Source file src/runtime/proc.go

     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.
     4  
     5  package runtime
     6  
     7  import (
     8  	"internal/abi"
     9  	"internal/cpu"
    10  	"internal/goarch"
    11  	"runtime/internal/atomic"
    12  	"runtime/internal/sys"
    13  	"unsafe"
    14  )
    15  
    16  // set using cmd/go/internal/modload.ModInfoProg
    17  var modinfo string
    18  
    19  // Goroutine scheduler
    20  // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
    21  //
    22  // The main concepts are:
    23  // G - goroutine.
    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.
    28  //
    29  // Design doc at https://golang.org/s/go11sched.
    30  
    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).
    39  //
    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
    51  //    any work to do.
    52  //
    53  // The current approach:
    54  //
    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.
    58  //
    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.
    62  //
    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.
    72  //
    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
    78  // utilization.
    79  //
    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.
    85  //
    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.
    90  //
    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.
    95  //
    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.
    99  //
   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
   107  //   wait.
   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
   111  //   workers.
   112  
   113  var (
   114  	m0           m
   115  	g0           g
   116  	mcache0      *mcache
   117  	raceprocctx0 uintptr
   118  )
   119  
   120  //go:linkname runtime_inittask runtime..inittask
   121  var runtime_inittask initTask
   122  
   123  //go:linkname main_inittask main..inittask
   124  var main_inittask initTask
   125  
   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
   131  
   132  //go:linkname main_main main.main
   133  func main_main()
   134  
   135  // mainStarted indicates that the main M has started.
   136  var mainStarted bool
   137  
   138  // runtimeInitTime is the nanotime() at which the runtime started.
   139  var runtimeInitTime int64
   140  
   141  // Value to use for signal mask for newly created M's.
   142  var initSigmask sigset
   143  
   144  // The main goroutine.
   145  func main() {
   146  	mp := getg().m
   147  
   148  	// Racectx of m0->g0 is used only as the parent of the main goroutine.
   149  	// It must not be used for anything else.
   150  	mp.g0.racectx = 0
   151  
   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
   157  	} else {
   158  		maxstacksize = 250000000
   159  	}
   160  
   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
   165  
   166  	// Allow newproc to start new Ms.
   167  	mainStarted = true
   168  
   169  	if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
   170  		systemstack(func() {
   171  			newm(sysmon, nil, -1)
   172  		})
   173  	}
   174  
   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.
   181  	lockOSThread()
   182  
   183  	if mp != &m0 {
   184  		throw("runtime.main not on m0")
   185  	}
   186  
   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")
   192  	}
   193  
   194  	if debug.inittrace != 0 {
   195  		inittrace.id = getg().goid
   196  		inittrace.active = true
   197  	}
   198  
   199  	doInit(&runtime_inittask) // Must be before defer.
   200  
   201  	// Defer unlock so that runtime.Goexit during init does the unlock too.
   202  	needUnlock := true
   203  	defer func() {
   204  		if needUnlock {
   205  			unlockOSThread()
   206  		}
   207  	}()
   208  
   209  	gcenable()
   210  
   211  	main_init_done = make(chan bool)
   212  	if iscgo {
   213  		if _cgo_thread_start == nil {
   214  			throw("_cgo_thread_start missing")
   215  		}
   216  		if GOOS != "windows" {
   217  			if _cgo_setenv == nil {
   218  				throw("_cgo_setenv missing")
   219  			}
   220  			if _cgo_unsetenv == nil {
   221  				throw("_cgo_unsetenv missing")
   222  			}
   223  		}
   224  		if _cgo_notify_runtime_init_done == nil {
   225  			throw("_cgo_notify_runtime_init_done missing")
   226  		}
   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)
   231  	}
   232  
   233  	doInit(&main_inittask)
   234  
   235  	// Disable init tracing after main init done to avoid overhead
   236  	// of collecting statistics in malloc and newproc
   237  	inittrace.active = false
   238  
   239  	close(main_init_done)
   240  
   241  	needUnlock = false
   242  	unlockOSThread()
   243  
   244  	if isarchive || islibrary {
   245  		// A program compiled with -buildmode=c-archive or c-shared
   246  		// has a main, but it is not executed.
   247  		return
   248  	}
   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
   250  	fn()
   251  	if raceenabled {
   252  		runExitHooks(0) // run hooks now, since racefini does not return
   253  		racefini()
   254  	}
   255  
   256  	// Make racy client program work: if panicking on
   257  	// another goroutine at the same time as main returns,
   258  	// let the other goroutine finish printing the panic trace.
   259  	// Once it does, it will exit. See issues 3934 and 20018.
   260  	if runningPanicDefers.Load() != 0 {
   261  		// Running deferred functions should not take long.
   262  		for c := 0; c < 1000; c++ {
   263  			if runningPanicDefers.Load() == 0 {
   264  				break
   265  			}
   266  			Gosched()
   267  		}
   268  	}
   269  	if panicking.Load() != 0 {
   270  		gopark(nil, nil, waitReasonPanicWait, traceEvGoStop, 1)
   271  	}
   272  	runExitHooks(0)
   273  
   274  	exit(0)
   275  	for {
   276  		var x *int32
   277  		*x = 0
   278  	}
   279  }
   280  
   281  // os_beforeExit is called from os.Exit(0).
   282  //
   283  //go:linkname os_beforeExit os.runtime_beforeExit
   284  func os_beforeExit(exitCode int) {
   285  	runExitHooks(exitCode)
   286  	if exitCode == 0 && raceenabled {
   287  		racefini()
   288  	}
   289  }
   290  
   291  // start forcegc helper goroutine
   292  func init() {
   293  	go forcegchelper()
   294  }
   295  
   296  func forcegchelper() {
   297  	forcegc.g = getg()
   298  	lockInit(&forcegc.lock, lockRankForcegc)
   299  	for {
   300  		lock(&forcegc.lock)
   301  		if forcegc.idle.Load() {
   302  			throw("forcegc: phase error")
   303  		}
   304  		forcegc.idle.Store(true)
   305  		goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceEvGoBlock, 1)
   306  		// this goroutine is explicitly resumed by sysmon
   307  		if debug.gctrace > 0 {
   308  			println("GC forced")
   309  		}
   310  		// Time-triggered, fully concurrent.
   311  		gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
   312  	}
   313  }
   314  
   315  //go:nosplit
   316  
   317  // Gosched yields the processor, allowing other goroutines to run. It does not
   318  // suspend the current goroutine, so execution resumes automatically.
   319  func Gosched() {
   320  	checkTimeouts()
   321  	mcall(gosched_m)
   322  }
   323  
   324  // goschedguarded yields the processor like gosched, but also checks
   325  // for forbidden states and opts out of the yield in those cases.
   326  //
   327  //go:nosplit
   328  func goschedguarded() {
   329  	mcall(goschedguarded_m)
   330  }
   331  
   332  // goschedIfBusy yields the processor like gosched, but only does so if
   333  // there are no idle Ps or if we're on the only P and there's nothing in
   334  // the run queue. In both cases, there is freely available idle time.
   335  //
   336  //go:nosplit
   337  func goschedIfBusy() {
   338  	gp := getg()
   339  	// Call gosched if gp.preempt is set; we may be in a tight loop that
   340  	// doesn't otherwise yield.
   341  	if !gp.preempt && sched.npidle.Load() > 0 {
   342  		return
   343  	}
   344  	mcall(gosched_m)
   345  }
   346  
   347  // Puts the current goroutine into a waiting state and calls unlockf on the
   348  // system stack.
   349  //
   350  // If unlockf returns false, the goroutine is resumed.
   351  //
   352  // unlockf must not access this G's stack, as it may be moved between
   353  // the call to gopark and the call to unlockf.
   354  //
   355  // Note that because unlockf is called after putting the G into a waiting
   356  // state, the G may have already been readied by the time unlockf is called
   357  // unless there is external synchronization preventing the G from being
   358  // readied. If unlockf returns false, it must guarantee that the G cannot be
   359  // externally readied.
   360  //
   361  // Reason explains why the goroutine has been parked. It is displayed in stack
   362  // traces and heap dumps. Reasons should be unique and descriptive. Do not
   363  // re-use reasons, add new ones.
   364  func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceEv byte, traceskip int) {
   365  	if reason != waitReasonSleep {
   366  		checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
   367  	}
   368  	mp := acquirem()
   369  	gp := mp.curg
   370  	status := readgstatus(gp)
   371  	if status != _Grunning && status != _Gscanrunning {
   372  		throw("gopark: bad g status")
   373  	}
   374  	mp.waitlock = lock
   375  	mp.waitunlockf = unlockf
   376  	gp.waitreason = reason
   377  	mp.waittraceev = traceEv
   378  	mp.waittraceskip = traceskip
   379  	releasem(mp)
   380  	// can't do anything that might move the G between Ms here.
   381  	mcall(park_m)
   382  }
   383  
   384  // Puts the current goroutine into a waiting state and unlocks the lock.
   385  // The goroutine can be made runnable again by calling goready(gp).
   386  func goparkunlock(lock *mutex, reason waitReason, traceEv byte, traceskip int) {
   387  	gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
   388  }
   389  
   390  func goready(gp *g, traceskip int) {
   391  	systemstack(func() {
   392  		ready(gp, traceskip, true)
   393  	})
   394  }
   395  
   396  //go:nosplit
   397  func acquireSudog() *sudog {
   398  	// Delicate dance: the semaphore implementation calls
   399  	// acquireSudog, acquireSudog calls new(sudog),
   400  	// new calls malloc, malloc can call the garbage collector,
   401  	// and the garbage collector calls the semaphore implementation
   402  	// in stopTheWorld.
   403  	// Break the cycle by doing acquirem/releasem around new(sudog).
   404  	// The acquirem/releasem increments m.locks during new(sudog),
   405  	// which keeps the garbage collector from being invoked.
   406  	mp := acquirem()
   407  	pp := mp.p.ptr()
   408  	if len(pp.sudogcache) == 0 {
   409  		lock(&sched.sudoglock)
   410  		// First, try to grab a batch from central cache.
   411  		for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
   412  			s := sched.sudogcache
   413  			sched.sudogcache = s.next
   414  			s.next = nil
   415  			pp.sudogcache = append(pp.sudogcache, s)
   416  		}
   417  		unlock(&sched.sudoglock)
   418  		// If the central cache is empty, allocate a new one.
   419  		if len(pp.sudogcache) == 0 {
   420  			pp.sudogcache = append(pp.sudogcache, new(sudog))
   421  		}
   422  	}
   423  	n := len(pp.sudogcache)
   424  	s := pp.sudogcache[n-1]
   425  	pp.sudogcache[n-1] = nil
   426  	pp.sudogcache = pp.sudogcache[:n-1]
   427  	if s.elem != nil {
   428  		throw("acquireSudog: found s.elem != nil in cache")
   429  	}
   430  	releasem(mp)
   431  	return s
   432  }
   433  
   434  //go:nosplit
   435  func releaseSudog(s *sudog) {
   436  	if s.elem != nil {
   437  		throw("runtime: sudog with non-nil elem")
   438  	}
   439  	if s.isSelect {
   440  		throw("runtime: sudog with non-false isSelect")
   441  	}
   442  	if s.next != nil {
   443  		throw("runtime: sudog with non-nil next")
   444  	}
   445  	if s.prev != nil {
   446  		throw("runtime: sudog with non-nil prev")
   447  	}
   448  	if s.waitlink != nil {
   449  		throw("runtime: sudog with non-nil waitlink")
   450  	}
   451  	if s.c != nil {
   452  		throw("runtime: sudog with non-nil c")
   453  	}
   454  	gp := getg()
   455  	if gp.param != nil {
   456  		throw("runtime: releaseSudog with non-nil gp.param")
   457  	}
   458  	mp := acquirem() // avoid rescheduling to another P
   459  	pp := mp.p.ptr()
   460  	if len(pp.sudogcache) == cap(pp.sudogcache) {
   461  		// Transfer half of local cache to the central cache.
   462  		var first, last *sudog
   463  		for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
   464  			n := len(pp.sudogcache)
   465  			p := pp.sudogcache[n-1]
   466  			pp.sudogcache[n-1] = nil
   467  			pp.sudogcache = pp.sudogcache[:n-1]
   468  			if first == nil {
   469  				first = p
   470  			} else {
   471  				last.next = p
   472  			}
   473  			last = p
   474  		}
   475  		lock(&sched.sudoglock)
   476  		last.next = sched.sudogcache
   477  		sched.sudogcache = first
   478  		unlock(&sched.sudoglock)
   479  	}
   480  	pp.sudogcache = append(pp.sudogcache, s)
   481  	releasem(mp)
   482  }
   483  
   484  // called from assembly.
   485  func badmcall(fn func(*g)) {
   486  	throw("runtime: mcall called on m->g0 stack")
   487  }
   488  
   489  func badmcall2(fn func(*g)) {
   490  	throw("runtime: mcall function returned")
   491  }
   492  
   493  func badreflectcall() {
   494  	panic(plainError("arg size to reflect.call more than 1GB"))
   495  }
   496  
   497  //go:nosplit
   498  //go:nowritebarrierrec
   499  func badmorestackg0() {
   500  	writeErrStr("fatal: morestack on g0\n")
   501  }
   502  
   503  //go:nosplit
   504  //go:nowritebarrierrec
   505  func badmorestackgsignal() {
   506  	writeErrStr("fatal: morestack on gsignal\n")
   507  }
   508  
   509  //go:nosplit
   510  func badctxt() {
   511  	throw("ctxt != 0")
   512  }
   513  
   514  func lockedOSThread() bool {
   515  	gp := getg()
   516  	return gp.lockedm != 0 && gp.m.lockedg != 0
   517  }
   518  
   519  var (
   520  	// allgs contains all Gs ever created (including dead Gs), and thus
   521  	// never shrinks.
   522  	//
   523  	// Access via the slice is protected by allglock or stop-the-world.
   524  	// Readers that cannot take the lock may (carefully!) use the atomic
   525  	// variables below.
   526  	allglock mutex
   527  	allgs    []*g
   528  
   529  	// allglen and allgptr are atomic variables that contain len(allgs) and
   530  	// &allgs[0] respectively. Proper ordering depends on totally-ordered
   531  	// loads and stores. Writes are protected by allglock.
   532  	//
   533  	// allgptr is updated before allglen. Readers should read allglen
   534  	// before allgptr to ensure that allglen is always <= len(allgptr). New
   535  	// Gs appended during the race can be missed. For a consistent view of
   536  	// all Gs, allglock must be held.
   537  	//
   538  	// allgptr copies should always be stored as a concrete type or
   539  	// unsafe.Pointer, not uintptr, to ensure that GC can still reach it
   540  	// even if it points to a stale array.
   541  	allglen uintptr
   542  	allgptr **g
   543  )
   544  
   545  func allgadd(gp *g) {
   546  	if readgstatus(gp) == _Gidle {
   547  		throw("allgadd: bad status Gidle")
   548  	}
   549  
   550  	lock(&allglock)
   551  	allgs = append(allgs, gp)
   552  	if &allgs[0] != allgptr {
   553  		atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
   554  	}
   555  	atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
   556  	unlock(&allglock)
   557  }
   558  
   559  // allGsSnapshot returns a snapshot of the slice of all Gs.
   560  //
   561  // The world must be stopped or allglock must be held.
   562  func allGsSnapshot() []*g {
   563  	assertWorldStoppedOrLockHeld(&allglock)
   564  
   565  	// Because the world is stopped or allglock is held, allgadd
   566  	// cannot happen concurrently with this. allgs grows
   567  	// monotonically and existing entries never change, so we can
   568  	// simply return a copy of the slice header. For added safety,
   569  	// we trim everything past len because that can still change.
   570  	return allgs[:len(allgs):len(allgs)]
   571  }
   572  
   573  // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
   574  func atomicAllG() (**g, uintptr) {
   575  	length := atomic.Loaduintptr(&allglen)
   576  	ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
   577  	return ptr, length
   578  }
   579  
   580  // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
   581  func atomicAllGIndex(ptr **g, i uintptr) *g {
   582  	return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
   583  }
   584  
   585  // forEachG calls fn on every G from allgs.
   586  //
   587  // forEachG takes a lock to exclude concurrent addition of new Gs.
