // again, if need be (at which point, this check will pass, if relevant).
3139 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3142 if !gcMarkWorkAvailable(nil) {
3146 // Work is available; we can start an idle GC worker only if there is
3147 // an available P and available worker G.
3149 // We can attempt to acquire these in either order, though both have
3150 // synchronization concerns (see below). Workers are almost always
3151 // available (see comment in findRunnableGCWorker for the one case
3152 // there may be none). Since we're slightly less likely to find a P,
3153 // check for that first.
3155 // Synchronization: note that we must hold sched.lock until we are
3156 // committed to keeping it. Otherwise we cannot put the unnecessary P
3157 // back in sched.pidle without performing the full set of idle
3158 // transition checks.
3160 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3161 // the assumption in gcControllerState.findRunnableGCWorker that an
3162 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3164 pp, now := pidlegetSpinning(0)
3170 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3171 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3177 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3181 gcController.removeIdleMarkWorker()
3187 return pp, node.gp.ptr()
3190 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3191 // going to wake up before the when argument; or it wakes an idle P to service
3192 // timers and the network poller if there isn't one already.
3193 func wakeNetPoller(when int64) {
3194 if sched.lastpoll.Load() == 0 {
3195 // In findrunnable we ensure that when polling the pollUntil
3196 // field is either zero or the time to which the current
3197 // poll is expected to run. This can have a spurious wakeup
3198 // but should never miss a wakeup.
3199 pollerPollUntil := sched.pollUntil.Load()
3200 if pollerPollUntil == 0 || pollerPollUntil > when {
3204 // There are no threads in the network poller, try to get
3205 // one there so it can handle new timers.
3206 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3212 func resetspinning() {
3215 throw("resetspinning: not a spinning m")
3217 gp.m.spinning = false
3218 nmspinning := sched.nmspinning.Add(-1)
3220 throw("findrunnable: negative nmspinning")
3222 // M wakeup policy is deliberately somewhat conservative, so check if we
3223 // need to wakeup another P here. See "Worker thread parking/unparking"
3224 // comment at the top of the file for details.
3228 // injectglist adds each runnable G on the list to some run queue,
3229 // and clears glist. If there is no current P, they are added to the
3230 // global queue, and up to npidle M's are started to run them.
3231 // Otherwise, for each idle P, this adds a G to the global queue
3232 // and starts an M. Any remaining G's are added to the current P's
3234 // This may temporarily acquire sched.lock.
3235 // Can run concurrently with GC.
3236 func injectglist(glist *gList) {
3241 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3242 traceGoUnpark(gp, 0)
3246 // Mark all the goroutines as runnable before we put them
3247 // on the run queues.
3248 head := glist.head.ptr()
3251 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3254 casgstatus(gp, _Gwaiting, _Grunnable)
3257 // Turn the gList into a gQueue.
3263 startIdle := func(n int) {
3264 for i := 0; i < n; i++ {
3265 mp := acquirem() // See comment in startm.
3268 pp, _ := pidlegetSpinning(0)
3281 pp := getg().m.p.ptr()
3284 globrunqputbatch(&q, int32(qsize))
3290 npidle := int(sched.npidle.Load())
3293 for n = 0; n < npidle && !q.empty(); n++ {
3299 globrunqputbatch(&globq, int32(n))
3306 runqputbatch(pp, &q, qsize)
3310 // One round of scheduler: find a runnable goroutine and execute it.
3316 throw("schedule: holding locks")
3319 if mp.lockedg != 0 {
3321 execute(mp.lockedg.ptr(), false) // Never returns.
3324 // We should not schedule away from a g that is executing a cgo call,
3325 // since the cgo call is using the m's g0 stack.
3327 throw("schedule: in cgo")
3334 // Safety check: if we are spinning, the run queue should be empty.
3335 // Check this before calling checkTimers, as that might call
3336 // goready to put a ready goroutine on the local run queue.
3337 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3338 throw("schedule: spinning with local work")
3341 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3343 // This thread is going to run a goroutine and is not spinning anymore,
3344 // so if it was marked as spinning we need to reset it now and potentially
3345 // start a new spinning M.
3350 if sched.disable.user && !schedEnabled(gp) {
3351 // Scheduling of this goroutine is disabled. Put it on
3352 // the list of pending runnable goroutines for when we
3353 // re-enable user scheduling and look again.
3355 if schedEnabled(gp) {
3356 // Something re-enabled scheduling while we
3357 // were acquiring the lock.
3360 sched.disable.runnable.pushBack(gp)
3367 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3368 // wake a P if there is one.
3372 if gp.lockedm != 0 {
3373 // Hands off own p to the locked m,
3374 // then blocks waiting for a new p.
3379 execute(gp, inheritTime)
3382 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3383 // Typically a caller sets gp's status away from Grunning and then
3384 // immediately calls dropg to finish the job. The caller is also responsible
3385 // for arranging that gp will be restarted using ready at an
3386 // appropriate time. After calling dropg and arranging for gp to be
3387 // readied later, the caller can do other work but eventually should
3388 // call schedule to restart the scheduling of goroutines on this m.
3392 setMNoWB(&gp.m.curg.m, nil)
3393 setGNoWB(&gp.m.curg, nil)
3396 // checkTimers runs any timers for the P that are ready.
3397 // If now is not 0 it is the current time.
3398 // It returns the passed time or the current time if now was passed as 0.
3399 // and the time when the next timer should run or 0 if there is no next timer,
3400 // and reports whether it ran any timers.
3401 // If the time when the next timer should run is not 0,
3402 // it is always larger than the returned time.
3403 // We pass now in and out to avoid extra calls of nanotime.
3405 //go:yeswritebarrierrec
3406 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3407 // If it's not yet time for the first timer, or the first adjusted
3408 // timer, then there is nothing to do.
3409 next := pp.timer0When.Load()
3410 nextAdj := pp.timerModifiedEarliest.Load()
3411 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3416 // No timers to run or adjust.
3417 return now, 0, false
3424 // Next timer is not ready to run, but keep going
3425 // if we would clear deleted timers.
3426 // This corresponds to the condition below where
3427 // we decide whether to call clearDeletedTimers.
3428 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3429 return now, next, false
3433 lock(&pp.timersLock)
3435 if len(pp.timers) > 0 {
3436 adjusttimers(pp, now)
3437 for len(pp.timers) > 0 {
3438 // Note that runtimer may temporarily unlock
3440 if tw := runtimer(pp, now); tw != 0 {
3450 // If this is the local P, and there are a lot of deleted timers,
3451 // clear them out. We only do this for the local P to reduce
3452 // lock contention on timersLock.
3453 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3454 clearDeletedTimers(pp)
3457 unlock(&pp.timersLock)
3459 return now, pollUntil, ran
3462 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3463 unlock((*mutex)(lock))
3467 // park continuation on g0.
3468 func park_m(gp *g) {
3472 traceGoPark(mp.waittraceev, mp.waittraceskip)
3475 // N.B. Not using casGToWaiting here because the waitreason is
3476 // set by park_m's caller.
3477 casgstatus(gp, _Grunning, _Gwaiting)
3480 if fn := mp.waitunlockf; fn != nil {
3481 ok := fn(gp, mp.waitlock)
3482 mp.waitunlockf = nil
3486 traceGoUnpark(gp, 2)
3488 casgstatus(gp, _Gwaiting, _Grunnable)
3489 execute(gp, true) // Schedule it back, never returns.
3495 func goschedImpl(gp *g) {
3496 status := readgstatus(gp)
3497 if status&^_Gscan != _Grunning {
3499 throw("bad g status")
3501 casgstatus(gp, _Grunning, _Grunnable)
3510 // Gosched continuation on g0.
3511 func gosched_m(gp *g) {
3518 // goschedguarded is a forbidden-states-avoided version of gosched_m
3519 func goschedguarded_m(gp *g) {
3521 if !canPreemptM(gp.m) {
3522 gogo(&gp.sched) // never return
3531 func gopreempt_m(gp *g) {
3538 // preemptPark parks gp and puts it in _Gpreempted.
3541 func preemptPark(gp *g) {
3543 traceGoPark(traceEvGoBlock, 0)
3545 status := readgstatus(gp)
3546 if status&^_Gscan != _Grunning {
3548 throw("bad g status")
3551 if gp.asyncSafePoint {
3552 // Double-check that async preemption does not
3553 // happen in SPWRITE assembly functions.
3554 // isAsyncSafePoint must exclude this case.
3555 f := findfunc(gp.sched.pc)
3557 throw("preempt at unknown pc")
3559 if f.flag&funcFlag_SPWRITE != 0 {
3560 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3561 throw("preempt SPWRITE")
3565 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3566 // be in _Grunning when we dropg because then we'd be running
3567 // without an M, but the moment we're in _Gpreempted,
3568 // something could claim this G before we've fully cleaned it
3569 // up. Hence, we set the scan bit to lock down further
3570 // transitions until we can dropg.
