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
255 // Make racy client program work: if panicking on
256 // another goroutine at the same time as main returns,
257 // let the other goroutine finish printing the panic trace.
258 // Once it does, it will exit. See issues 3934 and 20018.
259 if runningPanicDefers.Load() != 0 {
260 // Running deferred functions should not take long.
261 for c := 0; c < 1000; c++ {
262 if runningPanicDefers.Load() == 0 {
268 if panicking.Load() != 0 {
269 gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
279 // os_beforeExit is called from os.Exit(0).
281 //go:linkname os_beforeExit os.runtime_beforeExit
282 func os_beforeExit() {
288 // start forcegc helper goroutine
293 func forcegchelper() {
295 lockInit(&forcegc.lock, lockRankForcegc)
298 if forcegc.idle.Load() {
299 throw("forcegc: phase error")
301 forcegc.idle.Store(true)
302 goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
303 // this goroutine is explicitly resumed by sysmon
304 if debug.gctrace > 0 {
307 // Time-triggered, fully concurrent.
308 gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
314 // Gosched yields the processor, allowing other goroutines to run. It does not
315 // suspend the current goroutine, so execution resumes automatically.
321 // goschedguarded yields the processor like gosched, but also checks
322 // for forbidden states and opts out of the yield in those cases.
325 func goschedguarded() {
326 mcall(goschedguarded_m)
329 // Puts the current goroutine into a waiting state and calls unlockf on the
332 // If unlockf returns false, the goroutine is resumed.
334 // unlockf must not access this G's stack, as it may be moved between
335 // the call to gopark and the call to unlockf.
337 // Note that because unlockf is called after putting the G into a waiting
338 // state, the G may have already been readied by the time unlockf is called
339 // unless there is external synchronization preventing the G from being
340 // readied. If unlockf returns false, it must guarantee that the G cannot be
341 // externally readied.
343 // Reason explains why the goroutine has been parked. It is displayed in stack
344 // traces and heap dumps. Reasons should be unique and descriptive. Do not
345 // re-use reasons, add new ones.
346 func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
347 if reason != waitReasonSleep {
348 checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
352 status := readgstatus(gp)
353 if status != _Grunning && status != _Gscanrunning {
354 throw("gopark: bad g status")
357 mp.waitunlockf = unlockf
358 gp.waitreason = reason
359 mp.waittraceev = traceEv
360 mp.waittraceskip = traceskip
362 // can't do anything that might move the G between Ms here.
366 // Puts the current goroutine into a waiting state and unlocks the lock.
367 // The goroutine can be made runnable again by calling goready(gp).
368 func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
369 gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
372 func goready(gp *g, traceskip int) {
374 ready(gp, traceskip, true)
379 func acquireSudog() *sudog {
380 // Delicate dance: the semaphore implementation calls
381 // acquireSudog, acquireSudog calls new(sudog),
382 // new calls malloc, malloc can call the garbage collector,
383 // and the garbage collector calls the semaphore implementation
385 // Break the cycle by doing acquirem/releasem around new(sudog).
386 // The acquirem/releasem increments m.locks during new(sudog),
387 // which keeps the garbage collector from being invoked.
390 if len(pp.sudogcache) == 0 {
391 lock(&sched.sudoglock)
392 // First, try to grab a batch from central cache.
393 for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
394 s := sched.sudogcache
395 sched.sudogcache = s.next
397 pp.sudogcache = append(pp.sudogcache, s)
399 unlock(&sched.sudoglock)
400 // If the central cache is empty, allocate a new one.
401 if len(pp.sudogcache) == 0 {
402 pp.sudogcache = append(pp.sudogcache, new(sudog))
405 n := len(pp.sudogcache)
406 s := pp.sudogcache[n-1]
407 pp.sudogcache[n-1] = nil
408 pp.sudogcache = pp.sudogcache[:n-1]
410 throw("acquireSudog: found s.elem != nil in cache")
417 func releaseSudog(s *sudog) {
419 throw("runtime: sudog with non-nil elem")
422 throw("runtime: sudog with non-false isSelect")
425 throw("runtime: sudog with non-nil next")
428 throw("runtime: sudog with non-nil prev")
430 if s.waitlink != nil {
431 throw("runtime: sudog with non-nil waitlink")
434 throw("runtime: sudog with non-nil c")
438 throw("runtime: releaseSudog with non-nil gp.param")
440 mp := acquirem() // avoid rescheduling to another P
442 if len(pp.sudogcache) == cap(pp.sudogcache) {
443 // Transfer half of local cache to the central cache.
444 var first, last *sudog
445 for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
446 n := len(pp.sudogcache)
447 p := pp.sudogcache[n-1]
448 pp.sudogcache[n-1] = nil
449 pp.sudogcache = pp.sudogcache[:n-1]
457 lock(&sched.sudoglock)
458 last.next = sched.sudogcache
459 sched.sudogcache = first
460 unlock(&sched.sudoglock)
462 pp.sudogcache = append(pp.sudogcache, s)
466 // called from assembly
467 func badmcall(fn func(*g)) {
468 throw("runtime: mcall called on m->g0 stack")
471 func badmcall2(fn func(*g)) {
472 throw("runtime: mcall function returned")
475 func badreflectcall() {
476 panic(plainError("arg size to reflect.call more than 1GB"))
479 var badmorestackg0Msg = "fatal: morestack on g0\n"
482 //go:nowritebarrierrec
483 func badmorestackg0() {
484 sp := stringStructOf(&badmorestackg0Msg)
485 write(2, sp.str, int32(sp.len))
488 var badmorestackgsignalMsg = "fatal: morestack on gsignal\n"
491 //go:nowritebarrierrec
492 func badmorestackgsignal() {
493 sp := stringStructOf(&badmorestackgsignalMsg)
494 write(2, sp.str, int32(sp.len))
502 func lockedOSThread() bool {
504 return gp.lockedm != 0 && gp.m.lockedg != 0
508 // allgs contains all Gs ever created (including dead Gs), and thus
511 // Access via the slice is protected by allglock or stop-the-world.
512 // Readers that cannot take the lock may (carefully!) use the atomic
517 // allglen and allgptr are atomic variables that contain len(allgs) and
518 // &allgs[0] respectively. Proper ordering depends on totally-ordered
519 // loads and stores. Writes are protected by allglock.
521 // allgptr is updated before allglen. Readers should read allglen
522 // before allgptr to ensure that allglen is always <= len(allgptr). New
523 // Gs appended during the race can be missed. For a consistent view of
524 // all Gs, allglock must be held.
526 // allgptr copies should always be stored as a concrete type or
527 // unsafe.Pointer, not uintptr, to ensure that GC can still reach it
528 // even if it points to a stale array.
533 func allgadd(gp *g) {
534 if readgstatus(gp) == _Gidle {
535 throw("allgadd: bad status Gidle")
539 allgs = append(allgs, gp)
540 if &allgs[0] != allgptr {
541 atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
543 atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
547 // allGsSnapshot returns a snapshot of the slice of all Gs.
549 // The world must be stopped or allglock must be held.
550 func allGsSnapshot() []*g {
551 assertWorldStoppedOrLockHeld(&allglock)
553 // Because the world is stopped or allglock is held, allgadd
554 // cannot happen concurrently with this. allgs grows
555 // monotonically and existing entries never change, so we can
556 // simply return a copy of the slice header. For added safety,
557 // we trim everything past len because that can still change.
558 return allgs[:len(allgs):len(allgs)]
561 // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
562 func atomicAllG() (**g, uintptr) {
563 length := atomic.Loaduintptr(&allglen)
564 ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
568 // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
569 func atomicAllGIndex(ptr **g, i uintptr) *g {
570 return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
573 // forEachG calls fn on every G from allgs.
575 // forEachG takes a lock to exclude concurrent addition of new Gs.
576 func forEachG(fn func(gp *g)) {
578 for _, gp := range allgs {
584 // forEachGRace calls fn on every G from allgs.
586 // forEachGRace avoids locking, but does not exclude addition of new Gs during
587 // execution, which may be missed.
588 func forEachGRace(fn func(gp *g)) {
589 ptr, length := atomicAllG()
590 for i := uintptr(0); i < length; i++ {
591 gp := atomicAllGIndex(ptr, i)
598 // Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
599 // 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
603 // cpuinit extracts the environment variable GODEBUG from the environment on
604 // Unix-like operating systems and calls internal/cpu.Initialize.
606 const prefix = "GODEBUG="
610 case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
611 cpu.DebugOptions = true
613 // Similar to goenv_unix but extracts the environment value for
615 // TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
617 for argv_index(argv, argc+1+n) != nil {
621 for i := int32(0); i < n; i++ {
622 p := argv_index(argv, argc+1+i)
623 s := unsafe.String(p, findnull(p))
625 if hasPrefix(s, prefix) {
626 env = gostring(p)[len(prefix):]
634 // Support cpu feature variables are used in code generated by the compiler
635 // to guard execution of instructions that can not be assumed to be always supported.
638 x86HasPOPCNT = cpu.X86.HasPOPCNT
639 x86HasSSE41 = cpu.X86.HasSSE41
640 x86HasFMA = cpu.X86.HasFMA
643 armHasVFPv4 = cpu.ARM.HasVFPv4
646 arm64HasATOMICS = cpu.ARM64.HasATOMICS
650 // The bootstrap sequence is:
654 // make & queue new G
655 // call runtime·mstart
657 // The new G calls runtime·main.
659 lockInit(&sched.lock, lockRankSched)
660 lockInit(&sched.sysmonlock, lockRankSysmon)
661 lockInit(&sched.deferlock, lockRankDefer)
662 lockInit(&sched.sudoglock, lockRankSudog)
663 lockInit(&deadlock, lockRankDeadlock)
664 lockInit(&paniclk, lockRankPanic)
665 lockInit(&allglock, lockRankAllg)
666 lockInit(&allpLock, lockRankAllp)
667 lockInit(&reflectOffs.lock, lockRankReflectOffs)
668 lockInit(&finlock, lockRankFin)
669 lockInit(&trace.bufLock, lockRankTraceBuf)
670 lockInit(&trace.stringsLock, lockRankTraceStrings)
671 lockInit(&trace.lock, lockRankTrace)
672 lockInit(&cpuprof.lock, lockRankCpuprof)
673 lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
674 // Enforce that this lock is always a leaf lock.
675 // All of this lock's critical sections should be
677 lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
679 // raceinit must be the first call to race detector.
680 // In particular, it must be done before mallocinit below calls racemapshadow.
683 gp.racectx, raceprocctx0 = raceinit()
686 sched.maxmcount = 10000
688 // The world starts stopped.
694 cpuinit() // must run before alginit
695 alginit() // maps, hash, fastrand must not be used before this call
696 fastrandinit() // must run before mcommoninit
697 mcommoninit(gp.m, -1)
698 modulesinit() // provides activeModules
699 typelinksinit() // uses maps, activeModules
700 itabsinit() // uses activeModules
701 stkobjinit() // must run before GC starts
703 sigsave(&gp.m.sigmask)
704 initSigmask = gp.m.sigmask
712 sched.lastpoll.Store(nanotime())
714 if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
717 if procresize(procs) != nil {
718 throw("unknown runnable goroutine during bootstrap")
722 // World is effectively started now, as P's can run.
725 // For cgocheck > 1, we turn on the write barrier at all times
726 // and check all pointer writes. We can't do this until after
727 // procresize because the write barrier needs a P.
728 if debug.cgocheck > 1 {
729 writeBarrier.cgo = true
730 writeBarrier.enabled = true
731 for _, pp := range allp {
736 if buildVersion == "" {
737 // Condition should never trigger. This code just serves
738 // to ensure runtime·buildVersion is kept in the resulting binary.
739 buildVersion = "unknown"
741 if len(modinfo) == 1 {
742 // Condition should never trigger. This code just serves
743 // to ensure runtime·modinfo is kept in the resulting binary.
748 func dumpgstatus(gp *g) {
750 print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
751 print("runtime: getg: g=", thisg, ", goid=", thisg.goid, ", g->atomicstatus=", readgstatus(thisg), "\n")
754 // sched.lock must be held.
756 assertLockHeld(&sched.lock)
758 if mcount() > sched.maxmcount {
759 print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
760 throw("thread exhaustion")
764 // mReserveID returns the next ID to use for a new m. This new m is immediately
765 // considered 'running' by checkdead.
767 // sched.lock must be held.
768 func mReserveID() int64 {
769 assertLockHeld(&sched.lock)
771 if sched.mnext+1 < sched.mnext {
772 throw("runtime: thread ID overflow")
780 // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
781 func mcommoninit(mp *m, id int64) {
784 // g0 stack won't make sense for user (and is not necessary unwindable).
786 callers(1, mp.createstack[:])
797 lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
798 hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
802 // Same behavior as for 1.17.
803 // TODO: Simplify ths.
804 if goarch.BigEndian {
805 mp.fastrand = uint64(lo)<<32 | uint64(hi)
807 mp.fastrand = uint64(hi)<<32 | uint64(lo)
811 if mp.gsignal != nil {
812 mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
815 // Add to allm so garbage collector doesn't free g->m
816 // when it is just in a register or thread-local storage.
819 // NumCgoCall() iterates over allm w/o schedlock,
820 // so we need to publish it safely.
821 atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
824 // Allocate memory to hold a cgo traceback if the cgo call crashes.
825 if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
826 mp.cgoCallers = new(cgoCallers)
830 func (mp *m) becomeSpinning() {
832 sched.nmspinning.Add(1)
833 sched.needspinning.Store(0)
836 var fastrandseed uintptr
838 func fastrandinit() {
839 s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
843 // Mark gp ready to run.
844 func ready(gp *g, traceskip int, next bool) {
846 traceGoUnpark(gp, traceskip)
849 status := readgstatus(gp)
852 mp := acquirem() // disable preemption because it can be holding p in a local var
853 if status&^_Gscan != _Gwaiting {
855 throw("bad g->status in ready")
858 // status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
859 casgstatus(gp, _Gwaiting, _Grunnable)
860 runqput(mp.p.ptr(), gp, next)
865 // freezeStopWait is a large value that freezetheworld sets
866 // sched.stopwait to in order to request that all Gs permanently stop.
867 const freezeStopWait = 0x7fffffff
869 // freezing is set to non-zero if the runtime is trying to freeze the
871 var freezing atomic.Bool
873 // Similar to stopTheWorld but best-effort and can be called several times.
874 // There is no reverse operation, used during crashing.
875 // This function must not lock any mutexes.
876 func freezetheworld() {
878 // stopwait and preemption requests can be lost
879 // due to races with concurrently executing threads,
880 // so try several times
881 for i := 0; i < 5; i++ {
882 // this should tell the scheduler to not start any new goroutines
883 sched.stopwait = freezeStopWait
884 sched.gcwaiting.Store(true)
885 // this should stop running goroutines
887 break // no running goroutines
897 // All reads and writes of g's status go through readgstatus, casgstatus
898 // castogscanstatus, casfrom_Gscanstatus.
901 func readgstatus(gp *g) uint32 {
902 return gp.atomicstatus.Load()
905 // The Gscanstatuses are acting like locks and this releases them.
906 // If it proves to be a performance hit we should be able to make these
907 // simple atomic stores but for now we are going to throw if
908 // we see an inconsistent state.
909 func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
912 // Check that transition is valid.
915 print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
917 throw("casfrom_Gscanstatus:top gp->status is not in scan state")
923 if newval == oldval&^_Gscan {
924 success = gp.atomicstatus.CompareAndSwap(oldval, newval)
928 print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
930 throw("casfrom_Gscanstatus: gp->status is not in scan state")
932 releaseLockRank(lockRankGscan)
935 // This will return false if the gp is not in the expected status and the cas fails.
936 // This acts like a lock acquire while the casfromgstatus acts like a lock release.
937 func castogscanstatus(gp *g, oldval, newval uint32) bool {
943 if newval == oldval|_Gscan {
944 r := gp.atomicstatus.CompareAndSwap(oldval, newval)
946 acquireLockRank(lockRankGscan)
952 print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
953 throw("castogscanstatus")
957 // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
958 // various latencies on every transition instead of sampling them.
959 var casgstatusAlwaysTrack = false
961 // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
962 // and casfrom_Gscanstatus instead.
963 // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
964 // put it in the Gscan state is finished.
967 func casgstatus(gp *g, oldval, newval uint32) {
968 if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
970 print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
971 throw("casgstatus: bad incoming values")
975 acquireLockRank(lockRankGscan)
976 releaseLockRank(lockRankGscan)
978 // See https://golang.org/cl/21503 for justification of the yield delay.
979 const yieldDelay = 5 * 1000
982 // loop if gp->atomicstatus is in a scan state giving
983 // GC time to finish and change the state to oldval.
984 for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
985 if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
986 throw("casgstatus: waiting for Gwaiting but is Grunnable")
989 nextYield = nanotime() + yieldDelay
991 if nanotime() < nextYield {
992 for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
997 nextYield = nanotime() + yieldDelay/2
1001 if oldval == _Grunning {
1002 // Track every gTrackingPeriod time a goroutine transitions out of running.
1003 if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
1012 // Handle various kinds of tracking.
1015 // - Time spent in runnable.
1016 // - Time spent blocked on a sync.Mutex or sync.RWMutex.
1019 // We transitioned out of runnable, so measure how much
1020 // time we spent in this state and add it to
1023 gp.runnableTime += now - gp.trackingStamp
1024 gp.trackingStamp = 0
1026 if !gp.waitreason.isMutexWait() {
1027 // Not blocking on a lock.
1030 // Blocking on a lock, measure it. Note that because we're
1031 // sampling, we have to multiply by our sampling period to get
1032 // a more representative estimate of the absolute value.
1033 // gTrackingPeriod also represents an accurate sampling period
1034 // because we can only enter this state from _Grunning.
1036 sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
1037 gp.trackingStamp = 0
1041 if !gp.waitreason.isMutexWait() {
1042 // Not blocking on a lock.
1045 // Blocking on a lock. Write down the timestamp.
1047 gp.trackingStamp = now
1049 // We just transitioned into runnable, so record what
1050 // time that happened.