   588  func forEachG(fn func(gp *g)) {
   589  	lock(&allglock)
   590  	for _, gp := range allgs {
   591  		fn(gp)
   592  	}
   593  	unlock(&allglock)
   594  }
   595  
   596  // forEachGRace calls fn on every G from allgs.
   597  //
   598  // forEachGRace avoids locking, but does not exclude addition of new Gs during
   599  // execution, which may be missed.
   600  func forEachGRace(fn func(gp *g)) {
   601  	ptr, length := atomicAllG()
   602  	for i := uintptr(0); i < length; i++ {
   603  		gp := atomicAllGIndex(ptr, i)
   604  		fn(gp)
   605  	}
   606  	return
   607  }
   608  
   609  const (
   610  	// Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
   611  	// 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
   612  	_GoidCacheBatch = 16
   613  )
   614  
   615  // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
   616  // value of the GODEBUG environment variable.
   617  func cpuinit(env string) {
   618  	switch GOOS {
   619  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   620  		cpu.DebugOptions = true
   621  	}
   622  	cpu.Initialize(env)
   623  
   624  	// Support cpu feature variables are used in code generated by the compiler
   625  	// to guard execution of instructions that can not be assumed to be always supported.
   626  	switch GOARCH {
   627  	case "386", "amd64":
   628  		x86HasPOPCNT = cpu.X86.HasPOPCNT
   629  		x86HasSSE41 = cpu.X86.HasSSE41
   630  		x86HasFMA = cpu.X86.HasFMA
   631  
   632  	case "arm":
   633  		armHasVFPv4 = cpu.ARM.HasVFPv4
   634  
   635  	case "arm64":
   636  		arm64HasATOMICS = cpu.ARM64.HasATOMICS
   637  	}
   638  }
   639  
   640  // getGodebugEarly extracts the environment variable GODEBUG from the environment on
   641  // Unix-like operating systems and returns it. This function exists to extract GODEBUG
   642  // early before much of the runtime is initialized.
   643  func getGodebugEarly() string {
   644  	const prefix = "GODEBUG="
   645  	var env string
   646  	switch GOOS {
   647  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   648  		// Similar to goenv_unix but extracts the environment value for
   649  		// GODEBUG directly.
   650  		// TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
   651  		n := int32(0)
   652  		for argv_index(argv, argc+1+n) != nil {
   653  			n++
   654  		}
   655  
   656  		for i := int32(0); i < n; i++ {
   657  			p := argv_index(argv, argc+1+i)
   658  			s := unsafe.String(p, findnull(p))
   659  
   660  			if hasPrefix(s, prefix) {
   661  				env = gostring(p)[len(prefix):]
   662  				break
   663  			}
   664  		}
   665  	}
   666  	return env
   667  }
   668  
   669  // The bootstrap sequence is:
   670  //
   671  //	call osinit
   672  //	call schedinit
   673  //	make & queue new G
   674  //	call runtime·mstart
   675  //
   676  // The new G calls runtime·main.
   677  func schedinit() {
   678  	lockInit(&sched.lock, lockRankSched)
   679  	lockInit(&sched.sysmonlock, lockRankSysmon)
   680  	lockInit(&sched.deferlock, lockRankDefer)
   681  	lockInit(&sched.sudoglock, lockRankSudog)
   682  	lockInit(&deadlock, lockRankDeadlock)
   683  	lockInit(&paniclk, lockRankPanic)
   684  	lockInit(&allglock, lockRankAllg)
   685  	lockInit(&allpLock, lockRankAllp)
   686  	lockInit(&reflectOffs.lock, lockRankReflectOffs)
   687  	lockInit(&finlock, lockRankFin)
   688  	lockInit(&trace.bufLock, lockRankTraceBuf)
   689  	lockInit(&trace.stringsLock, lockRankTraceStrings)
   690  	lockInit(&trace.lock, lockRankTrace)
   691  	lockInit(&cpuprof.lock, lockRankCpuprof)
   692  	lockInit(&trace.stackTab.lock, lockRankTraceStackTab)
   693  	// Enforce that this lock is always a leaf lock.
   694  	// All of this lock's critical sections should be
   695  	// extremely short.
   696  	lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
   697  
   698  	// raceinit must be the first call to race detector.
   699  	// In particular, it must be done before mallocinit below calls racemapshadow.
   700  	gp := getg()
   701  	if raceenabled {
   702  		gp.racectx, raceprocctx0 = raceinit()
   703  	}
   704  
   705  	sched.maxmcount = 10000
   706  
   707  	// The world starts stopped.
   708  	worldStopped()
   709  
   710  	moduledataverify()
   711  	stackinit()
   712  	mallocinit()
   713  	godebug := getGodebugEarly()
   714  	initPageTrace(godebug) // must run after mallocinit but before anything allocates
   715  	cpuinit(godebug)       // must run before alginit
   716  	alginit()              // maps, hash, fastrand must not be used before this call
   717  	fastrandinit()         // must run before mcommoninit
   718  	mcommoninit(gp.m, -1)
   719  	modulesinit()   // provides activeModules
   720  	typelinksinit() // uses maps, activeModules
   721  	itabsinit()     // uses activeModules
   722  	stkobjinit()    // must run before GC starts
   723  
   724  	sigsave(&gp.m.sigmask)
   725  	initSigmask = gp.m.sigmask
   726  
   727  	goargs()
   728  	goenvs()
   729  	parsedebugvars()
   730  	gcinit()
   731  
   732  	// if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
   733  	// Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
   734  	// set to true by the linker, it means that nothing is consuming the profile, it is
   735  	// safe to set MemProfileRate to 0.
   736  	if disableMemoryProfiling {
   737  		MemProfileRate = 0
   738  	}
   739  
   740  	lock(&sched.lock)
   741  	sched.lastpoll.Store(nanotime())
   742  	procs := ncpu
   743  	if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
   744  		procs = n
   745  	}
   746  	if procresize(procs) != nil {
   747  		throw("unknown runnable goroutine during bootstrap")
   748  	}
   749  	unlock(&sched.lock)
   750  
   751  	// World is effectively started now, as P's can run.
   752  	worldStarted()
   753  
   754  	// For cgocheck > 1, we turn on the write barrier at all times
   755  	// and check all pointer writes. We can't do this until after
   756  	// procresize because the write barrier needs a P.
   757  	if debug.cgocheck > 1 {
   758  		writeBarrier.cgo = true
   759  		writeBarrier.enabled = true
   760  		for _, pp := range allp {
   761  			pp.wbBuf.reset()
   762  		}
   763  	}
   764  
   765  	if buildVersion == "" {
   766  		// Condition should never trigger. This code just serves
   767  		// to ensure runtime·buildVersion is kept in the resulting binary.
   768  		buildVersion = "unknown"
   769  	}
   770  	if len(modinfo) == 1 {
   771  		// Condition should never trigger. This code just serves
   772  		// to ensure runtime·modinfo is kept in the resulting binary.
   773  		modinfo = ""
   774  	}
   775  }
   776  
   777  func dumpgstatus(gp *g) {
   778  	thisg := getg()
   779  	print("runtime:   gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   780  	print("runtime: getg:  g=", thisg, ", goid=", thisg.goid, ",  g->atomicstatus=", readgstatus(thisg), "\n")
   781  }
   782  
   783  // sched.lock must be held.
   784  func checkmcount() {
   785  	assertLockHeld(&sched.lock)
   786  
   787  	if mcount() > sched.maxmcount {
   788  		print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
   789  		throw("thread exhaustion")
   790  	}
   791  }
   792  
   793  // mReserveID returns the next ID to use for a new m. This new m is immediately
   794  // considered 'running' by checkdead.
   795  //
   796  // sched.lock must be held.
   797  func mReserveID() int64 {
   798  	assertLockHeld(&sched.lock)
   799  
   800  	if sched.mnext+1 < sched.mnext {
   801  		throw("runtime: thread ID overflow")
   802  	}
   803  	id := sched.mnext
   804  	sched.mnext++
   805  	checkmcount()
   806  	return id
   807  }
   808  
   809  // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
   810  func mcommoninit(mp *m, id int64) {
   811  	gp := getg()
   812  
   813  	// g0 stack won't make sense for user (and is not necessary unwindable).
   814  	if gp != gp.m.g0 {
   815  		callers(1, mp.createstack[:])
   816  	}
   817  
   818  	lock(&sched.lock)
   819  
   820  	if id >= 0 {
   821  		mp.id = id
   822  	} else {
   823  		mp.id = mReserveID()
   824  	}
   825  
   826  	lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
   827  	hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
   828  	if lo|hi == 0 {
   829  		hi = 1
   830  	}
   831  	// Same behavior as for 1.17.
   832  	// TODO: Simplify ths.
   833  	if goarch.BigEndian {
   834  		mp.fastrand = uint64(lo)<<32 | uint64(hi)
   835  	} else {
   836  		mp.fastrand = uint64(hi)<<32 | uint64(lo)
   837  	}
   838  
   839  	mpreinit(mp)
   840  	if mp.gsignal != nil {
   841  		mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
   842  	}
   843  
   844  	// Add to allm so garbage collector doesn't free g->m
   845  	// when it is just in a register or thread-local storage.
   846  	mp.alllink = allm
   847  
   848  	// NumCgoCall() iterates over allm w/o schedlock,
   849  	// so we need to publish it safely.
   850  	atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
   851  	unlock(&sched.lock)
   852  
   853  	// Allocate memory to hold a cgo traceback if the cgo call crashes.
   854  	if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
   855  		mp.cgoCallers = new(cgoCallers)
   856  	}
   857  }
   858  
   859  func (mp *m) becomeSpinning() {
   860  	mp.spinning = true
   861  	sched.nmspinning.Add(1)
   862  	sched.needspinning.Store(0)
   863  }
   864  
   865  var fastrandseed uintptr
   866  
   867  func fastrandinit() {
   868  	s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
   869  	getRandomData(s)
   870  }
   871  
   872  // Mark gp ready to run.
   873  func ready(gp *g, traceskip int, next bool) {
   874  	if trace.enabled {
   875  		traceGoUnpark(gp, traceskip)
   876  	}
   877  
   878  	status := readgstatus(gp)
   879  
   880  	// Mark runnable.
   881  	mp := acquirem() // disable preemption because it can be holding p in a local var
   882  	if status&^_Gscan != _Gwaiting {
   883  		dumpgstatus(gp)
   884  		throw("bad g->status in ready")
   885  	}
   886  
   887  	// status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
   888  	casgstatus(gp, _Gwaiting, _Grunnable)
   889  	runqput(mp.p.ptr(), gp, next)
   890  	wakep()
   891  	releasem(mp)
   892  }
   893  
   894  // freezeStopWait is a large value that freezetheworld sets
   895  // sched.stopwait to in order to request that all Gs permanently stop.
   896  const freezeStopWait = 0x7fffffff
   897  
   898  // freezing is set to non-zero if the runtime is trying to freeze the
   899  // world.
   900  var freezing atomic.Bool
   901  
   902  // Similar to stopTheWorld but best-effort and can be called several times.
   903  // There is no reverse operation, used during crashing.
   904  // This function must not lock any mutexes.
   905  func freezetheworld() {
   906  	freezing.Store(true)
   907  	// stopwait and preemption requests can be lost
   908  	// due to races with concurrently executing threads,
   909  	// so try several times
   910  	for i := 0; i < 5; i++ {
   911  		// this should tell the scheduler to not start any new goroutines
   912  		sched.stopwait = freezeStopWait
   913  		sched.gcwaiting.Store(true)
   914  		// this should stop running goroutines
   915  		if !preemptall() {
   916  			break // no running goroutines
   917  		}
   918  		usleep(1000)
   919  	}
   920  	// to be sure
   921  	usleep(1000)
   922  	preemptall()
   923  	usleep(1000)
   924  }
   925  
   926  // All reads and writes of g's status go through readgstatus, casgstatus
   927  // castogscanstatus, casfrom_Gscanstatus.
   928  //
   929  //go:nosplit
   930  func readgstatus(gp *g) uint32 {
   931  	return gp.atomicstatus.Load()
   932  }
   933  
   934  // The Gscanstatuses are acting like locks and this releases them.
   935  // If it proves to be a performance hit we should be able to make these
   936  // simple atomic stores but for now we are going to throw if
   937  // we see an inconsistent state.
   938  func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
   939  	success := false
   940  
   941  	// Check that transition is valid.
   942  	switch oldval {
   943  	default:
   944  		print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
   945  		dumpgstatus(gp)
   946  		throw("casfrom_Gscanstatus:top gp->status is not in scan state")
   947  	case _Gscanrunnable,
   948  		_Gscanwaiting,
   949  		_Gscanrunning,
   950  		_Gscansyscall,
   951  		_Gscanpreempted:
   952  		if newval == oldval&^_Gscan {
   953  			success = gp.atomicstatus.CompareAndSwap(oldval, newval)
   954  		}
   955  	}
   956  	if !success {
   957  		print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
   958  		dumpgstatus(gp)
   959  		throw("casfrom_Gscanstatus: gp->status is not in scan state")
   960  	}
   961  	releaseLockRank(lockRankGscan)
   962  }
   963  
   964  // This will return false if the gp is not in the expected status and the cas fails.
   965  // This acts like a lock acquire while the casfromgstatus acts like a lock release.
   966  func castogscanstatus(gp *g, oldval, newval uint32) bool {
   967  	switch oldval {
   968  	case _Grunnable,
   969  		_Grunning,
   970  		_Gwaiting,
   971  		_Gsyscall:
   972  		if newval == oldval|_Gscan {
   973  			r := gp.atomicstatus.CompareAndSwap(oldval, newval)
   974  			if r {
   975  				acquireLockRank(lockRankGscan)
   976  			}
   977  			return r
   978  
   979  		}
   980  	}
   981  	print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
   982  	throw("castogscanstatus")
   983  	panic("not reached")
   984  }
   985  
   986  // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
   987  // various latencies on every transition instead of sampling them.
   988  var casgstatusAlwaysTrack = false
   989  
   990  // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
   991  // and casfrom_Gscanstatus instead.
   992  // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
   993  // put it in the Gscan state is finished.
   994  //
   995  //go:nosplit
   996  func casgstatus(gp *g, oldval, newval uint32) {
   997  	if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
   998  		systemstack(func() {
   999  			print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1000  			throw("casgstatus: bad incoming values")
  1001  		})
  1002  	}
  1003  
  1004  	acquireLockRank(lockRankGscan)
  1005  	releaseLockRank(lockRankGscan)
  1006  
  1007  	// See https://golang.org/cl/21503 for justification of the yield delay.
  1008  	const yieldDelay = 5 * 1000
  1009  	var nextYield int64
  1010  
  1011  	// loop if gp->atomicstatus is in a scan state giving
  1012  	// GC time to finish and change the state to oldval.
  1013  	for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
  1014  		if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
  1015  			throw("casgstatus: waiting for Gwaiting but is Grunnable")
  1016  		}
  1017  		if i == 0 {
  1018  			nextYield = nanotime() + yieldDelay
  1019  		}
  1020  		if nanotime() < nextYield {
  1021  			for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
  1022  				procyield(1)
  1023  			}
  1024  		} else {
  1025  			osyield()
  1026  			nextYield = nanotime() + yieldDelay/2
  1027  		}
  1028  	}
  1029  
  1030  	if oldval == _Grunning {
  1031  		// Track every gTrackingPeriod time a goroutine transitions out of running.
  1032  		if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
  1033  			gp.tracking = true
  1034  		}
  1035  		gp.trackingSeq++
  1036  	}
  1037  	if !gp.tracking {
  1038  		return
  1039  	}
  1040  
  1041  	// Handle various kinds of tracking.
  1042  	//
  1043  	// Currently:
  1044  	// - Time spent in runnable.
  1045  	// - Time spent blocked on a sync.Mutex or sync.RWMutex.
  1046  	switch oldval {
  1047  	case _Grunnable:
  1048  		// We transitioned out of runnable, so measure how much
  1049  		// time we spent in this state and add it to
  1050  		// runnableTime.
  1051  		now := nanotime()
  1052  		gp.runnableTime += now - gp.trackingStamp
  1053  		gp.trackingStamp = 0
  1054  	case _Gwaiting:
  1055  		if !gp.waitreason.isMutexWait() {
  1056  			// Not blocking on a lock.
  1057  			break
  1058  		}
  1059  		// Blocking on a lock, measure it. Note that because we're
  1060  		// sampling, we have to multiply by our sampling period to get
  1061  		// a more representative estimate of the absolute value.