3571 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3573 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3577 // goyield is like Gosched, but it:
3578 // - emits a GoPreempt trace event instead of a GoSched trace event
3579 // - puts the current G on the runq of the current P instead of the globrunq
3585 func goyield_m(gp *g) {
3590 casgstatus(gp, _Grunning, _Grunnable)
3592 runqput(pp, gp, false)
3596 // Finishes execution of the current goroutine.
3607 // goexit continuation on g0.
3608 func goexit0(gp *g) {
3612 casgstatus(gp, _Grunning, _Gdead)
3613 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3614 if isSystemGoroutine(gp, false) {
3618 locked := gp.lockedm != 0
3621 gp.preemptStop = false
3622 gp.paniconfault = false
3623 gp._defer = nil // should be true already but just in case.
3624 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3626 gp.waitreason = waitReasonZero
3631 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3632 // Flush assist credit to the global pool. This gives
3633 // better information to pacing if the application is
3634 // rapidly creating an exiting goroutines.
3635 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3636 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3637 gcController.bgScanCredit.Add(scanCredit)
3638 gp.gcAssistBytes = 0
3643 if GOARCH == "wasm" { // no threads yet on wasm
3645 schedule() // never returns
3648 if mp.lockedInt != 0 {
3649 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3650 throw("internal lockOSThread error")
3654 // The goroutine may have locked this thread because
3655 // it put it in an unusual kernel state. Kill it
3656 // rather than returning it to the thread pool.
3658 // Return to mstart, which will release the P and exit
3660 if GOOS != "plan9" { // See golang.org/issue/22227.
3663 // Clear lockedExt on plan9 since we may end up re-using
3671 // save updates getg().sched to refer to pc and sp so that a following
3672 // gogo will restore pc and sp.
3674 // save must not have write barriers because invoking a write barrier
3675 // can clobber getg().sched.
3678 //go:nowritebarrierrec
3679 func save(pc, sp uintptr) {
3682 if gp == gp.m.g0 || gp == gp.m.gsignal {
3683 // m.g0.sched is special and must describe the context
3684 // for exiting the thread. mstart1 writes to it directly.
3685 // m.gsignal.sched should not be used at all.
3686 // This check makes sure save calls do not accidentally
3687 // run in contexts where they'd write to system g's.
3688 throw("save on system g not allowed")
3695 // We need to ensure ctxt is zero, but can't have a write
3696 // barrier here. However, it should always already be zero.
3698 if gp.sched.ctxt != nil {
3703 // The goroutine g is about to enter a system call.
3704 // Record that it's not using the cpu anymore.
3705 // This is called only from the go syscall library and cgocall,
3706 // not from the low-level system calls used by the runtime.
3708 // Entersyscall cannot split the stack: the save must
3709 // make g->sched refer to the caller's stack segment, because
3710 // entersyscall is going to return immediately after.
3712 // Nothing entersyscall calls can split the stack either.
3713 // We cannot safely move the stack during an active call to syscall,
3714 // because we do not know which of the uintptr arguments are
3715 // really pointers (back into the stack).
3716 // In practice, this means that we make the fast path run through
3717 // entersyscall doing no-split things, and the slow path has to use systemstack
3718 // to run bigger things on the system stack.
3720 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3721 // saved SP and PC are restored. This is needed when exitsyscall will be called
3722 // from a function further up in the call stack than the parent, as g->syscallsp
3723 // must always point to a valid stack frame. entersyscall below is the normal
3724 // entry point for syscalls, which obtains the SP and PC from the caller.
3727 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3728 // If the syscall does not block, that is it, we do not emit any other events.
3729 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3730 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3731 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3732 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3733 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3734 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3735 // and we wait for the increment before emitting traceGoSysExit.
3736 // Note that the increment is done even if tracing is not enabled,
3737 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3740 func reentersyscall(pc, sp uintptr) {
3743 // Disable preemption because during this function g is in Gsyscall status,
3744 // but can have inconsistent g->sched, do not let GC observe it.
3747 // Entersyscall must not call any function that might split/grow the stack.
3748 // (See details in comment above.)
3749 // Catch calls that might, by replacing the stack guard with something that
3750 // will trip any stack check and leaving a flag to tell newstack to die.
3751 gp.stackguard0 = stackPreempt
3752 gp.throwsplit = true
3754 // Leave SP around for GC and traceback.
3758 casgstatus(gp, _Grunning, _Gsyscall)
3759 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3760 systemstack(func() {
3761 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3762 throw("entersyscall")
3767 systemstack(traceGoSysCall)
3768 // systemstack itself clobbers g.sched.{pc,sp} and we might
3769 // need them later when the G is genuinely blocked in a
3774 if sched.sysmonwait.Load() {
3775 systemstack(entersyscall_sysmon)
3779 if gp.m.p.ptr().runSafePointFn != 0 {
3780 // runSafePointFn may stack split if run on this stack
3781 systemstack(runSafePointFn)
3785 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3786 gp.sysblocktraced = true
3791 atomic.Store(&pp.status, _Psyscall)
3792 if sched.gcwaiting.Load() {
3793 systemstack(entersyscall_gcwait)
3800 // Standard syscall entry used by the go syscall library and normal cgo calls.
3802 // This is exported via linkname to assembly in the syscall package and x/sys.
3805 //go:linkname entersyscall
3806 func entersyscall() {
3807 reentersyscall(getcallerpc(), getcallersp())
3810 func entersyscall_sysmon() {
3812 if sched.sysmonwait.Load() {
3813 sched.sysmonwait.Store(false)
3814 notewakeup(&sched.sysmonnote)
3819 func entersyscall_gcwait() {
3821 pp := gp.m.oldp.ptr()
3824 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3830 if sched.stopwait--; sched.stopwait == 0 {
3831 notewakeup(&sched.stopnote)
3837 // The same as entersyscall(), but with a hint that the syscall is blocking.
3840 func entersyscallblock() {
3843 gp.m.locks++ // see comment in entersyscall
3844 gp.throwsplit = true
3845 gp.stackguard0 = stackPreempt // see comment in entersyscall
3846 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3847 gp.sysblocktraced = true
3848 gp.m.p.ptr().syscalltick++
3850 // Leave SP around for GC and traceback.
3854 gp.syscallsp = gp.sched.sp
3855 gp.syscallpc = gp.sched.pc
3856 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3860 systemstack(func() {
3861 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3862 throw("entersyscallblock")
3865 casgstatus(gp, _Grunning, _Gsyscall)
3866 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3867 systemstack(func() {
3868 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3869 throw("entersyscallblock")
3873 systemstack(entersyscallblock_handoff)
3875 // Resave for traceback during blocked call.
3876 save(getcallerpc(), getcallersp())
3881 func entersyscallblock_handoff() {
3884 traceGoSysBlock(getg().m.p.ptr())
3886 handoffp(releasep())
3889 // The goroutine g exited its system call.
3890 // Arrange for it to run on a cpu again.
3891 // This is called only from the go syscall library, not
3892 // from the low-level system calls used by the runtime.
3894 // Write barriers are not allowed because our P may have been stolen.
3896 // This is exported via linkname to assembly in the syscall package.
3899 //go:nowritebarrierrec
3900 //go:linkname exitsyscall
3901 func exitsyscall() {
3904 gp.m.locks++ // see comment in entersyscall
3905 if getcallersp() > gp.syscallsp {
3906 throw("exitsyscall: syscall frame is no longer valid")
3910 oldp := gp.m.oldp.ptr()
3912 if exitsyscallfast(oldp) {
3913 // When exitsyscallfast returns success, we have a P so can now use
3915 if goroutineProfile.active {
3916 // Make sure that gp has had its stack written out to the goroutine
3917 // profile, exactly as it was when the goroutine profiler first
3918 // stopped the world.
3919 systemstack(func() {
3920 tryRecordGoroutineProfileWB(gp)
3924 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3925 systemstack(traceGoStart)
3928 // There's a cpu for us, so we can run.
3929 gp.m.p.ptr().syscalltick++
3930 // We need to cas the status and scan before resuming...
3931 casgstatus(gp, _Gsyscall, _Grunning)
3933 // Garbage collector isn't running (since we are),
3934 // so okay to clear syscallsp.
3938 // restore the preemption request in case we've cleared it in newstack
3939 gp.stackguard0 = stackPreempt
3941 // otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
3942 gp.stackguard0 = gp.stack.lo + _StackGuard
3944 gp.throwsplit = false
3946 if sched.disable.user && !schedEnabled(gp) {
3947 // Scheduling of this goroutine is disabled.
3956 // Wait till traceGoSysBlock event is emitted.
3957 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3958 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
3961 // We can't trace syscall exit right now because we don't have a P.
3962 // Tracing code can invoke write barriers that cannot run without a P.
3963 // So instead we remember the syscall exit time and emit the event
3964 // in execute when we have a P.
3965 gp.sysexitticks = cputicks()
3970 // Call the scheduler.
3973 // Scheduler returned, so we're allowed to run now.
3974 // Delete the syscallsp information that we left for
3975 // the garbage collector during the system call.
3976 // Must wait until now because until gosched returns
3977 // we don't know for sure that the garbage collector
3980 gp.m.p.ptr().syscalltick++
3981 gp.throwsplit = false
3985 func exitsyscallfast(oldp *p) bool {
3988 // Freezetheworld sets stopwait but does not retake P's.