1052 gp.trackingStamp = now
1054 // We're transitioning into running, so turn off
1055 // tracking and record how much time we spent in
1058 sched.timeToRun.record(gp.runnableTime)
1063 // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
1065 // Use this over casgstatus when possible to ensure that a waitreason is set.
1066 func casGToWaiting(gp *g, old uint32, reason waitReason) {
1067 // Set the wait reason before calling casgstatus, because casgstatus will use it.
1068 gp.waitreason = reason
1069 casgstatus(gp, old, _Gwaiting)
1072 // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
1073 // Returns old status. Cannot call casgstatus directly, because we are racing with an
1074 // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
1075 // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
1076 // it would loop waiting for the status to go back to Gwaiting, which it never will.
1079 func casgcopystack(gp *g) uint32 {
1081 oldstatus := readgstatus(gp) &^ _Gscan
1082 if oldstatus != _Gwaiting && oldstatus != _Grunnable {
1083 throw("copystack: bad status, not Gwaiting or Grunnable")
1085 if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
1091 // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
1093 // TODO(austin): This is the only status operation that both changes
1094 // the status and locks the _Gscan bit. Rethink this.
1095 func casGToPreemptScan(gp *g, old, new uint32) {
1096 if old != _Grunning || new != _Gscan|_Gpreempted {
1097 throw("bad g transition")
1099 acquireLockRank(lockRankGscan)
1100 for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
1104 // casGFromPreempted attempts to transition gp from _Gpreempted to
1105 // _Gwaiting. If successful, the caller is responsible for
1106 // re-scheduling gp.
1107 func casGFromPreempted(gp *g, old, new uint32) bool {
1108 if old != _Gpreempted || new != _Gwaiting {
1109 throw("bad g transition")
1111 gp.waitreason = waitReasonPreempted
1112 return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
1115 // stopTheWorld stops all P's from executing goroutines, interrupting
1116 // all goroutines at GC safe points and records reason as the reason
1117 // for the stop. On return, only the current goroutine's P is running.
1118 // stopTheWorld must not be called from a system stack and the caller
1119 // must not hold worldsema. The caller must call startTheWorld when
1120 // other P's should resume execution.
1122 // stopTheWorld is safe for multiple goroutines to call at the
1123 // same time. Each will execute its own stop, and the stops will
1126 // This is also used by routines that do stack dumps. If the system is
1127 // in panic or being exited, this may not reliably stop all
1129 func stopTheWorld(reason string) {
1130 semacquire(&worldsema)
1132 gp.m.preemptoff = reason
1133 systemstack(func() {
1134 // Mark the goroutine which called stopTheWorld preemptible so its
1135 // stack may be scanned.
1136 // This lets a mark worker scan us while we try to stop the world
1137 // since otherwise we could get in a mutual preemption deadlock.
1138 // We must not modify anything on the G stack because a stack shrink
1139 // may occur. A stack shrink is otherwise OK though because in order
1140 // to return from this function (and to leave the system stack) we
1141 // must have preempted all goroutines, including any attempting
1142 // to scan our stack, in which case, any stack shrinking will
1143 // have already completed by the time we exit.
1144 // Don't provide a wait reason because we're still executing.
1145 casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
1146 stopTheWorldWithSema()
1147 casgstatus(gp, _Gwaiting, _Grunning)
1151 // startTheWorld undoes the effects of stopTheWorld.
1152 func startTheWorld() {
1153 systemstack(func() { startTheWorldWithSema(false) })
1155 // worldsema must be held over startTheWorldWithSema to ensure
1156 // gomaxprocs cannot change while worldsema is held.
1158 // Release worldsema with direct handoff to the next waiter, but
1159 // acquirem so that semrelease1 doesn't try to yield our time.
1161 // Otherwise if e.g. ReadMemStats is being called in a loop,
1162 // it might stomp on other attempts to stop the world, such as
1163 // for starting or ending GC. The operation this blocks is
1164 // so heavy-weight that we should just try to be as fair as
1167 // We don't want to just allow us to get preempted between now
1168 // and releasing the semaphore because then we keep everyone
1169 // (including, for example, GCs) waiting longer.
1172 semrelease1(&worldsema, true, 0)
1176 // stopTheWorldGC has the same effect as stopTheWorld, but blocks
1177 // until the GC is not running. It also blocks a GC from starting
1178 // until startTheWorldGC is called.
1179 func stopTheWorldGC(reason string) {
1181 stopTheWorld(reason)
1184 // startTheWorldGC undoes the effects of stopTheWorldGC.
1185 func startTheWorldGC() {
1190 // Holding worldsema grants an M the right to try to stop the world.
1191 var worldsema uint32 = 1
1193 // Holding gcsema grants the M the right to block a GC, and blocks
1194 // until the current GC is done. In particular, it prevents gomaxprocs
1195 // from changing concurrently.
1197 // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
1198 // being changed/enabled during a GC, remove this.
1199 var gcsema uint32 = 1
1201 // stopTheWorldWithSema is the core implementation of stopTheWorld.
1202 // The caller is responsible for acquiring worldsema and disabling
1203 // preemption first and then should stopTheWorldWithSema on the system
1206 // semacquire(&worldsema, 0)
1207 // m.preemptoff = "reason"
1208 // systemstack(stopTheWorldWithSema)
1210 // When finished, the caller must either call startTheWorld or undo
1211 // these three operations separately:
1213 // m.preemptoff = ""
1214 // systemstack(startTheWorldWithSema)
1215 // semrelease(&worldsema)
1217 // It is allowed to acquire worldsema once and then execute multiple
1218 // startTheWorldWithSema/stopTheWorldWithSema pairs.
1219 // Other P's are able to execute between successive calls to
1220 // startTheWorldWithSema and stopTheWorldWithSema.
1221 // Holding worldsema causes any other goroutines invoking
1222 // stopTheWorld to block.
1223 func stopTheWorldWithSema() {
1226 // If we hold a lock, then we won't be able to stop another M
1227 // that is blocked trying to acquire the lock.
1229 throw("stopTheWorld: holding locks")
1233 sched.stopwait = gomaxprocs
1234 sched.gcwaiting.Store(true)
1237 gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
1239 // try to retake all P's in Psyscall status
1240 for _, pp := range allp {
1242 if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
1254 pp, _ := pidleget(now)
1258 pp.status = _Pgcstop
1261 wait := sched.stopwait > 0
1264 // wait for remaining P's to stop voluntarily
1267 // wait for 100us, then try to re-preempt in case of any races
1268 if notetsleep(&sched.stopnote, 100*1000) {
1269 noteclear(&sched.stopnote)
1278 if sched.stopwait != 0 {
1279 bad = "stopTheWorld: not stopped (stopwait != 0)"
1281 for _, pp := range allp {
1282 if pp.status != _Pgcstop {
1283 bad = "stopTheWorld: not stopped (status != _Pgcstop)"
1287 if freezing.Load() {
1288 // Some other thread is panicking. This can cause the
1289 // sanity checks above to fail if the panic happens in
1290 // the signal handler on a stopped thread. Either way,
1291 // we should halt this thread.
1302 func startTheWorldWithSema(emitTraceEvent bool) int64 {
1303 assertWorldStopped()
1305 mp := acquirem() // disable preemption because it can be holding p in a local var
1306 if netpollinited() {
1307 list := netpoll(0) // non-blocking
1317 p1 := procresize(procs)
1318 sched.gcwaiting.Store(false)
1319 if sched.sysmonwait.Load() {
1320 sched.sysmonwait.Store(false)
1321 notewakeup(&sched.sysmonnote)
1334 throw("startTheWorld: inconsistent mp->nextp")
1337 notewakeup(&mp.park)
1339 // Start M to run P. Do not start another M below.
1344 // Capture start-the-world time before doing clean-up tasks.
1345 startTime := nanotime()
1350 // Wakeup an additional proc in case we have excessive runnable goroutines
1351 // in local queues or in the global queue. If we don't, the proc will park itself.
1352 // If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
1360 // usesLibcall indicates whether this runtime performs system calls
1362 func usesLibcall() bool {
1364 case "aix", "darwin", "illumos", "ios", "solaris", "windows":
1367 return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
1372 // mStackIsSystemAllocated indicates whether this runtime starts on a
1373 // system-allocated stack.
1374 func mStackIsSystemAllocated() bool {
1376 case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
1380 case "386", "amd64", "arm", "arm64":
1387 // mstart is the entry-point for new Ms.
1388 // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
1391 // mstart0 is the Go entry-point for new Ms.
1392 // This must not split the stack because we may not even have stack
1393 // bounds set up yet.
1395 // May run during STW (because it doesn't have a P yet), so write
1396 // barriers are not allowed.
1399 //go:nowritebarrierrec
1403 osStack := gp.stack.lo == 0
1405 // Initialize stack bounds from system stack.
1406 // Cgo may have left stack size in stack.hi.
1407 // minit may update the stack bounds.
1409 // Note: these bounds may not be very accurate.
1410 // We set hi to &size, but there are things above
1411 // it. The 1024 is supposed to compensate this,
1412 // but is somewhat arbitrary.
1415 size = 8192 * sys.StackGuardMultiplier
1417 gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
1418 gp.stack.lo = gp.stack.hi - size + 1024
1420 // Initialize stack guard so that we can start calling regular
1422 gp.stackguard0 = gp.stack.lo + _StackGuard
1423 // This is the g0, so we can also call go:systemstack
1424 // functions, which check stackguard1.
1425 gp.stackguard1 = gp.stackguard0
1428 // Exit this thread.
1429 if mStackIsSystemAllocated() {
1430 // Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
1431 // the stack, but put it in gp.stack before mstart,
1432 // so the logic above hasn't set osStack yet.
1438 // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
1439 // so that we can set up g0.sched to return to the call of mstart1 above.
1446 throw("bad runtime·mstart")
1449 // Set up m.g0.sched as a label returning to just
1450 // after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
1451 // We're never coming back to mstart1 after we call schedule,
1452 // so other calls can reuse the current frame.
1453 // And goexit0 does a gogo that needs to return from mstart1
1454 // and let mstart0 exit the thread.
1455 gp.sched.g = guintptr(unsafe.Pointer(gp))
1456 gp.sched.pc = getcallerpc()
1457 gp.sched.sp = getcallersp()
1462 // Install signal handlers; after minit so that minit can
1463 // prepare the thread to be able to handle the signals.
1468 if fn := gp.m.mstartfn; fn != nil {
1473 acquirep(gp.m.nextp.ptr())
1479 // mstartm0 implements part of mstart1 that only runs on the m0.
1481 // Write barriers are allowed here because we know the GC can't be
1482 // running yet, so they'll be no-ops.
1484 //go:yeswritebarrierrec
1486 // Create an extra M for callbacks on threads not created by Go.
1487 // An extra M is also needed on Windows for callbacks created by
1488 // syscall.NewCallback. See issue #6751 for details.
1489 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1496 // mPark causes a thread to park itself, returning once woken.
1501 notesleep(&gp.m.park)
1502 noteclear(&gp.m.park)
1505 // mexit tears down and exits the current thread.
1507 // Don't call this directly to exit the thread, since it must run at
1508 // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
1509 // unwind the stack to the point that exits the thread.
1511 // It is entered with m.p != nil, so write barriers are allowed. It
1512 // will release the P before exiting.
1514 //go:yeswritebarrierrec
1515 func mexit(osStack bool) {
1519 // This is the main thread. Just wedge it.
1521 // On Linux, exiting the main thread puts the process
1522 // into a non-waitable zombie state. On Plan 9,
1523 // exiting the main thread unblocks wait even though
1524 // other threads are still running. On Solaris we can
1525 // neither exitThread nor return from mstart. Other
1526 // bad things probably happen on other platforms.
1528 // We could try to clean up this M more before wedging
1529 // it, but that complicates signal handling.
1530 handoffp(releasep())
1536 throw("locked m0 woke up")
1542 // Free the gsignal stack.
1543 if mp.gsignal != nil {
1544 stackfree(mp.gsignal.stack)
1545 // On some platforms, when calling into VDSO (e.g. nanotime)
1546 // we store our g on the gsignal stack, if there is one.
1547 // Now the stack is freed, unlink it from the m, so we
1548 // won't write to it when calling VDSO code.
1552 // Remove m from allm.
1554 for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
1560 throw("m not found in allm")
1563 // Delay reaping m until it's done with the stack.
1565 // If this is using an OS stack, the OS will free it
1566 // so there's no need for reaping.
1567 atomic.Store(&mp.freeWait, 1)
1568 // Put m on the free list, though it will not be reaped until
1569 // freeWait is 0. Note that the free list must not be linked
1570 // through alllink because some functions walk allm without
1571 // locking, so may be using alllink.
1572 mp.freelink = sched.freem
1577 atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
1580 handoffp(releasep())
1581 // After this point we must not have write barriers.
1583 // Invoke the deadlock detector. This must happen after
1584 // handoffp because it may have started a new M to take our
1591 if GOOS == "darwin" || GOOS == "ios" {
1592 // Make sure pendingPreemptSignals is correct when an M exits.
1594 if mp.signalPending.Load() != 0 {
1595 pendingPreemptSignals.Add(-1)
1599 // Destroy all allocated resources. After this is called, we may no
1600 // longer take any locks.
1604 // Return from mstart and let the system thread
1605 // library free the g0 stack and terminate the thread.
1609 // mstart is the thread's entry point, so there's nothing to
1610 // return to. Exit the thread directly. exitThread will clear
1611 // m.freeWait when it's done with the stack and the m can be
1613 exitThread(&mp.freeWait)
1616 // forEachP calls fn(p) for every P p when p reaches a GC safe point.
1617 // If a P is currently executing code, this will bring the P to a GC
1618 // safe point and execute fn on that P. If the P is not executing code
1619 // (it is idle or in a syscall), this will call fn(p) directly while
1620 // preventing the P from exiting its state. This does not ensure that
1621 // fn will run on every CPU executing Go code, but it acts as a global
1622 // memory barrier. GC uses this as a "ragged barrier."
1624 // The caller must hold worldsema.
1627 func forEachP(fn func(*p)) {
1629 pp := getg().m.p.ptr()
1632 if sched.safePointWait != 0 {
1633 throw("forEachP: sched.safePointWait != 0")
1635 sched.safePointWait = gomaxprocs - 1
1636 sched.safePointFn = fn
1638 // Ask all Ps to run the safe point function.
1639 for _, p2 := range allp {
1641 atomic.Store(&p2.runSafePointFn, 1)
1646 // Any P entering _Pidle or _Psyscall from now on will observe
1647 // p.runSafePointFn == 1 and will call runSafePointFn when
1648 // changing its status to _Pidle/_Psyscall.
1650 // Run safe point function for all idle Ps. sched.pidle will
1651 // not change because we hold sched.lock.
1652 for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
1653 if atomic.Cas(&p.runSafePointFn, 1, 0) {
1655 sched.safePointWait--
1659 wait := sched.safePointWait > 0
1662 // Run fn for the current P.
1665 // Force Ps currently in _Psyscall into _Pidle and hand them
1666 // off to induce safe point function execution.
1667 for _, p2 := range allp {
1669 if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
1679 // Wait for remaining Ps to run fn.
1682 // Wait for 100us, then try to re-preempt in
1683 // case of any races.
1685 // Requires system stack.
1686 if notetsleep(&sched.safePointNote, 100*1000) {
1687 noteclear(&sched.safePointNote)
1693 if sched.safePointWait != 0 {
1694 throw("forEachP: not done")
1696 for _, p2 := range allp {
1697 if p2.runSafePointFn != 0 {
1698 throw("forEachP: P did not run fn")
1703 sched.safePointFn = nil
1708 // runSafePointFn runs the safe point function, if any, for this P.
1709 // This should be called like
1711 // if getg().m.p.runSafePointFn != 0 {
1715 // runSafePointFn must be checked on any transition in to _Pidle or
1716 // _Psyscall to avoid a race where forEachP sees that the P is running
1717 // just before the P goes into _Pidle/_Psyscall and neither forEachP
1718 // nor the P run the safe-point function.
1719 func runSafePointFn() {
1720 p := getg().m.p.ptr()
1721 // Resolve the race between forEachP running the safe-point
1722 // function on this P's behalf and this P running the
1723 // safe-point function directly.
1724 if !atomic.Cas(&p.runSafePointFn, 1, 0) {
1727 sched.safePointFn(p)
1729 sched.safePointWait--
1730 if sched.safePointWait == 0 {
1731 notewakeup(&sched.safePointNote)
1736 // When running with cgo, we call _cgo_thread_start
1737 // to start threads for us so that we can play nicely with
1739 var cgoThreadStart unsafe.Pointer
1741 type cgothreadstart struct {
1747 // Allocate a new m unassociated with any thread.
1748 // Can use p for allocation context if needed.
1749 // fn is recorded as the new m's m.mstartfn.
1750 // id is optional pre-allocated m ID. Omit by passing -1.
1752 // This function is allowed to have write barriers even if the caller
1753 // isn't because it borrows pp.
1755 //go:yeswritebarrierrec
1756 func allocm(pp *p, fn func(), id int64) *m {
1759 // The caller owns pp, but we may borrow (i.e., acquirep) it. We must
1760 // disable preemption to ensure it is not stolen, which would make the
1761 // caller lose ownership.
1766 acquirep(pp) // temporarily borrow p for mallocs in this function
1769 // Release the free M list. We need to do this somewhere and
1770 // this may free up a stack we can use.
1771 if sched.freem != nil {
1774 for freem := sched.freem; freem != nil; {
1775 if freem.freeWait != 0 {
1776 next := freem.freelink
1777 freem.freelink = newList
1782 // stackfree must be on the system stack, but allocm is
1783 // reachable off the system stack transitively from
1785 systemstack(func() {
1786 stackfree(freem.g0.stack)
1788 freem = freem.freelink
1790 sched.freem = newList
1798 // In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
1799 // Windows and Plan 9 will layout sched stack on OS stack.