  1062  		// gTrackingPeriod also represents an accurate sampling period
  1063  		// because we can only enter this state from _Grunning.
  1064  		now := nanotime()
  1065  		sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
  1066  		gp.trackingStamp = 0
  1067  	}
  1068  	switch newval {
  1069  	case _Gwaiting:
  1070  		if !gp.waitreason.isMutexWait() {
  1071  			// Not blocking on a lock.
  1072  			break
  1073  		}
  1074  		// Blocking on a lock. Write down the timestamp.
  1075  		now := nanotime()
  1076  		gp.trackingStamp = now
  1077  	case _Grunnable:
  1078  		// We just transitioned into runnable, so record what
  1079  		// time that happened.
  1080  		now := nanotime()
  1081  		gp.trackingStamp = now
  1082  	case _Grunning:
  1083  		// We're transitioning into running, so turn off
  1084  		// tracking and record how much time we spent in
  1085  		// runnable.
  1086  		gp.tracking = false
  1087  		sched.timeToRun.record(gp.runnableTime)
  1088  		gp.runnableTime = 0
  1089  	}
  1090  }
  1091  
  1092  // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
  1093  //
  1094  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1095  func casGToWaiting(gp *g, old uint32, reason waitReason) {
  1096  	// Set the wait reason before calling casgstatus, because casgstatus will use it.
  1097  	gp.waitreason = reason
  1098  	casgstatus(gp, old, _Gwaiting)
  1099  }
  1100  
  1101  // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
  1102  // Returns old status. Cannot call casgstatus directly, because we are racing with an
  1103  // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
  1104  // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
  1105  // it would loop waiting for the status to go back to Gwaiting, which it never will.
  1106  //
  1107  //go:nosplit
  1108  func casgcopystack(gp *g) uint32 {
  1109  	for {
  1110  		oldstatus := readgstatus(gp) &^ _Gscan
  1111  		if oldstatus != _Gwaiting && oldstatus != _Grunnable {
  1112  			throw("copystack: bad status, not Gwaiting or Grunnable")
  1113  		}
  1114  		if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
  1115  			return oldstatus
  1116  		}
  1117  	}
  1118  }
  1119  
  1120  // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
  1121  //
  1122  // TODO(austin): This is the only status operation that both changes
  1123  // the status and locks the _Gscan bit. Rethink this.
  1124  func casGToPreemptScan(gp *g, old, new uint32) {
  1125  	if old != _Grunning || new != _Gscan|_Gpreempted {
  1126  		throw("bad g transition")
  1127  	}
  1128  	acquireLockRank(lockRankGscan)
  1129  	for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
  1130  	}
  1131  }
  1132  
  1133  // casGFromPreempted attempts to transition gp from _Gpreempted to
  1134  // _Gwaiting. If successful, the caller is responsible for
  1135  // re-scheduling gp.
  1136  func casGFromPreempted(gp *g, old, new uint32) bool {
  1137  	if old != _Gpreempted || new != _Gwaiting {
  1138  		throw("bad g transition")
  1139  	}
  1140  	gp.waitreason = waitReasonPreempted
  1141  	return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
  1142  }
  1143  
  1144  // stopTheWorld stops all P's from executing goroutines, interrupting
  1145  // all goroutines at GC safe points and records reason as the reason
  1146  // for the stop. On return, only the current goroutine's P is running.
  1147  // stopTheWorld must not be called from a system stack and the caller
  1148  // must not hold worldsema. The caller must call startTheWorld when
  1149  // other P's should resume execution.
  1150  //
  1151  // stopTheWorld is safe for multiple goroutines to call at the
  1152  // same time. Each will execute its own stop, and the stops will
  1153  // be serialized.
  1154  //
  1155  // This is also used by routines that do stack dumps. If the system is
  1156  // in panic or being exited, this may not reliably stop all
  1157  // goroutines.
  1158  func stopTheWorld(reason string) {
  1159  	semacquire(&worldsema)
  1160  	gp := getg()
  1161  	gp.m.preemptoff = reason
  1162  	systemstack(func() {
  1163  		// Mark the goroutine which called stopTheWorld preemptible so its
  1164  		// stack may be scanned.
  1165  		// This lets a mark worker scan us while we try to stop the world
  1166  		// since otherwise we could get in a mutual preemption deadlock.
  1167  		// We must not modify anything on the G stack because a stack shrink
  1168  		// may occur. A stack shrink is otherwise OK though because in order
  1169  		// to return from this function (and to leave the system stack) we
  1170  		// must have preempted all goroutines, including any attempting
  1171  		// to scan our stack, in which case, any stack shrinking will
  1172  		// have already completed by the time we exit.
  1173  		// Don't provide a wait reason because we're still executing.
  1174  		casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
  1175  		stopTheWorldWithSema()
  1176  		casgstatus(gp, _Gwaiting, _Grunning)
  1177  	})
  1178  }
  1179  
  1180  // startTheWorld undoes the effects of stopTheWorld.
  1181  func startTheWorld() {
  1182  	systemstack(func() { startTheWorldWithSema(false) })
  1183  
  1184  	// worldsema must be held over startTheWorldWithSema to ensure
  1185  	// gomaxprocs cannot change while worldsema is held.
  1186  	//
  1187  	// Release worldsema with direct handoff to the next waiter, but
  1188  	// acquirem so that semrelease1 doesn't try to yield our time.
  1189  	//
  1190  	// Otherwise if e.g. ReadMemStats is being called in a loop,
  1191  	// it might stomp on other attempts to stop the world, such as
  1192  	// for starting or ending GC. The operation this blocks is
  1193  	// so heavy-weight that we should just try to be as fair as
  1194  	// possible here.
  1195  	//
  1196  	// We don't want to just allow us to get preempted between now
  1197  	// and releasing the semaphore because then we keep everyone
  1198  	// (including, for example, GCs) waiting longer.
  1199  	mp := acquirem()
  1200  	mp.preemptoff = ""
  1201  	semrelease1(&worldsema, true, 0)
  1202  	releasem(mp)
  1203  }
  1204  
  1205  // stopTheWorldGC has the same effect as stopTheWorld, but blocks
  1206  // until the GC is not running. It also blocks a GC from starting
  1207  // until startTheWorldGC is called.
  1208  func stopTheWorldGC(reason string) {
  1209  	semacquire(&gcsema)
  1210  	stopTheWorld(reason)
  1211  }
  1212  
  1213  // startTheWorldGC undoes the effects of stopTheWorldGC.
  1214  func startTheWorldGC() {
  1215  	startTheWorld()
  1216  	semrelease(&gcsema)
  1217  }
  1218  
  1219  // Holding worldsema grants an M the right to try to stop the world.
  1220  var worldsema uint32 = 1
  1221  
  1222  // Holding gcsema grants the M the right to block a GC, and blocks
  1223  // until the current GC is done. In particular, it prevents gomaxprocs
  1224  // from changing concurrently.
  1225  //
  1226  // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
  1227  // being changed/enabled during a GC, remove this.
  1228  var gcsema uint32 = 1
  1229  
  1230  // stopTheWorldWithSema is the core implementation of stopTheWorld.
  1231  // The caller is responsible for acquiring worldsema and disabling
  1232  // preemption first and then should stopTheWorldWithSema on the system
  1233  // stack:
  1234  //
  1235  //	semacquire(&worldsema, 0)
  1236  //	m.preemptoff = "reason"
  1237  //	systemstack(stopTheWorldWithSema)
  1238  //
  1239  // When finished, the caller must either call startTheWorld or undo
  1240  // these three operations separately:
  1241  //
  1242  //	m.preemptoff = ""
  1243  //	systemstack(startTheWorldWithSema)
  1244  //	semrelease(&worldsema)
  1245  //
  1246  // It is allowed to acquire worldsema once and then execute multiple
  1247  // startTheWorldWithSema/stopTheWorldWithSema pairs.
  1248  // Other P's are able to execute between successive calls to
  1249  // startTheWorldWithSema and stopTheWorldWithSema.
  1250  // Holding worldsema causes any other goroutines invoking
  1251  // stopTheWorld to block.
  1252  func stopTheWorldWithSema() {
  1253  	gp := getg()
  1254  
  1255  	// If we hold a lock, then we won't be able to stop another M
  1256  	// that is blocked trying to acquire the lock.
  1257  	if gp.m.locks > 0 {
  1258  		throw("stopTheWorld: holding locks")
  1259  	}
  1260  
  1261  	lock(&sched.lock)
  1262  	sched.stopwait = gomaxprocs
  1263  	sched.gcwaiting.Store(true)
  1264  	preemptall()
  1265  	// stop current P
  1266  	gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
  1267  	sched.stopwait--
  1268  	// try to retake all P's in Psyscall status
  1269  	for _, pp := range allp {
  1270  		s := pp.status
  1271  		if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
  1272  			if trace.enabled {
  1273  				traceGoSysBlock(pp)
  1274  				traceProcStop(pp)
  1275  			}
  1276  			pp.syscalltick++
  1277  			sched.stopwait--
  1278  		}
  1279  	}
  1280  	// stop idle P's
  1281  	now := nanotime()
  1282  	for {
  1283  		pp, _ := pidleget(now)
  1284  		if pp == nil {
  1285  			break
  1286  		}
  1287  		pp.status = _Pgcstop
  1288  		sched.stopwait--
  1289  	}
  1290  	wait := sched.stopwait > 0
  1291  	unlock(&sched.lock)
  1292  
  1293  	// wait for remaining P's to stop voluntarily
  1294  	if wait {
  1295  		for {
  1296  			// wait for 100us, then try to re-preempt in case of any races
  1297  			if notetsleep(&sched.stopnote, 100*1000) {
  1298  				noteclear(&sched.stopnote)
  1299  				break
  1300  			}
  1301  			preemptall()
  1302  		}
  1303  	}
  1304  
  1305  	// sanity checks
  1306  	bad := ""
  1307  	if sched.stopwait != 0 {
  1308  		bad = "stopTheWorld: not stopped (stopwait != 0)"
  1309  	} else {
  1310  		for _, pp := range allp {
  1311  			if pp.status != _Pgcstop {
  1312  				bad = "stopTheWorld: not stopped (status != _Pgcstop)"
  1313  			}
  1314  		}
  1315  	}
  1316  	if freezing.Load() {
  1317  		// Some other thread is panicking. This can cause the
  1318  		// sanity checks above to fail if the panic happens in
  1319  		// the signal handler on a stopped thread. Either way,
  1320  		// we should halt this thread.
  1321  		lock(&deadlock)
  1322  		lock(&deadlock)
  1323  	}
  1324  	if bad != "" {
  1325  		throw(bad)
  1326  	}
  1327  
  1328  	worldStopped()
  1329  }
  1330  
  1331  func startTheWorldWithSema(emitTraceEvent bool) int64 {
  1332  	assertWorldStopped()
  1333  
  1334  	mp := acquirem() // disable preemption because it can be holding p in a local var
  1335  	if netpollinited() {
  1336  		list := netpoll(0) // non-blocking
  1337  		injectglist(&list)
  1338  	}
  1339  	lock(&sched.lock)
  1340  
  1341  	procs := gomaxprocs
  1342  	if newprocs != 0 {
  1343  		procs = newprocs
  1344  		newprocs = 0
  1345  	}
  1346  	p1 := procresize(procs)
  1347  	sched.gcwaiting.Store(false)
  1348  	if sched.sysmonwait.Load() {
  1349  		sched.sysmonwait.Store(false)
  1350  		notewakeup(&sched.sysmonnote)
  1351  	}
  1352  	unlock(&sched.lock)
  1353  
  1354  	worldStarted()
  1355  
  1356  	for p1 != nil {
  1357  		p := p1
  1358  		p1 = p1.link.ptr()
  1359  		if p.m != 0 {
  1360  			mp := p.m.ptr()
  1361  			p.m = 0
  1362  			if mp.nextp != 0 {
  1363  				throw("startTheWorld: inconsistent mp->nextp")
  1364  			}
  1365  			mp.nextp.set(p)
  1366  			notewakeup(&mp.park)
  1367  		} else {
  1368  			// Start M to run P.  Do not start another M below.
  1369  			newm(nil, p, -1)
  1370  		}
  1371  	}
  1372  
  1373  	// Capture start-the-world time before doing clean-up tasks.
  1374  	startTime := nanotime()
  1375  	if emitTraceEvent {
  1376  		traceGCSTWDone()
  1377  	}
  1378  
  1379  	// Wakeup an additional proc in case we have excessive runnable goroutines
  1380  	// in local queues or in the global queue. If we don't, the proc will park itself.
  1381  	// If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
  1382  	wakep()
  1383  
  1384  	releasem(mp)
  1385  
  1386  	return startTime
  1387  }
  1388  
  1389  // usesLibcall indicates whether this runtime performs system calls
  1390  // via libcall.
  1391  func usesLibcall() bool {
  1392  	switch GOOS {
  1393  	case "aix", "darwin", "illumos", "ios", "solaris", "windows":
  1394  		return true
  1395  	case "openbsd":
  1396  		return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
  1397  	}
  1398  	return false
  1399  }
  1400  
  1401  // mStackIsSystemAllocated indicates whether this runtime starts on a
  1402  // system-allocated stack.
  1403  func mStackIsSystemAllocated() bool {
  1404  	switch GOOS {
  1405  	case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
  1406  		return true
  1407  	case "openbsd":
  1408  		switch GOARCH {
  1409  		case "386", "amd64", "arm", "arm64":
  1410  			return true
  1411  		}
  1412  	}
  1413  	return false
  1414  }
  1415  
  1416  // mstart is the entry-point for new Ms.
  1417  // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
  1418  func mstart()
  1419  
  1420  // mstart0 is the Go entry-point for new Ms.
  1421  // This must not split the stack because we may not even have stack
  1422  // bounds set up yet.
  1423  //
  1424  // May run during STW (because it doesn't have a P yet), so write
  1425  // barriers are not allowed.
  1426  //
  1427  //go:nosplit
  1428  //go:nowritebarrierrec
  1429  func mstart0() {
  1430  	gp := getg()
  1431  
  1432  	osStack := gp.stack.lo == 0
  1433  	if osStack {
  1434  		// Initialize stack bounds from system stack.
  1435  		// Cgo may have left stack size in stack.hi.
  1436  		// minit may update the stack bounds.
  1437  		//
  1438  		// Note: these bounds may not be very accurate.
  1439  		// We set hi to &size, but there are things above
  1440  		// it. The 1024 is supposed to compensate this,
  1441  		// but is somewhat arbitrary.
  1442  		size := gp.stack.hi
  1443  		if size == 0 {
  1444  			size = 8192 * sys.StackGuardMultiplier
  1445  		}
  1446  		gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
  1447  		gp.stack.lo = gp.stack.hi - size + 1024
  1448  	}
  1449  	// Initialize stack guard so that we can start calling regular
  1450  	// Go code.
  1451  	gp.stackguard0 = gp.stack.lo + _StackGuard
  1452  	// This is the g0, so we can also call go:systemstack
  1453  	// functions, which check stackguard1.
  1454  	gp.stackguard1 = gp.stackguard0
  1455  	mstart1()
  1456  
  1457  	// Exit this thread.
  1458  	if mStackIsSystemAllocated() {
  1459  		// Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
  1460  		// the stack, but put it in gp.stack before mstart,
  1461  		// so the logic above hasn't set osStack yet.
  1462  		osStack = true
  1463  	}
  1464  	mexit(osStack)
  1465  }
  1466  
  1467  // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
  1468  // so that we can set up g0.sched to return to the call of mstart1 above.
  1469  //
  1470  //go:noinline
  1471  func mstart1() {
  1472  	gp := getg()
  1473  
  1474  	if gp != gp.m.g0 {
  1475  		throw("bad runtime·mstart")
  1476  	}
  1477  
  1478  	// Set up m.g0.sched as a label returning to just
  1479  	// after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
  1480  	// We're never coming back to mstart1 after we call schedule,
  1481  	// so other calls can reuse the current frame.
  1482  	// And goexit0 does a gogo that needs to return from mstart1
  1483  	// and let mstart0 exit the thread.
  1484  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1485  	gp.sched.pc = getcallerpc()
  1486  	gp.sched.sp = getcallersp()
  1487  
  1488  	asminit()
  1489  	minit()
  1490  
  1491  	// Install signal handlers; after minit so that minit can
  1492  	// prepare the thread to be able to handle the signals.