3989 if sched.stopwait == freezeStopWait {
3993 // Try to re-acquire the last P.
3994 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
3995 // There's a cpu for us, so we can run.
3997 exitsyscallfast_reacquired()
4001 // Try to get any other idle P.
4002 if sched.pidle != 0 {
4004 systemstack(func() {
4005 ok = exitsyscallfast_pidle()
4006 if ok && trace.enabled {
4008 // Wait till traceGoSysBlock event is emitted.
4009 // This ensures consistency of the trace (the goroutine is started after it is blocked).
4010 for oldp.syscalltick == gp.m.syscalltick {
4024 // exitsyscallfast_reacquired is the exitsyscall path on which this G
4025 // has successfully reacquired the P it was running on before the
4029 func exitsyscallfast_reacquired() {
4031 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4033 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4034 // traceGoSysBlock for this syscall was already emitted,
4035 // but here we effectively retake the p from the new syscall running on the same p.
4036 systemstack(func() {
4037 // Denote blocking of the new syscall.
4038 traceGoSysBlock(gp.m.p.ptr())
4039 // Denote completion of the current syscall.
4043 gp.m.p.ptr().syscalltick++
4047 func exitsyscallfast_pidle() bool {
4049 pp, _ := pidleget(0)
4050 if pp != nil && sched.sysmonwait.Load() {
4051 sched.sysmonwait.Store(false)
4052 notewakeup(&sched.sysmonnote)
4062 // exitsyscall slow path on g0.
4063 // Failed to acquire P, enqueue gp as runnable.
4065 // Called via mcall, so gp is the calling g from this M.
4067 //go:nowritebarrierrec
4068 func exitsyscall0(gp *g) {
4069 casgstatus(gp, _Gsyscall, _Grunnable)
4073 if schedEnabled(gp) {
4080 // Below, we stoplockedm if gp is locked. globrunqput releases
4081 // ownership of gp, so we must check if gp is locked prior to
4082 // committing the release by unlocking sched.lock, otherwise we
4083 // could race with another M transitioning gp from unlocked to
4085 locked = gp.lockedm != 0
4086 } else if sched.sysmonwait.Load() {
4087 sched.sysmonwait.Store(false)
4088 notewakeup(&sched.sysmonnote)
4093 execute(gp, false) // Never returns.
4096 // Wait until another thread schedules gp and so m again.
4098 // N.B. lockedm must be this M, as this g was running on this M
4099 // before entersyscall.
4101 execute(gp, false) // Never returns.
4104 schedule() // Never returns.
4107 // Called from syscall package before fork.
4109 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4111 func syscall_runtime_BeforeFork() {
4114 // Block signals during a fork, so that the child does not run
4115 // a signal handler before exec if a signal is sent to the process
4116 // group. See issue #18600.
4118 sigsave(&gp.m.sigmask)
4121 // This function is called before fork in syscall package.
4122 // Code between fork and exec must not allocate memory nor even try to grow stack.
4123 // Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
4124 // runtime_AfterFork will undo this in parent process, but not in child.
4125 gp.stackguard0 = stackFork
4128 // Called from syscall package after fork in parent.
4130 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4132 func syscall_runtime_AfterFork() {
4135 // See the comments in beforefork.
4136 gp.stackguard0 = gp.stack.lo + _StackGuard
4138 msigrestore(gp.m.sigmask)
4143 // inForkedChild is true while manipulating signals in the child process.
4144 // This is used to avoid calling libc functions in case we are using vfork.
4145 var inForkedChild bool
4147 // Called from syscall package after fork in child.
4148 // It resets non-sigignored signals to the default handler, and
4149 // restores the signal mask in preparation for the exec.
4151 // Because this might be called during a vfork, and therefore may be
4152 // temporarily sharing address space with the parent process, this must
4153 // not change any global variables or calling into C code that may do so.
4155 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4157 //go:nowritebarrierrec
4158 func syscall_runtime_AfterForkInChild() {
4159 // It's OK to change the global variable inForkedChild here
4160 // because we are going to change it back. There is no race here,
4161 // because if we are sharing address space with the parent process,
4162 // then the parent process can not be running concurrently.
4163 inForkedChild = true
4165 clearSignalHandlers()
4167 // When we are the child we are the only thread running,
4168 // so we know that nothing else has changed gp.m.sigmask.
4169 msigrestore(getg().m.sigmask)
4171 inForkedChild = false
4174 // pendingPreemptSignals is the number of preemption signals
4175 // that have been sent but not received. This is only used on Darwin.
4177 var pendingPreemptSignals atomic.Int32
4179 // Called from syscall package before Exec.
4181 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4182 func syscall_runtime_BeforeExec() {
4183 // Prevent thread creation during exec.
4186 // On Darwin, wait for all pending preemption signals to
4187 // be received. See issue #41702.
4188 if GOOS == "darwin" || GOOS == "ios" {
4189 for pendingPreemptSignals.Load() > 0 {
4195 // Called from syscall package after Exec.
4197 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4198 func syscall_runtime_AfterExec() {
4202 // Allocate a new g, with a stack big enough for stacksize bytes.
4203 func malg(stacksize int32) *g {
4206 stacksize = round2(_StackSystem + stacksize)
4207 systemstack(func() {
4208 newg.stack = stackalloc(uint32(stacksize))
4210 newg.stackguard0 = newg.stack.lo + _StackGuard
4211 newg.stackguard1 = ^uintptr(0)
4212 // Clear the bottom word of the stack. We record g
4213 // there on gsignal stack during VDSO on ARM and ARM64.
4214 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4219 // Create a new g running fn.
4220 // Put it on the queue of g's waiting to run.
4221 // The compiler turns a go statement into a call to this.
4222 func newproc(fn *funcval) {
4225 systemstack(func() {
4226 newg := newproc1(fn, gp, pc)
4228 pp := getg().m.p.ptr()
4229 runqput(pp, newg, true)
4237 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4238 // address of the go statement that created this. The caller is responsible
4239 // for adding the new g to the scheduler.
4240 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4242 fatal("go of nil func value")
4245 mp := acquirem() // disable preemption because we hold M and P in local vars.
4249 newg = malg(_StackMin)
4250 casgstatus(newg, _Gidle, _Gdead)
4251 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4253 if newg.stack.hi == 0 {
4254 throw("newproc1: newg missing stack")
4257 if readgstatus(newg) != _Gdead {
4258 throw("newproc1: new g is not Gdead")
4261 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4262 totalSize = alignUp(totalSize, sys.StackAlign)
4263 sp := newg.stack.hi - totalSize
4267 *(*uintptr)(unsafe.Pointer(sp)) = 0
4269 spArg += sys.MinFrameSize
4272 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4275 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4276 newg.sched.g = guintptr(unsafe.Pointer(newg))
4277 gostartcallfn(&newg.sched, fn)
4278 newg.gopc = callerpc
4279 newg.ancestors = saveAncestors(callergp)
4280 newg.startpc = fn.fn
4281 if isSystemGoroutine(newg, false) {
4284 // Only user goroutines inherit pprof labels.
4286 newg.labels = mp.curg.labels
4288 if goroutineProfile.active {
4289 // A concurrent goroutine profile is running. It should include
4290 // exactly the set of goroutines that were alive when the goroutine
4291 // profiler first stopped the world. That does not include newg, so
4292 // mark it as not needing a profile before transitioning it from
4294 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4297 // Track initial transition?
4298 newg.trackingSeq = uint8(fastrand())
4299 if newg.trackingSeq%gTrackingPeriod == 0 {
4300 newg.tracking = true
4302 casgstatus(newg, _Gdead, _Grunnable)
4303 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4305 if pp.goidcache == pp.goidcacheend {
4306 // Sched.goidgen is the last allocated id,
4307 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4308 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4309 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4310 pp.goidcache -= _GoidCacheBatch - 1
4311 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4313 newg.goid = pp.goidcache
4316 newg.racectx = racegostart(callerpc)
4317 if newg.labels != nil {
4318 // See note in proflabel.go on labelSync's role in synchronizing
4319 // with the reads in the signal handler.
4320 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4324 traceGoCreate(newg, newg.startpc)
4331 // saveAncestors copies previous ancestors of the given caller g and
4332 // includes infor for the current caller into a new set of tracebacks for
4333 // a g being created.
4334 func saveAncestors(callergp *g) *[]ancestorInfo {
4335 // Copy all prior info, except for the root goroutine (goid 0).
4336 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4339 var callerAncestors []ancestorInfo
4340 if callergp.ancestors != nil {
4341 callerAncestors = *callergp.ancestors
4343 n := int32(len(callerAncestors)) + 1
4344 if n > debug.tracebackancestors {
4345 n = debug.tracebackancestors
4347 ancestors := make([]ancestorInfo, n)
4348 copy(ancestors[1:], callerAncestors)
4350 var pcs [_TracebackMaxFrames]uintptr
4351 npcs := gcallers(callergp, 0, pcs[:])
4352 ipcs := make([]uintptr, npcs)
4354 ancestors[0] = ancestorInfo{
4356 goid: callergp.goid,
4357 gopc: callergp.gopc,
4360 ancestorsp := new([]ancestorInfo)
4361 *ancestorsp = ancestors
4365 // Put on gfree list.