1800 if iscgo || mStackIsSystemAllocated() {
1803 mp.g0 = malg(8192 * sys.StackGuardMultiplier)
1807 if pp == gp.m.p.ptr() {
1812 allocmLock.runlock()
1816 // needm is called when a cgo callback happens on a
1817 // thread without an m (a thread not created by Go).
1818 // In this case, needm is expected to find an m to use
1819 // and return with m, g initialized correctly.
1820 // Since m and g are not set now (likely nil, but see below)
1821 // needm is limited in what routines it can call. In particular
1822 // it can only call nosplit functions (textflag 7) and cannot
1823 // do any scheduling that requires an m.
1825 // In order to avoid needing heavy lifting here, we adopt
1826 // the following strategy: there is a stack of available m's
1827 // that can be stolen. Using compare-and-swap
1828 // to pop from the stack has ABA races, so we simulate
1829 // a lock by doing an exchange (via Casuintptr) to steal the stack
1830 // head and replace the top pointer with MLOCKED (1).
1831 // This serves as a simple spin lock that we can use even
1832 // without an m. The thread that locks the stack in this way
1833 // unlocks the stack by storing a valid stack head pointer.
1835 // In order to make sure that there is always an m structure
1836 // available to be stolen, we maintain the invariant that there
1837 // is always one more than needed. At the beginning of the
1838 // program (if cgo is in use) the list is seeded with a single m.
1839 // If needm finds that it has taken the last m off the list, its job
1840 // is - once it has installed its own m so that it can do things like
1841 // allocate memory - to create a spare m and put it on the list.
1843 // Each of these extra m's also has a g0 and a curg that are
1844 // pressed into service as the scheduling stack and current
1845 // goroutine for the duration of the cgo callback.
1847 // When the callback is done with the m, it calls dropm to
1848 // put the m back on the list.
1852 if (iscgo || GOOS == "windows") && !cgoHasExtraM {
1853 // Can happen if C/C++ code calls Go from a global ctor.
1854 // Can also happen on Windows if a global ctor uses a
1855 // callback created by syscall.NewCallback. See issue #6751
1858 // Can not throw, because scheduler is not initialized yet.
1859 write(2, unsafe.Pointer(&earlycgocallback[0]), int32(len(earlycgocallback)))
1863 // Save and block signals before getting an M.
1864 // The signal handler may call needm itself,
1865 // and we must avoid a deadlock. Also, once g is installed,
1866 // any incoming signals will try to execute,
1867 // but we won't have the sigaltstack settings and other data
1868 // set up appropriately until the end of minit, which will
1869 // unblock the signals. This is the same dance as when
1870 // starting a new m to run Go code via newosproc.
1875 // Lock extra list, take head, unlock popped list.
1876 // nilokay=false is safe here because of the invariant above,
1877 // that the extra list always contains or will soon contain
1879 mp := lockextra(false)
1881 // Set needextram when we've just emptied the list,
1882 // so that the eventual call into cgocallbackg will
1883 // allocate a new m for the extra list. We delay the
1884 // allocation until then so that it can be done
1885 // after exitsyscall makes sure it is okay to be
1886 // running at all (that is, there's no garbage collection
1887 // running right now).
1888 mp.needextram = mp.schedlink == 0
1890 unlockextra(mp.schedlink.ptr())
1892 // Store the original signal mask for use by minit.
1893 mp.sigmask = sigmask
1895 // Install TLS on some platforms (previously setg
1896 // would do this if necessary).
1899 // Install g (= m->g0) and set the stack bounds
1900 // to match the current stack. We don't actually know
1901 // how big the stack is, like we don't know how big any
1902 // scheduling stack is, but we assume there's at least 32 kB,
1903 // which is more than enough for us.
1906 gp.stack.hi = getcallersp() + 1024
1907 gp.stack.lo = getcallersp() - 32*1024
1908 gp.stackguard0 = gp.stack.lo + _StackGuard
1910 // Initialize this thread to use the m.
1914 // mp.curg is now a real goroutine.
1915 casgstatus(mp.curg, _Gdead, _Gsyscall)
1919 var earlycgocallback = []byte("fatal error: cgo callback before cgo call\n")
1921 // newextram allocates m's and puts them on the extra list.
1922 // It is called with a working local m, so that it can do things
1923 // like call schedlock and allocate.
1925 c := extraMWaiters.Swap(0)
1927 for i := uint32(0); i < c; i++ {
1931 // Make sure there is at least one extra M.
1932 mp := lockextra(true)
1940 // oneNewExtraM allocates an m and puts it on the extra list.
1941 func oneNewExtraM() {
1942 // Create extra goroutine locked to extra m.
1943 // The goroutine is the context in which the cgo callback will run.
1944 // The sched.pc will never be returned to, but setting it to
1945 // goexit makes clear to the traceback routines where
1946 // the goroutine stack ends.
1947 mp := allocm(nil, nil, -1)
1949 gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
1950 gp.sched.sp = gp.stack.hi
1951 gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
1953 gp.sched.g = guintptr(unsafe.Pointer(gp))
1954 gp.syscallpc = gp.sched.pc
1955 gp.syscallsp = gp.sched.sp
1956 gp.stktopsp = gp.sched.sp
1957 // malg returns status as _Gidle. Change to _Gdead before
1958 // adding to allg where GC can see it. We use _Gdead to hide
1959 // this from tracebacks and stack scans since it isn't a
1960 // "real" goroutine until needm grabs it.
1961 casgstatus(gp, _Gidle, _Gdead)
1967 gp.goid = sched.goidgen.Add(1)
1969 gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
1971 // put on allg for garbage collector
1974 // gp is now on the allg list, but we don't want it to be
1975 // counted by gcount. It would be more "proper" to increment
1976 // sched.ngfree, but that requires locking. Incrementing ngsys
1977 // has the same effect.
1980 // Add m to the extra list.
1981 mnext := lockextra(true)
1982 mp.schedlink.set(mnext)
1987 // dropm is called when a cgo callback has called needm but is now
1988 // done with the callback and returning back into the non-Go thread.
1989 // It puts the current m back onto the extra list.
1991 // The main expense here is the call to signalstack to release the
1992 // m's signal stack, and then the call to needm on the next callback
1993 // from this thread. It is tempting to try to save the m for next time,
1994 // which would eliminate both these costs, but there might not be
1995 // a next time: the current thread (which Go does not control) might exit.
1996 // If we saved the m for that thread, there would be an m leak each time
1997 // such a thread exited. Instead, we acquire and release an m on each
1998 // call. These should typically not be scheduling operations, just a few
1999 // atomics, so the cost should be small.
2001 // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
2002 // variable using pthread_key_create. Unlike the pthread keys we already use
2003 // on OS X, this dummy key would never be read by Go code. It would exist
2004 // only so that we could register at thread-exit-time destructor.
2005 // That destructor would put the m back onto the extra list.
2006 // This is purely a performance optimization. The current version,
2007 // in which dropm happens on each cgo call, is still correct too.
2008 // We may have to keep the current version on systems with cgo
2009 // but without pthreads, like Windows.
2011 // Clear m and g, and return m to the extra list.
2012 // After the call to setg we can only call nosplit functions
2013 // with no pointer manipulation.
2016 // Return mp.curg to dead state.
2017 casgstatus(mp.curg, _Gsyscall, _Gdead)
2018 mp.curg.preemptStop = false
2021 // Block signals before unminit.
2022 // Unminit unregisters the signal handling stack (but needs g on some systems).
2023 // Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
2024 // It's important not to try to handle a signal between those two steps.
2025 sigmask := mp.sigmask
2029 mnext := lockextra(true)
2031 mp.schedlink.set(mnext)
2035 // Commit the release of mp.
2038 msigrestore(sigmask)
2041 // A helper function for EnsureDropM.
2042 func getm() uintptr {
2043 return uintptr(unsafe.Pointer(getg().m))
2046 var extram atomic.Uintptr
2047 var extraMCount uint32 // Protected by lockextra
2048 var extraMWaiters atomic.Uint32
2050 // lockextra locks the extra list and returns the list head.
2051 // The caller must unlock the list by storing a new list head
2052 // to extram. If nilokay is true, then lockextra will
2053 // return a nil list head if that's what it finds. If nilokay is false,
2054 // lockextra will keep waiting until the list head is no longer nil.
2057 func lockextra(nilokay bool) *m {
2062 old := extram.Load()
2067 if old == 0 && !nilokay {
2069 // Add 1 to the number of threads
2070 // waiting for an M.
2071 // This is cleared by newextram.
2072 extraMWaiters.Add(1)
2078 if extram.CompareAndSwap(old, locked) {
2079 return (*m)(unsafe.Pointer(old))
2087 func unlockextra(mp *m) {
2088 extram.Store(uintptr(unsafe.Pointer(mp)))
2092 // allocmLock is locked for read when creating new Ms in allocm and their
2093 // addition to allm. Thus acquiring this lock for write blocks the
2094 // creation of new Ms.
2097 // execLock serializes exec and clone to avoid bugs or unspecified
2098 // behaviour around exec'ing while creating/destroying threads. See
2103 // newmHandoff contains a list of m structures that need new OS threads.
2104 // This is used by newm in situations where newm itself can't safely
2105 // start an OS thread.
2106 var newmHandoff struct {
2109 // newm points to a list of M structures that need new OS
2110 // threads. The list is linked through m.schedlink.
2113 // waiting indicates that wake needs to be notified when an m
2114 // is put on the list.
2118 // haveTemplateThread indicates that the templateThread has
2119 // been started. This is not protected by lock. Use cas to set
2121 haveTemplateThread uint32
2124 // Create a new m. It will start off with a call to fn, or else the scheduler.
2125 // fn needs to be static and not a heap allocated closure.
2126 // May run with m.p==nil, so write barriers are not allowed.
2128 // id is optional pre-allocated m ID. Omit by passing -1.
2130 //go:nowritebarrierrec
2131 func newm(fn func(), pp *p, id int64) {
2132 // allocm adds a new M to allm, but they do not start until created by
2133 // the OS in newm1 or the template thread.
2135 // doAllThreadsSyscall requires that every M in allm will eventually
2136 // start and be signal-able, even with a STW.
2138 // Disable preemption here until we start the thread to ensure that
2139 // newm is not preempted between allocm and starting the new thread,
2140 // ensuring that anything added to allm is guaranteed to eventually
2144 mp := allocm(pp, fn, id)
2146 mp.sigmask = initSigmask
2147 if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
2148 // We're on a locked M or a thread that may have been
2149 // started by C. The kernel state of this thread may
2150 // be strange (the user may have locked it for that
2151 // purpose). We don't want to clone that into another
2152 // thread. Instead, ask a known-good thread to create
2153 // the thread for us.
2155 // This is disabled on Plan 9. See golang.org/issue/22227.
2157 // TODO: This may be unnecessary on Windows, which
2158 // doesn't model thread creation off fork.
2159 lock(&newmHandoff.lock)
2160 if newmHandoff.haveTemplateThread == 0 {
2161 throw("on a locked thread with no template thread")
2163 mp.schedlink = newmHandoff.newm
2164 newmHandoff.newm.set(mp)
2165 if newmHandoff.waiting {
2166 newmHandoff.waiting = false
2167 notewakeup(&newmHandoff.wake)
2169 unlock(&newmHandoff.lock)
2170 // The M has not started yet, but the template thread does not
2171 // participate in STW, so it will always process queued Ms and
2172 // it is safe to releasem.
2182 var ts cgothreadstart
2183 if _cgo_thread_start == nil {
2184 throw("_cgo_thread_start missing")
2187 ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
2188 ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
2190 msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2193 asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
2195 execLock.rlock() // Prevent process clone.
2196 asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
2200 execLock.rlock() // Prevent process clone.
2205 // startTemplateThread starts the template thread if it is not already
2208 // The calling thread must itself be in a known-good state.
2209 func startTemplateThread() {
2210 if GOARCH == "wasm" { // no threads on wasm yet
2214 // Disable preemption to guarantee that the template thread will be
2215 // created before a park once haveTemplateThread is set.
2217 if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
2221 newm(templateThread, nil, -1)
2225 // templateThread is a thread in a known-good state that exists solely
2226 // to start new threads in known-good states when the calling thread
2227 // may not be in a good state.
2229 // Many programs never need this, so templateThread is started lazily
2230 // when we first enter a state that might lead to running on a thread
2231 // in an unknown state.
2233 // templateThread runs on an M without a P, so it must not have write
2236 //go:nowritebarrierrec
2237 func templateThread() {
2244 lock(&newmHandoff.lock)
2245 for newmHandoff.newm != 0 {
2246 newm := newmHandoff.newm.ptr()
2247 newmHandoff.newm = 0
2248 unlock(&newmHandoff.lock)
2250 next := newm.schedlink.ptr()
2255 lock(&newmHandoff.lock)
2257 newmHandoff.waiting = true
2258 noteclear(&newmHandoff.wake)
2259 unlock(&newmHandoff.lock)
2260 notesleep(&newmHandoff.wake)
2264 // Stops execution of the current m until new work is available.
2265 // Returns with acquired P.
2269 if gp.m.locks != 0 {
2270 throw("stopm holding locks")
2273 throw("stopm holding p")
2276 throw("stopm spinning")
2283 acquirep(gp.m.nextp.ptr())
2288 // startm's caller incremented nmspinning. Set the new M's spinning.
2289 getg().m.spinning = true
2292 // Schedules some M to run the p (creates an M if necessary).
2293 // If p==nil, tries to get an idle P, if no idle P's does nothing.
2294 // May run with m.p==nil, so write barriers are not allowed.
2295 // If spinning is set, the caller has incremented nmspinning and must provide a
2296 // P. startm will set m.spinning in the newly started M.
2298 // Callers passing a non-nil P must call from a non-preemptible context. See
2299 // comment on acquirem below.
2301 // Must not have write barriers because this may be called without a P.
2303 //go:nowritebarrierrec
2304 func startm(pp *p, spinning bool) {
2305 // Disable preemption.
2307 // Every owned P must have an owner that will eventually stop it in the
2308 // event of a GC stop request. startm takes transient ownership of a P
2309 // (either from argument or pidleget below) and transfers ownership to
2310 // a started M, which will be responsible for performing the stop.
2312 // Preemption must be disabled during this transient ownership,
2313 // otherwise the P this is running on may enter GC stop while still
2314 // holding the transient P, leaving that P in limbo and deadlocking the
2317 // Callers passing a non-nil P must already be in non-preemptible
2318 // context, otherwise such preemption could occur on function entry to
2319 // startm. Callers passing a nil P may be preemptible, so we must
2320 // disable preemption before acquiring a P from pidleget below.
2325 // TODO(prattmic): All remaining calls to this function
2326 // with _p_ == nil could be cleaned up to find a P
2327 // before calling startm.
2328 throw("startm: P required for spinning=true")
2339 // No M is available, we must drop sched.lock and call newm.
2340 // However, we already own a P to assign to the M.
2342 // Once sched.lock is released, another G (e.g., in a syscall),
2343 // could find no idle P while checkdead finds a runnable G but
2344 // no running M's because this new M hasn't started yet, thus
2345 // throwing in an apparent deadlock.
2347 // Avoid this situation by pre-allocating the ID for the new M,
2348 // thus marking it as 'running' before we drop sched.lock. This
2349 // new M will eventually run the scheduler to execute any
2356 // The caller incremented nmspinning, so set m.spinning in the new M.
2360 // Ownership transfer of pp committed by start in newm.
2361 // Preemption is now safe.
2367 throw("startm: m is spinning")
2370 throw("startm: m has p")
2372 if spinning && !runqempty(pp) {
2373 throw("startm: p has runnable gs")
2375 // The caller incremented nmspinning, so set m.spinning in the new M.
2376 nmp.spinning = spinning
2378 notewakeup(&nmp.park)
2379 // Ownership transfer of pp committed by wakeup. Preemption is now
2384 // Hands off P from syscall or locked M.
2385 // Always runs without a P, so write barriers are not allowed.
2387 //go:nowritebarrierrec
2388 func handoffp(pp *p) {
2389 // handoffp must start an M in any situation where
2390 // findrunnable would return a G to run on pp.
2392 // if it has local work, start it straight away
2393 if !runqempty(pp) || sched.runqsize != 0 {
2397 // if there's trace work to do, start it straight away
2398 if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
2402 // if it has GC work, start it straight away
2403 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
2407 // no local work, check that there are no spinning/idle M's,
2408 // otherwise our help is not required
2409 if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
2410 sched.needspinning.Store(0)
2415 if sched.gcwaiting.Load() {
2416 pp.status = _Pgcstop
2418 if sched.stopwait == 0 {
2419 notewakeup(&sched.stopnote)
2424 if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
2425 sched.safePointFn(pp)
2426 sched.safePointWait--
2427 if sched.safePointWait == 0 {
2428 notewakeup(&sched.safePointNote)
2431 if sched.runqsize != 0 {
2436 // If this is the last running P and nobody is polling network,
2437 // need to wakeup another M to poll network.
2438 if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
2444 // The scheduler lock cannot be held when calling wakeNetPoller below
2445 // because wakeNetPoller may call wakep which may call startm.
2446 when := nobarrierWakeTime(pp)
2455 // Tries to add one more P to execute G's.
2456 // Called when a G is made runnable (newproc, ready).
2457 // Must be called with a P.
2459 // Be conservative about spinning threads, only start one if none exist
2461 if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
2465 // Disable preemption until ownership of pp transfers to the next M in
2466 // startm. Otherwise preemption here would leave pp stuck waiting to
2469 // See preemption comment on acquirem in startm for more details.
2474 pp, _ = pidlegetSpinning(0)
2476 if sched.nmspinning.Add(-1) < 0 {
2477 throw("wakep: negative nmspinning")
2483 // Since we always have a P, the race in the "No M is available"
2484 // comment in startm doesn't apply during the small window between the
2485 // unlock here and lock in startm. A checkdead in between will always
2486 // see at least one running M (ours).
2494 // Stops execution of the current m that is locked to a g until the g is runnable again.
2495 // Returns with acquired P.
2496 func stoplockedm() {
2499 if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
2500 throw("stoplockedm: inconsistent locking")
2503 // Schedule another M to run this p.