  1493  	if gp.m == &m0 {
  1494  		mstartm0()
  1495  	}
  1496  
  1497  	if fn := gp.m.mstartfn; fn != nil {
  1498  		fn()
  1499  	}
  1500  
  1501  	if gp.m != &m0 {
  1502  		acquirep(gp.m.nextp.ptr())
  1503  		gp.m.nextp = 0
  1504  	}
  1505  	schedule()
  1506  }
  1507  
  1508  // mstartm0 implements part of mstart1 that only runs on the m0.
  1509  //
  1510  // Write barriers are allowed here because we know the GC can't be
  1511  // running yet, so they'll be no-ops.
  1512  //
  1513  //go:yeswritebarrierrec
  1514  func mstartm0() {
  1515  	// Create an extra M for callbacks on threads not created by Go.
  1516  	// An extra M is also needed on Windows for callbacks created by
  1517  	// syscall.NewCallback. See issue #6751 for details.
  1518  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1519  		cgoHasExtraM = true
  1520  		newextram()
  1521  	}
  1522  	initsig(false)
  1523  }
  1524  
  1525  // mPark causes a thread to park itself, returning once woken.
  1526  //
  1527  //go:nosplit
  1528  func mPark() {
  1529  	gp := getg()
  1530  	notesleep(&gp.m.park)
  1531  	noteclear(&gp.m.park)
  1532  }
  1533  
  1534  // mexit tears down and exits the current thread.
  1535  //
  1536  // Don't call this directly to exit the thread, since it must run at
  1537  // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
  1538  // unwind the stack to the point that exits the thread.
  1539  //
  1540  // It is entered with m.p != nil, so write barriers are allowed. It
  1541  // will release the P before exiting.
  1542  //
  1543  //go:yeswritebarrierrec
  1544  func mexit(osStack bool) {
  1545  	mp := getg().m
  1546  
  1547  	if mp == &m0 {
  1548  		// This is the main thread. Just wedge it.
  1549  		//
  1550  		// On Linux, exiting the main thread puts the process
  1551  		// into a non-waitable zombie state. On Plan 9,
  1552  		// exiting the main thread unblocks wait even though
  1553  		// other threads are still running. On Solaris we can
  1554  		// neither exitThread nor return from mstart. Other
  1555  		// bad things probably happen on other platforms.
  1556  		//
  1557  		// We could try to clean up this M more before wedging
  1558  		// it, but that complicates signal handling.
  1559  		handoffp(releasep())
  1560  		lock(&sched.lock)
  1561  		sched.nmfreed++
  1562  		checkdead()
  1563  		unlock(&sched.lock)
  1564  		mPark()
  1565  		throw("locked m0 woke up")
  1566  	}
  1567  
  1568  	sigblock(true)
  1569  	unminit()
  1570  
  1571  	// Free the gsignal stack.
  1572  	if mp.gsignal != nil {
  1573  		stackfree(mp.gsignal.stack)
  1574  		// On some platforms, when calling into VDSO (e.g. nanotime)
  1575  		// we store our g on the gsignal stack, if there is one.
  1576  		// Now the stack is freed, unlink it from the m, so we
  1577  		// won't write to it when calling VDSO code.
  1578  		mp.gsignal = nil
  1579  	}
  1580  
  1581  	// Remove m from allm.
  1582  	lock(&sched.lock)
  1583  	for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
  1584  		if *pprev == mp {
  1585  			*pprev = mp.alllink
  1586  			goto found
  1587  		}
  1588  	}
  1589  	throw("m not found in allm")
  1590  found:
  1591  	// Delay reaping m until it's done with the stack.
  1592  	//
  1593  	// Put mp on the free list, though it will not be reaped while freeWait
  1594  	// is freeMWait. mp is no longer reachable via allm, so even if it is
  1595  	// on an OS stack, we must keep a reference to mp alive so that the GC
  1596  	// doesn't free mp while we are still using it.
  1597  	//
  1598  	// Note that the free list must not be linked through alllink because
  1599  	// some functions walk allm without locking, so may be using alllink.
  1600  	mp.freeWait.Store(freeMWait)
  1601  	mp.freelink = sched.freem
  1602  	sched.freem = mp
  1603  	unlock(&sched.lock)
  1604  
  1605  	atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
  1606  
  1607  	// Release the P.
  1608  	handoffp(releasep())
  1609  	// After this point we must not have write barriers.
  1610  
  1611  	// Invoke the deadlock detector. This must happen after
  1612  	// handoffp because it may have started a new M to take our
  1613  	// P's work.
  1614  	lock(&sched.lock)
  1615  	sched.nmfreed++
  1616  	checkdead()
  1617  	unlock(&sched.lock)
  1618  
  1619  	if GOOS == "darwin" || GOOS == "ios" {
  1620  		// Make sure pendingPreemptSignals is correct when an M exits.
  1621  		// For #41702.
  1622  		if mp.signalPending.Load() != 0 {
  1623  			pendingPreemptSignals.Add(-1)
  1624  		}
  1625  	}
  1626  
  1627  	// Destroy all allocated resources. After this is called, we may no
  1628  	// longer take any locks.
  1629  	mdestroy(mp)
  1630  
  1631  	if osStack {
  1632  		// No more uses of mp, so it is safe to drop the reference.
  1633  		mp.freeWait.Store(freeMRef)
  1634  
  1635  		// Return from mstart and let the system thread
  1636  		// library free the g0 stack and terminate the thread.
  1637  		return
  1638  	}
  1639  
  1640  	// mstart is the thread's entry point, so there's nothing to
  1641  	// return to. Exit the thread directly. exitThread will clear
  1642  	// m.freeWait when it's done with the stack and the m can be
  1643  	// reaped.
  1644  	exitThread(&mp.freeWait)
  1645  }
  1646  
  1647  // forEachP calls fn(p) for every P p when p reaches a GC safe point.
  1648  // If a P is currently executing code, this will bring the P to a GC
  1649  // safe point and execute fn on that P. If the P is not executing code
  1650  // (it is idle or in a syscall), this will call fn(p) directly while
  1651  // preventing the P from exiting its state. This does not ensure that
  1652  // fn will run on every CPU executing Go code, but it acts as a global
  1653  // memory barrier. GC uses this as a "ragged barrier."
  1654  //
  1655  // The caller must hold worldsema.
  1656  //
  1657  //go:systemstack
  1658  func forEachP(fn func(*p)) {
  1659  	mp := acquirem()
  1660  	pp := getg().m.p.ptr()
  1661  
  1662  	lock(&sched.lock)
  1663  	if sched.safePointWait != 0 {
  1664  		throw("forEachP: sched.safePointWait != 0")
  1665  	}
  1666  	sched.safePointWait = gomaxprocs - 1
  1667  	sched.safePointFn = fn
  1668  
  1669  	// Ask all Ps to run the safe point function.
  1670  	for _, p2 := range allp {
  1671  		if p2 != pp {
  1672  			atomic.Store(&p2.runSafePointFn, 1)
  1673  		}
  1674  	}
  1675  	preemptall()
  1676  
  1677  	// Any P entering _Pidle or _Psyscall from now on will observe
  1678  	// p.runSafePointFn == 1 and will call runSafePointFn when
  1679  	// changing its status to _Pidle/_Psyscall.
  1680  
  1681  	// Run safe point function for all idle Ps. sched.pidle will
  1682  	// not change because we hold sched.lock.
  1683  	for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
  1684  		if atomic.Cas(&p.runSafePointFn, 1, 0) {
  1685  			fn(p)
  1686  			sched.safePointWait--
  1687  		}
  1688  	}
  1689  
  1690  	wait := sched.safePointWait > 0
  1691  	unlock(&sched.lock)
  1692  
  1693  	// Run fn for the current P.
  1694  	fn(pp)
  1695  
  1696  	// Force Ps currently in _Psyscall into _Pidle and hand them
  1697  	// off to induce safe point function execution.
  1698  	for _, p2 := range allp {
  1699  		s := p2.status
  1700  		if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
  1701  			if trace.enabled {
  1702  				traceGoSysBlock(p2)
  1703  				traceProcStop(p2)
  1704  			}
  1705  			p2.syscalltick++
  1706  			handoffp(p2)
  1707  		}
  1708  	}
  1709  
  1710  	// Wait for remaining Ps to run fn.
  1711  	if wait {
  1712  		for {
  1713  			// Wait for 100us, then try to re-preempt in
  1714  			// case of any races.
  1715  			//
  1716  			// Requires system stack.
  1717  			if notetsleep(&sched.safePointNote, 100*1000) {
  1718  				noteclear(&sched.safePointNote)
  1719  				break
  1720  			}
  1721  			preemptall()
  1722  		}
  1723  	}
  1724  	if sched.safePointWait != 0 {
  1725  		throw("forEachP: not done")
  1726  	}
  1727  	for _, p2 := range allp {
  1728  		if p2.runSafePointFn != 0 {
  1729  			throw("forEachP: P did not run fn")
  1730  		}
  1731  	}
  1732  
  1733  	lock(&sched.lock)
  1734  	sched.safePointFn = nil
  1735  	unlock(&sched.lock)
  1736  	releasem(mp)
  1737  }
  1738  
  1739  // runSafePointFn runs the safe point function, if any, for this P.
  1740  // This should be called like
  1741  //
  1742  //	if getg().m.p.runSafePointFn != 0 {
  1743  //	    runSafePointFn()
  1744  //	}
  1745  //
  1746  // runSafePointFn must be checked on any transition in to _Pidle or
  1747  // _Psyscall to avoid a race where forEachP sees that the P is running
  1748  // just before the P goes into _Pidle/_Psyscall and neither forEachP
  1749  // nor the P run the safe-point function.
  1750  func runSafePointFn() {
  1751  	p := getg().m.p.ptr()
  1752  	// Resolve the race between forEachP running the safe-point
  1753  	// function on this P's behalf and this P running the
  1754  	// safe-point function directly.
  1755  	if !atomic.Cas(&p.runSafePointFn, 1, 0) {
  1756  		return
  1757  	}
  1758  	sched.safePointFn(p)
  1759  	lock(&sched.lock)
  1760  	sched.safePointWait--
  1761  	if sched.safePointWait == 0 {
  1762  		notewakeup(&sched.safePointNote)
  1763  	}
  1764  	unlock(&sched.lock)
  1765  }
  1766  
  1767  // When running with cgo, we call _cgo_thread_start
  1768  // to start threads for us so that we can play nicely with
  1769  // foreign code.
  1770  var cgoThreadStart unsafe.Pointer
  1771  
  1772  type cgothreadstart struct {
  1773  	g   guintptr
  1774  	tls *uint64
  1775  	fn  unsafe.Pointer
  1776  }
  1777  
  1778  // Allocate a new m unassociated with any thread.
  1779  // Can use p for allocation context if needed.
  1780  // fn is recorded as the new m's m.mstartfn.
  1781  // id is optional pre-allocated m ID. Omit by passing -1.
  1782  //
  1783  // This function is allowed to have write barriers even if the caller
  1784  // isn't because it borrows pp.
  1785  //
  1786  //go:yeswritebarrierrec
  1787  func allocm(pp *p, fn func(), id int64) *m {
  1788  	allocmLock.rlock()
  1789  
  1790  	// The caller owns pp, but we may borrow (i.e., acquirep) it. We must
  1791  	// disable preemption to ensure it is not stolen, which would make the
  1792  	// caller lose ownership.
  1793  	acquirem()
  1794  
  1795  	gp := getg()
  1796  	if gp.m.p == 0 {
  1797  		acquirep(pp) // temporarily borrow p for mallocs in this function
  1798  	}
  1799  
  1800  	// Release the free M list. We need to do this somewhere and
  1801  	// this may free up a stack we can use.
  1802  	if sched.freem != nil {
  1803  		lock(&sched.lock)
  1804  		var newList *m
  1805  		for freem := sched.freem; freem != nil; {
  1806  			wait := freem.freeWait.Load()
  1807  			if wait == freeMWait {
  1808  				next := freem.freelink
  1809  				freem.freelink = newList
  1810  				newList = freem
  1811  				freem = next
  1812  				continue
  1813  			}
  1814  			// Free the stack if needed. For freeMRef, there is
  1815  			// nothing to do except drop freem from the sched.freem
  1816  			// list.
  1817  			if wait == freeMStack {
  1818  				// stackfree must be on the system stack, but allocm is
  1819  				// reachable off the system stack transitively from
  1820  				// startm.
  1821  				systemstack(func() {
  1822  					stackfree(freem.g0.stack)
  1823  				})
  1824  			}
  1825  			freem = freem.freelink
  1826  		}
  1827  		sched.freem = newList
  1828  		unlock(&sched.lock)
  1829  	}
  1830  
  1831  	mp := new(m)
  1832  	mp.mstartfn = fn
  1833  	mcommoninit(mp, id)
  1834  
  1835  	// In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
  1836  	// Windows and Plan 9 will layout sched stack on OS stack.
  1837  	if iscgo || mStackIsSystemAllocated() {
  1838  		mp.g0 = malg(-1)
  1839  	} else {
  1840  		mp.g0 = malg(8192 * sys.StackGuardMultiplier)
  1841  	}
  1842  	mp.g0.m = mp
  1843  
  1844  	if pp == gp.m.p.ptr() {
  1845  		releasep()
  1846  	}
  1847  
  1848  	releasem(gp.m)
  1849  	allocmLock.runlock()
  1850  	return mp
  1851  }
  1852  
  1853  // needm is called when a cgo callback happens on a
  1854  // thread without an m (a thread not created by Go).
  1855  // In this case, needm is expected to find an m to use
  1856  // and return with m, g initialized correctly.
  1857  // Since m and g are not set now (likely nil, but see below)
  1858  // needm is limited in what routines it can call. In particular
  1859  // it can only call nosplit functions (textflag 7) and cannot
  1860  // do any scheduling that requires an m.
  1861  //
  1862  // In order to avoid needing heavy lifting here, we adopt
  1863  // the following strategy: there is a stack of available m's
  1864  // that can be stolen. Using compare-and-swap
  1865  // to pop from the stack has ABA races, so we simulate
  1866  // a lock by doing an exchange (via Casuintptr) to steal the stack
  1867  // head and replace the top pointer with MLOCKED (1).
  1868  // This serves as a simple spin lock that we can use even
  1869  // without an m. The thread that locks the stack in this way
  1870  // unlocks the stack by storing a valid stack head pointer.
  1871  //
  1872  // In order to make sure that there is always an m structure
  1873  // available to be stolen, we maintain the invariant that there
  1874  // is always one more than needed. At the beginning of the
  1875  // program (if cgo is in use) the list is seeded with a single m.
  1876  // If needm finds that it has taken the last m off the list, its job
  1877  // is - once it has installed its own m so that it can do things like
  1878  // allocate memory - to create a spare m and put it on the list.
  1879  //
  1880  // Each of these extra m's also has a g0 and a curg that are
  1881  // pressed into service as the scheduling stack and current
  1882  // goroutine for the duration of the cgo callback.
  1883  //
  1884  // When the callback is done with the m, it calls dropm to
  1885  // put the m back on the list.
  1886  //
  1887  //go:nosplit
  1888  func needm() {
  1889  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1890  		// Can happen if C/C++ code calls Go from a global ctor.
  1891  		// Can also happen on Windows if a global ctor uses a
  1892  		// callback created by syscall.NewCallback. See issue #6751
  1893  		// for details.
  1894  		//
  1895  		// Can not throw, because scheduler is not initialized yet.
  1896  		writeErrStr("fatal error: cgo callback before cgo call\n")
  1897  		exit(1)
  1898  	}
  1899  
  1900  	// Save and block signals before getting an M.
  1901  	// The signal handler may call needm itself,
  1902  	// and we must avoid a deadlock. Also, once g is installed,
  1903  	// any incoming signals will try to execute,
  1904  	// but we won't have the sigaltstack settings and other data
  1905  	// set up appropriately until the end of minit, which will
  1906  	// unblock the signals. This is the same dance as when
  1907  	// starting a new m to run Go code via newosproc.
  1908  	var sigmask sigset
  1909  	sigsave(&sigmask)
  1910  	sigblock(false)
  1911  
  1912  	// Lock extra list, take head, unlock popped list.