4366 // If local list is too long, transfer a batch to the global list.
4367 func gfput(pp *p, gp *g) {
4368 if readgstatus(gp) != _Gdead {
4369 throw("gfput: bad status (not Gdead)")
4372 stksize := gp.stack.hi - gp.stack.lo
4374 if stksize != uintptr(startingStackSize) {
4375 // non-standard stack size - free it.
4384 if pp.gFree.n >= 64 {
4390 for pp.gFree.n >= 32 {
4391 gp := pp.gFree.pop()
4393 if gp.stack.lo == 0 {
4400 lock(&sched.gFree.lock)
4401 sched.gFree.noStack.pushAll(noStackQ)
4402 sched.gFree.stack.pushAll(stackQ)
4403 sched.gFree.n += inc
4404 unlock(&sched.gFree.lock)
4408 // Get from gfree list.
4409 // If local list is empty, grab a batch from global list.
4410 func gfget(pp *p) *g {
4412 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4413 lock(&sched.gFree.lock)
4414 // Move a batch of free Gs to the P.
4415 for pp.gFree.n < 32 {
4416 // Prefer Gs with stacks.
4417 gp := sched.gFree.stack.pop()
4419 gp = sched.gFree.noStack.pop()
4428 unlock(&sched.gFree.lock)
4431 gp := pp.gFree.pop()
4436 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4437 // Deallocate old stack. We kept it in gfput because it was the
4438 // right size when the goroutine was put on the free list, but
4439 // the right size has changed since then.
4440 systemstack(func() {
4447 if gp.stack.lo == 0 {
4448 // Stack was deallocated in gfput or just above. Allocate a new one.
4449 systemstack(func() {
4450 gp.stack = stackalloc(startingStackSize)
4452 gp.stackguard0 = gp.stack.lo + _StackGuard
4455 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4458 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4461 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4467 // Purge all cached G's from gfree list to the global list.
4468 func gfpurge(pp *p) {
4474 for !pp.gFree.empty() {
4475 gp := pp.gFree.pop()
4477 if gp.stack.lo == 0 {
4484 lock(&sched.gFree.lock)
4485 sched.gFree.noStack.pushAll(noStackQ)
4486 sched.gFree.stack.pushAll(stackQ)
4487 sched.gFree.n += inc
4488 unlock(&sched.gFree.lock)
4491 // Breakpoint executes a breakpoint trap.
4496 // dolockOSThread is called by LockOSThread and lockOSThread below
4497 // after they modify m.locked. Do not allow preemption during this call,
4498 // or else the m might be different in this function than in the caller.
4501 func dolockOSThread() {
4502 if GOARCH == "wasm" {
4503 return // no threads on wasm yet
4506 gp.m.lockedg.set(gp)
4507 gp.lockedm.set(gp.m)
4512 // LockOSThread wires the calling goroutine to its current operating system thread.
4513 // The calling goroutine will always execute in that thread,
4514 // and no other goroutine will execute in it,
4515 // until the calling goroutine has made as many calls to
4516 // UnlockOSThread as to LockOSThread.
4517 // If the calling goroutine exits without unlocking the thread,
4518 // the thread will be terminated.
4520 // All init functions are run on the startup thread. Calling LockOSThread
4521 // from an init function will cause the main function to be invoked on
4524 // A goroutine should call LockOSThread before calling OS services or
4525 // non-Go library functions that depend on per-thread state.
4526 func LockOSThread() {
4527 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4528 // If we need to start a new thread from the locked
4529 // thread, we need the template thread. Start it now
4530 // while we're in a known-good state.
4531 startTemplateThread()
4535 if gp.m.lockedExt == 0 {
4537 panic("LockOSThread nesting overflow")
4543 func lockOSThread() {
4544 getg().m.lockedInt++
4548 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4549 // after they update m->locked. Do not allow preemption during this call,
4550 // or else the m might be in different in this function than in the caller.
4553 func dounlockOSThread() {
4554 if GOARCH == "wasm" {
4555 return // no threads on wasm yet
4558 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4567 // UnlockOSThread undoes an earlier call to LockOSThread.
4568 // If this drops the number of active LockOSThread calls on the
4569 // calling goroutine to zero, it unwires the calling goroutine from
4570 // its fixed operating system thread.
4571 // If there are no active LockOSThread calls, this is a no-op.
4573 // Before calling UnlockOSThread, the caller must ensure that the OS
4574 // thread is suitable for running other goroutines. If the caller made
4575 // any permanent changes to the state of the thread that would affect
4576 // other goroutines, it should not call this function and thus leave
4577 // the goroutine locked to the OS thread until the goroutine (and
4578 // hence the thread) exits.
4579 func UnlockOSThread() {
4581 if gp.m.lockedExt == 0 {
4589 func unlockOSThread() {
4591 if gp.m.lockedInt == 0 {
4592 systemstack(badunlockosthread)
4598 func badunlockosthread() {
4599 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4602 func gcount() int32 {
4603 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4604 for _, pp := range allp {
4608 // All these variables can be changed concurrently, so the result can be inconsistent.
4609 // But at least the current goroutine is running.
4616 func mcount() int32 {
4617 return int32(sched.mnext - sched.nmfreed)
4621 signalLock atomic.Uint32
4623 // Must hold signalLock to write. Reads may be lock-free, but
4624 // signalLock should be taken to synchronize with changes.
4628 func _System() { _System() }
4629 func _ExternalCode() { _ExternalCode() }
4630 func _LostExternalCode() { _LostExternalCode() }
4631 func _GC() { _GC() }
4632 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4633 func _VDSO() { _VDSO() }
4635 // Called if we receive a SIGPROF signal.
4636 // Called by the signal handler, may run during STW.
4638 //go:nowritebarrierrec
4639 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4640 if prof.hz.Load() == 0 {
4644 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4645 // We must check this to avoid a deadlock between setcpuprofilerate
4646 // and the call to cpuprof.add, below.
4647 if mp != nil && mp.profilehz == 0 {
4651 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4652 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4653 // the critical section, it creates a deadlock (when writing the sample).
4654 // As a workaround, create a counter of SIGPROFs while in critical section
4655 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4656 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4657 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4658 if f := findfunc(pc); f.valid() {
4659 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4660 cpuprof.lostAtomic++
4664 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4665 // runtime/internal/atomic functions call into kernel
4666 // helpers on arm < 7. See
4667 // runtime/internal/atomic/sys_linux_arm.s.
4668 cpuprof.lostAtomic++
4673 // Profiling runs concurrently with GC, so it must not allocate.
4674 // Set a trap in case the code does allocate.
4675 // Note that on windows, one thread takes profiles of all the
4676 // other threads, so mp is usually not getg().m.
4677 // In fact mp may not even be stopped.
4678 // See golang.org/issue/17165.
4679 getg().m.mallocing++
4681 var stk [maxCPUProfStack]uintptr
4683 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4685 // Check cgoCallersUse to make sure that we are not
4686 // interrupting other code that is fiddling with
4687 // cgoCallers. We are running in a signal handler
4688 // with all signals blocked, so we don't have to worry
4689 // about any other code interrupting us.
4690 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4691 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4694 copy(stk[:], mp.cgoCallers[:cgoOff])
4695 mp.cgoCallers[0] = 0
4698 // Collect Go stack that leads to the cgo call.
4699 n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
4704 n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4708 // Normal traceback is impossible or has failed.
4709 // See if it falls into several common cases.
4711 if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4712 // Libcall, i.e. runtime syscall on windows.
4713 // Collect Go stack that leads to the call.
4714 n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[0], len(stk), nil, nil, 0)
4716 if n == 0 && mp != nil && mp.vdsoSP != 0 {
4717 n = gentraceback(mp.vdsoPC, mp.vdsoSP, 0, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4720 // If all of the above has failed, account it against abstract "System" or "GC".
4723 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4724 } else if pc > firstmoduledata.etext {
4725 // "ExternalCode" is better than "etext".
4726 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4729 if mp.preemptoff != "" {
4730 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4732 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4737 if prof.hz.Load() != 0 {
4738 // Note: it can happen on Windows that we interrupted a system thread
4739 // with no g, so gp could nil. The other nil checks are done out of
4740 // caution, but not expected to be nil in practice.
4741 var tagPtr *unsafe.Pointer
4742 if gp != nil && gp.m != nil && gp.m.curg != nil {
4743 tagPtr = &gp.m.curg.labels
4745 cpuprof.add(tagPtr, stk[:n])
4749 if gp != nil && gp.m != nil {
4750 if gp.m.curg != nil {
4755 traceCPUSample(gprof, pp, stk[:n])
4757 getg().m.mallocing--
4760 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4761 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4762 func setcpuprofilerate(hz int32) {
4763 // Force sane arguments.
4768 // Disable preemption, otherwise we can be rescheduled to another thread
4769 // that has profiling enabled.