2508 // Wait until another thread schedules lockedg again.
2510 status := readgstatus(gp.m.lockedg.ptr())
2511 if status&^_Gscan != _Grunnable {
2512 print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
2513 dumpgstatus(gp.m.lockedg.ptr())
2514 throw("stoplockedm: not runnable")
2516 acquirep(gp.m.nextp.ptr())
2520 // Schedules the locked m to run the locked gp.
2521 // May run during STW, so write barriers are not allowed.
2523 //go:nowritebarrierrec
2524 func startlockedm(gp *g) {
2525 mp := gp.lockedm.ptr()
2527 throw("startlockedm: locked to me")
2530 throw("startlockedm: m has p")
2532 // directly handoff current P to the locked m
2536 notewakeup(&mp.park)
2540 // Stops the current m for stopTheWorld.
2541 // Returns when the world is restarted.
2545 if !sched.gcwaiting.Load() {
2546 throw("gcstopm: not waiting for gc")
2549 gp.m.spinning = false
2550 // OK to just drop nmspinning here,
2551 // startTheWorld will unpark threads as necessary.
2552 if sched.nmspinning.Add(-1) < 0 {
2553 throw("gcstopm: negative nmspinning")
2558 pp.status = _Pgcstop
2560 if sched.stopwait == 0 {
2561 notewakeup(&sched.stopnote)
2567 // Schedules gp to run on the current M.
2568 // If inheritTime is true, gp inherits the remaining time in the
2569 // current time slice. Otherwise, it starts a new time slice.
2572 // Write barriers are allowed because this is called immediately after
2573 // acquiring a P in several places.
2575 //go:yeswritebarrierrec
2576 func execute(gp *g, inheritTime bool) {
2579 if goroutineProfile.active {
2580 // Make sure that gp has had its stack written out to the goroutine
2581 // profile, exactly as it was when the goroutine profiler first stopped
2583 tryRecordGoroutineProfile(gp, osyield)
2586 // Assign gp.m before entering _Grunning so running Gs have an
2590 casgstatus(gp, _Grunnable, _Grunning)
2593 gp.stackguard0 = gp.stack.lo + _StackGuard
2595 mp.p.ptr().schedtick++
2598 // Check whether the profiler needs to be turned on or off.
2599 hz := sched.profilehz
2600 if mp.profilehz != hz {
2601 setThreadCPUProfiler(hz)
2605 // GoSysExit has to happen when we have a P, but before GoStart.
2606 // So we emit it here.
2607 if gp.syscallsp != 0 && gp.sysblocktraced {
2608 traceGoSysExit(gp.sysexitticks)
2616 // Finds a runnable goroutine to execute.
2617 // Tries to steal from other P's, get g from local or global queue, poll network.
2618 // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
2619 // reader) so the caller should try to wake a P.
2620 func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
2623 // The conditions here and in handoffp must agree: if
2624 // findrunnable would return a G to run, handoffp must start
2629 if sched.gcwaiting.Load() {
2633 if pp.runSafePointFn != 0 {
2637 // now and pollUntil are saved for work stealing later,
2638 // which may steal timers. It's important that between now
2639 // and then, nothing blocks, so these numbers remain mostly
2641 now, pollUntil, _ := checkTimers(pp, 0)
2643 // Try to schedule the trace reader.
2644 if trace.enabled || trace.shutdown {
2647 casgstatus(gp, _Gwaiting, _Grunnable)
2648 traceGoUnpark(gp, 0)
2649 return gp, false, true
2653 // Try to schedule a GC worker.
2654 if gcBlackenEnabled != 0 {
2655 gp, tnow := gcController.findRunnableGCWorker(pp, now)
2657 return gp, false, true
2662 // Check the global runnable queue once in a while to ensure fairness.
2663 // Otherwise two goroutines can completely occupy the local runqueue
2664 // by constantly respawning each other.
2665 if pp.schedtick%61 == 0 && sched.runqsize > 0 {
2667 gp := globrunqget(pp, 1)
2670 return gp, false, false
2674 // Wake up the finalizer G.
2675 if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
2676 if gp := wakefing(); gp != nil {
2680 if *cgo_yield != nil {
2681 asmcgocall(*cgo_yield, nil)
2685 if gp, inheritTime := runqget(pp); gp != nil {
2686 return gp, inheritTime, false
2690 if sched.runqsize != 0 {
2692 gp := globrunqget(pp, 0)
2695 return gp, false, false
2700 // This netpoll is only an optimization before we resort to stealing.
2701 // We can safely skip it if there are no waiters or a thread is blocked
2702 // in netpoll already. If there is any kind of logical race with that
2703 // blocked thread (e.g. it has already returned from netpoll, but does
2704 // not set lastpoll yet), this thread will do blocking netpoll below
2706 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2707 if list := netpoll(0); !list.empty() { // non-blocking
2710 casgstatus(gp, _Gwaiting, _Grunnable)
2712 traceGoUnpark(gp, 0)
2714 return gp, false, false
2718 // Spinning Ms: steal work from other Ps.
2720 // Limit the number of spinning Ms to half the number of busy Ps.
2721 // This is necessary to prevent excessive CPU consumption when
2722 // GOMAXPROCS>>1 but the program parallelism is low.
2723 if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
2728 gp, inheritTime, tnow, w, newWork := stealWork(now)
2730 // Successfully stole.
2731 return gp, inheritTime, false
2734 // There may be new timer or GC work; restart to
2740 if w != 0 && (pollUntil == 0 || w < pollUntil) {
2741 // Earlier timer to wait for.
2746 // We have nothing to do.
2748 // If we're in the GC mark phase, can safely scan and blacken objects,
2749 // and have work to do, run idle-time marking rather than give up the P.
2750 if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
2751 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
2753 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2755 casgstatus(gp, _Gwaiting, _Grunnable)
2757 traceGoUnpark(gp, 0)
2759 return gp, false, false
2761 gcController.removeIdleMarkWorker()
2765 // If a callback returned and no other goroutine is awake,
2766 // then wake event handler goroutine which pauses execution
2767 // until a callback was triggered.
2768 gp, otherReady := beforeIdle(now, pollUntil)
2770 casgstatus(gp, _Gwaiting, _Grunnable)
2772 traceGoUnpark(gp, 0)
2774 return gp, false, false
2780 // Before we drop our P, make a snapshot of the allp slice,
2781 // which can change underfoot once we no longer block
2782 // safe-points. We don't need to snapshot the contents because
2783 // everything up to cap(allp) is immutable.
2784 allpSnapshot := allp
2785 // Also snapshot masks. Value changes are OK, but we can't allow
2786 // len to change out from under us.
2787 idlepMaskSnapshot := idlepMask
2788 timerpMaskSnapshot := timerpMask
2790 // return P and block
2792 if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
2796 if sched.runqsize != 0 {
2797 gp := globrunqget(pp, 0)
2799 return gp, false, false
2801 if !mp.spinning && sched.needspinning.Load() == 1 {
2802 // See "Delicate dance" comment below.
2807 if releasep() != pp {
2808 throw("findrunnable: wrong p")
2810 now = pidleput(pp, now)
2813 // Delicate dance: thread transitions from spinning to non-spinning
2814 // state, potentially concurrently with submission of new work. We must
2815 // drop nmspinning first and then check all sources again (with
2816 // #StoreLoad memory barrier in between). If we do it the other way
2817 // around, another thread can submit work after we've checked all
2818 // sources but before we drop nmspinning; as a result nobody will
2819 // unpark a thread to run the work.
2821 // This applies to the following sources of work:
2823 // * Goroutines added to a per-P run queue.
2824 // * New/modified-earlier timers on a per-P timer heap.
2825 // * Idle-priority GC work (barring golang.org/issue/19112).
2827 // If we discover new work below, we need to restore m.spinning as a
2828 // signal for resetspinning to unpark a new worker thread (because
2829 // there can be more than one starving goroutine).
2831 // However, if after discovering new work we also observe no idle Ps
2832 // (either here or in resetspinning), we have a problem. We may be
2833 // racing with a non-spinning M in the block above, having found no
2834 // work and preparing to release its P and park. Allowing that P to go
2835 // idle will result in loss of work conservation (idle P while there is
2836 // runnable work). This could result in complete deadlock in the
2837 // unlikely event that we discover new work (from netpoll) right as we
2838 // are racing with _all_ other Ps going idle.
2840 // We use sched.needspinning to synchronize with non-spinning Ms going
2841 // idle. If needspinning is set when they are about to drop their P,
2842 // they abort the drop and instead become a new spinning M on our
2843 // behalf. If we are not racing and the system is truly fully loaded
2844 // then no spinning threads are required, and the next thread to
2845 // naturally become spinning will clear the flag.
2847 // Also see "Worker thread parking/unparking" comment at the top of the
2849 wasSpinning := mp.spinning
2852 if sched.nmspinning.Add(-1) < 0 {
2853 throw("findrunnable: negative nmspinning")
2856 // Note the for correctness, only the last M transitioning from
2857 // spinning to non-spinning must perform these rechecks to
2858 // ensure no missed work. However, the runtime has some cases
2859 // of transient increments of nmspinning that are decremented
2860 // without going through this path, so we must be conservative
2861 // and perform the check on all spinning Ms.
2863 // See https://go.dev/issue/43997.
2865 // Check all runqueues once again.
2866 pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
2873 // Check for idle-priority GC work again.
2874 pp, gp := checkIdleGCNoP()
2879 // Run the idle worker.
2880 pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
2881 casgstatus(gp, _Gwaiting, _Grunnable)
2883 traceGoUnpark(gp, 0)
2885 return gp, false, false
2888 // Finally, check for timer creation or expiry concurrently with
2889 // transitioning from spinning to non-spinning.
2891 // Note that we cannot use checkTimers here because it calls
2892 // adjusttimers which may need to allocate memory, and that isn't
2893 // allowed when we don't have an active P.
2894 pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
2897 // Poll network until next timer.
2898 if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
2899 sched.pollUntil.Store(pollUntil)
2901 throw("findrunnable: netpoll with p")
2904 throw("findrunnable: netpoll with spinning")
2910 delay = pollUntil - now
2916 // When using fake time, just poll.
2919 list := netpoll(delay) // block until new work is available
2920 sched.pollUntil.Store(0)
2921 sched.lastpoll.Store(now)
2922 if faketime != 0 && list.empty() {
2923 // Using fake time and nothing is ready; stop M.
2924 // When all M's stop, checkdead will call timejump.
2929 pp, _ := pidleget(now)
2938 casgstatus(gp, _Gwaiting, _Grunnable)
2940 traceGoUnpark(gp, 0)
2942 return gp, false, false
2949 } else if pollUntil != 0 && netpollinited() {
2950 pollerPollUntil := sched.pollUntil.Load()
2951 if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
2959 // pollWork reports whether there is non-background work this P could
2960 // be doing. This is a fairly lightweight check to be used for
2961 // background work loops, like idle GC. It checks a subset of the
2962 // conditions checked by the actual scheduler.
2963 func pollWork() bool {
2964 if sched.runqsize != 0 {
2967 p := getg().m.p.ptr()
2971 if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
2972 if list := netpoll(0); !list.empty() {
2980 // stealWork attempts to steal a runnable goroutine or timer from any P.
2982 // If newWork is true, new work may have been readied.
2984 // If now is not 0 it is the current time. stealWork returns the passed time or
2985 // the current time if now was passed as 0.
2986 func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
2987 pp := getg().m.p.ptr()
2991 const stealTries = 4
2992 for i := 0; i < stealTries; i++ {
2993 stealTimersOrRunNextG := i == stealTries-1
2995 for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
2996 if sched.gcwaiting.Load() {
2997 // GC work may be available.
2998 return nil, false, now, pollUntil, true
3000 p2 := allp[enum.position()]
3005 // Steal timers from p2. This call to checkTimers is the only place
3006 // where we might hold a lock on a different P's timers. We do this
3007 // once on the last pass before checking runnext because stealing
3008 // from the other P's runnext should be the last resort, so if there
3009 // are timers to steal do that first.
3011 // We only check timers on one of the stealing iterations because
3012 // the time stored in now doesn't change in this loop and checking
3013 // the timers for each P more than once with the same value of now
3014 // is probably a waste of time.
3016 // timerpMask tells us whether the P may have timers at all. If it
3017 // can't, no need to check at all.
3018 if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
3019 tnow, w, ran := checkTimers(p2, now)
3021 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3025 // Running the timers may have
3026 // made an arbitrary number of G's
3027 // ready and added them to this P's
3028 // local run queue. That invalidates
3029 // the assumption of runqsteal
3030 // that it always has room to add
3031 // stolen G's. So check now if there
3032 // is a local G to run.
3033 if gp, inheritTime := runqget(pp); gp != nil {
3034 return gp, inheritTime, now, pollUntil, ranTimer
3040 // Don't bother to attempt to steal if p2 is idle.
3041 if !idlepMask.read(enum.position()) {
3042 if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
3043 return gp, false, now, pollUntil, ranTimer
3049 // No goroutines found to steal. Regardless, running a timer may have
3050 // made some goroutine ready that we missed. Indicate the next timer to
3052 return nil, false, now, pollUntil, ranTimer
3055 // Check all Ps for a runnable G to steal.
3057 // On entry we have no P. If a G is available to steal and a P is available,
3058 // the P is returned which the caller should acquire and attempt to steal the
3060 func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
3061 for id, p2 := range allpSnapshot {
3062 if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
3064 pp, _ := pidlegetSpinning(0)
3066 // Can't get a P, don't bother checking remaining Ps.
3075 // No work available.
3079 // Check all Ps for a timer expiring sooner than pollUntil.
3081 // Returns updated pollUntil value.
3082 func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
3083 for id, p2 := range allpSnapshot {
3084 if timerpMaskSnapshot.read(uint32(id)) {
3085 w := nobarrierWakeTime(p2)
3086 if w != 0 && (pollUntil == 0 || w < pollUntil) {
3095 // Check for idle-priority GC, without a P on entry.
3097 // If some GC work, a P, and a worker G are all available, the P and G will be
3098 // returned. The returned P has not been wired yet.
3099 func checkIdleGCNoP() (*p, *g) {
3100 // N.B. Since we have no P, gcBlackenEnabled may change at any time; we
3101 // must check again after acquiring a P. As an optimization, we also check
3102 // if an idle mark worker is needed at all. This is OK here, because if we
3103 // observe that one isn't needed, at least one is currently running. Even if
3104 // it stops running, its own journey into the scheduler should schedule it
3105 // again, if need be (at which point, this check will pass, if relevant).
3106 if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
3109 if !gcMarkWorkAvailable(nil) {
3113 // Work is available; we can start an idle GC worker only if there is
3114 // an available P and available worker G.
3116 // We can attempt to acquire these in either order, though both have
3117 // synchronization concerns (see below). Workers are almost always
3118 // available (see comment in findRunnableGCWorker for the one case
3119 // there may be none). Since we're slightly less likely to find a P,
3120 // check for that first.
3122 // Synchronization: note that we must hold sched.lock until we are
3123 // committed to keeping it. Otherwise we cannot put the unnecessary P
3124 // back in sched.pidle without performing the full set of idle
3125 // transition checks.
3127 // If we were to check gcBgMarkWorkerPool first, we must somehow handle
3128 // the assumption in gcControllerState.findRunnableGCWorker that an
3129 // empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
3131 pp, now := pidlegetSpinning(0)
3137 // Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
3138 if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
3144 node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
3148 gcController.removeIdleMarkWorker()
3154 return pp, node.gp.ptr()
3157 // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
3158 // going to wake up before the when argument; or it wakes an idle P to service
3159 // timers and the network poller if there isn't one already.
3160 func wakeNetPoller(when int64) {
3161 if sched.lastpoll.Load() == 0 {
3162 // In findrunnable we ensure that when polling the pollUntil
3163 // field is either zero or the time to which the current
3164 // poll is expected to run. This can have a spurious wakeup
3165 // but should never miss a wakeup.
3166 pollerPollUntil := sched.pollUntil.Load()
3167 if pollerPollUntil == 0 || pollerPollUntil > when {
3171 // There are no threads in the network poller, try to get
3172 // one there so it can handle new timers.
3173 if GOOS != "plan9" { // Temporary workaround - see issue #42303.
3179 func resetspinning() {
3182 throw("resetspinning: not a spinning m")
3184 gp.m.spinning = false
3185 nmspinning := sched.nmspinning.Add(-1)
3187 throw("findrunnable: negative nmspinning")
3189 // M wakeup policy is deliberately somewhat conservative, so check if we
3190 // need to wakeup another P here. See "Worker thread parking/unparking"
3191 // comment at the top of the file for details.
3195 // injectglist adds each runnable G on the list to some run queue,
3196 // and clears glist. If there is no current P, they are added to the
3197 // global queue, and up to npidle M's are started to run them.
3198 // Otherwise, for each idle P, this adds a G to the global queue
3199 // and starts an M. Any remaining G's are added to the current P's
3201 // This may temporarily acquire sched.lock.
3202 // Can run concurrently with GC.
3203 func injectglist(glist *gList) {
3208 for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
3209 traceGoUnpark(gp, 0)
3213 // Mark all the goroutines as runnable before we put them
3214 // on the run queues.
3215 head := glist.head.ptr()
3218 for gp := head; gp != nil; gp = gp.schedlink.ptr() {
3221 casgstatus(gp, _Gwaiting, _Grunnable)
3224 // Turn the gList into a gQueue.
3230 startIdle := func(n int) {
3231 for i := 0; i < n; i++ {
3232 mp := acquirem() // See comment in startm.
3235 pp, _ := pidlegetSpinning(0)
3248 pp := getg().m.p.ptr()
3251 globrunqputbatch(&q, int32(qsize))
3257 npidle := int(sched.npidle.Load())
3260 for n = 0; n < npidle && !q.empty(); n++ {
3266 globrunqputbatch(&globq, int32(n))
3273 runqputbatch(pp, &q, qsize)
3277 // One round of scheduler: find a runnable goroutine and execute it.