  1913  	// nilokay=false is safe here because of the invariant above,
  1914  	// that the extra list always contains or will soon contain
  1915  	// at least one m.
  1916  	mp := lockextra(false)
  1917  
  1918  	// Set needextram when we've just emptied the list,
  1919  	// so that the eventual call into cgocallbackg will
  1920  	// allocate a new m for the extra list. We delay the
  1921  	// allocation until then so that it can be done
  1922  	// after exitsyscall makes sure it is okay to be
  1923  	// running at all (that is, there's no garbage collection
  1924  	// running right now).
  1925  	mp.needextram = mp.schedlink == 0
  1926  	extraMCount--
  1927  	unlockextra(mp.schedlink.ptr())
  1928  
  1929  	// Store the original signal mask for use by minit.
  1930  	mp.sigmask = sigmask
  1931  
  1932  	// Install TLS on some platforms (previously setg
  1933  	// would do this if necessary).
  1934  	osSetupTLS(mp)
  1935  
  1936  	// Install g (= m->g0) and set the stack bounds
  1937  	// to match the current stack. We don't actually know
  1938  	// how big the stack is, like we don't know how big any
  1939  	// scheduling stack is, but we assume there's at least 32 kB,
  1940  	// which is more than enough for us.
  1941  	setg(mp.g0)
  1942  	gp := getg()
  1943  	gp.stack.hi = getcallersp() + 1024
  1944  	gp.stack.lo = getcallersp() - 32*1024
  1945  	gp.stackguard0 = gp.stack.lo + _StackGuard
  1946  
  1947  	// Initialize this thread to use the m.
  1948  	asminit()
  1949  	minit()
  1950  
  1951  	// mp.curg is now a real goroutine.
  1952  	casgstatus(mp.curg, _Gdead, _Gsyscall)
  1953  	sched.ngsys.Add(-1)
  1954  }
  1955  
  1956  // newextram allocates m's and puts them on the extra list.
  1957  // It is called with a working local m, so that it can do things
  1958  // like call schedlock and allocate.
  1959  func newextram() {
  1960  	c := extraMWaiters.Swap(0)
  1961  	if c > 0 {
  1962  		for i := uint32(0); i < c; i++ {
  1963  			oneNewExtraM()
  1964  		}
  1965  	} else {
  1966  		// Make sure there is at least one extra M.
  1967  		mp := lockextra(true)
  1968  		unlockextra(mp)
  1969  		if mp == nil {
  1970  			oneNewExtraM()
  1971  		}
  1972  	}
  1973  }
  1974  
  1975  // oneNewExtraM allocates an m and puts it on the extra list.
  1976  func oneNewExtraM() {
  1977  	// Create extra goroutine locked to extra m.
  1978  	// The goroutine is the context in which the cgo callback will run.
  1979  	// The sched.pc will never be returned to, but setting it to
  1980  	// goexit makes clear to the traceback routines where
  1981  	// the goroutine stack ends.
  1982  	mp := allocm(nil, nil, -1)
  1983  	gp := malg(4096)
  1984  	gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
  1985  	gp.sched.sp = gp.stack.hi
  1986  	gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
  1987  	gp.sched.lr = 0
  1988  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1989  	gp.syscallpc = gp.sched.pc
  1990  	gp.syscallsp = gp.sched.sp
  1991  	gp.stktopsp = gp.sched.sp
  1992  	// malg returns status as _Gidle. Change to _Gdead before
  1993  	// adding to allg where GC can see it. We use _Gdead to hide
  1994  	// this from tracebacks and stack scans since it isn't a
  1995  	// "real" goroutine until needm grabs it.
  1996  	casgstatus(gp, _Gidle, _Gdead)
  1997  	gp.m = mp
  1998  	mp.curg = gp
  1999  	mp.isextra = true
  2000  	mp.lockedInt++
  2001  	mp.lockedg.set(gp)
  2002  	gp.lockedm.set(mp)
  2003  	gp.goid = sched.goidgen.Add(1)
  2004  	gp.sysblocktraced = true
  2005  	if raceenabled {
  2006  		gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
  2007  	}
  2008  	if trace.enabled {
  2009  		// Trigger two trace events for the locked g in the extra m,
  2010  		// since the next event of the g will be traceEvGoSysExit in exitsyscall,
  2011  		// while calling from C thread to Go.
  2012  		traceGoCreate(gp, 0) // no start pc
  2013  		gp.traceseq++
  2014  		traceEvent(traceEvGoInSyscall, -1, gp.goid)
  2015  	}
  2016  	// put on allg for garbage collector
  2017  	allgadd(gp)
  2018  
  2019  	// gp is now on the allg list, but we don't want it to be
  2020  	// counted by gcount. It would be more "proper" to increment
  2021  	// sched.ngfree, but that requires locking. Incrementing ngsys
  2022  	// has the same effect.
  2023  	sched.ngsys.Add(1)
  2024  
  2025  	// Add m to the extra list.
  2026  	mnext := lockextra(true)
  2027  	mp.schedlink.set(mnext)
  2028  	extraMCount++
  2029  	unlockextra(mp)
  2030  }
  2031  
  2032  // dropm is called when a cgo callback has called needm but is now
  2033  // done with the callback and returning back into the non-Go thread.
  2034  // It puts the current m back onto the extra list.
  2035  //
  2036  // The main expense here is the call to signalstack to release the
  2037  // m's signal stack, and then the call to needm on the next callback
  2038  // from this thread. It is tempting to try to save the m for next time,
  2039  // which would eliminate both these costs, but there might not be
  2040  // a next time: the current thread (which Go does not control) might exit.
  2041  // If we saved the m for that thread, there would be an m leak each time
  2042  // such a thread exited. Instead, we acquire and release an m on each
  2043  // call. These should typically not be scheduling operations, just a few
  2044  // atomics, so the cost should be small.
  2045  //
  2046  // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
  2047  // variable using pthread_key_create. Unlike the pthread keys we already use
  2048  // on OS X, this dummy key would never be read by Go code. It would exist
  2049  // only so that we could register at thread-exit-time destructor.
  2050  // That destructor would put the m back onto the extra list.
  2051  // This is purely a performance optimization. The current version,
  2052  // in which dropm happens on each cgo call, is still correct too.
  2053  // We may have to keep the current version on systems with cgo
  2054  // but without pthreads, like Windows.
  2055  func dropm() {
  2056  	// Clear m and g, and return m to the extra list.
  2057  	// After the call to setg we can only call nosplit functions
  2058  	// with no pointer manipulation.
  2059  	mp := getg().m
  2060  
  2061  	// Return mp.curg to dead state.
  2062  	casgstatus(mp.curg, _Gsyscall, _Gdead)
  2063  	mp.curg.preemptStop = false
  2064  	sched.ngsys.Add(1)
  2065  
  2066  	// Block signals before unminit.
  2067  	// Unminit unregisters the signal handling stack (but needs g on some systems).
  2068  	// Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
  2069  	// It's important not to try to handle a signal between those two steps.
  2070  	sigmask := mp.sigmask
  2071  	sigblock(false)
  2072  	unminit()
  2073  
  2074  	mnext := lockextra(true)
  2075  	extraMCount++
  2076  	mp.schedlink.set(mnext)
  2077  
  2078  	setg(nil)
  2079  
  2080  	// Commit the release of mp.
  2081  	unlockextra(mp)
  2082  
  2083  	msigrestore(sigmask)
  2084  }
  2085  
  2086  // A helper function for EnsureDropM.
  2087  func getm() uintptr {
  2088  	return uintptr(unsafe.Pointer(getg().m))
  2089  }
  2090  
  2091  var extram atomic.Uintptr
  2092  var extraMCount uint32 // Protected by lockextra
  2093  var extraMWaiters atomic.Uint32
  2094  
  2095  // lockextra locks the extra list and returns the list head.
  2096  // The caller must unlock the list by storing a new list head
  2097  // to extram. If nilokay is true, then lockextra will
  2098  // return a nil list head if that's what it finds. If nilokay is false,
  2099  // lockextra will keep waiting until the list head is no longer nil.
  2100  //
  2101  //go:nosplit
  2102  func lockextra(nilokay bool) *m {
  2103  	const locked = 1
  2104  
  2105  	incr := false
  2106  	for {
  2107  		old := extram.Load()
  2108  		if old == locked {
  2109  			osyield_no_g()
  2110  			continue
  2111  		}
  2112  		if old == 0 && !nilokay {
  2113  			if !incr {
  2114  				// Add 1 to the number of threads
  2115  				// waiting for an M.
  2116  				// This is cleared by newextram.
  2117  				extraMWaiters.Add(1)
  2118  				incr = true
  2119  			}
  2120  			usleep_no_g(1)
  2121  			continue
  2122  		}
  2123  		if extram.CompareAndSwap(old, locked) {
  2124  			return (*m)(unsafe.Pointer(old))
  2125  		}
  2126  		osyield_no_g()
  2127  		continue
  2128  	}
  2129  }
  2130  
  2131  //go:nosplit
  2132  func unlockextra(mp *m) {
  2133  	extram.Store(uintptr(unsafe.Pointer(mp)))
  2134  }
  2135  
  2136  var (
  2137  	// allocmLock is locked for read when creating new Ms in allocm and their
  2138  	// addition to allm. Thus acquiring this lock for write blocks the
  2139  	// creation of new Ms.
  2140  	allocmLock rwmutex
  2141  
  2142  	// execLock serializes exec and clone to avoid bugs or unspecified
  2143  	// behaviour around exec'ing while creating/destroying threads. See
  2144  	// issue #19546.
  2145  	execLock rwmutex
  2146  )
  2147  
  2148  // These errors are reported (via writeErrStr) by some OS-specific
  2149  // versions of newosproc and newosproc0.
  2150  const (
  2151  	failthreadcreate  = "runtime: failed to create new OS thread\n"
  2152  	failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
  2153  )
  2154  
  2155  // newmHandoff contains a list of m structures that need new OS threads.
  2156  // This is used by newm in situations where newm itself can't safely
  2157  // start an OS thread.
  2158  var newmHandoff struct {
  2159  	lock mutex
  2160  
  2161  	// newm points to a list of M structures that need new OS
  2162  	// threads. The list is linked through m.schedlink.
  2163  	newm muintptr
  2164  
  2165  	// waiting indicates that wake needs to be notified when an m
  2166  	// is put on the list.
  2167  	waiting bool
  2168  	wake    note
  2169  
  2170  	// haveTemplateThread indicates that the templateThread has
  2171  	// been started. This is not protected by lock. Use cas to set
  2172  	// to 1.
  2173  	haveTemplateThread uint32
  2174  }
  2175  
  2176  // Create a new m. It will start off with a call to fn, or else the scheduler.
  2177  // fn needs to be static and not a heap allocated closure.
  2178  // May run with m.p==nil, so write barriers are not allowed.
  2179  //
  2180  // id is optional pre-allocated m ID. Omit by passing -1.
  2181  //
  2182  //go:nowritebarrierrec
  2183  func newm(fn func(), pp *p, id int64) {
  2184  	// allocm adds a new M to allm, but they do not start until created by
  2185  	// the OS in newm1 or the template thread.
  2186  	//
  2187  	// doAllThreadsSyscall requires that every M in allm will eventually
  2188  	// start and be signal-able, even with a STW.
  2189  	//
  2190  	// Disable preemption here until we start the thread to ensure that
  2191  	// newm is not preempted between allocm and starting the new thread,
  2192  	// ensuring that anything added to allm is guaranteed to eventually
  2193  	// start.
  2194  	acquirem()
  2195  
  2196  	mp := allocm(pp, fn, id)
  2197  	mp.nextp.set(pp)
  2198  	mp.sigmask = initSigmask
  2199  	if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
  2200  		// We're on a locked M or a thread that may have been
  2201  		// started by C. The kernel state of this thread may
  2202  		// be strange (the user may have locked it for that
  2203  		// purpose). We don't want to clone that into another
  2204  		// thread. Instead, ask a known-good thread to create
  2205  		// the thread for us.
  2206  		//
  2207  		// This is disabled on Plan 9. See golang.org/issue/22227.
  2208  		//
  2209  		// TODO: This may be unnecessary on Windows, which
  2210  		// doesn't model thread creation off fork.
  2211  		lock(&newmHandoff.lock)
  2212  		if newmHandoff.haveTemplateThread == 0 {
  2213  			throw("on a locked thread with no template thread")
  2214  		}
  2215  		mp.schedlink = newmHandoff.newm
  2216  		newmHandoff.newm.set(mp)
  2217  		if newmHandoff.waiting {
  2218  			newmHandoff.waiting = false
  2219  			notewakeup(&newmHandoff.wake)
  2220  		}
  2221  		unlock(&newmHandoff.lock)
  2222  		// The M has not started yet, but the template thread does not
  2223  		// participate in STW, so it will always process queued Ms and
  2224  		// it is safe to releasem.
  2225  		releasem(getg().m)
  2226  		return
  2227  	}
  2228  	newm1(mp)
  2229  	releasem(getg().m)
  2230  }
  2231  
  2232  func newm1(mp *m) {
  2233  	if iscgo {
  2234  		var ts cgothreadstart
  2235  		if _cgo_thread_start == nil {
  2236  			throw("_cgo_thread_start missing")
  2237  		}
  2238  		ts.g.set(mp.g0)
  2239  		ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
  2240  		ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
  2241  		if msanenabled {
  2242  			msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2243  		}
  2244  		if asanenabled {
  2245  			asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2246  		}
  2247  		execLock.rlock() // Prevent process clone.
  2248  		asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
  2249  		execLock.runlock()
  2250  		return
  2251  	}
  2252  	execLock.rlock() // Prevent process clone.
  2253  	newosproc(mp)
  2254  	execLock.runlock()
  2255  }
  2256  
  2257  // startTemplateThread starts the template thread if it is not already
  2258  // running.
  2259  //
  2260  // The calling thread must itself be in a known-good state.
  2261  func startTemplateThread() {
  2262  	if GOARCH == "wasm" { // no threads on wasm yet
  2263  		return
  2264  	}
  2265  
  2266  	// Disable preemption to guarantee that the template thread will be
  2267  	// created before a park once haveTemplateThread is set.
  2268  	mp := acquirem()
  2269  	if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
  2270  		releasem(mp)
  2271  		return
  2272  	}
  2273  	newm(templateThread, nil, -1)
  2274  	releasem(mp)
  2275  }
  2276  
  2277  // templateThread is a thread in a known-good state that exists solely
  2278  // to start new threads in known-good states when the calling thread
  2279  // may not be in a good state.
  2280  //
  2281  // Many programs never need this, so templateThread is started lazily
  2282  // when we first enter a state that might lead to running on a thread
  2283  // in an unknown state.
  2284  //
  2285  // templateThread runs on an M without a P, so it must not have write
  2286  // barriers.
  2287  //
  2288  //go:nowritebarrierrec
  2289  func templateThread() {
  2290  	lock(&sched.lock)
  2291  	sched.nmsys++
  2292  	checkdead()
  2293  	unlock(&sched.lock)
  2294  
  2295  	for {
  2296  		lock(&newmHandoff.lock)
  2297  		for newmHandoff.newm != 0 {
  2298  			newm := newmHandoff.newm.ptr()
  2299  			newmHandoff.newm = 0
  2300  			unlock(&newmHandoff.lock)
  2301  			for newm != nil {
  2302  				next := newm.schedlink.ptr()
  2303  				newm.schedlink = 0
  2304  				newm1(newm)
  2305  				newm = next
  2306  			}
  2307  			lock(&newmHandoff.lock)
  2308  		}
  2309  		newmHandoff.waiting = true
  2310  		noteclear(&newmHandoff.wake)
  2311  		unlock(&newmHandoff.lock)
  2312  		notesleep(&newmHandoff.wake)
  2313  	}
  2314  }
  2315  
  2316  // Stops execution of the current m until new work is available.
  2317  // Returns with acquired P.
  2318  func stopm() {
  2319  	gp := getg()
  2320  
  2321  	if gp.m.locks != 0 {
  2322  		throw("stopm holding locks")
  2323  	}
  2324  	if gp.m.p != 0 {
  2325  		throw("stopm holding p")
  2326  	}
  2327  	if gp.m.spinning {
  2328  		throw("stopm spinning")
  2329  	}
  2330  
  2331  	lock(&sched.lock)
  2332  	mput(gp.m)
  2333  	unlock(&sched.lock)
  2334  	mPark()
  2335  	acquirep(gp.m.nextp.ptr())
  2336  	gp.m.nextp = 0
  2337  }
  2338  
  2339  func mspinning() {
  2340  	// startm's caller incremented nmspinning. Set the new M's spinning.