4773 // Stop profiler on this thread so that it is safe to lock prof.
4774 // if a profiling signal came in while we had prof locked,
4775 // it would deadlock.
4776 setThreadCPUProfiler(0)
4778 for !prof.signalLock.CompareAndSwap(0, 1) {
4781 if prof.hz.Load() != hz {
4782 setProcessCPUProfiler(hz)
4785 prof.signalLock.Store(0)
4788 sched.profilehz = hz
4792 setThreadCPUProfiler(hz)
4798 // init initializes pp, which may be a freshly allocated p or a
4799 // previously destroyed p, and transitions it to status _Pgcstop.
4800 func (pp *p) init(id int32) {
4802 pp.status = _Pgcstop
4803 pp.sudogcache = pp.sudogbuf[:0]
4804 pp.deferpool = pp.deferpoolbuf[:0]
4806 if pp.mcache == nil {
4809 throw("missing mcache?")
4811 // Use the bootstrap mcache0. Only one P will get
4812 // mcache0: the one with ID 0.
4815 pp.mcache = allocmcache()
4818 if raceenabled && pp.raceprocctx == 0 {
4820 pp.raceprocctx = raceprocctx0
4821 raceprocctx0 = 0 // bootstrap
4823 pp.raceprocctx = raceproccreate()
4826 lockInit(&pp.timersLock, lockRankTimers)
4828 // This P may get timers when it starts running. Set the mask here
4829 // since the P may not go through pidleget (notably P 0 on startup).
4831 // Similarly, we may not go through pidleget before this P starts
4832 // running if it is P 0 on startup.
4836 // destroy releases all of the resources associated with pp and
4837 // transitions it to status _Pdead.
4839 // sched.lock must be held and the world must be stopped.
4840 func (pp *p) destroy() {
4841 assertLockHeld(&sched.lock)
4842 assertWorldStopped()
4844 // Move all runnable goroutines to the global queue
4845 for pp.runqhead != pp.runqtail {
4846 // Pop from tail of local queue
4848 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4849 // Push onto head of global queue
4852 if pp.runnext != 0 {
4853 globrunqputhead(pp.runnext.ptr())
4856 if len(pp.timers) > 0 {
4857 plocal := getg().m.p.ptr()
4858 // The world is stopped, but we acquire timersLock to
4859 // protect against sysmon calling timeSleepUntil.
4860 // This is the only case where we hold the timersLock of
4861 // more than one P, so there are no deadlock concerns.
4862 lock(&plocal.timersLock)
4863 lock(&pp.timersLock)
4864 moveTimers(plocal, pp.timers)
4866 pp.numTimers.Store(0)
4867 pp.deletedTimers.Store(0)
4868 pp.timer0When.Store(0)
4869 unlock(&pp.timersLock)
4870 unlock(&plocal.timersLock)
4872 // Flush p's write barrier buffer.
4873 if gcphase != _GCoff {
4877 for i := range pp.sudogbuf {
4878 pp.sudogbuf[i] = nil
4880 pp.sudogcache = pp.sudogbuf[:0]
4881 for j := range pp.deferpoolbuf {
4882 pp.deferpoolbuf[j] = nil
4884 pp.deferpool = pp.deferpoolbuf[:0]
4885 systemstack(func() {
4886 for i := 0; i < pp.mspancache.len; i++ {
4887 // Safe to call since the world is stopped.
4888 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4890 pp.mspancache.len = 0
4892 pp.pcache.flush(&mheap_.pages)
4893 unlock(&mheap_.lock)
4895 freemcache(pp.mcache)
4900 if pp.timerRaceCtx != 0 {
4901 // The race detector code uses a callback to fetch
4902 // the proc context, so arrange for that callback
4903 // to see the right thing.
4904 // This hack only works because we are the only
4910 racectxend(pp.timerRaceCtx)
4915 raceprocdestroy(pp.raceprocctx)
4922 // Change number of processors.
4924 // sched.lock must be held, and the world must be stopped.
4926 // gcworkbufs must not be being modified by either the GC or the write barrier
4927 // code, so the GC must not be running if the number of Ps actually changes.
4929 // Returns list of Ps with local work, they need to be scheduled by the caller.
4930 func procresize(nprocs int32) *p {
4931 assertLockHeld(&sched.lock)
4932 assertWorldStopped()
4935 if old < 0 || nprocs <= 0 {
4936 throw("procresize: invalid arg")
4939 traceGomaxprocs(nprocs)
4942 // update statistics
4944 if sched.procresizetime != 0 {
4945 sched.totaltime += int64(old) * (now - sched.procresizetime)
4947 sched.procresizetime = now
4949 maskWords := (nprocs + 31) / 32
4951 // Grow allp if necessary.
4952 if nprocs > int32(len(allp)) {
4953 // Synchronize with retake, which could be running
4954 // concurrently since it doesn't run on a P.
4956 if nprocs <= int32(cap(allp)) {
4957 allp = allp[:nprocs]
4959 nallp := make([]*p, nprocs)
4960 // Copy everything up to allp's cap so we
4961 // never lose old allocated Ps.
4962 copy(nallp, allp[:cap(allp)])
4966 if maskWords <= int32(cap(idlepMask)) {
4967 idlepMask = idlepMask[:maskWords]
4968 timerpMask = timerpMask[:maskWords]
4970 nidlepMask := make([]uint32, maskWords)
4971 // No need to copy beyond len, old Ps are irrelevant.
4972 copy(nidlepMask, idlepMask)
4973 idlepMask = nidlepMask
4975 ntimerpMask := make([]uint32, maskWords)
4976 copy(ntimerpMask, timerpMask)
4977 timerpMask = ntimerpMask
4982 // initialize new P's
4983 for i := old; i < nprocs; i++ {
4989 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
4993 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
4994 // continue to use the current P
4995 gp.m.p.ptr().status = _Prunning
4996 gp.m.p.ptr().mcache.prepareForSweep()
4998 // release the current P and acquire allp[0].
5000 // We must do this before destroying our current P
5001 // because p.destroy itself has write barriers, so we
5002 // need to do that from a valid P.
5005 // Pretend that we were descheduled
5006 // and then scheduled again to keep
5009 traceProcStop(gp.m.p.ptr())
5023 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
5026 // release resources from unused P's
5027 for i := nprocs; i < old; i++ {
5030 // can't free P itself because it can be referenced by an M in syscall
5034 if int32(len(allp)) != nprocs {
5036 allp = allp[:nprocs]
5037 idlepMask = idlepMask[:maskWords]
5038 timerpMask = timerpMask[:maskWords]
5043 for i := nprocs - 1; i >= 0; i-- {
5045 if gp.m.p.ptr() == pp {
5053 pp.link.set(runnablePs)
5057 stealOrder.reset(uint32(nprocs))
5058 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5059 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5061 // Notify the limiter that the amount of procs has changed.
5062 gcCPULimiter.resetCapacity(now, nprocs)
5067 // Associate p and the current m.
5069 // This function is allowed to have write barriers even if the caller
5070 // isn't because it immediately acquires pp.
5072 //go:yeswritebarrierrec
5073 func acquirep(pp *p) {
5074 // Do the part that isn't allowed to have write barriers.
5077 // Have p; write barriers now allowed.
5079 // Perform deferred mcache flush before this P can allocate
5080 // from a potentially stale mcache.
5081 pp.mcache.prepareForSweep()
5088 // wirep is the first step of acquirep, which actually associates the
5089 // current M to pp. This is broken out so we can disallow write
5090 // barriers for this part, since we don't yet have a P.
5092 //go:nowritebarrierrec
5098 throw("wirep: already in go")
5100 if pp.m != 0 || pp.status != _Pidle {
5105 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5106 throw("wirep: invalid p state")
5110 pp.status = _Prunning
5113 // Disassociate p and the current m.
5114 func releasep() *p {
5118 throw("releasep: invalid arg")
5121 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5122 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5123 throw("releasep: invalid p state")
5126 traceProcStop(gp.m.p.ptr())
5134 func incidlelocked(v int32) {
5136 sched.nmidlelocked += v
5143 // Check for deadlock situation.
5144 // The check is based on number of running M's, if 0 -> deadlock.
5145 // sched.lock must be held.
5147 assertLockHeld(&sched.lock)
5149 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5150 // there are no running goroutines. The calling program is
5151 // assumed to be running.
5152 if islibrary || isarchive {
5156 // If we are dying because of a signal caught on an already idle thread,
5157 // freezetheworld will cause all running threads to block.
5158 // And runtime will essentially enter into deadlock state,
5159 // except that there is a thread that will call exit soon.
5160 if panicking.Load() > 0 {
5164 // If we are not running under cgo, but we have an extra M then account
5165 // for it. (It is possible to have an extra M on Windows without cgo to
5166 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5169 if !iscgo && cgoHasExtraM {
5170 mp := lockextra(true)
5171 haveExtraM := extraMCount > 0
5178 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5183 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5184 throw("checkdead: inconsistent counts")
5188 forEachG(func(gp *g) {
5189 if isSystemGoroutine(gp, false) {
5192 s := readgstatus(gp)
5193 switch s &^ _Gscan {
5200 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5201 throw("checkdead: runnable g")
5204 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5205 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5206 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5209 // Maybe jump time forward for playground.