3283 throw("schedule: holding locks")
3286 if mp.lockedg != 0 {
3288 execute(mp.lockedg.ptr(), false) // Never returns.
3291 // We should not schedule away from a g that is executing a cgo call,
3292 // since the cgo call is using the m's g0 stack.
3294 throw("schedule: in cgo")
3301 // Safety check: if we are spinning, the run queue should be empty.
3302 // Check this before calling checkTimers, as that might call
3303 // goready to put a ready goroutine on the local run queue.
3304 if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
3305 throw("schedule: spinning with local work")
3308 gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
3310 // This thread is going to run a goroutine and is not spinning anymore,
3311 // so if it was marked as spinning we need to reset it now and potentially
3312 // start a new spinning M.
3317 if sched.disable.user && !schedEnabled(gp) {
3318 // Scheduling of this goroutine is disabled. Put it on
3319 // the list of pending runnable goroutines for when we
3320 // re-enable user scheduling and look again.
3322 if schedEnabled(gp) {
3323 // Something re-enabled scheduling while we
3324 // were acquiring the lock.
3327 sched.disable.runnable.pushBack(gp)
3334 // If about to schedule a not-normal goroutine (a GCworker or tracereader),
3335 // wake a P if there is one.
3339 if gp.lockedm != 0 {
3340 // Hands off own p to the locked m,
3341 // then blocks waiting for a new p.
3346 execute(gp, inheritTime)
3349 // dropg removes the association between m and the current goroutine m->curg (gp for short).
3350 // Typically a caller sets gp's status away from Grunning and then
3351 // immediately calls dropg to finish the job. The caller is also responsible
3352 // for arranging that gp will be restarted using ready at an
3353 // appropriate time. After calling dropg and arranging for gp to be
3354 // readied later, the caller can do other work but eventually should
3355 // call schedule to restart the scheduling of goroutines on this m.
3359 setMNoWB(&gp.m.curg.m, nil)
3360 setGNoWB(&gp.m.curg, nil)
3363 // checkTimers runs any timers for the P that are ready.
3364 // If now is not 0 it is the current time.
3365 // It returns the passed time or the current time if now was passed as 0.
3366 // and the time when the next timer should run or 0 if there is no next timer,
3367 // and reports whether it ran any timers.
3368 // If the time when the next timer should run is not 0,
3369 // it is always larger than the returned time.
3370 // We pass now in and out to avoid extra calls of nanotime.
3372 //go:yeswritebarrierrec
3373 func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
3374 // If it's not yet time for the first timer, or the first adjusted
3375 // timer, then there is nothing to do.
3376 next := pp.timer0When.Load()
3377 nextAdj := pp.timerModifiedEarliest.Load()
3378 if next == 0 || (nextAdj != 0 && nextAdj < next) {
3383 // No timers to run or adjust.
3384 return now, 0, false
3391 // Next timer is not ready to run, but keep going
3392 // if we would clear deleted timers.
3393 // This corresponds to the condition below where
3394 // we decide whether to call clearDeletedTimers.
3395 if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
3396 return now, next, false
3400 lock(&pp.timersLock)
3402 if len(pp.timers) > 0 {
3403 adjusttimers(pp, now)
3404 for len(pp.timers) > 0 {
3405 // Note that runtimer may temporarily unlock
3407 if tw := runtimer(pp, now); tw != 0 {
3417 // If this is the local P, and there are a lot of deleted timers,
3418 // clear them out. We only do this for the local P to reduce
3419 // lock contention on timersLock.
3420 if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
3421 clearDeletedTimers(pp)
3424 unlock(&pp.timersLock)
3426 return now, pollUntil, ran
3429 func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
3430 unlock((*mutex)(lock))
3434 // park continuation on g0.
3435 func park_m(gp *g) {
3439 traceGoPark(mp.waittraceev, mp.waittraceskip)
3442 // N.B. Not using casGToWaiting here because the waitreason is
3443 // set by park_m's caller.
3444 casgstatus(gp, _Grunning, _Gwaiting)
3447 if fn := mp.waitunlockf; fn != nil {
3448 ok := fn(gp, mp.waitlock)
3449 mp.waitunlockf = nil
3453 traceGoUnpark(gp, 2)
3455 casgstatus(gp, _Gwaiting, _Grunnable)
3456 execute(gp, true) // Schedule it back, never returns.
3462 func goschedImpl(gp *g) {
3463 status := readgstatus(gp)
3464 if status&^_Gscan != _Grunning {
3466 throw("bad g status")
3468 casgstatus(gp, _Grunning, _Grunnable)
3477 // Gosched continuation on g0.
3478 func gosched_m(gp *g) {
3485 // goschedguarded is a forbidden-states-avoided version of gosched_m
3486 func goschedguarded_m(gp *g) {
3488 if !canPreemptM(gp.m) {
3489 gogo(&gp.sched) // never return
3498 func gopreempt_m(gp *g) {
3505 // preemptPark parks gp and puts it in _Gpreempted.
3508 func preemptPark(gp *g) {
3510 traceGoPark(traceEvGoBlock, 0)
3512 status := readgstatus(gp)
3513 if status&^_Gscan != _Grunning {
3515 throw("bad g status")
3518 if gp.asyncSafePoint {
3519 // Double-check that async preemption does not
3520 // happen in SPWRITE assembly functions.
3521 // isAsyncSafePoint must exclude this case.
3522 f := findfunc(gp.sched.pc)
3524 throw("preempt at unknown pc")
3526 if f.flag&funcFlag_SPWRITE != 0 {
3527 println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
3528 throw("preempt SPWRITE")
3532 // Transition from _Grunning to _Gscan|_Gpreempted. We can't
3533 // be in _Grunning when we dropg because then we'd be running
3534 // without an M, but the moment we're in _Gpreempted,
3535 // something could claim this G before we've fully cleaned it
3536 // up. Hence, we set the scan bit to lock down further
3537 // transitions until we can dropg.
3538 casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
3540 casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
3544 // goyield is like Gosched, but it:
3545 // - emits a GoPreempt trace event instead of a GoSched trace event
3546 // - puts the current G on the runq of the current P instead of the globrunq
3552 func goyield_m(gp *g) {
3557 casgstatus(gp, _Grunning, _Grunnable)
3559 runqput(pp, gp, false)
3563 // Finishes execution of the current goroutine.
3574 // goexit continuation on g0.
3575 func goexit0(gp *g) {
3579 casgstatus(gp, _Grunning, _Gdead)
3580 gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
3581 if isSystemGoroutine(gp, false) {
3585 locked := gp.lockedm != 0
3588 gp.preemptStop = false
3589 gp.paniconfault = false
3590 gp._defer = nil // should be true already but just in case.
3591 gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
3593 gp.waitreason = waitReasonZero
3598 if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
3599 // Flush assist credit to the global pool. This gives
3600 // better information to pacing if the application is
3601 // rapidly creating an exiting goroutines.
3602 assistWorkPerByte := gcController.assistWorkPerByte.Load()
3603 scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
3604 gcController.bgScanCredit.Add(scanCredit)
3605 gp.gcAssistBytes = 0
3610 if GOARCH == "wasm" { // no threads yet on wasm
3612 schedule() // never returns
3615 if mp.lockedInt != 0 {
3616 print("invalid m->lockedInt = ", mp.lockedInt, "\n")
3617 throw("internal lockOSThread error")
3621 // The goroutine may have locked this thread because
3622 // it put it in an unusual kernel state. Kill it
3623 // rather than returning it to the thread pool.
3625 // Return to mstart, which will release the P and exit
3627 if GOOS != "plan9" { // See golang.org/issue/22227.
3630 // Clear lockedExt on plan9 since we may end up re-using
3638 // save updates getg().sched to refer to pc and sp so that a following
3639 // gogo will restore pc and sp.
3641 // save must not have write barriers because invoking a write barrier
3642 // can clobber getg().sched.
3645 //go:nowritebarrierrec
3646 func save(pc, sp uintptr) {
3649 if gp == gp.m.g0 || gp == gp.m.gsignal {
3650 // m.g0.sched is special and must describe the context
3651 // for exiting the thread. mstart1 writes to it directly.
3652 // m.gsignal.sched should not be used at all.
3653 // This check makes sure save calls do not accidentally
3654 // run in contexts where they'd write to system g's.
3655 throw("save on system g not allowed")
3662 // We need to ensure ctxt is zero, but can't have a write
3663 // barrier here. However, it should always already be zero.
3665 if gp.sched.ctxt != nil {
3670 // The goroutine g is about to enter a system call.
3671 // Record that it's not using the cpu anymore.
3672 // This is called only from the go syscall library and cgocall,
3673 // not from the low-level system calls used by the runtime.
3675 // Entersyscall cannot split the stack: the save must
3676 // make g->sched refer to the caller's stack segment, because
3677 // entersyscall is going to return immediately after.
3679 // Nothing entersyscall calls can split the stack either.
3680 // We cannot safely move the stack during an active call to syscall,
3681 // because we do not know which of the uintptr arguments are
3682 // really pointers (back into the stack).
3683 // In practice, this means that we make the fast path run through
3684 // entersyscall doing no-split things, and the slow path has to use systemstack
3685 // to run bigger things on the system stack.
3687 // reentersyscall is the entry point used by cgo callbacks, where explicitly
3688 // saved SP and PC are restored. This is needed when exitsyscall will be called
3689 // from a function further up in the call stack than the parent, as g->syscallsp
3690 // must always point to a valid stack frame. entersyscall below is the normal
3691 // entry point for syscalls, which obtains the SP and PC from the caller.
3694 // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
3695 // If the syscall does not block, that is it, we do not emit any other events.
3696 // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
3697 // when syscall returns we emit traceGoSysExit and when the goroutine starts running
3698 // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
3699 // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
3700 // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
3701 // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
3702 // and we wait for the increment before emitting traceGoSysExit.
3703 // Note that the increment is done even if tracing is not enabled,
3704 // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
3707 func reentersyscall(pc, sp uintptr) {
3710 // Disable preemption because during this function g is in Gsyscall status,
3711 // but can have inconsistent g->sched, do not let GC observe it.
3714 // Entersyscall must not call any function that might split/grow the stack.
3715 // (See details in comment above.)
3716 // Catch calls that might, by replacing the stack guard with something that
3717 // will trip any stack check and leaving a flag to tell newstack to die.
3718 gp.stackguard0 = stackPreempt
3719 gp.throwsplit = true
3721 // Leave SP around for GC and traceback.
3725 casgstatus(gp, _Grunning, _Gsyscall)
3726 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3727 systemstack(func() {
3728 print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3729 throw("entersyscall")
3734 systemstack(traceGoSysCall)
3735 // systemstack itself clobbers g.sched.{pc,sp} and we might
3736 // need them later when the G is genuinely blocked in a
3741 if sched.sysmonwait.Load() {
3742 systemstack(entersyscall_sysmon)
3746 if gp.m.p.ptr().runSafePointFn != 0 {
3747 // runSafePointFn may stack split if run on this stack
3748 systemstack(runSafePointFn)
3752 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3753 gp.sysblocktraced = true
3758 atomic.Store(&pp.status, _Psyscall)
3759 if sched.gcwaiting.Load() {
3760 systemstack(entersyscall_gcwait)
3767 // Standard syscall entry used by the go syscall library and normal cgo calls.
3769 // This is exported via linkname to assembly in the syscall package and x/sys.
3772 //go:linkname entersyscall
3773 func entersyscall() {
3774 reentersyscall(getcallerpc(), getcallersp())
3777 func entersyscall_sysmon() {
3779 if sched.sysmonwait.Load() {
3780 sched.sysmonwait.Store(false)
3781 notewakeup(&sched.sysmonnote)
3786 func entersyscall_gcwait() {
3788 pp := gp.m.oldp.ptr()
3791 if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
3797 if sched.stopwait--; sched.stopwait == 0 {
3798 notewakeup(&sched.stopnote)
3804 // The same as entersyscall(), but with a hint that the syscall is blocking.
3807 func entersyscallblock() {
3810 gp.m.locks++ // see comment in entersyscall
3811 gp.throwsplit = true
3812 gp.stackguard0 = stackPreempt // see comment in entersyscall
3813 gp.m.syscalltick = gp.m.p.ptr().syscalltick
3814 gp.sysblocktraced = true
3815 gp.m.p.ptr().syscalltick++
3817 // Leave SP around for GC and traceback.
3821 gp.syscallsp = gp.sched.sp
3822 gp.syscallpc = gp.sched.pc
3823 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3827 systemstack(func() {
3828 print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3829 throw("entersyscallblock")
3832 casgstatus(gp, _Grunning, _Gsyscall)
3833 if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
3834 systemstack(func() {
3835 print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
3836 throw("entersyscallblock")
3840 systemstack(entersyscallblock_handoff)
3842 // Resave for traceback during blocked call.
3843 save(getcallerpc(), getcallersp())
3848 func entersyscallblock_handoff() {
3851 traceGoSysBlock(getg().m.p.ptr())
3853 handoffp(releasep())
3856 // The goroutine g exited its system call.
3857 // Arrange for it to run on a cpu again.
3858 // This is called only from the go syscall library, not
3859 // from the low-level system calls used by the runtime.
3861 // Write barriers are not allowed because our P may have been stolen.
3863 // This is exported via linkname to assembly in the syscall package.
3866 //go:nowritebarrierrec
3867 //go:linkname exitsyscall
3868 func exitsyscall() {
3871 gp.m.locks++ // see comment in entersyscall
3872 if getcallersp() > gp.syscallsp {
3873 throw("exitsyscall: syscall frame is no longer valid")
3877 oldp := gp.m.oldp.ptr()
3879 if exitsyscallfast(oldp) {
3880 // When exitsyscallfast returns success, we have a P so can now use
3882 if goroutineProfile.active {
3883 // Make sure that gp has had its stack written out to the goroutine
3884 // profile, exactly as it was when the goroutine profiler first
3885 // stopped the world.
3886 systemstack(func() {
3887 tryRecordGoroutineProfileWB(gp)
3891 if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
3892 systemstack(traceGoStart)
3895 // There's a cpu for us, so we can run.
3896 gp.m.p.ptr().syscalltick++
3897 // We need to cas the status and scan before resuming...
3898 casgstatus(gp, _Gsyscall, _Grunning)
3900 // Garbage collector isn't running (since we are),
3901 // so okay to clear syscallsp.
3905 // restore the preemption request in case we've cleared it in newstack
3906 gp.stackguard0 = stackPreempt
3908 // otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
3909 gp.stackguard0 = gp.stack.lo + _StackGuard
3911 gp.throwsplit = false
3913 if sched.disable.user && !schedEnabled(gp) {
3914 // Scheduling of this goroutine is disabled.
3923 // Wait till traceGoSysBlock event is emitted.
3924 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3925 for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
3928 // We can't trace syscall exit right now because we don't have a P.
3929 // Tracing code can invoke write barriers that cannot run without a P.
3930 // So instead we remember the syscall exit time and emit the event
3931 // in execute when we have a P.
3932 gp.sysexitticks = cputicks()
3937 // Call the scheduler.
3940 // Scheduler returned, so we're allowed to run now.
3941 // Delete the syscallsp information that we left for
3942 // the garbage collector during the system call.
3943 // Must wait until now because until gosched returns
3944 // we don't know for sure that the garbage collector
3947 gp.m.p.ptr().syscalltick++
3948 gp.throwsplit = false
3952 func exitsyscallfast(oldp *p) bool {
3955 // Freezetheworld sets stopwait but does not retake P's.
3956 if sched.stopwait == freezeStopWait {
3960 // Try to re-acquire the last P.
3961 if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
3962 // There's a cpu for us, so we can run.
3964 exitsyscallfast_reacquired()
3968 // Try to get any other idle P.
3969 if sched.pidle != 0 {
3971 systemstack(func() {
3972 ok = exitsyscallfast_pidle()
3973 if ok && trace.enabled {
3975 // Wait till traceGoSysBlock event is emitted.
3976 // This ensures consistency of the trace (the goroutine is started after it is blocked).
3977 for oldp.syscalltick == gp.m.syscalltick {
3991 // exitsyscallfast_reacquired is the exitsyscall path on which this G
3992 // has successfully reacquired the P it was running on before the
3996 func exitsyscallfast_reacquired() {
3998 if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
4000 // The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
4001 // traceGoSysBlock for this syscall was already emitted,
4002 // but here we effectively retake the p from the new syscall running on the same p.
4003 systemstack(func() {
4004 // Denote blocking of the new syscall.
4005 traceGoSysBlock(gp.m.p.ptr())
4006 // Denote completion of the current syscall.
4010 gp.m.p.ptr().syscalltick++
4014 func exitsyscallfast_pidle() bool {
4016 pp, _ := pidleget(0)
4017 if pp != nil && sched.sysmonwait.Load() {
4018 sched.sysmonwait.Store(false)
4019 notewakeup(&sched.sysmonnote)
4029 // exitsyscall slow path on g0.
4030 // Failed to acquire P, enqueue gp as runnable.
4032 // Called via mcall, so gp is the calling g from this M.
4034 //go:nowritebarrierrec
4035 func exitsyscall0(gp *g) {
4036 casgstatus(gp, _Gsyscall, _Grunnable)
4040 if schedEnabled(gp) {
4047 // Below, we stoplockedm if gp is locked. globrunqput releases
4048 // ownership of gp, so we must check if gp is locked prior to
4049 // committing the release by unlocking sched.lock, otherwise we
4050 // could race with another M transitioning gp from unlocked to
4052 locked = gp.lockedm != 0
4053 } else if sched.sysmonwait.Load() {
4054 sched.sysmonwait.Store(false)
4055 notewakeup(&sched.sysmonnote)
4060 execute(gp, false) // Never returns.
4063 // Wait until another thread schedules gp and so m again.
4065 // N.B. lockedm must be this M, as this g was running on this M
4066 // before entersyscall.
4068 execute(gp, false) // Never returns.
4071 schedule() // Never returns.
4074 // Called from syscall package before fork.