  2341  	getg().m.spinning = true
  2342  }
  2343  
  2344  // Schedules some M to run the p (creates an M if necessary).
  2345  // If p==nil, tries to get an idle P, if no idle P's does nothing.
  2346  // May run with m.p==nil, so write barriers are not allowed.
  2347  // If spinning is set, the caller has incremented nmspinning and must provide a
  2348  // P. startm will set m.spinning in the newly started M.
  2349  //
  2350  // Callers passing a non-nil P must call from a non-preemptible context. See
  2351  // comment on acquirem below.
  2352  //
  2353  // Must not have write barriers because this may be called without a P.
  2354  //
  2355  //go:nowritebarrierrec
  2356  func startm(pp *p, spinning bool) {
  2357  	// Disable preemption.
  2358  	//
  2359  	// Every owned P must have an owner that will eventually stop it in the
  2360  	// event of a GC stop request. startm takes transient ownership of a P
  2361  	// (either from argument or pidleget below) and transfers ownership to
  2362  	// a started M, which will be responsible for performing the stop.
  2363  	//
  2364  	// Preemption must be disabled during this transient ownership,
  2365  	// otherwise the P this is running on may enter GC stop while still
  2366  	// holding the transient P, leaving that P in limbo and deadlocking the
  2367  	// STW.
  2368  	//
  2369  	// Callers passing a non-nil P must already be in non-preemptible
  2370  	// context, otherwise such preemption could occur on function entry to
  2371  	// startm. Callers passing a nil P may be preemptible, so we must
  2372  	// disable preemption before acquiring a P from pidleget below.
  2373  	mp := acquirem()
  2374  	lock(&sched.lock)
  2375  	if pp == nil {
  2376  		if spinning {
  2377  			// TODO(prattmic): All remaining calls to this function
  2378  			// with _p_ == nil could be cleaned up to find a P
  2379  			// before calling startm.
  2380  			throw("startm: P required for spinning=true")
  2381  		}
  2382  		pp, _ = pidleget(0)
  2383  		if pp == nil {
  2384  			unlock(&sched.lock)
  2385  			releasem(mp)
  2386  			return
  2387  		}
  2388  	}
  2389  	nmp := mget()
  2390  	if nmp == nil {
  2391  		// No M is available, we must drop sched.lock and call newm.
  2392  		// However, we already own a P to assign to the M.
  2393  		//
  2394  		// Once sched.lock is released, another G (e.g., in a syscall),
  2395  		// could find no idle P while checkdead finds a runnable G but
  2396  		// no running M's because this new M hasn't started yet, thus
  2397  		// throwing in an apparent deadlock.
  2398  		//
  2399  		// Avoid this situation by pre-allocating the ID for the new M,
  2400  		// thus marking it as 'running' before we drop sched.lock. This
  2401  		// new M will eventually run the scheduler to execute any
  2402  		// queued G's.
  2403  		id := mReserveID()
  2404  		unlock(&sched.lock)
  2405  
  2406  		var fn func()
  2407  		if spinning {
  2408  			// The caller incremented nmspinning, so set m.spinning in the new M.
  2409  			fn = mspinning
  2410  		}
  2411  		newm(fn, pp, id)
  2412  		// Ownership transfer of pp committed by start in newm.
  2413  		// Preemption is now safe.
  2414  		releasem(mp)
  2415  		return
  2416  	}
  2417  	unlock(&sched.lock)
  2418  	if nmp.spinning {
  2419  		throw("startm: m is spinning")
  2420  	}
  2421  	if nmp.nextp != 0 {
  2422  		throw("startm: m has p")
  2423  	}
  2424  	if spinning && !runqempty(pp) {
  2425  		throw("startm: p has runnable gs")
  2426  	}
  2427  	// The caller incremented nmspinning, so set m.spinning in the new M.
  2428  	nmp.spinning = spinning
  2429  	nmp.nextp.set(pp)
  2430  	notewakeup(&nmp.park)
  2431  	// Ownership transfer of pp committed by wakeup. Preemption is now
  2432  	// safe.
  2433  	releasem(mp)
  2434  }
  2435  
  2436  // Hands off P from syscall or locked M.
  2437  // Always runs without a P, so write barriers are not allowed.
  2438  //
  2439  //go:nowritebarrierrec
  2440  func handoffp(pp *p) {
  2441  	// handoffp must start an M in any situation where
  2442  	// findrunnable would return a G to run on pp.
  2443  
  2444  	// if it has local work, start it straight away
  2445  	if !runqempty(pp) || sched.runqsize != 0 {
  2446  		startm(pp, false)
  2447  		return
  2448  	}
  2449  	// if there's trace work to do, start it straight away
  2450  	if (trace.enabled || trace.shutdown) && traceReaderAvailable() != nil {
  2451  		startm(pp, false)
  2452  		return
  2453  	}
  2454  	// if it has GC work, start it straight away
  2455  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
  2456  		startm(pp, false)
  2457  		return
  2458  	}
  2459  	// no local work, check that there are no spinning/idle M's,
  2460  	// otherwise our help is not required
  2461  	if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
  2462  		sched.needspinning.Store(0)
  2463  		startm(pp, true)
  2464  		return
  2465  	}
  2466  	lock(&sched.lock)
  2467  	if sched.gcwaiting.Load() {
  2468  		pp.status = _Pgcstop
  2469  		sched.stopwait--
  2470  		if sched.stopwait == 0 {
  2471  			notewakeup(&sched.stopnote)
  2472  		}
  2473  		unlock(&sched.lock)
  2474  		return
  2475  	}
  2476  	if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
  2477  		sched.safePointFn(pp)
  2478  		sched.safePointWait--
  2479  		if sched.safePointWait == 0 {
  2480  			notewakeup(&sched.safePointNote)
  2481  		}
  2482  	}
  2483  	if sched.runqsize != 0 {
  2484  		unlock(&sched.lock)
  2485  		startm(pp, false)
  2486  		return
  2487  	}
  2488  	// If this is the last running P and nobody is polling network,
  2489  	// need to wakeup another M to poll network.
  2490  	if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
  2491  		unlock(&sched.lock)
  2492  		startm(pp, false)
  2493  		return
  2494  	}
  2495  
  2496  	// The scheduler lock cannot be held when calling wakeNetPoller below
  2497  	// because wakeNetPoller may call wakep which may call startm.
  2498  	when := nobarrierWakeTime(pp)
  2499  	pidleput(pp, 0)
  2500  	unlock(&sched.lock)
  2501  
  2502  	if when != 0 {
  2503  		wakeNetPoller(when)
  2504  	}
  2505  }
  2506  
  2507  // Tries to add one more P to execute G's.
  2508  // Called when a G is made runnable (newproc, ready).
  2509  // Must be called with a P.
  2510  func wakep() {
  2511  	// Be conservative about spinning threads, only start one if none exist
  2512  	// already.
  2513  	if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
  2514  		return
  2515  	}
  2516  
  2517  	// Disable preemption until ownership of pp transfers to the next M in
  2518  	// startm. Otherwise preemption here would leave pp stuck waiting to
  2519  	// enter _Pgcstop.
  2520  	//
  2521  	// See preemption comment on acquirem in startm for more details.
  2522  	mp := acquirem()
  2523  
  2524  	var pp *p
  2525  	lock(&sched.lock)
  2526  	pp, _ = pidlegetSpinning(0)
  2527  	if pp == nil {
  2528  		if sched.nmspinning.Add(-1) < 0 {
  2529  			throw("wakep: negative nmspinning")
  2530  		}
  2531  		unlock(&sched.lock)
  2532  		releasem(mp)
  2533  		return
  2534  	}
  2535  	// Since we always have a P, the race in the "No M is available"
  2536  	// comment in startm doesn't apply during the small window between the
  2537  	// unlock here and lock in startm. A checkdead in between will always
  2538  	// see at least one running M (ours).
  2539  	unlock(&sched.lock)
  2540  
  2541  	startm(pp, true)
  2542  
  2543  	releasem(mp)
  2544  }
  2545  
  2546  // Stops execution of the current m that is locked to a g until the g is runnable again.
  2547  // Returns with acquired P.
  2548  func stoplockedm() {
  2549  	gp := getg()
  2550  
  2551  	if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
  2552  		throw("stoplockedm: inconsistent locking")
  2553  	}
  2554  	if gp.m.p != 0 {
  2555  		// Schedule another M to run this p.
  2556  		pp := releasep()
  2557  		handoffp(pp)
  2558  	}
  2559  	incidlelocked(1)
  2560  	// Wait until another thread schedules lockedg again.
  2561  	mPark()
  2562  	status := readgstatus(gp.m.lockedg.ptr())
  2563  	if status&^_Gscan != _Grunnable {
  2564  		print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
  2565  		dumpgstatus(gp.m.lockedg.ptr())
  2566  		throw("stoplockedm: not runnable")
  2567  	}
  2568  	acquirep(gp.m.nextp.ptr())
  2569  	gp.m.nextp = 0
  2570  }
  2571  
  2572  // Schedules the locked m to run the locked gp.
  2573  // May run during STW, so write barriers are not allowed.
  2574  //
  2575  //go:nowritebarrierrec
  2576  func startlockedm(gp *g) {
  2577  	mp := gp.lockedm.ptr()
  2578  	if mp == getg().m {
  2579  		throw("startlockedm: locked to me")
  2580  	}
  2581  	if mp.nextp != 0 {
  2582  		throw("startlockedm: m has p")
  2583  	}
  2584  	// directly handoff current P to the locked m
  2585  	incidlelocked(-1)
  2586  	pp := releasep()
  2587  	mp.nextp.set(pp)
  2588  	notewakeup(&mp.park)
  2589  	stopm()
  2590  }
  2591  
  2592  // Stops the current m for stopTheWorld.
  2593  // Returns when the world is restarted.
  2594  func gcstopm() {
  2595  	gp := getg()
  2596  
  2597  	if !sched.gcwaiting.Load() {
  2598  		throw("gcstopm: not waiting for gc")
  2599  	}
  2600  	if gp.m.spinning {
  2601  		gp.m.spinning = false
  2602  		// OK to just drop nmspinning here,
  2603  		// startTheWorld will unpark threads as necessary.
  2604  		if sched.nmspinning.Add(-1) < 0 {
  2605  			throw("gcstopm: negative nmspinning")
  2606  		}
  2607  	}
  2608  	pp := releasep()
  2609  	lock(&sched.lock)
  2610  	pp.status = _Pgcstop
  2611  	sched.stopwait--
  2612  	if sched.stopwait == 0 {
  2613  		notewakeup(&sched.stopnote)
  2614  	}
  2615  	unlock(&sched.lock)
  2616  	stopm()
  2617  }
  2618  
  2619  // Schedules gp to run on the current M.
  2620  // If inheritTime is true, gp inherits the remaining time in the
  2621  // current time slice. Otherwise, it starts a new time slice.
  2622  // Never returns.
  2623  //
  2624  // Write barriers are allowed because this is called immediately after
  2625  // acquiring a P in several places.
  2626  //
  2627  //go:yeswritebarrierrec
  2628  func execute(gp *g, inheritTime bool) {
  2629  	mp := getg().m
  2630  
  2631  	if goroutineProfile.active {
  2632  		// Make sure that gp has had its stack written out to the goroutine
  2633  		// profile, exactly as it was when the goroutine profiler first stopped
  2634  		// the world.
  2635  		tryRecordGoroutineProfile(gp, osyield)
  2636  	}
  2637  
  2638  	// Assign gp.m before entering _Grunning so running Gs have an
  2639  	// M.
  2640  	mp.curg = gp
  2641  	gp.m = mp
  2642  	casgstatus(gp, _Grunnable, _Grunning)
  2643  	gp.waitsince = 0
  2644  	gp.preempt = false
  2645  	gp.stackguard0 = gp.stack.lo + _StackGuard
  2646  	if !inheritTime {
  2647  		mp.p.ptr().schedtick++
  2648  	}
  2649  
  2650  	// Check whether the profiler needs to be turned on or off.
  2651  	hz := sched.profilehz
  2652  	if mp.profilehz != hz {
  2653  		setThreadCPUProfiler(hz)
  2654  	}
  2655  
  2656  	if trace.enabled {
  2657  		// GoSysExit has to happen when we have a P, but before GoStart.
  2658  		// So we emit it here.
  2659  		if gp.syscallsp != 0 && gp.sysblocktraced {
  2660  			traceGoSysExit(gp.sysexitticks)
  2661  		}
  2662  		traceGoStart()
  2663  	}
  2664  
  2665  	gogo(&gp.sched)
  2666  }
  2667  
  2668  // Finds a runnable goroutine to execute.
  2669  // Tries to steal from other P's, get g from local or global queue, poll network.
  2670  // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
  2671  // reader) so the caller should try to wake a P.
  2672  func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
  2673  	mp := getg().m
  2674  
  2675  	// The conditions here and in handoffp must agree: if
  2676  	// findrunnable would return a G to run, handoffp must start
  2677  	// an M.
  2678  
  2679  top:
  2680  	pp := mp.p.ptr()
  2681  	if sched.gcwaiting.Load() {
  2682  		gcstopm()
  2683  		goto top
  2684  	}
  2685  	if pp.runSafePointFn != 0 {
  2686  		runSafePointFn()
  2687  	}
  2688  
  2689  	// now and pollUntil are saved for work stealing later,
  2690  	// which may steal timers. It's important that between now
  2691  	// and then, nothing blocks, so these numbers remain mostly
  2692  	// relevant.
  2693  	now, pollUntil, _ := checkTimers(pp, 0)
  2694  
  2695  	// Try to schedule the trace reader.
  2696  	if trace.enabled || trace.shutdown {
  2697  		gp := traceReader()
  2698  		if gp != nil {
  2699  			casgstatus(gp, _Gwaiting, _Grunnable)
  2700  			traceGoUnpark(gp, 0)
  2701  			return gp, false, true
  2702  		}
  2703  	}
  2704  
  2705  	// Try to schedule a GC worker.
  2706  	if gcBlackenEnabled != 0 {
  2707  		gp, tnow := gcController.findRunnableGCWorker(pp, now)
  2708  		if gp != nil {
  2709  			return gp, false, true
  2710  		}
  2711  		now = tnow
  2712  	}
  2713  
  2714  	// Check the global runnable queue once in a while to ensure fairness.
  2715  	// Otherwise two goroutines can completely occupy the local runqueue
  2716  	// by constantly respawning each other.
  2717  	if pp.schedtick%61 == 0 && sched.runqsize > 0 {
  2718  		lock(&sched.lock)
  2719  		gp := globrunqget(pp, 1)
  2720  		unlock(&sched.lock)
  2721  		if gp != nil {
  2722  			return gp, false, false
  2723  		}
  2724  	}
  2725  
  2726  	// Wake up the finalizer G.
  2727  	if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
  2728  		if gp := wakefing(); gp != nil {
  2729  			ready(gp, 0, true)
  2730  		}
  2731  	}
  2732  	if *cgo_yield != nil {
  2733  		asmcgocall(*cgo_yield, nil)
  2734  	}
  2735  
  2736  	// local runq
  2737  	if gp, inheritTime := runqget(pp); gp != nil {
  2738  		return gp, inheritTime, false
  2739  	}
  2740  
  2741  	// global runq
  2742  	if sched.runqsize != 0 {
  2743  		lock(&sched.lock)
  2744  		gp := globrunqget(pp, 0)
  2745  		unlock(&sched.lock)
  2746  		if gp != nil {
  2747  			return gp, false, false
  2748  		}
  2749  	}
  2750  
  2751  	// Poll network.
  2752  	// This netpoll is only an optimization before we resort to stealing.
  2753  	// We can safely skip it if there are no waiters or a thread is blocked
  2754  	// in netpoll already. If there is any kind of logical race with that
  2755  	// blocked thread (e.g. it has already returned from netpoll, but does
  2756  	// not set lastpoll yet), this thread will do blocking netpoll below
  2757  	// anyway.