5211 if when := timeSleepUntil(); when < maxWhen {
5214 // Start an M to steal the timer.
5215 pp, _ := pidleget(faketime)
5217 // There should always be a free P since
5218 // nothing is running.
5219 throw("checkdead: no p for timer")
5223 // There should always be a free M since
5224 // nothing is running.
5225 throw("checkdead: no m for timer")
5227 // M must be spinning to steal. We set this to be
5228 // explicit, but since this is the only M it would
5229 // become spinning on its own anyways.
5230 sched.nmspinning.Add(1)
5233 notewakeup(&mp.park)
5238 // There are no goroutines running, so we can look at the P's.
5239 for _, pp := range allp {
5240 if len(pp.timers) > 0 {
5245 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5246 fatal("all goroutines are asleep - deadlock!")
5249 // forcegcperiod is the maximum time in nanoseconds between garbage
5250 // collections. If we go this long without a garbage collection, one
5251 // is forced to run.
5253 // This is a variable for testing purposes. It normally doesn't change.
5254 var forcegcperiod int64 = 2 * 60 * 1e9
5256 // needSysmonWorkaround is true if the workaround for
5257 // golang.org/issue/42515 is needed on NetBSD.
5258 var needSysmonWorkaround bool = false
5260 // Always runs without a P, so write barriers are not allowed.
5262 //go:nowritebarrierrec
5269 lasttrace := int64(0)
5270 idle := 0 // how many cycles in succession we had not wokeup somebody
5274 if idle == 0 { // start with 20us sleep...
5276 } else if idle > 50 { // start doubling the sleep after 1ms...
5279 if delay > 10*1000 { // up to 10ms
5284 // sysmon should not enter deep sleep if schedtrace is enabled so that
5285 // it can print that information at the right time.
5287 // It should also not enter deep sleep if there are any active P's so
5288 // that it can retake P's from syscalls, preempt long running G's, and
5289 // poll the network if all P's are busy for long stretches.
5291 // It should wakeup from deep sleep if any P's become active either due
5292 // to exiting a syscall or waking up due to a timer expiring so that it
5293 // can resume performing those duties. If it wakes from a syscall it
5294 // resets idle and delay as a bet that since it had retaken a P from a
5295 // syscall before, it may need to do it again shortly after the
5296 // application starts work again. It does not reset idle when waking
5297 // from a timer to avoid adding system load to applications that spend
5298 // most of their time sleeping.
5300 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5302 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5303 syscallWake := false
5304 next := timeSleepUntil()
5306 sched.sysmonwait.Store(true)
5308 // Make wake-up period small enough
5309 // for the sampling to be correct.
5310 sleep := forcegcperiod / 2
5311 if next-now < sleep {
5314 shouldRelax := sleep >= osRelaxMinNS
5318 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5323 sched.sysmonwait.Store(false)
5324 noteclear(&sched.sysmonnote)
5334 lock(&sched.sysmonlock)
5335 // Update now in case we blocked on sysmonnote or spent a long time
5336 // blocked on schedlock or sysmonlock above.
5339 // trigger libc interceptors if needed
5340 if *cgo_yield != nil {
5341 asmcgocall(*cgo_yield, nil)
5343 // poll network if not polled for more than 10ms
5344 lastpoll := sched.lastpoll.Load()
5345 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5346 sched.lastpoll.CompareAndSwap(lastpoll, now)
5347 list := netpoll(0) // non-blocking - returns list of goroutines
5349 // Need to decrement number of idle locked M's
5350 // (pretending that one more is running) before injectglist.
5351 // Otherwise it can lead to the following situation:
5352 // injectglist grabs all P's but before it starts M's to run the P's,
5353 // another M returns from syscall, finishes running its G,
5354 // observes that there is no work to do and no other running M's
5355 // and reports deadlock.
5361 if GOOS == "netbsd" && needSysmonWorkaround {
5362 // netpoll is responsible for waiting for timer
5363 // expiration, so we typically don't have to worry
5364 // about starting an M to service timers. (Note that
5365 // sleep for timeSleepUntil above simply ensures sysmon
5366 // starts running again when that timer expiration may
5367 // cause Go code to run again).
5369 // However, netbsd has a kernel bug that sometimes
5370 // misses netpollBreak wake-ups, which can lead to
5371 // unbounded delays servicing timers. If we detect this
5372 // overrun, then startm to get something to handle the
5375 // See issue 42515 and
5376 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5377 if next := timeSleepUntil(); next < now {
5381 if scavenger.sysmonWake.Load() != 0 {
5382 // Kick the scavenger awake if someone requested it.
5385 // retake P's blocked in syscalls
5386 // and preempt long running G's
5387 if retake(now) != 0 {
5392 // check if we need to force a GC
5393 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5395 forcegc.idle.Store(false)
5397 list.push(forcegc.g)
5399 unlock(&forcegc.lock)
5401 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5403 schedtrace(debug.scheddetail > 0)
5405 unlock(&sched.sysmonlock)
5409 type sysmontick struct {
5416 // forcePreemptNS is the time slice given to a G before it is
5418 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5420 func retake(now int64) uint32 {
5422 // Prevent allp slice changes. This lock will be completely
5423 // uncontended unless we're already stopping the world.
5425 // We can't use a range loop over allp because we may
5426 // temporarily drop the allpLock. Hence, we need to re-fetch
5427 // allp each time around the loop.
5428 for i := 0; i < len(allp); i++ {
5431 // This can happen if procresize has grown
5432 // allp but not yet created new Ps.
5435 pd := &pp.sysmontick
5438 if s == _Prunning || s == _Psyscall {
5439 // Preempt G if it's running for too long.
5440 t := int64(pp.schedtick)
5441 if int64(pd.schedtick) != t {
5442 pd.schedtick = uint32(t)
5444 } else if pd.schedwhen+forcePreemptNS <= now {
5446 // In case of syscall, preemptone() doesn't
5447 // work, because there is no M wired to P.
5452 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5453 t := int64(pp.syscalltick)
5454 if !sysretake && int64(pd.syscalltick) != t {
5455 pd.syscalltick = uint32(t)
5456 pd.syscallwhen = now
5459 // On the one hand we don't want to retake Ps if there is no other work to do,
5460 // but on the other hand we want to retake them eventually
5461 // because they can prevent the sysmon thread from deep sleep.
5462 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5465 // Drop allpLock so we can take sched.lock.
5467 // Need to decrement number of idle locked M's
5468 // (pretending that one more is running) before the CAS.
5469 // Otherwise the M from which we retake can exit the syscall,
5470 // increment nmidle and report deadlock.
5472 if atomic.Cas(&pp.status, s, _Pidle) {
5489 // Tell all goroutines that they have been preempted and they should stop.
5490 // This function is purely best-effort. It can fail to inform a goroutine if a
5491 // processor just started running it.
5492 // No locks need to be held.
5493 // Returns true if preemption request was issued to at least one goroutine.
5494 func preemptall() bool {
5496 for _, pp := range allp {
5497 if pp.status != _Prunning {
5507 // Tell the goroutine running on processor P to stop.
5508 // This function is purely best-effort. It can incorrectly fail to inform the
5509 // goroutine. It can inform the wrong goroutine. Even if it informs the
5510 // correct goroutine, that goroutine might ignore the request if it is
5511 // simultaneously executing newstack.
5512 // No lock needs to be held.
5513 // Returns true if preemption request was issued.
5514 // The actual preemption will happen at some point in the future
5515 // and will be indicated by the gp->status no longer being
5517 func preemptone(pp *p) bool {
5519 if mp == nil || mp == getg().m {
5523 if gp == nil || gp == mp.g0 {
5529 // Every call in a goroutine checks for stack overflow by
5530 // comparing the current stack pointer to gp->stackguard0.
5531 // Setting gp->stackguard0 to StackPreempt folds
5532 // preemption into the normal stack overflow check.
5533 gp.stackguard0 = stackPreempt
5535 // Request an async preemption of this P.
5536 if preemptMSupported && debug.asyncpreemptoff == 0 {
5546 func schedtrace(detailed bool) {
5553 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)
5555 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5557 // We must be careful while reading data from P's, M's and G's.
5558 // Even if we hold schedlock, most data can be changed concurrently.
5559 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5560 for i, pp := range allp {
5562 h := atomic.Load(&pp.runqhead)
5563 t := atomic.Load(&pp.runqtail)
5565 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5571 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5573 // In non-detailed mode format lengths of per-P run queues as:
5574 // [len1 len2 len3 len4]
5580 if i == len(allp)-1 {
5591 for mp := allm; mp != nil; mp = mp.alllink {
5593 print(" M", mp.id, ": p=")
5605 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5606 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5614 forEachG(func(gp *g) {
5615 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5622 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5632 // schedEnableUser enables or disables the scheduling of user
5635 // This does not stop already running user goroutines, so the caller
5636 // should first stop the world when disabling user goroutines.