4076 //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
4078 func syscall_runtime_BeforeFork() {
4081 // Block signals during a fork, so that the child does not run
4082 // a signal handler before exec if a signal is sent to the process
4083 // group. See issue #18600.
4085 sigsave(&gp.m.sigmask)
4088 // This function is called before fork in syscall package.
4089 // Code between fork and exec must not allocate memory nor even try to grow stack.
4090 // Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
4091 // runtime_AfterFork will undo this in parent process, but not in child.
4092 gp.stackguard0 = stackFork
4095 // Called from syscall package after fork in parent.
4097 //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
4099 func syscall_runtime_AfterFork() {
4102 // See the comments in beforefork.
4103 gp.stackguard0 = gp.stack.lo + _StackGuard
4105 msigrestore(gp.m.sigmask)
4110 // inForkedChild is true while manipulating signals in the child process.
4111 // This is used to avoid calling libc functions in case we are using vfork.
4112 var inForkedChild bool
4114 // Called from syscall package after fork in child.
4115 // It resets non-sigignored signals to the default handler, and
4116 // restores the signal mask in preparation for the exec.
4118 // Because this might be called during a vfork, and therefore may be
4119 // temporarily sharing address space with the parent process, this must
4120 // not change any global variables or calling into C code that may do so.
4122 //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
4124 //go:nowritebarrierrec
4125 func syscall_runtime_AfterForkInChild() {
4126 // It's OK to change the global variable inForkedChild here
4127 // because we are going to change it back. There is no race here,
4128 // because if we are sharing address space with the parent process,
4129 // then the parent process can not be running concurrently.
4130 inForkedChild = true
4132 clearSignalHandlers()
4134 // When we are the child we are the only thread running,
4135 // so we know that nothing else has changed gp.m.sigmask.
4136 msigrestore(getg().m.sigmask)
4138 inForkedChild = false
4141 // pendingPreemptSignals is the number of preemption signals
4142 // that have been sent but not received. This is only used on Darwin.
4144 var pendingPreemptSignals atomic.Int32
4146 // Called from syscall package before Exec.
4148 //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
4149 func syscall_runtime_BeforeExec() {
4150 // Prevent thread creation during exec.
4153 // On Darwin, wait for all pending preemption signals to
4154 // be received. See issue #41702.
4155 if GOOS == "darwin" || GOOS == "ios" {
4156 for pendingPreemptSignals.Load() > 0 {
4162 // Called from syscall package after Exec.
4164 //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
4165 func syscall_runtime_AfterExec() {
4169 // Allocate a new g, with a stack big enough for stacksize bytes.
4170 func malg(stacksize int32) *g {
4173 stacksize = round2(_StackSystem + stacksize)
4174 systemstack(func() {
4175 newg.stack = stackalloc(uint32(stacksize))
4177 newg.stackguard0 = newg.stack.lo + _StackGuard
4178 newg.stackguard1 = ^uintptr(0)
4179 // Clear the bottom word of the stack. We record g
4180 // there on gsignal stack during VDSO on ARM and ARM64.
4181 *(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
4186 // Create a new g running fn.
4187 // Put it on the queue of g's waiting to run.
4188 // The compiler turns a go statement into a call to this.
4189 func newproc(fn *funcval) {
4192 systemstack(func() {
4193 newg := newproc1(fn, gp, pc)
4195 pp := getg().m.p.ptr()
4196 runqput(pp, newg, true)
4204 // Create a new g in state _Grunnable, starting at fn. callerpc is the
4205 // address of the go statement that created this. The caller is responsible
4206 // for adding the new g to the scheduler.
4207 func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
4209 fatal("go of nil func value")
4212 mp := acquirem() // disable preemption because we hold M and P in local vars.
4216 newg = malg(_StackMin)
4217 casgstatus(newg, _Gidle, _Gdead)
4218 allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
4220 if newg.stack.hi == 0 {
4221 throw("newproc1: newg missing stack")
4224 if readgstatus(newg) != _Gdead {
4225 throw("newproc1: new g is not Gdead")
4228 totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
4229 totalSize = alignUp(totalSize, sys.StackAlign)
4230 sp := newg.stack.hi - totalSize
4234 *(*uintptr)(unsafe.Pointer(sp)) = 0
4236 spArg += sys.MinFrameSize
4239 memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
4242 newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
4243 newg.sched.g = guintptr(unsafe.Pointer(newg))
4244 gostartcallfn(&newg.sched, fn)
4245 newg.gopc = callerpc
4246 newg.ancestors = saveAncestors(callergp)
4247 newg.startpc = fn.fn
4248 if isSystemGoroutine(newg, false) {
4251 // Only user goroutines inherit pprof labels.
4253 newg.labels = mp.curg.labels
4255 if goroutineProfile.active {
4256 // A concurrent goroutine profile is running. It should include
4257 // exactly the set of goroutines that were alive when the goroutine
4258 // profiler first stopped the world. That does not include newg, so
4259 // mark it as not needing a profile before transitioning it from
4261 newg.goroutineProfiled.Store(goroutineProfileSatisfied)
4264 // Track initial transition?
4265 newg.trackingSeq = uint8(fastrand())
4266 if newg.trackingSeq%gTrackingPeriod == 0 {
4267 newg.tracking = true
4269 casgstatus(newg, _Gdead, _Grunnable)
4270 gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
4272 if pp.goidcache == pp.goidcacheend {
4273 // Sched.goidgen is the last allocated id,
4274 // this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
4275 // At startup sched.goidgen=0, so main goroutine receives goid=1.
4276 pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
4277 pp.goidcache -= _GoidCacheBatch - 1
4278 pp.goidcacheend = pp.goidcache + _GoidCacheBatch
4280 newg.goid = pp.goidcache
4283 newg.racectx = racegostart(callerpc)
4284 if newg.labels != nil {
4285 // See note in proflabel.go on labelSync's role in synchronizing
4286 // with the reads in the signal handler.
4287 racereleasemergeg(newg, unsafe.Pointer(&labelSync))
4291 traceGoCreate(newg, newg.startpc)
4298 // saveAncestors copies previous ancestors of the given caller g and
4299 // includes infor for the current caller into a new set of tracebacks for
4300 // a g being created.
4301 func saveAncestors(callergp *g) *[]ancestorInfo {
4302 // Copy all prior info, except for the root goroutine (goid 0).
4303 if debug.tracebackancestors <= 0 || callergp.goid == 0 {
4306 var callerAncestors []ancestorInfo
4307 if callergp.ancestors != nil {
4308 callerAncestors = *callergp.ancestors
4310 n := int32(len(callerAncestors)) + 1
4311 if n > debug.tracebackancestors {
4312 n = debug.tracebackancestors
4314 ancestors := make([]ancestorInfo, n)
4315 copy(ancestors[1:], callerAncestors)
4317 var pcs [_TracebackMaxFrames]uintptr
4318 npcs := gcallers(callergp, 0, pcs[:])
4319 ipcs := make([]uintptr, npcs)
4321 ancestors[0] = ancestorInfo{
4323 goid: callergp.goid,
4324 gopc: callergp.gopc,
4327 ancestorsp := new([]ancestorInfo)
4328 *ancestorsp = ancestors
4332 // Put on gfree list.
4333 // If local list is too long, transfer a batch to the global list.
4334 func gfput(pp *p, gp *g) {
4335 if readgstatus(gp) != _Gdead {
4336 throw("gfput: bad status (not Gdead)")
4339 stksize := gp.stack.hi - gp.stack.lo
4341 if stksize != uintptr(startingStackSize) {
4342 // non-standard stack size - free it.
4351 if pp.gFree.n >= 64 {
4357 for pp.gFree.n >= 32 {
4358 gp := pp.gFree.pop()
4360 if gp.stack.lo == 0 {
4367 lock(&sched.gFree.lock)
4368 sched.gFree.noStack.pushAll(noStackQ)
4369 sched.gFree.stack.pushAll(stackQ)
4370 sched.gFree.n += inc
4371 unlock(&sched.gFree.lock)
4375 // Get from gfree list.
4376 // If local list is empty, grab a batch from global list.
4377 func gfget(pp *p) *g {
4379 if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
4380 lock(&sched.gFree.lock)
4381 // Move a batch of free Gs to the P.
4382 for pp.gFree.n < 32 {
4383 // Prefer Gs with stacks.
4384 gp := sched.gFree.stack.pop()
4386 gp = sched.gFree.noStack.pop()
4395 unlock(&sched.gFree.lock)
4398 gp := pp.gFree.pop()
4403 if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
4404 // Deallocate old stack. We kept it in gfput because it was the
4405 // right size when the goroutine was put on the free list, but
4406 // the right size has changed since then.
4407 systemstack(func() {
4414 if gp.stack.lo == 0 {
4415 // Stack was deallocated in gfput or just above. Allocate a new one.
4416 systemstack(func() {
4417 gp.stack = stackalloc(startingStackSize)
4419 gp.stackguard0 = gp.stack.lo + _StackGuard
4422 racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4425 msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4428 asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
4434 // Purge all cached G's from gfree list to the global list.
4435 func gfpurge(pp *p) {
4441 for !pp.gFree.empty() {
4442 gp := pp.gFree.pop()
4444 if gp.stack.lo == 0 {
4451 lock(&sched.gFree.lock)
4452 sched.gFree.noStack.pushAll(noStackQ)
4453 sched.gFree.stack.pushAll(stackQ)
4454 sched.gFree.n += inc
4455 unlock(&sched.gFree.lock)
4458 // Breakpoint executes a breakpoint trap.
4463 // dolockOSThread is called by LockOSThread and lockOSThread below
4464 // after they modify m.locked. Do not allow preemption during this call,
4465 // or else the m might be different in this function than in the caller.
4468 func dolockOSThread() {
4469 if GOARCH == "wasm" {
4470 return // no threads on wasm yet
4473 gp.m.lockedg.set(gp)
4474 gp.lockedm.set(gp.m)
4479 // LockOSThread wires the calling goroutine to its current operating system thread.
4480 // The calling goroutine will always execute in that thread,
4481 // and no other goroutine will execute in it,
4482 // until the calling goroutine has made as many calls to
4483 // UnlockOSThread as to LockOSThread.
4484 // If the calling goroutine exits without unlocking the thread,
4485 // the thread will be terminated.
4487 // All init functions are run on the startup thread. Calling LockOSThread
4488 // from an init function will cause the main function to be invoked on
4491 // A goroutine should call LockOSThread before calling OS services or
4492 // non-Go library functions that depend on per-thread state.
4493 func LockOSThread() {
4494 if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
4495 // If we need to start a new thread from the locked
4496 // thread, we need the template thread. Start it now
4497 // while we're in a known-good state.
4498 startTemplateThread()
4502 if gp.m.lockedExt == 0 {
4504 panic("LockOSThread nesting overflow")
4510 func lockOSThread() {
4511 getg().m.lockedInt++
4515 // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
4516 // after they update m->locked. Do not allow preemption during this call,
4517 // or else the m might be in different in this function than in the caller.
4520 func dounlockOSThread() {
4521 if GOARCH == "wasm" {
4522 return // no threads on wasm yet
4525 if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
4534 // UnlockOSThread undoes an earlier call to LockOSThread.
4535 // If this drops the number of active LockOSThread calls on the
4536 // calling goroutine to zero, it unwires the calling goroutine from
4537 // its fixed operating system thread.
4538 // If there are no active LockOSThread calls, this is a no-op.
4540 // Before calling UnlockOSThread, the caller must ensure that the OS
4541 // thread is suitable for running other goroutines. If the caller made
4542 // any permanent changes to the state of the thread that would affect
4543 // other goroutines, it should not call this function and thus leave
4544 // the goroutine locked to the OS thread until the goroutine (and
4545 // hence the thread) exits.
4546 func UnlockOSThread() {
4548 if gp.m.lockedExt == 0 {
4556 func unlockOSThread() {
4558 if gp.m.lockedInt == 0 {
4559 systemstack(badunlockosthread)
4565 func badunlockosthread() {
4566 throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
4569 func gcount() int32 {
4570 n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
4571 for _, pp := range allp {
4575 // All these variables can be changed concurrently, so the result can be inconsistent.
4576 // But at least the current goroutine is running.
4583 func mcount() int32 {
4584 return int32(sched.mnext - sched.nmfreed)
4588 signalLock atomic.Uint32
4590 // Must hold signalLock to write. Reads may be lock-free, but
4591 // signalLock should be taken to synchronize with changes.
4595 func _System() { _System() }
4596 func _ExternalCode() { _ExternalCode() }
4597 func _LostExternalCode() { _LostExternalCode() }
4598 func _GC() { _GC() }
4599 func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
4600 func _VDSO() { _VDSO() }
4602 // Called if we receive a SIGPROF signal.
4603 // Called by the signal handler, may run during STW.
4605 //go:nowritebarrierrec
4606 func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
4607 if prof.hz.Load() == 0 {
4611 // If mp.profilehz is 0, then profiling is not enabled for this thread.
4612 // We must check this to avoid a deadlock between setcpuprofilerate
4613 // and the call to cpuprof.add, below.
4614 if mp != nil && mp.profilehz == 0 {
4618 // On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
4619 // runtime/internal/atomic. If SIGPROF arrives while the program is inside
4620 // the critical section, it creates a deadlock (when writing the sample).
4621 // As a workaround, create a counter of SIGPROFs while in critical section
4622 // to store the count, and pass it to sigprof.add() later when SIGPROF is
4623 // received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
4624 if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
4625 if f := findfunc(pc); f.valid() {
4626 if hasPrefix(funcname(f), "runtime/internal/atomic") {
4627 cpuprof.lostAtomic++
4631 if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
4632 // runtime/internal/atomic functions call into kernel
4633 // helpers on arm < 7. See
4634 // runtime/internal/atomic/sys_linux_arm.s.
4635 cpuprof.lostAtomic++
4640 // Profiling runs concurrently with GC, so it must not allocate.
4641 // Set a trap in case the code does allocate.
4642 // Note that on windows, one thread takes profiles of all the
4643 // other threads, so mp is usually not getg().m.
4644 // In fact mp may not even be stopped.
4645 // See golang.org/issue/17165.
4646 getg().m.mallocing++
4648 var stk [maxCPUProfStack]uintptr
4650 if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
4652 // Check cgoCallersUse to make sure that we are not
4653 // interrupting other code that is fiddling with
4654 // cgoCallers. We are running in a signal handler
4655 // with all signals blocked, so we don't have to worry
4656 // about any other code interrupting us.
4657 if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
4658 for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
4661 copy(stk[:], mp.cgoCallers[:cgoOff])
4662 mp.cgoCallers[0] = 0
4665 // Collect Go stack that leads to the cgo call.
4666 n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
4671 n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4675 // Normal traceback is impossible or has failed.
4676 // See if it falls into several common cases.
4678 if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
4679 // Libcall, i.e. runtime syscall on windows.
4680 // Collect Go stack that leads to the call.
4681 n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[0], len(stk), nil, nil, 0)
4683 if n == 0 && mp != nil && mp.vdsoSP != 0 {
4684 n = gentraceback(mp.vdsoPC, mp.vdsoSP, 0, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
4687 // If all of the above has failed, account it against abstract "System" or "GC".
4690 pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
4691 } else if pc > firstmoduledata.etext {
4692 // "ExternalCode" is better than "etext".
4693 pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
4696 if mp.preemptoff != "" {
4697 stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
4699 stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
4704 if prof.hz.Load() != 0 {
4705 // Note: it can happen on Windows that we interrupted a system thread
4706 // with no g, so gp could nil. The other nil checks are done out of
4707 // caution, but not expected to be nil in practice.
4708 var tagPtr *unsafe.Pointer
4709 if gp != nil && gp.m != nil && gp.m.curg != nil {
4710 tagPtr = &gp.m.curg.labels
4712 cpuprof.add(tagPtr, stk[:n])
4716 if gp != nil && gp.m != nil {
4717 if gp.m.curg != nil {
4722 traceCPUSample(gprof, pp, stk[:n])
4724 getg().m.mallocing--
4727 // setcpuprofilerate sets the CPU profiling rate to hz times per second.
4728 // If hz <= 0, setcpuprofilerate turns off CPU profiling.
4729 func setcpuprofilerate(hz int32) {
4730 // Force sane arguments.
4735 // Disable preemption, otherwise we can be rescheduled to another thread
4736 // that has profiling enabled.
4740 // Stop profiler on this thread so that it is safe to lock prof.
4741 // if a profiling signal came in while we had prof locked,
4742 // it would deadlock.
4743 setThreadCPUProfiler(0)
4745 for !prof.signalLock.CompareAndSwap(0, 1) {
4748 if prof.hz.Load() != hz {
4749 setProcessCPUProfiler(hz)
4752 prof.signalLock.Store(0)
4755 sched.profilehz = hz
4759 setThreadCPUProfiler(hz)
4765 // init initializes pp, which may be a freshly allocated p or a
4766 // previously destroyed p, and transitions it to status _Pgcstop.
4767 func (pp *p) init(id int32) {
4769 pp.status = _Pgcstop
4770 pp.sudogcache = pp.sudogbuf[:0]
4771 pp.deferpool = pp.deferpoolbuf[:0]
4773 if pp.mcache == nil {
4776 throw("missing mcache?")
4778 // Use the bootstrap mcache0. Only one P will get
4779 // mcache0: the one with ID 0.
4782 pp.mcache = allocmcache()
4785 if raceenabled && pp.raceprocctx == 0 {
4787 pp.raceprocctx = raceprocctx0
4788 raceprocctx0 = 0 // bootstrap
4790 pp.raceprocctx = raceproccreate()
4793 lockInit(&pp.timersLock, lockRankTimers)
4795 // This P may get timers when it starts running. Set the mask here
4796 // since the P may not go through pidleget (notably P 0 on startup).
4798 // Similarly, we may not go through pidleget before this P starts
4799 // running if it is P 0 on startup.
4803 // destroy releases all of the resources associated with pp and
4804 // transitions it to status _Pdead.
4806 // sched.lock must be held and the world must be stopped.