  2758  	if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
  2759  		if list := netpoll(0); !list.empty() { // non-blocking
  2760  			gp := list.pop()
  2761  			injectglist(&list)
  2762  			casgstatus(gp, _Gwaiting, _Grunnable)
  2763  			if trace.enabled {
  2764  				traceGoUnpark(gp, 0)
  2765  			}
  2766  			return gp, false, false
  2767  		}
  2768  	}
  2769  
  2770  	// Spinning Ms: steal work from other Ps.
  2771  	//
  2772  	// Limit the number of spinning Ms to half the number of busy Ps.
  2773  	// This is necessary to prevent excessive CPU consumption when
  2774  	// GOMAXPROCS>>1 but the program parallelism is low.
  2775  	if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
  2776  		if !mp.spinning {
  2777  			mp.becomeSpinning()
  2778  		}
  2779  
  2780  		gp, inheritTime, tnow, w, newWork := stealWork(now)
  2781  		if gp != nil {
  2782  			// Successfully stole.
  2783  			return gp, inheritTime, false
  2784  		}
  2785  		if newWork {
  2786  			// There may be new timer or GC work; restart to
  2787  			// discover.
  2788  			goto top
  2789  		}
  2790  
  2791  		now = tnow
  2792  		if w != 0 && (pollUntil == 0 || w < pollUntil) {
  2793  			// Earlier timer to wait for.
  2794  			pollUntil = w
  2795  		}
  2796  	}
  2797  
  2798  	// We have nothing to do.
  2799  	//
  2800  	// If we're in the GC mark phase, can safely scan and blacken objects,
  2801  	// and have work to do, run idle-time marking rather than give up the P.
  2802  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
  2803  		node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  2804  		if node != nil {
  2805  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  2806  			gp := node.gp.ptr()
  2807  			casgstatus(gp, _Gwaiting, _Grunnable)
  2808  			if trace.enabled {
  2809  				traceGoUnpark(gp, 0)
  2810  			}
  2811  			return gp, false, false
  2812  		}
  2813  		gcController.removeIdleMarkWorker()
  2814  	}
  2815  
  2816  	// wasm only:
  2817  	// If a callback returned and no other goroutine is awake,
  2818  	// then wake event handler goroutine which pauses execution
  2819  	// until a callback was triggered.
  2820  	gp, otherReady := beforeIdle(now, pollUntil)
  2821  	if gp != nil {
  2822  		casgstatus(gp, _Gwaiting, _Grunnable)
  2823  		if trace.enabled {
  2824  			traceGoUnpark(gp, 0)
  2825  		}
  2826  		return gp, false, false
  2827  	}
  2828  	if otherReady {
  2829  		goto top
  2830  	}
  2831  
  2832  	// Before we drop our P, make a snapshot of the allp slice,
  2833  	// which can change underfoot once we no longer block
  2834  	// safe-points. We don't need to snapshot the contents because
  2835  	// everything up to cap(allp) is immutable.
  2836  	allpSnapshot := allp
  2837  	// Also snapshot masks. Value changes are OK, but we can't allow
  2838  	// len to change out from under us.
  2839  	idlepMaskSnapshot := idlepMask
  2840  	timerpMaskSnapshot := timerpMask
  2841  
  2842  	// return P and block
  2843  	lock(&sched.lock)
  2844  	if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
  2845  		unlock(&sched.lock)
  2846  		goto top
  2847  	}
  2848  	if sched.runqsize != 0 {
  2849  		gp := globrunqget(pp, 0)
  2850  		unlock(&sched.lock)
  2851  		return gp, false, false
  2852  	}
  2853  	if !mp.spinning && sched.needspinning.Load() == 1 {
  2854  		// See "Delicate dance" comment below.
  2855  		mp.becomeSpinning()
  2856  		unlock(&sched.lock)
  2857  		goto top
  2858  	}
  2859  	if releasep() != pp {
  2860  		throw("findrunnable: wrong p")
  2861  	}
  2862  	now = pidleput(pp, now)
  2863  	unlock(&sched.lock)
  2864  
  2865  	// Delicate dance: thread transitions from spinning to non-spinning
  2866  	// state, potentially concurrently with submission of new work. We must
  2867  	// drop nmspinning first and then check all sources again (with
  2868  	// #StoreLoad memory barrier in between). If we do it the other way
  2869  	// around, another thread can submit work after we've checked all
  2870  	// sources but before we drop nmspinning; as a result nobody will
  2871  	// unpark a thread to run the work.
  2872  	//
  2873  	// This applies to the following sources of work:
  2874  	//
  2875  	// * Goroutines added to a per-P run queue.
  2876  	// * New/modified-earlier timers on a per-P timer heap.
  2877  	// * Idle-priority GC work (barring golang.org/issue/19112).
  2878  	//
  2879  	// If we discover new work below, we need to restore m.spinning as a
  2880  	// signal for resetspinning to unpark a new worker thread (because
  2881  	// there can be more than one starving goroutine).
  2882  	//
  2883  	// However, if after discovering new work we also observe no idle Ps
  2884  	// (either here or in resetspinning), we have a problem. We may be
  2885  	// racing with a non-spinning M in the block above, having found no
  2886  	// work and preparing to release its P and park. Allowing that P to go
  2887  	// idle will result in loss of work conservation (idle P while there is
  2888  	// runnable work). This could result in complete deadlock in the
  2889  	// unlikely event that we discover new work (from netpoll) right as we
  2890  	// are racing with _all_ other Ps going idle.
  2891  	//
  2892  	// We use sched.needspinning to synchronize with non-spinning Ms going
  2893  	// idle. If needspinning is set when they are about to drop their P,
  2894  	// they abort the drop and instead become a new spinning M on our
  2895  	// behalf. If we are not racing and the system is truly fully loaded
  2896  	// then no spinning threads are required, and the next thread to
  2897  	// naturally become spinning will clear the flag.
  2898  	//
  2899  	// Also see "Worker thread parking/unparking" comment at the top of the
  2900  	// file.
  2901  	wasSpinning := mp.spinning
  2902  	if mp.spinning {
  2903  		mp.spinning = false
  2904  		if sched.nmspinning.Add(-1) < 0 {
  2905  			throw("findrunnable: negative nmspinning")
  2906  		}
  2907  
  2908  		// Note the for correctness, only the last M transitioning from
  2909  		// spinning to non-spinning must perform these rechecks to
  2910  		// ensure no missed work. However, the runtime has some cases
  2911  		// of transient increments of nmspinning that are decremented
  2912  		// without going through this path, so we must be conservative
  2913  		// and perform the check on all spinning Ms.
  2914  		//
  2915  		// See https://go.dev/issue/43997.
  2916  
  2917  		// Check all runqueues once again.
  2918  		pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
  2919  		if pp != nil {
  2920  			acquirep(pp)
  2921  			mp.becomeSpinning()
  2922  			goto top
  2923  		}
  2924  
  2925  		// Check for idle-priority GC work again.
  2926  		pp, gp := checkIdleGCNoP()
  2927  		if pp != nil {
  2928  			acquirep(pp)
  2929  			mp.becomeSpinning()
  2930  
  2931  			// Run the idle worker.
  2932  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  2933  			casgstatus(gp, _Gwaiting, _Grunnable)
  2934  			if trace.enabled {
  2935  				traceGoUnpark(gp, 0)
  2936  			}
  2937  			return gp, false, false
  2938  		}
  2939  
  2940  		// Finally, check for timer creation or expiry concurrently with
  2941  		// transitioning from spinning to non-spinning.
  2942  		//
  2943  		// Note that we cannot use checkTimers here because it calls
  2944  		// adjusttimers which may need to allocate memory, and that isn't
  2945  		// allowed when we don't have an active P.
  2946  		pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
  2947  	}
  2948  
  2949  	// Poll network until next timer.
  2950  	if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
  2951  		sched.pollUntil.Store(pollUntil)
  2952  		if mp.p != 0 {
  2953  			throw("findrunnable: netpoll with p")
  2954  		}
  2955  		if mp.spinning {
  2956  			throw("findrunnable: netpoll with spinning")
  2957  		}
  2958  		// Refresh now.
  2959  		now = nanotime()
  2960  		delay := int64(-1)
  2961  		if pollUntil != 0 {
  2962  			delay = pollUntil - now
  2963  			if delay < 0 {
  2964  				delay = 0
  2965  			}
  2966  		}
  2967  		if faketime != 0 {
  2968  			// When using fake time, just poll.
  2969  			delay = 0
  2970  		}
  2971  		list := netpoll(delay) // block until new work is available
  2972  		sched.pollUntil.Store(0)
  2973  		sched.lastpoll.Store(now)
  2974  		if faketime != 0 && list.empty() {
  2975  			// Using fake time and nothing is ready; stop M.
  2976  			// When all M's stop, checkdead will call timejump.
  2977  			stopm()
  2978  			goto top
  2979  		}
  2980  		lock(&sched.lock)
  2981  		pp, _ := pidleget(now)
  2982  		unlock(&sched.lock)
  2983  		if pp == nil {
  2984  			injectglist(&list)
  2985  		} else {
  2986  			acquirep(pp)
  2987  			if !list.empty() {
  2988  				gp := list.pop()
  2989  				injectglist(&list)
  2990  				casgstatus(gp, _Gwaiting, _Grunnable)
  2991  				if trace.enabled {
  2992  					traceGoUnpark(gp, 0)
  2993  				}
  2994  				return gp, false, false
  2995  			}
  2996  			if wasSpinning {
  2997  				mp.becomeSpinning()
  2998  			}
  2999  			goto top
  3000  		}
  3001  	} else if pollUntil != 0 && netpollinited() {
  3002  		pollerPollUntil := sched.pollUntil.Load()
  3003  		if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
  3004  			netpollBreak()
  3005  		}
  3006  	}
  3007  	stopm()
  3008  	goto top
  3009  }
  3010  
  3011  // pollWork reports whether there is non-background work this P could
  3012  // be doing. This is a fairly lightweight check to be used for
  3013  // background work loops, like idle GC. It checks a subset of the
  3014  // conditions checked by the actual scheduler.
  3015  func pollWork() bool {
  3016  	if sched.runqsize != 0 {
  3017  		return true
  3018  	}
  3019  	p := getg().m.p.ptr()
  3020  	if !runqempty(p) {
  3021  		return true
  3022  	}
  3023  	if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
  3024  		if list := netpoll(0); !list.empty() {
  3025  			injectglist(&list)
  3026  			return true
  3027  		}
  3028  	}
  3029  	return false
  3030  }
  3031  
  3032  // stealWork attempts to steal a runnable goroutine or timer from any P.
  3033  //
  3034  // If newWork is true, new work may have been readied.
  3035  //
  3036  // If now is not 0 it is the current time. stealWork returns the passed time or
  3037  // the current time if now was passed as 0.
  3038  func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
  3039  	pp := getg().m.p.ptr()
  3040  
  3041  	ranTimer := false
  3042  
  3043  	const stealTries = 4
  3044  	for i := 0; i < stealTries; i++ {
  3045  		stealTimersOrRunNextG := i == stealTries-1
  3046  
  3047  		for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
  3048  			if sched.gcwaiting.Load() {
  3049  				// GC work may be available.
  3050  				return nil, false, now, pollUntil, true
  3051  			}
  3052  			p2 := allp[enum.position()]
  3053  			if pp == p2 {
  3054  				continue
  3055  			}
  3056  
  3057  			// Steal timers from p2. This call to checkTimers is the only place
  3058  			// where we might hold a lock on a different P's timers. We do this
  3059  			// once on the last pass before checking runnext because stealing
  3060  			// from the other P's runnext should be the last resort, so if there
  3061  			// are timers to steal do that first.
  3062  			//
  3063  			// We only check timers on one of the stealing iterations because
  3064  			// the time stored in now doesn't change in this loop and checking
  3065  			// the timers for each P more than once with the same value of now
  3066  			// is probably a waste of time.
  3067  			//
  3068  			// timerpMask tells us whether the P may have timers at all. If it
  3069  			// can't, no need to check at all.
  3070  			if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
  3071  				tnow, w, ran := checkTimers(p2, now)
  3072  				now = tnow
  3073  				if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3074  					pollUntil = w
  3075  				}
  3076  				if ran {
  3077  					// Running the timers may have
  3078  					// made an arbitrary number of G's
  3079  					// ready and added them to this P's
  3080  					// local run queue. That invalidates
  3081  					// the assumption of runqsteal
  3082  					// that it always has room to add
  3083  					// stolen G's. So check now if there
  3084  					// is a local G to run.
  3085  					if gp, inheritTime := runqget(pp); gp != nil {
  3086  						return gp, inheritTime, now, pollUntil, ranTimer
  3087  					}
  3088  					ranTimer = true
  3089  				}
  3090  			}
  3091  
  3092  			// Don't bother to attempt to steal if p2 is idle.
  3093  			if !idlepMask.read(enum.position()) {
  3094  				if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
  3095  					return gp, false, now, pollUntil, ranTimer
  3096  				}
  3097  			}
  3098  		}
  3099  	}
  3100  
  3101  	// No goroutines found to steal. Regardless, running a timer may have
  3102  	// made some goroutine ready that we missed. Indicate the next timer to
  3103  	// wait for.
  3104  	return nil, false, now, pollUntil, ranTimer
  3105  }
  3106  
  3107  // Check all Ps for a runnable G to steal.
  3108  //
  3109  // On entry we have no P. If a G is available to steal and a P is available,
  3110  // the P is returned which the caller should acquire and attempt to steal the
  3111  // work to.
  3112  func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
  3113  	for id, p2 := range allpSnapshot {
  3114  		if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
  3115  			lock(&sched.lock)
  3116  			pp, _ := pidlegetSpinning(0)
  3117  			if pp == nil {
  3118  				// Can't get a P, don't bother checking remaining Ps.
  3119  				unlock(&sched.lock)
  3120  				return nil
  3121  			}
  3122  			unlock(&sched.lock)
  3123  			return pp
  3124  		}
  3125  	}
  3126  
  3127  	// No work available.
  3128  	return nil
  3129  }
  3130  
  3131  // Check all Ps for a timer expiring sooner than pollUntil.
  3132  //
  3133  // Returns updated pollUntil value.
  3134  func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
  3135  	for id, p2 := range allpSnapshot {
  3136  		if timerpMaskSnapshot.read(uint32(id)) {
  3137  			w := nobarrierWakeTime(p2)
  3138  			if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3139  				pollUntil = w
  3140  			}
  3141  		}
  3142  	}
  3143  
  3144  	return pollUntil
  3145  }
  3146  
  3147  // Check for idle-priority GC, without a P on entry.
  3148  //
  3149  // If some GC work, a P, and a worker G are all available, the P and G will be
  3150  // returned. The returned P has not been wired yet.
  3151  func checkIdleGCNoP() (*p, *g) {
  3152  	// N.B. Since we have no P, gcBlackenEnabled may change at any time; we
  3153  	// must check again after acquiring a P. As an optimization, we also check
  3154  	// if an idle mark worker is needed at all. This is OK here, because if we
  3155  	// observe that one isn't needed, at least one is currently running. Even if
  3156  	// it stops running, its own journey into the scheduler should schedule it
  3157  	// again, if need be (at which point, this check will pass, if relevant).
  3158  	if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
  3159  		return nil, nil
  3160  	}
  3161  	if !gcMarkWorkAvailable(nil) {
  3162  		return nil, nil
  3163  	}
  3164  
  3165  	// Work is available; we can start an idle GC worker only if there is
  3166  	// an available P and available worker G.
  3167  	//
  3168  	// We can attempt to acquire these in either order, though both have
  3169  	// synchronization concerns (see below). Workers are almost always
  3170  	// available (see comment in findRunnableGCWorker for the one case
  3171  	// there may be none). Since we're slightly less likely to find a P,
  3172  	// check for that first.
  3173  	//
  3174  	// Synchronization: note that we must hold sched.lock until we are
  3175  	// committed to keeping it. Otherwise we cannot put the unnecessary P
  3176  	// back in sched.pidle without performing the full set of idle
  3177  	// transition checks.