5637 func schedEnableUser(enable bool) {
5639 if sched.disable.user == !enable {
5643 sched.disable.user = !enable
5645 n := sched.disable.n
5647 globrunqputbatch(&sched.disable.runnable, n)
5649 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5657 // schedEnabled reports whether gp should be scheduled. It returns
5658 // false is scheduling of gp is disabled.
5660 // sched.lock must be held.
5661 func schedEnabled(gp *g) bool {
5662 assertLockHeld(&sched.lock)
5664 if sched.disable.user {
5665 return isSystemGoroutine(gp, true)
5670 // Put mp on midle list.
5671 // sched.lock must be held.
5672 // May run during STW, so write barriers are not allowed.
5674 //go:nowritebarrierrec
5676 assertLockHeld(&sched.lock)
5678 mp.schedlink = sched.midle
5684 // Try to get an m from midle list.
5685 // sched.lock must be held.
5686 // May run during STW, so write barriers are not allowed.
5688 //go:nowritebarrierrec
5690 assertLockHeld(&sched.lock)
5692 mp := sched.midle.ptr()
5694 sched.midle = mp.schedlink
5700 // Put gp on the global runnable queue.
5701 // sched.lock must be held.
5702 // May run during STW, so write barriers are not allowed.
5704 //go:nowritebarrierrec
5705 func globrunqput(gp *g) {
5706 assertLockHeld(&sched.lock)
5708 sched.runq.pushBack(gp)
5712 // Put gp at the head of the global runnable queue.
5713 // sched.lock must be held.
5714 // May run during STW, so write barriers are not allowed.
5716 //go:nowritebarrierrec
5717 func globrunqputhead(gp *g) {
5718 assertLockHeld(&sched.lock)
5724 // Put a batch of runnable goroutines on the global runnable queue.
5725 // This clears *batch.
5726 // sched.lock must be held.
5727 // May run during STW, so write barriers are not allowed.
5729 //go:nowritebarrierrec
5730 func globrunqputbatch(batch *gQueue, n int32) {
5731 assertLockHeld(&sched.lock)
5733 sched.runq.pushBackAll(*batch)
5738 // Try get a batch of G's from the global runnable queue.
5739 // sched.lock must be held.
5740 func globrunqget(pp *p, max int32) *g {
5741 assertLockHeld(&sched.lock)
5743 if sched.runqsize == 0 {
5747 n := sched.runqsize/gomaxprocs + 1
5748 if n > sched.runqsize {
5751 if max > 0 && n > max {
5754 if n > int32(len(pp.runq))/2 {
5755 n = int32(len(pp.runq)) / 2
5760 gp := sched.runq.pop()
5763 gp1 := sched.runq.pop()
5764 runqput(pp, gp1, false)
5769 // pMask is an atomic bitstring with one bit per P.
5772 // read returns true if P id's bit is set.
5773 func (p pMask) read(id uint32) bool {
5775 mask := uint32(1) << (id % 32)
5776 return (atomic.Load(&p[word]) & mask) != 0
5779 // set sets P id's bit.
5780 func (p pMask) set(id int32) {
5782 mask := uint32(1) << (id % 32)
5783 atomic.Or(&p[word], mask)
5786 // clear clears P id's bit.
5787 func (p pMask) clear(id int32) {
5789 mask := uint32(1) << (id % 32)
5790 atomic.And(&p[word], ^mask)
5793 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5795 // Ideally, the timer mask would be kept immediately consistent on any timer
5796 // operations. Unfortunately, updating a shared global data structure in the
5797 // timer hot path adds too much overhead in applications frequently switching
5798 // between no timers and some timers.
5800 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5801 // running P (returned by pidleget) may add a timer at any time, so its mask
5802 // must be set. An idle P (passed to pidleput) cannot add new timers while
5803 // idle, so if it has no timers at that time, its mask may be cleared.
5805 // Thus, we get the following effects on timer-stealing in findrunnable:
5807 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5808 // (for work- or timer-stealing; this is the ideal case).
5809 // - Running Ps must always be checked.
5810 // - Idle Ps whose timers are stolen must continue to be checked until they run
5811 // again, even after timer expiration.
5813 // When the P starts running again, the mask should be set, as a timer may be
5814 // added at any time.
5816 // TODO(prattmic): Additional targeted updates may improve the above cases.
5817 // e.g., updating the mask when stealing a timer.
5818 func updateTimerPMask(pp *p) {
5819 if pp.numTimers.Load() > 0 {
5823 // Looks like there are no timers, however another P may transiently
5824 // decrement numTimers when handling a timerModified timer in
5825 // checkTimers. We must take timersLock to serialize with these changes.
5826 lock(&pp.timersLock)
5827 if pp.numTimers.Load() == 0 {
5828 timerpMask.clear(pp.id)
5830 unlock(&pp.timersLock)
5833 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5834 // to nanotime or zero. Returns now or the current time if now was zero.
5836 // This releases ownership of p. Once sched.lock is released it is no longer
5839 // sched.lock must be held.
5841 // May run during STW, so write barriers are not allowed.
5843 //go:nowritebarrierrec
5844 func pidleput(pp *p, now int64) int64 {
5845 assertLockHeld(&sched.lock)
5848 throw("pidleput: P has non-empty run queue")
5853 updateTimerPMask(pp) // clear if there are no timers.
5854 idlepMask.set(pp.id)
5855 pp.link = sched.pidle
5858 if !pp.limiterEvent.start(limiterEventIdle, now) {
5859 throw("must be able to track idle limiter event")
5864 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5866 // sched.lock must be held.
5868 // May run during STW, so write barriers are not allowed.
5870 //go:nowritebarrierrec
5871 func pidleget(now int64) (*p, int64) {
5872 assertLockHeld(&sched.lock)
5874 pp := sched.pidle.ptr()
5876 // Timer may get added at any time now.
5880 timerpMask.set(pp.id)
5881 idlepMask.clear(pp.id)
5882 sched.pidle = pp.link
5883 sched.npidle.Add(-1)
5884 pp.limiterEvent.stop(limiterEventIdle, now)
5889 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5890 // This is called by spinning Ms (or callers than need a spinning M) that have
5891 // found work. If no P is available, this must synchronized with non-spinning
5892 // Ms that may be preparing to drop their P without discovering this work.
5894 // sched.lock must be held.
5896 // May run during STW, so write barriers are not allowed.
5898 //go:nowritebarrierrec
5899 func pidlegetSpinning(now int64) (*p, int64) {
5900 assertLockHeld(&sched.lock)
5902 pp, now := pidleget(now)
5904 // See "Delicate dance" comment in findrunnable. We found work
5905 // that we cannot take, we must synchronize with non-spinning
5906 // Ms that may be preparing to drop their P.
5907 sched.needspinning.Store(1)
5914 // runqempty reports whether pp has no Gs on its local run queue.
5915 // It never returns true spuriously.
5916 func runqempty(pp *p) bool {
5917 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5918 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5919 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5920 // does not mean the queue is empty.
5922 head := atomic.Load(&pp.runqhead)
5923 tail := atomic.Load(&pp.runqtail)
5924 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5925 if tail == atomic.Load(&pp.runqtail) {
5926 return head == tail && runnext == 0
5931 // To shake out latent assumptions about scheduling order,
5932 // we introduce some randomness into scheduling decisions
5933 // when running with the race detector.
5934 // The need for this was made obvious by changing the
5935 // (deterministic) scheduling order in Go 1.5 and breaking
5936 // many poorly-written tests.
5937 // With the randomness here, as long as the tests pass
5938 // consistently with -race, they shouldn't have latent scheduling
5940 const randomizeScheduler = raceenabled
5942 // runqput tries to put g on the local runnable queue.
5943 // If next is false, runqput adds g to the tail of the runnable queue.
5944 // If next is true, runqput puts g in the pp.runnext slot.
5945 // If the run queue is full, runnext puts g on the global queue.
5946 // Executed only by the owner P.
5947 func runqput(pp *p, gp *g, next bool) {
5948 if randomizeScheduler && next && fastrandn(2) == 0 {
5954 oldnext := pp.runnext
5955 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
5961 // Kick the old runnext out to the regular run queue.
5966 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
5968 if t-h < uint32(len(pp.runq)) {
5969 pp.runq[t%uint32(len(pp.runq))].set(gp)
5970 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
5973 if runqputslow(pp, gp, h, t) {
5976 // the queue is not full, now the put above must succeed
5980 // Put g and a batch of work from local runnable queue on global queue.
5981 // Executed only by the owner P.
5982 func runqputslow(pp *p, gp *g, h, t uint32) bool {
5983 var batch [len(pp.runq)/2 + 1]*g
5985 // First, grab a batch from local queue.
5988 if n != uint32(len(pp.runq)/2) {
5989 throw("runqputslow: queue is not full")
5991 for i := uint32(0); i < n; i++ {
5992 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
5994 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
5999 if randomizeScheduler {
6000 for i := uint32(1); i <= n; i++ {
6001 j := fastrandn(i + 1)
6002 batch[i], batch[j] = batch[j], batch[i]
6006 // Link the goroutines.