4807 func (pp *p) destroy() {
4808 assertLockHeld(&sched.lock)
4809 assertWorldStopped()
4811 // Move all runnable goroutines to the global queue
4812 for pp.runqhead != pp.runqtail {
4813 // Pop from tail of local queue
4815 gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
4816 // Push onto head of global queue
4819 if pp.runnext != 0 {
4820 globrunqputhead(pp.runnext.ptr())
4823 if len(pp.timers) > 0 {
4824 plocal := getg().m.p.ptr()
4825 // The world is stopped, but we acquire timersLock to
4826 // protect against sysmon calling timeSleepUntil.
4827 // This is the only case where we hold the timersLock of
4828 // more than one P, so there are no deadlock concerns.
4829 lock(&plocal.timersLock)
4830 lock(&pp.timersLock)
4831 moveTimers(plocal, pp.timers)
4833 pp.numTimers.Store(0)
4834 pp.deletedTimers.Store(0)
4835 pp.timer0When.Store(0)
4836 unlock(&pp.timersLock)
4837 unlock(&plocal.timersLock)
4839 // Flush p's write barrier buffer.
4840 if gcphase != _GCoff {
4844 for i := range pp.sudogbuf {
4845 pp.sudogbuf[i] = nil
4847 pp.sudogcache = pp.sudogbuf[:0]
4848 for j := range pp.deferpoolbuf {
4849 pp.deferpoolbuf[j] = nil
4851 pp.deferpool = pp.deferpoolbuf[:0]
4852 systemstack(func() {
4853 for i := 0; i < pp.mspancache.len; i++ {
4854 // Safe to call since the world is stopped.
4855 mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
4857 pp.mspancache.len = 0
4859 pp.pcache.flush(&mheap_.pages)
4860 unlock(&mheap_.lock)
4862 freemcache(pp.mcache)
4867 if pp.timerRaceCtx != 0 {
4868 // The race detector code uses a callback to fetch
4869 // the proc context, so arrange for that callback
4870 // to see the right thing.
4871 // This hack only works because we are the only
4877 racectxend(pp.timerRaceCtx)
4882 raceprocdestroy(pp.raceprocctx)
4889 // Change number of processors.
4891 // sched.lock must be held, and the world must be stopped.
4893 // gcworkbufs must not be being modified by either the GC or the write barrier
4894 // code, so the GC must not be running if the number of Ps actually changes.
4896 // Returns list of Ps with local work, they need to be scheduled by the caller.
4897 func procresize(nprocs int32) *p {
4898 assertLockHeld(&sched.lock)
4899 assertWorldStopped()
4902 if old < 0 || nprocs <= 0 {
4903 throw("procresize: invalid arg")
4906 traceGomaxprocs(nprocs)
4909 // update statistics
4911 if sched.procresizetime != 0 {
4912 sched.totaltime += int64(old) * (now - sched.procresizetime)
4914 sched.procresizetime = now
4916 maskWords := (nprocs + 31) / 32
4918 // Grow allp if necessary.
4919 if nprocs > int32(len(allp)) {
4920 // Synchronize with retake, which could be running
4921 // concurrently since it doesn't run on a P.
4923 if nprocs <= int32(cap(allp)) {
4924 allp = allp[:nprocs]
4926 nallp := make([]*p, nprocs)
4927 // Copy everything up to allp's cap so we
4928 // never lose old allocated Ps.
4929 copy(nallp, allp[:cap(allp)])
4933 if maskWords <= int32(cap(idlepMask)) {
4934 idlepMask = idlepMask[:maskWords]
4935 timerpMask = timerpMask[:maskWords]
4937 nidlepMask := make([]uint32, maskWords)
4938 // No need to copy beyond len, old Ps are irrelevant.
4939 copy(nidlepMask, idlepMask)
4940 idlepMask = nidlepMask
4942 ntimerpMask := make([]uint32, maskWords)
4943 copy(ntimerpMask, timerpMask)
4944 timerpMask = ntimerpMask
4949 // initialize new P's
4950 for i := old; i < nprocs; i++ {
4956 atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
4960 if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
4961 // continue to use the current P
4962 gp.m.p.ptr().status = _Prunning
4963 gp.m.p.ptr().mcache.prepareForSweep()
4965 // release the current P and acquire allp[0].
4967 // We must do this before destroying our current P
4968 // because p.destroy itself has write barriers, so we
4969 // need to do that from a valid P.
4972 // Pretend that we were descheduled
4973 // and then scheduled again to keep
4976 traceProcStop(gp.m.p.ptr())
4990 // g.m.p is now set, so we no longer need mcache0 for bootstrapping.
4993 // release resources from unused P's
4994 for i := nprocs; i < old; i++ {
4997 // can't free P itself because it can be referenced by an M in syscall
5001 if int32(len(allp)) != nprocs {
5003 allp = allp[:nprocs]
5004 idlepMask = idlepMask[:maskWords]
5005 timerpMask = timerpMask[:maskWords]
5010 for i := nprocs - 1; i >= 0; i-- {
5012 if gp.m.p.ptr() == pp {
5020 pp.link.set(runnablePs)
5024 stealOrder.reset(uint32(nprocs))
5025 var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
5026 atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
5028 // Notify the limiter that the amount of procs has changed.
5029 gcCPULimiter.resetCapacity(now, nprocs)
5034 // Associate p and the current m.
5036 // This function is allowed to have write barriers even if the caller
5037 // isn't because it immediately acquires pp.
5039 //go:yeswritebarrierrec
5040 func acquirep(pp *p) {
5041 // Do the part that isn't allowed to have write barriers.
5044 // Have p; write barriers now allowed.
5046 // Perform deferred mcache flush before this P can allocate
5047 // from a potentially stale mcache.
5048 pp.mcache.prepareForSweep()
5055 // wirep is the first step of acquirep, which actually associates the
5056 // current M to pp. This is broken out so we can disallow write
5057 // barriers for this part, since we don't yet have a P.
5059 //go:nowritebarrierrec
5065 throw("wirep: already in go")
5067 if pp.m != 0 || pp.status != _Pidle {
5072 print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
5073 throw("wirep: invalid p state")
5077 pp.status = _Prunning
5080 // Disassociate p and the current m.
5081 func releasep() *p {
5085 throw("releasep: invalid arg")
5088 if pp.m.ptr() != gp.m || pp.status != _Prunning {
5089 print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
5090 throw("releasep: invalid p state")
5093 traceProcStop(gp.m.p.ptr())
5101 func incidlelocked(v int32) {
5103 sched.nmidlelocked += v
5110 // Check for deadlock situation.
5111 // The check is based on number of running M's, if 0 -> deadlock.
5112 // sched.lock must be held.
5114 assertLockHeld(&sched.lock)
5116 // For -buildmode=c-shared or -buildmode=c-archive it's OK if
5117 // there are no running goroutines. The calling program is
5118 // assumed to be running.
5119 if islibrary || isarchive {
5123 // If we are dying because of a signal caught on an already idle thread,
5124 // freezetheworld will cause all running threads to block.
5125 // And runtime will essentially enter into deadlock state,
5126 // except that there is a thread that will call exit soon.
5127 if panicking.Load() > 0 {
5131 // If we are not running under cgo, but we have an extra M then account
5132 // for it. (It is possible to have an extra M on Windows without cgo to
5133 // accommodate callbacks created by syscall.NewCallback. See issue #6751
5136 if !iscgo && cgoHasExtraM {
5137 mp := lockextra(true)
5138 haveExtraM := extraMCount > 0
5145 run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
5150 print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
5151 throw("checkdead: inconsistent counts")
5155 forEachG(func(gp *g) {
5156 if isSystemGoroutine(gp, false) {
5159 s := readgstatus(gp)
5160 switch s &^ _Gscan {
5167 print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
5168 throw("checkdead: runnable g")
5171 if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
5172 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5173 fatal("no goroutines (main called runtime.Goexit) - deadlock!")
5176 // Maybe jump time forward for playground.
5178 if when := timeSleepUntil(); when < maxWhen {
5181 // Start an M to steal the timer.
5182 pp, _ := pidleget(faketime)
5184 // There should always be a free P since
5185 // nothing is running.
5186 throw("checkdead: no p for timer")
5190 // There should always be a free M since
5191 // nothing is running.
5192 throw("checkdead: no m for timer")
5194 // M must be spinning to steal. We set this to be
5195 // explicit, but since this is the only M it would
5196 // become spinning on its own anyways.
5197 sched.nmspinning.Add(1)
5200 notewakeup(&mp.park)
5205 // There are no goroutines running, so we can look at the P's.
5206 for _, pp := range allp {
5207 if len(pp.timers) > 0 {
5212 unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
5213 fatal("all goroutines are asleep - deadlock!")
5216 // forcegcperiod is the maximum time in nanoseconds between garbage
5217 // collections. If we go this long without a garbage collection, one
5218 // is forced to run.
5220 // This is a variable for testing purposes. It normally doesn't change.
5221 var forcegcperiod int64 = 2 * 60 * 1e9
5223 // needSysmonWorkaround is true if the workaround for
5224 // golang.org/issue/42515 is needed on NetBSD.
5225 var needSysmonWorkaround bool = false
5227 // Always runs without a P, so write barriers are not allowed.
5229 //go:nowritebarrierrec
5236 lasttrace := int64(0)
5237 idle := 0 // how many cycles in succession we had not wokeup somebody
5241 if idle == 0 { // start with 20us sleep...
5243 } else if idle > 50 { // start doubling the sleep after 1ms...
5246 if delay > 10*1000 { // up to 10ms
5251 // sysmon should not enter deep sleep if schedtrace is enabled so that
5252 // it can print that information at the right time.
5254 // It should also not enter deep sleep if there are any active P's so
5255 // that it can retake P's from syscalls, preempt long running G's, and
5256 // poll the network if all P's are busy for long stretches.
5258 // It should wakeup from deep sleep if any P's become active either due
5259 // to exiting a syscall or waking up due to a timer expiring so that it
5260 // can resume performing those duties. If it wakes from a syscall it
5261 // resets idle and delay as a bet that since it had retaken a P from a
5262 // syscall before, it may need to do it again shortly after the
5263 // application starts work again. It does not reset idle when waking
5264 // from a timer to avoid adding system load to applications that spend
5265 // most of their time sleeping.
5267 if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
5269 if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
5270 syscallWake := false
5271 next := timeSleepUntil()
5273 sched.sysmonwait.Store(true)
5275 // Make wake-up period small enough
5276 // for the sampling to be correct.
5277 sleep := forcegcperiod / 2
5278 if next-now < sleep {
5281 shouldRelax := sleep >= osRelaxMinNS
5285 syscallWake = notetsleep(&sched.sysmonnote, sleep)
5290 sched.sysmonwait.Store(false)
5291 noteclear(&sched.sysmonnote)
5301 lock(&sched.sysmonlock)
5302 // Update now in case we blocked on sysmonnote or spent a long time
5303 // blocked on schedlock or sysmonlock above.
5306 // trigger libc interceptors if needed
5307 if *cgo_yield != nil {
5308 asmcgocall(*cgo_yield, nil)
5310 // poll network if not polled for more than 10ms
5311 lastpoll := sched.lastpoll.Load()
5312 if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
5313 sched.lastpoll.CompareAndSwap(lastpoll, now)
5314 list := netpoll(0) // non-blocking - returns list of goroutines
5316 // Need to decrement number of idle locked M's
5317 // (pretending that one more is running) before injectglist.
5318 // Otherwise it can lead to the following situation:
5319 // injectglist grabs all P's but before it starts M's to run the P's,
5320 // another M returns from syscall, finishes running its G,
5321 // observes that there is no work to do and no other running M's
5322 // and reports deadlock.
5328 if GOOS == "netbsd" && needSysmonWorkaround {
5329 // netpoll is responsible for waiting for timer
5330 // expiration, so we typically don't have to worry
5331 // about starting an M to service timers. (Note that
5332 // sleep for timeSleepUntil above simply ensures sysmon
5333 // starts running again when that timer expiration may
5334 // cause Go code to run again).
5336 // However, netbsd has a kernel bug that sometimes
5337 // misses netpollBreak wake-ups, which can lead to
5338 // unbounded delays servicing timers. If we detect this
5339 // overrun, then startm to get something to handle the
5342 // See issue 42515 and
5343 // https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
5344 if next := timeSleepUntil(); next < now {
5348 if scavenger.sysmonWake.Load() != 0 {
5349 // Kick the scavenger awake if someone requested it.
5352 // retake P's blocked in syscalls
5353 // and preempt long running G's
5354 if retake(now) != 0 {
5359 // check if we need to force a GC
5360 if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
5362 forcegc.idle.Store(false)
5364 list.push(forcegc.g)
5366 unlock(&forcegc.lock)
5368 if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
5370 schedtrace(debug.scheddetail > 0)
5372 unlock(&sched.sysmonlock)
5376 type sysmontick struct {
5383 // forcePreemptNS is the time slice given to a G before it is
5385 const forcePreemptNS = 10 * 1000 * 1000 // 10ms
5387 func retake(now int64) uint32 {
5389 // Prevent allp slice changes. This lock will be completely
5390 // uncontended unless we're already stopping the world.
5392 // We can't use a range loop over allp because we may
5393 // temporarily drop the allpLock. Hence, we need to re-fetch
5394 // allp each time around the loop.
5395 for i := 0; i < len(allp); i++ {
5398 // This can happen if procresize has grown
5399 // allp but not yet created new Ps.
5402 pd := &pp.sysmontick
5405 if s == _Prunning || s == _Psyscall {
5406 // Preempt G if it's running for too long.
5407 t := int64(pp.schedtick)
5408 if int64(pd.schedtick) != t {
5409 pd.schedtick = uint32(t)
5411 } else if pd.schedwhen+forcePreemptNS <= now {
5413 // In case of syscall, preemptone() doesn't
5414 // work, because there is no M wired to P.
5419 // Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
5420 t := int64(pp.syscalltick)
5421 if !sysretake && int64(pd.syscalltick) != t {
5422 pd.syscalltick = uint32(t)
5423 pd.syscallwhen = now
5426 // On the one hand we don't want to retake Ps if there is no other work to do,
5427 // but on the other hand we want to retake them eventually
5428 // because they can prevent the sysmon thread from deep sleep.
5429 if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
5432 // Drop allpLock so we can take sched.lock.
5434 // Need to decrement number of idle locked M's
5435 // (pretending that one more is running) before the CAS.
5436 // Otherwise the M from which we retake can exit the syscall,
5437 // increment nmidle and report deadlock.
5439 if atomic.Cas(&pp.status, s, _Pidle) {
5456 // Tell all goroutines that they have been preempted and they should stop.
5457 // This function is purely best-effort. It can fail to inform a goroutine if a
5458 // processor just started running it.
5459 // No locks need to be held.
5460 // Returns true if preemption request was issued to at least one goroutine.
5461 func preemptall() bool {
5463 for _, pp := range allp {
5464 if pp.status != _Prunning {
5474 // Tell the goroutine running on processor P to stop.
5475 // This function is purely best-effort. It can incorrectly fail to inform the
5476 // goroutine. It can inform the wrong goroutine. Even if it informs the
5477 // correct goroutine, that goroutine might ignore the request if it is
5478 // simultaneously executing newstack.
5479 // No lock needs to be held.
5480 // Returns true if preemption request was issued.
5481 // The actual preemption will happen at some point in the future
5482 // and will be indicated by the gp->status no longer being
5484 func preemptone(pp *p) bool {
5486 if mp == nil || mp == getg().m {
5490 if gp == nil || gp == mp.g0 {
5496 // Every call in a goroutine checks for stack overflow by
5497 // comparing the current stack pointer to gp->stackguard0.
5498 // Setting gp->stackguard0 to StackPreempt folds
5499 // preemption into the normal stack overflow check.
5500 gp.stackguard0 = stackPreempt
5502 // Request an async preemption of this P.
5503 if preemptMSupported && debug.asyncpreemptoff == 0 {
5513 func schedtrace(detailed bool) {
5520 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)
5522 print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
5524 // We must be careful while reading data from P's, M's and G's.
5525 // Even if we hold schedlock, most data can be changed concurrently.
5526 // E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
5527 for i, pp := range allp {
5529 h := atomic.Load(&pp.runqhead)
5530 t := atomic.Load(&pp.runqtail)
5532 print(" P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
5538 print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
5540 // In non-detailed mode format lengths of per-P run queues as:
5541 // [len1 len2 len3 len4]
5547 if i == len(allp)-1 {
5558 for mp := allm; mp != nil; mp = mp.alllink {
5560 print(" M", mp.id, ": p=")
5572 print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
5573 if lockedg := mp.lockedg.ptr(); lockedg != nil {
5581 forEachG(func(gp *g) {
5582 print(" G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
5589 if lockedm := gp.lockedm.ptr(); lockedm != nil {
5599 // schedEnableUser enables or disables the scheduling of user
5602 // This does not stop already running user goroutines, so the caller
5603 // should first stop the world when disabling user goroutines.
5604 func schedEnableUser(enable bool) {
5606 if sched.disable.user == !enable {
5610 sched.disable.user = !enable
5612 n := sched.disable.n
5614 globrunqputbatch(&sched.disable.runnable, n)
5616 for ; n != 0 && sched.npidle.Load() != 0; n-- {
5624 // schedEnabled reports whether gp should be scheduled. It returns
5625 // false is scheduling of gp is disabled.
5627 // sched.lock must be held.
5628 func schedEnabled(gp *g) bool {
5629 assertLockHeld(&sched.lock)
5631 if sched.disable.user {
5632 return isSystemGoroutine(gp, true)
5637 // Put mp on midle list.
5638 // sched.lock must be held.
5639 // May run during STW, so write barriers are not allowed.
5641 //go:nowritebarrierrec
5643 assertLockHeld(&sched.lock)
5645 mp.schedlink = sched.midle
5651 // Try to get an m from midle list.
5652 // sched.lock must be held.
5653 // May run during STW, so write barriers are not allowed.
5655 //go:nowritebarrierrec
5657 assertLockHeld(&sched.lock)
5659 mp := sched.midle.ptr()
5661 sched.midle = mp.schedlink
5667 // Put gp on the global runnable queue.
5668 // sched.lock must be held.
5669 // May run during STW, so write barriers are not allowed.
5671 //go:nowritebarrierrec
5672 func globrunqput(gp *g) {
5673 assertLockHeld(&sched.lock)
5675 sched.runq.pushBack(gp)
5679 // Put gp at the head of the global runnable queue.
5680 // sched.lock must be held.
5681 // May run during STW, so write barriers are not allowed.
5683 //go:nowritebarrierrec
5684 func globrunqputhead(gp *g) {
5685 assertLockHeld(&sched.lock)
5691 // Put a batch of runnable goroutines on the global runnable queue.
5692 // This clears *batch.
5693 // sched.lock must be held.
5694 // May run during STW, so write barriers are not allowed.
5696 //go:nowritebarrierrec
5697 func globrunqputbatch(batch *gQueue, n int32) {
5698 assertLockHeld(&sched.lock)
5700 sched.runq.pushBackAll(*batch)
5705 // Try get a batch of G's from the global runnable queue.
5706 // sched.lock must be held.
5707 func globrunqget(pp *p, max int32) *g {
5708 assertLockHeld(&sched.lock)
5710 if sched.runqsize == 0 {
5714 n := sched.runqsize/gomaxprocs + 1
5715 if n > sched.runqsize {
5718 if max > 0 && n > max {
5721 if n > int32(len(pp.runq))/2 {
5722 n = int32(len(pp.runq)) / 2
5727 gp := sched.runq.pop()
5730 gp1 := sched.runq.pop()
5731 runqput(pp, gp1, false)
5736 // pMask is an atomic bitstring with one bit per P.
5739 // read returns true if P id's bit is set.
5740 func (p pMask) read(id uint32) bool {
5742 mask := uint32(1) << (id % 32)
5743 return (atomic.Load(&p[word]) & mask) != 0
5746 // set sets P id's bit.
5747 func (p pMask) set(id int32) {
5749 mask := uint32(1) << (id % 32)
5750 atomic.Or(&p[word], mask)
5753 // clear clears P id's bit.
5754 func (p pMask) clear(id int32) {
5756 mask := uint32(1) << (id % 32)
5757 atomic.And(&p[word], ^mask)
5760 // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
5762 // Ideally, the timer mask would be kept immediately consistent on any timer
5763 // operations. Unfortunately, updating a shared global data structure in the
5764 // timer hot path adds too much overhead in applications frequently switching
5765 // between no timers and some timers.
5767 // As a compromise, the timer mask is updated only on pidleget / pidleput. A
5768 // running P (returned by pidleget) may add a timer at any time, so its mask
5769 // must be set. An idle P (passed to pidleput) cannot add new timers while
5770 // idle, so if it has no timers at that time, its mask may be cleared.
5772 // Thus, we get the following effects on timer-stealing in findrunnable:
5774 // - Idle Ps with no timers when they go idle are never checked in findrunnable
5775 // (for work- or timer-stealing; this is the ideal case).
5776 // - Running Ps must always be checked.
5777 // - Idle Ps whose timers are stolen must continue to be checked until they run
5778 // again, even after timer expiration.
5780 // When the P starts running again, the mask should be set, as a timer may be
5781 // added at any time.
5783 // TODO(prattmic): Additional targeted updates may improve the above cases.
5784 // e.g., updating the mask when stealing a timer.
5785 func updateTimerPMask(pp *p) {
5786 if pp.numTimers.Load() > 0 {
5790 // Looks like there are no timers, however another P may transiently
5791 // decrement numTimers when handling a timerModified timer in
5792 // checkTimers. We must take timersLock to serialize with these changes.
5793 lock(&pp.timersLock)
5794 if pp.numTimers.Load() == 0 {
5795 timerpMask.clear(pp.id)
5797 unlock(&pp.timersLock)
5800 // pidleput puts p on the _Pidle list. now must be a relatively recent call
5801 // to nanotime or zero. Returns now or the current time if now was zero.
5803 // This releases ownership of p. Once sched.lock is released it is no longer
5806 // sched.lock must be held.
5808 // May run during STW, so write barriers are not allowed.
5810 //go:nowritebarrierrec
5811 func pidleput(pp *p, now int64) int64 {
5812 assertLockHeld(&sched.lock)
5815 throw("pidleput: P has non-empty run queue")
5820 updateTimerPMask(pp) // clear if there are no timers.
5821 idlepMask.set(pp.id)
5822 pp.link = sched.pidle
5825 if !pp.limiterEvent.start(limiterEventIdle, now) {
5826 throw("must be able to track idle limiter event")
5831 // pidleget tries to get a p from the _Pidle list, acquiring ownership.
5833 // sched.lock must be held.
5835 // May run during STW, so write barriers are not allowed.
5837 //go:nowritebarrierrec
5838 func pidleget(now int64) (*p, int64) {
5839 assertLockHeld(&sched.lock)
5841 pp := sched.pidle.ptr()
5843 // Timer may get added at any time now.
5847 timerpMask.set(pp.id)
5848 idlepMask.clear(pp.id)
5849 sched.pidle = pp.link
5850 sched.npidle.Add(-1)
5851 pp.limiterEvent.stop(limiterEventIdle, now)
5856 // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
5857 // This is called by spinning Ms (or callers than need a spinning M) that have
5858 // found work. If no P is available, this must synchronized with non-spinning
5859 // Ms that may be preparing to drop their P without discovering this work.
5861 // sched.lock must be held.
5863 // May run during STW, so write barriers are not allowed.
5865 //go:nowritebarrierrec
5866 func pidlegetSpinning(now int64) (*p, int64) {
5867 assertLockHeld(&sched.lock)
5869 pp, now := pidleget(now)
5871 // See "Delicate dance" comment in findrunnable. We found work
5872 // that we cannot take, we must synchronize with non-spinning
5873 // Ms that may be preparing to drop their P.
5874 sched.needspinning.Store(1)
5881 // runqempty reports whether pp has no Gs on its local run queue.
5882 // It never returns true spuriously.
5883 func runqempty(pp *p) bool {
5884 // Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
5885 // 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
5886 // Simply observing that runqhead == runqtail and then observing that runqnext == nil
5887 // does not mean the queue is empty.
5889 head := atomic.Load(&pp.runqhead)
5890 tail := atomic.Load(&pp.runqtail)
5891 runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
5892 if tail == atomic.Load(&pp.runqtail) {
5893 return head == tail && runnext == 0
5898 // To shake out latent assumptions about scheduling order,
5899 // we introduce some randomness into scheduling decisions
5900 // when running with the race detector.
5901 // The need for this was made obvious by changing the
5902 // (deterministic) scheduling order in Go 1.5 and breaking
5903 // many poorly-written tests.
5904 // With the randomness here, as long as the tests pass
5905 // consistently with -race, they shouldn't have latent scheduling
5907 const randomizeScheduler = raceenabled
5909 // runqput tries to put g on the local runnable queue.
5910 // If next is false, runqput adds g to the tail of the runnable queue.
5911 // If next is true, runqput puts g in the pp.runnext slot.
5912 // If the run queue is full, runnext puts g on the global queue.
5913 // Executed only by the owner P.
5914 func runqput(pp *p, gp *g, next bool) {
5915 if randomizeScheduler && next && fastrandn(2) == 0 {
5921 oldnext := pp.runnext
5922 if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
5928 // Kick the old runnext out to the regular run queue.
5933 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
5935 if t-h < uint32(len(pp.runq)) {
5936 pp.runq[t%uint32(len(pp.runq))].set(gp)
5937 atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
5940 if runqputslow(pp, gp, h, t) {
5943 // the queue is not full, now the put above must succeed
5947 // Put g and a batch of work from local runnable queue on global queue.
5948 // Executed only by the owner P.
5949 func runqputslow(pp *p, gp *g, h, t uint32) bool {
5950 var batch [len(pp.runq)/2 + 1]*g
5952 // First, grab a batch from local queue.
5955 if n != uint32(len(pp.runq)/2) {
5956 throw("runqputslow: queue is not full")
5958 for i := uint32(0); i < n; i++ {
5959 batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
5961 if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
5966 if randomizeScheduler {
5967 for i := uint32(1); i <= n; i++ {
5968 j := fastrandn(i + 1)
5969 batch[i], batch[j] = batch[j], batch[i]
5973 // Link the goroutines.
5974 for i := uint32(0); i < n; i++ {
5975 batch[i].schedlink.set(batch[i+1])
5978 q.head.set(batch[0])
5979 q.tail.set(batch[n])
5981 // Now put the batch on global queue.
5983 globrunqputbatch(&q, int32(n+1))
5988 // runqputbatch tries to put all the G's on q on the local runnable queue.
5989 // If the queue is full, they are put on the global queue; in that case
5990 // this will temporarily acquire the scheduler lock.
5991 // Executed only by the owner P.
5992 func runqputbatch(pp *p, q *gQueue, qsize int) {
5993 h := atomic.LoadAcq(&pp.runqhead)
5996 for !q.empty() && t-h < uint32(len(pp.runq)) {
5998 pp.runq[t%uint32(len(pp.runq))].set(gp)
6004 if randomizeScheduler {
6005 off := func(o uint32) uint32 {
6006 return (pp.runqtail + o) % uint32(len(pp.runq))
6008 for i := uint32(1); i < n; i++ {
6009 j := fastrandn(i + 1)
6010 pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
6014 atomic.StoreRel(&pp.runqtail, t)
6017 globrunqputbatch(q, int32(qsize))
6022 // Get g from local runnable queue.
6023 // If inheritTime is true, gp should inherit the remaining time in the
6024 // current time slice. Otherwise, it should start a new time slice.
6025 // Executed only by the owner P.
6026 func runqget(pp *p) (gp *g, inheritTime bool) {
6027 // If there's a runnext, it's the next G to run.
6029 // If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
6030 // because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
6031 // Hence, there's no need to retry this CAS if it fails.
6032 if next != 0 && pp.runnext.cas(next, 0) {
6033 return next.ptr(), true
6037 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6042 gp := pp.runq[h%uint32(len(pp.runq))].ptr()
6043 if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
6049 // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
6050 // Executed only by the owner P.
6051 func runqdrain(pp *p) (drainQ gQueue, n uint32) {
6052 oldNext := pp.runnext
6053 if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
6054 drainQ.pushBack(oldNext.ptr())
6059 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6065 if qn > uint32(len(pp.runq)) { // read inconsistent h and t
6069 if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
6073 // We've inverted the order in which it gets G's from the local P's runnable queue
6074 // and then advances the head pointer because we don't want to mess up the statuses of G's
6075 // while runqdrain() and runqsteal() are running in parallel.
6076 // Thus we should advance the head pointer before draining the local P into a gQueue,
6077 // so that we can update any gp.schedlink only after we take the full ownership of G,
6078 // meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
6079 // See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
6080 for i := uint32(0); i < qn; i++ {
6081 gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
6088 // Grabs a batch of goroutines from pp's runnable queue into batch.
6089 // Batch is a ring buffer starting at batchHead.
6090 // Returns number of grabbed goroutines.
6091 // Can be executed by any P.
6092 func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
6094 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
6095 t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
6100 // Try to steal from pp.runnext.
6101 if next := pp.runnext; next != 0 {
6102 if pp.status == _Prunning {
6103 // Sleep to ensure that pp isn't about to run the g
6104 // we are about to steal.
6105 // The important use case here is when the g running
6106 // on pp ready()s another g and then almost
6107 // immediately blocks. Instead of stealing runnext
6108 // in this window, back off to give pp a chance to
6109 // schedule runnext. This will avoid thrashing gs
6110 // between different Ps.
6111 // A sync chan send/recv takes ~50ns as of time of
6112 // writing, so 3us gives ~50x overshoot.
6113 if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
6116 // On some platforms system timer granularity is
6117 // 1-15ms, which is way too much for this
6118 // optimization. So just yield.
6122 if !pp.runnext.cas(next, 0) {
6125 batch[batchHead%uint32(len(batch))] = next
6131 if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
6134 for i := uint32(0); i < n; i++ {
6135 g := pp.runq[(h+i)%uint32(len(pp.runq))]
6136 batch[(batchHead+i)%uint32(len(batch))] = g
6138 if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
6144 // Steal half of elements from local runnable queue of p2
6145 // and put onto local runnable queue of p.
6146 // Returns one of the stolen elements (or nil if failed).
6147 func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
6149 n := runqgrab(p2, &pp.runq, t, stealRunNextG)
6154 gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
6158 h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
6159 if t-h+n >= uint32(len(pp.runq)) {
6160 throw("runqsteal: runq overflow")
6162 atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
6166 // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
6167 // be on one gQueue or gList at a time.
6168 type gQueue struct {
6173 // empty reports whether q is empty.
6174 func (q *gQueue) empty() bool {
6178 // push adds gp to the head of q.
6179 func (q *gQueue) push(gp *g) {
6180 gp.schedlink = q.head
6187 // pushBack adds gp to the tail of q.
6188 func (q *gQueue) pushBack(gp *g) {
6191 q.tail.ptr().schedlink.set(gp)
6198 // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
6200 func (q *gQueue) pushBackAll(q2 gQueue) {
6204 q2.tail.ptr().schedlink = 0
6206 q.tail.ptr().schedlink = q2.head
6213 // pop removes and returns the head of queue q. It returns nil if
6215 func (q *gQueue) pop() *g {
6218 q.head = gp.schedlink
6226 // popList takes all Gs in q and returns them as a gList.
6227 func (q *gQueue) popList() gList {
6228 stack := gList{q.head}
6233 // A gList is a list of Gs linked through g.schedlink. A G can only be
6234 // on one gQueue or gList at a time.
6239 // empty reports whether l is empty.
6240 func (l *gList) empty() bool {
6244 // push adds gp to the head of l.
6245 func (l *gList) push(gp *g) {
6246 gp.schedlink = l.head
6250 // pushAll prepends all Gs in q to l.
6251 func (l *gList) pushAll(q gQueue) {
6253 q.tail.ptr().schedlink = l.head
6258 // pop removes and returns the head of l. If l is empty, it returns nil.
6259 func (l *gList) pop() *g {
6262 l.head = gp.schedlink
6267 //go:linkname setMaxThreads runtime/debug.setMaxThreads
6268 func setMaxThreads(in int) (out int) {
6270 out = int(sched.maxmcount)
6271 if in > 0x7fffffff { // MaxInt32
6272 sched.maxmcount = 0x7fffffff
6274 sched.maxmcount = int32(in)
6282 func procPin() int {
6287 return int(mp.p.ptr().id)
6296 //go:linkname sync_runtime_procPin sync.runtime_procPin
6298 func sync_runtime_procPin() int {
6302 //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
6304 func sync_runtime_procUnpin() {
6308 //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
6310 func sync_atomic_runtime_procPin() int {
6314 //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
6316 func sync_atomic_runtime_procUnpin() {
6320 // Active spinning for sync.Mutex.
6322 //go:linkname sync_runtime_canSpin sync.runtime_canSpin
6324 func sync_runtime_canSpin(i int) bool {
6325 // sync.Mutex is cooperative, so we are conservative with spinning.
6326 // Spin only few times and only if running on a multicore machine and
6327 // GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
6328 // As opposed to runtime mutex we don't do passive spinning here,
6329 // because there can be work on global runq or on other Ps.
6330 if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
6333 if p := getg().m.p.ptr(); !runqempty(p) {
6339 //go:linkname sync_runtime_doSpin sync.runtime_doSpin
6341 func sync_runtime_doSpin() {
6342 procyield(active_spin_cnt)
6345 var stealOrder randomOrder
6347 // randomOrder/randomEnum are helper types for randomized work stealing.
6348 // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
6349 // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
6350 // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
6351 type randomOrder struct {
6356 type randomEnum struct {
6363 func (ord *randomOrder) reset(count uint32) {
6365 ord.coprimes = ord.coprimes[:0]
6366 for i := uint32(1); i <= count; i++ {
6367 if gcd(i, count) == 1 {
6368 ord.coprimes = append(ord.coprimes, i)
6373 func (ord *randomOrder) start(i uint32) randomEnum {
6377 inc: ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
6381 func (enum *randomEnum) done() bool {
6382 return enum.i == enum.count
6385 func (enum *randomEnum) next() {
6387 enum.pos = (enum.pos + enum.inc) % enum.count
6390 func (enum *randomEnum) position() uint32 {
6394 func gcd(a, b uint32) uint32 {
6401 // An initTask represents the set of initializations that need to be done for a package.
6402 // Keep in sync with ../../test/initempty.go:initTask
6403 type initTask struct {
6404 // TODO: pack the first 3 fields more tightly?
6405 state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
6408 // followed by ndeps instances of an *initTask, one per package depended on
6409 // followed by nfns pcs, one per init function to run
6412 // inittrace stores statistics for init functions which are
6413 // updated by malloc and newproc when active is true.
6414 var inittrace tracestat
6416 type tracestat struct {
6417 active bool // init tracing activation status
6418 id uint64 // init goroutine id
6419 allocs uint64 // heap allocations
6420 bytes uint64 // heap allocated bytes
6423 func doInit(t *initTask) {
6425 case 2: // fully initialized
6427 case 1: // initialization in progress
6428 throw("recursive call during initialization - linker skew")
6429 default: // not initialized yet
6430 t.state = 1 // initialization in progress
6432 for i := uintptr(0); i < t.ndeps; i++ {
6433 p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
6434 t2 := *(**initTask)(p)
6439 t.state = 2 // initialization done
6448 if inittrace.active {
6450 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6454 firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
6455 for i := uintptr(0); i < t.nfns; i++ {
6456 p := add(firstFunc, i*goarch.PtrSize)
6457 f := *(*func())(unsafe.Pointer(&p))
6461 if inittrace.active {
6463 // Load stats non-atomically since tracinit is updated only by this init goroutine.
6466 f := *(*func())(unsafe.Pointer(&firstFunc))
6467 pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
6470 print("init ", pkg, " @")
6471 print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
6472 print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
6473 print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
6474 print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
6478 t.state = 2 // initialization done