6007 for i := uint32(0); i < n; i++ {
6008 batch[i].schedlink.set(batch[i+1])
6011 q.head.set(batch[0])
6012 q.tail.set(batch[n])
6014 // Now put the batch on global queue.
6016 globrunqputbatch(&q, int32(n+1))
6021 // runqputbatch tries to put all the G's on q on the local runnable queue.
6022 // If the queue is full, they are put on the global queue; in that case
6023 // this will temporarily acquire the scheduler lock.
6024 // Executed only by the owner P.
6025 func runqputbatch(pp *p, q *gQueue, qsize int) {
6026 h := atomic.LoadAcq(&pp.runqhead)
6029 for !q.empty() && t-h < uint32(len(pp.runq)) {
6031 pp.runq[t%uint32(len(pp.runq))].set(gp)
6037 if randomizeScheduler {
6038 off := func(o uint32) uint32 {
6039 return (pp.runqtail + o) % uint32(len(pp.runq))
6041 for i := uint32(1); i < n; i++ {
6042 j := fastrandn(i + 1)
6043 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6047 atomic.StoreRel(&pp.runqtail, t)
6050 globrunqputbatch(q, int32(qsize))
6055 // Get g from local runnable queue.
6056 // If inheritTime is true, gp should inherit the remaining time in the
6057 // current time slice. Otherwise, it should start a new time slice.
6058 // Executed only by the owner P.
6059 func runqget(pp *p) (gp *g, inheritTime bool) {
6060 // If there's a runnext, it's the next G to run.
6062 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6063 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6064 // Hence, there's no need to retry this CAS if it fails.
6065 if next != 0 && pp.runnext.cas(next, 0) {
6066 return next.ptr(), true
6070 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6075 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6076 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6082 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6083 // Executed only by the owner P.
6084 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6085 oldNext := pp.runnext
6086 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6087 drainQ.pushBack(oldNext.ptr())
6092 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6098 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6102 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6106 // We've inverted the order in which it gets G's from the local P's runnable queue
6107 // and then advances the head pointer because we don't want to mess up the statuses of G's
6108 // while runqdrain() and runqsteal() are running in parallel.
6109 // Thus we should advance the head pointer before draining the local P into a gQueue,
6110 // so that we can update any gp.schedlink only after we take the full ownership of G,
6111 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6112 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6113 for i := uint32(0); i < qn; i++ {
6114 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6121 // Grabs a batch of goroutines from pp's runnable queue into batch.
6122 // Batch is a ring buffer starting at batchHead.
6123 // Returns number of grabbed goroutines.
6124 // Can be executed by any P.
6125 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6127 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6128 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6133 // Try to steal from pp.runnext.
6134 if next := pp.runnext; next != 0 {
6135 if pp.status == _Prunning {
6136 // Sleep to ensure that pp isn't about to run the g
6137 // we are about to steal.
6138 // The important use case here is when the g running
6139 // on pp ready()s another g and then almost
6140 // immediately blocks. Instead of stealing runnext
6141 // in this window, back off to give pp a chance to
6142 // schedule runnext. This will avoid thrashing gs
6143 // between different Ps.
6144 // A sync chan send/recv takes ~50ns as of time of
6145 // writing, so 3us gives ~50x overshoot.
6146 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6149 // On some platforms system timer granularity is
6150 // 1-15ms, which is way too much for this
6151 // optimization. So just yield.
6155 if !pp.runnext.cas(next, 0) {
6158 batch[batchHead%uint32(len(batch))] = next
6164 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6167 for i := uint32(0); i < n; i++ {
6168 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6169 batch[(batchHead+i)%uint32(len(batch))] = g
6171 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6177 // Steal half of elements from local runnable queue of p2
6178 // and put onto local runnable queue of p.
6179 // Returns one of the stolen elements (or nil if failed).
6180 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6182 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6187 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6191 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6192 if t-h+n >= uint32(len(pp.runq)) {
6193 throw("runqsteal: runq overflow")
6195 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6199 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6200 // be on one gQueue or gList at a time.
6201 type gQueue struct {
6206 // empty reports whether q is empty.
6207 func (q *gQueue) empty() bool {
6211 // push adds gp to the head of q.
6212 func (q *gQueue) push(gp *g) {
6213 gp.schedlink = q.head
6220 // pushBack adds gp to the tail of q.
6221 func (q *gQueue) pushBack(gp *g) {
6224 q.tail.ptr().schedlink.set(gp)
6231 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6233 func (q *gQueue) pushBackAll(q2 gQueue) {
6237 q2.tail.ptr().schedlink = 0
6239 q.tail.ptr().schedlink = q2.head
6246 // pop removes and returns the head of queue q. It returns nil if
6248 func (q *gQueue) pop() *g {
6251 q.head = gp.schedlink
6259 // popList takes all Gs in q and returns them as a gList.
6260 func (q *gQueue) popList() gList {
6261 stack := gList{q.head}
6266 // A gList is a list of Gs linked through g.schedlink. A G can only be
6267 // on one gQueue or gList at a time.
6272 // empty reports whether l is empty.
6273 func (l *gList) empty() bool {
6277 // push adds gp to the head of l.
6278 func (l *gList) push(gp *g) {
6279 gp.schedlink = l.head
6283 // pushAll prepends all Gs in q to l.
6284 func (l *gList) pushAll(q gQueue) {
6286 q.tail.ptr().schedlink = l.head
6291 // pop removes and returns the head of l. If l is empty, it returns nil.
6292 func (l *gList) pop() *g {
6295 l.head = gp.schedlink
6300 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6301 func setMaxThreads(in int) (out int) {
6303 out = int(sched.maxmcount)
6304 if in > 0x7fffffff { // MaxInt32
6305 sched.maxmcount = 0x7fffffff
6307 sched.maxmcount = int32(in)
6315 func procPin() int {
6320 return int(mp.p.ptr().id)
6329 //go:linkname sync_runtime_procPin sync.runtime_procPin
6331 func sync_runtime_procPin() int {
6335 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6337 func sync_runtime_procUnpin() {
6341 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6343 func sync_atomic_runtime_procPin() int {
6347 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6349 func sync_atomic_runtime_procUnpin() {
6353 // Active spinning for sync.Mutex.
6355 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6357 func sync_runtime_canSpin(i int) bool {
6358 // sync.Mutex is cooperative, so we are conservative with spinning.
6359 // Spin only few times and only if running on a multicore machine and
6360 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6361 // As opposed to runtime mutex we don't do passive spinning here,
6362 // because there can be work on global runq or on other Ps.
6363 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6366 if p := getg().m.p.ptr(); !runqempty(p) {
6372 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6374 func sync_runtime_doSpin() {
6375 procyield(active_spin_cnt)
6378 var stealOrder randomOrder
6380 // randomOrder/randomEnum are helper types for randomized work stealing.
6381 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6382 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6383 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6384 type randomOrder struct {
6389 type randomEnum struct {
6396 func (ord *randomOrder) reset(count uint32) {
6398 ord.coprimes = ord.coprimes[:0]
6399 for i := uint32(1); i <= count; i++ {
6400 if gcd(i, count) == 1 {
6401 ord.coprimes = append(ord.coprimes, i)
6406 func (ord *randomOrder) start(i uint32) randomEnum {
6410 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6414 func (enum *randomEnum) done() bool {
6415 return enum.i == enum.count
6418 func (enum *randomEnum) next() {
6420 enum.pos = (enum.pos + enum.inc) % enum.count
6423 func (enum *randomEnum) position() uint32 {
6427 func gcd(a, b uint32) uint32 {
6434 // An initTask represents the set of initializations that need to be done for a package.
6435 // Keep in sync with ../../test/initempty.go:initTask
6436 type initTask struct {
6437 // TODO: pack the first 3 fields more tightly?
6438 state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
6441 // followed by ndeps instances of an *initTask, one per package depended on
6442 // followed by nfns pcs, one per init function to run
6445 // inittrace stores statistics for init functions which are
6446 // updated by malloc and newproc when active is true.
6447 var inittrace tracestat
6449 type tracestat struct {
6450 active bool // init tracing activation status
6451 id uint64 // init goroutine id
6452 allocs uint64 // heap allocations
6453 bytes uint64 // heap allocated bytes
6456 func doInit(t *initTask) {
6458 case 2: // fully initialized
6460 case 1: // initialization in progress
6461 throw("recursive call during initialization - linker skew")
6462 default: // not initialized yet
6463 t.state = 1 // initialization in progress
6465 for i := uintptr(0); i < t.ndeps; i++ {
6466 p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
6467 t2 := *(**initTask)(p)
6472 t.state = 2 // initialization done
6481 if inittrace.active {
6483 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6487 firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
6488 for i := uintptr(0); i < t.nfns; i++ {
6489 p := add(firstFunc, i*goarch.PtrSize)
6490 f := *(*func())(unsafe.Pointer(&p))
6494 if inittrace.active {
6496 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6499 f := *(*func())(unsafe.Pointer(&firstFunc))
6500 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6503 print("init ", pkg, " @")
6504 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6505 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6506 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6507 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6511 t.state = 2 // initialization done