  3178  	//
  3179  	// If we were to check gcBgMarkWorkerPool first, we must somehow handle
  3180  	// the assumption in gcControllerState.findRunnableGCWorker that an
  3181  	// empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
  3182  	lock(&sched.lock)
  3183  	pp, now := pidlegetSpinning(0)
  3184  	if pp == nil {
  3185  		unlock(&sched.lock)
  3186  		return nil, nil
  3187  	}
  3188  
  3189  	// Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
  3190  	if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
  3191  		pidleput(pp, now)
  3192  		unlock(&sched.lock)
  3193  		return nil, nil
  3194  	}
  3195  
  3196  	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3197  	if node == nil {
  3198  		pidleput(pp, now)
  3199  		unlock(&sched.lock)
  3200  		gcController.removeIdleMarkWorker()
  3201  		return nil, nil
  3202  	}
  3203  
  3204  	unlock(&sched.lock)
  3205  
  3206  	return pp, node.gp.ptr()
  3207  }
  3208  
  3209  // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
  3210  // going to wake up before the when argument; or it wakes an idle P to service
  3211  // timers and the network poller if there isn't one already.
  3212  func wakeNetPoller(when int64) {
  3213  	if sched.lastpoll.Load() == 0 {
  3214  		// In findrunnable we ensure that when polling the pollUntil
  3215  		// field is either zero or the time to which the current
  3216  		// poll is expected to run. This can have a spurious wakeup
  3217  		// but should never miss a wakeup.
  3218  		pollerPollUntil := sched.pollUntil.Load()
  3219  		if pollerPollUntil == 0 || pollerPollUntil > when {
  3220  			netpollBreak()
  3221  		}
  3222  	} else {
  3223  		// There are no threads in the network poller, try to get
  3224  		// one there so it can handle new timers.
  3225  		if GOOS != "plan9" { // Temporary workaround - see issue #42303.
  3226  			wakep()
  3227  		}
  3228  	}
  3229  }
  3230  
  3231  func resetspinning() {
  3232  	gp := getg()
  3233  	if !gp.m.spinning {
  3234  		throw("resetspinning: not a spinning m")
  3235  	}
  3236  	gp.m.spinning = false
  3237  	nmspinning := sched.nmspinning.Add(-1)
  3238  	if nmspinning < 0 {
  3239  		throw("findrunnable: negative nmspinning")
  3240  	}
  3241  	// M wakeup policy is deliberately somewhat conservative, so check if we
  3242  	// need to wakeup another P here. See "Worker thread parking/unparking"
  3243  	// comment at the top of the file for details.
  3244  	wakep()
  3245  }
  3246  
  3247  // injectglist adds each runnable G on the list to some run queue,
  3248  // and clears glist. If there is no current P, they are added to the
  3249  // global queue, and up to npidle M's are started to run them.
  3250  // Otherwise, for each idle P, this adds a G to the global queue
  3251  // and starts an M. Any remaining G's are added to the current P's
  3252  // local run queue.
  3253  // This may temporarily acquire sched.lock.
  3254  // Can run concurrently with GC.
  3255  func injectglist(glist *gList) {
  3256  	if glist.empty() {
  3257  		return
  3258  	}
  3259  	if trace.enabled {
  3260  		for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
  3261  			traceGoUnpark(gp, 0)
  3262  		}
  3263  	}
  3264  
  3265  	// Mark all the goroutines as runnable before we put them
  3266  	// on the run queues.
  3267  	head := glist.head.ptr()
  3268  	var tail *g
  3269  	qsize := 0
  3270  	for gp := head; gp != nil; gp = gp.schedlink.ptr() {
  3271  		tail = gp
  3272  		qsize++
  3273  		casgstatus(gp, _Gwaiting, _Grunnable)
  3274  	}
  3275  
  3276  	// Turn the gList into a gQueue.
  3277  	var q gQueue
  3278  	q.head.set(head)
  3279  	q.tail.set(tail)
  3280  	*glist = gList{}
  3281  
  3282  	startIdle := func(n int) {
  3283  		for i := 0; i < n; i++ {
  3284  			mp := acquirem() // See comment in startm.
  3285  			lock(&sched.lock)
  3286  
  3287  			pp, _ := pidlegetSpinning(0)
  3288  			if pp == nil {
  3289  				unlock(&sched.lock)
  3290  				releasem(mp)
  3291  				break
  3292  			}
  3293  
  3294  			unlock(&sched.lock)
  3295  			startm(pp, false)
  3296  			releasem(mp)
  3297  		}
  3298  	}
  3299  
  3300  	pp := getg().m.p.ptr()
  3301  	if pp == nil {
  3302  		lock(&sched.lock)
  3303  		globrunqputbatch(&q, int32(qsize))
  3304  		unlock(&sched.lock)
  3305  		startIdle(qsize)
  3306  		return
  3307  	}
  3308  
  3309  	npidle := int(sched.npidle.Load())
  3310  	var globq gQueue
  3311  	var n int
  3312  	for n = 0; n < npidle && !q.empty(); n++ {
  3313  		g := q.pop()
  3314  		globq.pushBack(g)
  3315  	}
  3316  	if n > 0 {
  3317  		lock(&sched.lock)
  3318  		globrunqputbatch(&globq, int32(n))
  3319  		unlock(&sched.lock)
  3320  		startIdle(n)
  3321  		qsize -= n
  3322  	}
  3323  
  3324  	if !q.empty() {
  3325  		runqputbatch(pp, &q, qsize)
  3326  	}
  3327  }
  3328  
  3329  // One round of scheduler: find a runnable goroutine and execute it.
  3330  // Never returns.
  3331  func schedule() {
  3332  	mp := getg().m
  3333  
  3334  	if mp.locks != 0 {
  3335  		throw("schedule: holding locks")
  3336  	}
  3337  
  3338  	if mp.lockedg != 0 {
  3339  		stoplockedm()
  3340  		execute(mp.lockedg.ptr(), false) // Never returns.
  3341  	}
  3342  
  3343  	// We should not schedule away from a g that is executing a cgo call,
  3344  	// since the cgo call is using the m's g0 stack.
  3345  	if mp.incgo {
  3346  		throw("schedule: in cgo")
  3347  	}
  3348  
  3349  top:
  3350  	pp := mp.p.ptr()
  3351  	pp.preempt = false
  3352  
  3353  	// Safety check: if we are spinning, the run queue should be empty.
  3354  	// Check this before calling checkTimers, as that might call
  3355  	// goready to put a ready goroutine on the local run queue.
  3356  	if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
  3357  		throw("schedule: spinning with local work")
  3358  	}
  3359  
  3360  	gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
  3361  
  3362  	// This thread is going to run a goroutine and is not spinning anymore,
  3363  	// so if it was marked as spinning we need to reset it now and potentially
  3364  	// start a new spinning M.
  3365  	if mp.spinning {
  3366  		resetspinning()
  3367  	}
  3368  
  3369  	if sched.disable.user && !schedEnabled(gp) {
  3370  		// Scheduling of this goroutine is disabled. Put it on
  3371  		// the list of pending runnable goroutines for when we
  3372  		// re-enable user scheduling and look again.
  3373  		lock(&sched.lock)
  3374  		if schedEnabled(gp) {
  3375  			// Something re-enabled scheduling while we
  3376  			// were acquiring the lock.
  3377  			unlock(&sched.lock)
  3378  		} else {
  3379  			sched.disable.runnable.pushBack(gp)
  3380  			sched.disable.n++
  3381  			unlock(&sched.lock)
  3382  			goto top
  3383  		}
  3384  	}
  3385  
  3386  	// If about to schedule a not-normal goroutine (a GCworker or tracereader),
  3387  	// wake a P if there is one.
  3388  	if tryWakeP {
  3389  		wakep()
  3390  	}
  3391  	if gp.lockedm != 0 {
  3392  		// Hands off own p to the locked m,
  3393  		// then blocks waiting for a new p.
  3394  		startlockedm(gp)
  3395  		goto top
  3396  	}
  3397  
  3398  	execute(gp, inheritTime)
  3399  }
  3400  
  3401  // dropg removes the association between m and the current goroutine m->curg (gp for short).
  3402  // Typically a caller sets gp's status away from Grunning and then
  3403  // immediately calls dropg to finish the job. The caller is also responsible
  3404  // for arranging that gp will be restarted using ready at an
  3405  // appropriate time. After calling dropg and arranging for gp to be
  3406  // readied later, the caller can do other work but eventually should
  3407  // call schedule to restart the scheduling of goroutines on this m.
  3408  func dropg() {
  3409  	gp := getg()
  3410  
  3411  	setMNoWB(&gp.m.curg.m, nil)
  3412  	setGNoWB(&gp.m.curg, nil)
  3413  }
  3414  
  3415  // checkTimers runs any timers for the P that are ready.
  3416  // If now is not 0 it is the current time.
  3417  // It returns the passed time or the current time if now was passed as 0.
  3418  // and the time when the next timer should run or 0 if there is no next timer,
  3419  // and reports whether it ran any timers.
  3420  // If the time when the next timer should run is not 0,
  3421  // it is always larger than the returned time.
  3422  // We pass now in and out to avoid extra calls of nanotime.
  3423  //
  3424  //go:yeswritebarrierrec
  3425  func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
  3426  	// If it's not yet time for the first timer, or the first adjusted
  3427  	// timer, then there is nothing to do.
  3428  	next := pp.timer0When.Load()
  3429  	nextAdj := pp.timerModifiedEarliest.Load()
  3430  	if next == 0 || (nextAdj != 0 && nextAdj < next) {
  3431  		next = nextAdj
  3432  	}
  3433  
  3434  	if next == 0 {
  3435  		// No timers to run or adjust.
  3436  		return now, 0, false
  3437  	}
  3438  
  3439  	if now == 0 {
  3440  		now = nanotime()
  3441  	}
  3442  	if now < next {
  3443  		// Next timer is not ready to run, but keep going
  3444  		// if we would clear deleted timers.
  3445  		// This corresponds to the condition below where
  3446  		// we decide whether to call clearDeletedTimers.
  3447  		if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
  3448  			return now, next, false
  3449  		}
  3450  	}
  3451  
  3452  	lock(&pp.timersLock)
  3453  
  3454  	if len(pp.timers) > 0 {
  3455  		adjusttimers(pp, now)
  3456  		for len(pp.timers) > 0 {
  3457  			// Note that runtimer may temporarily unlock
  3458  			// pp.timersLock.
  3459  			if tw := runtimer(pp, now); tw != 0 {
  3460  				if tw > 0 {
  3461  					pollUntil = tw
  3462  				}
  3463  				break
  3464  			}
  3465  			ran = true
  3466  		}
  3467  	}
  3468  
  3469  	// If this is the local P, and there are a lot of deleted timers,
  3470  	// clear them out. We only do this for the local P to reduce
  3471  	// lock contention on timersLock.
  3472  	if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
  3473  		clearDeletedTimers(pp)
  3474  	}
  3475  
  3476  	unlock(&pp.timersLock)
  3477  
  3478  	return now, pollUntil, ran
  3479  }
  3480  
  3481  func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
  3482  	unlock((*mutex)(lock))
  3483  	return true
  3484  }
  3485  
  3486  // park continuation on g0.
  3487  func park_m(gp *g) {
  3488  	mp := getg().m
  3489  
  3490  	if trace.enabled {
  3491  		traceGoPark(mp.waittraceev, mp.waittraceskip)
  3492  	}
  3493  
  3494  	// N.B. Not using casGToWaiting here because the waitreason is
  3495  	// set by park_m's caller.
  3496  	casgstatus(gp, _Grunning, _Gwaiting)
  3497  	dropg()
  3498  
  3499  	if fn := mp.waitunlockf; fn != nil {
  3500  		ok := fn(gp, mp.waitlock)
  3501  		mp.waitunlockf = nil
  3502  		mp.waitlock = nil
  3503  		if !ok {
  3504  			if trace.enabled {
  3505  				traceGoUnpark(gp, 2)
  3506  			}
  3507  			casgstatus(gp, _Gwaiting, _Grunnable)
  3508  			execute(gp, true) // Schedule it back, never returns.
  3509  		}
  3510  	}
  3511  	schedule()
  3512  }
  3513  
  3514  func goschedImpl(gp *g) {
  3515  	status := readgstatus(gp)
  3516  	if status&^_Gscan != _Grunning {
  3517  		dumpgstatus(gp)
  3518  		throw("bad g status")
  3519  	}
  3520  	casgstatus(gp, _Grunning, _Grunnable)
  3521  	dropg()
  3522  	lock(&sched.lock)
  3523  	globrunqput(gp)
  3524  	unlock(&sched.lock)
  3525  
  3526  	schedule()
  3527  }
  3528  
  3529  // Gosched continuation on g0.
  3530  func gosched_m(gp *g) {
  3531  	if trace.enabled {
  3532  		traceGoSched()
  3533  	}
  3534  	goschedImpl(gp)
  3535  }
  3536  
  3537  // goschedguarded is a forbidden-states-avoided version of gosched_m.
  3538  func goschedguarded_m(gp *g) {
  3539  
  3540  	if !canPreemptM(gp.m) {
  3541  		gogo(&gp.sched) // never return
  3542  	}
  3543  
  3544  	if trace.enabled {
  3545  		traceGoSched()
  3546  	}
  3547  	goschedImpl(gp)
  3548  }
  3549  
  3550  func gopreempt_m(gp *g) {
  3551  	if trace.enabled {
  3552  		traceGoPreempt()
  3553  	}
  3554  	goschedImpl(gp)
  3555  }
  3556  
  3557  // preemptPark parks gp and puts it in _Gpreempted.
  3558  //
  3559  //go:systemstack
  3560  func preemptPark(gp *g) {
  3561  	if trace.enabled {
  3562  		traceGoPark(traceEvGoBlock, 0)
  3563  	}
  3564  	status := readgstatus(gp)
  3565  	if status&^_Gscan != _Grunning {
  3566  		dumpgstatus(gp)
  3567  		throw("bad g status")
  3568  	}
  3569  
  3570  	if gp.asyncSafePoint {
  3571  		// Double-check that async preemption does not
  3572  		// happen in SPWRITE assembly functions.
  3573  		// isAsyncSafePoint must exclude this case.
  3574  		f := findfunc(gp.sched.pc)
  3575  		if !f.valid() {
  3576  			throw("preempt at unknown pc")
  3577  		}
  3578  		if f.flag&funcFlag_SPWRITE != 0 {
  3579  			println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
  3580  			throw("preempt SPWRITE")
  3581  		}
  3582  	}
  3583  
  3584  	// Transition from _Grunning to _Gscan|_Gpreempted. We can't
  3585  	// be in _Grunning when we dropg because then we'd be running
  3586  	// without an M, but the moment we're in _Gpreempted,
  3587  	// something could claim this G before we've fully cleaned it
  3588  	// up. Hence, we set the scan bit to lock down further
  3589  	// transitions until we can dropg.
  3590  	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
  3591  	dropg()
  3592  	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
  3593  	schedule()
  3594  }
  3595  
  3596  // goyield is like Gosched, but it:
  3597  // - emits a GoPreempt trace event instead of a GoSched trace event
  3598  // - puts the current G on the runq of the current P instead of the globrunq
  3599  func goyield() {
  3600  	checkTimeouts()
  3601  	mcall(goyield_m)
  3602  }
  3603  
  3604  func goyield_m(gp *g) {
  3605  	if trace.enabled {
  3606  		traceGoPreempt()
  3607  	}
  3608  	pp := gp.m.p.ptr()
  3609  	casgstatus(gp, _Grunning, _Grunnable)
  3610  	dropg()
  3611  	runqput(pp, gp, false)
  3612  	schedule()
  3613  }
  3614  
  3615  // Finishes execution of the current goroutine.
  3616  func goexit1() {
  3617  	if raceenabled {
  3618  		racegoend()
  3619  	}
  3620  	if trace.enabled {
  3621  		traceGoEnd()
  3622  	}
  3623  	mcall(goexit0)
  3624  }
  3625  
  3626  // goexit continuation on g0.
  3627  func goexit0(gp *g) {
  3628  	mp := getg().m
  3629  	pp := mp.p.ptr()
  3630  
  3631  	casgstatus(gp, _Grunning, _Gdead)
  3632  	gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
  3633  	if isSystemGoroutine(gp, false) {
  3634  		sched.ngsys.Add(-1)
  3635  	}
  3636  	gp.m = nil
  3637  	locked := gp.lockedm != 0
  3638  	gp.lockedm = 0
  3639  	mp.lockedg = 0
  3640  	gp.preemptStop = false
  3641  	gp.paniconfault = false
  3642  	gp._defer = nil // should be true already but just in case.
  3643  	gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
  3644  	gp.writebuf = nil
  3645  	gp.waitreason = waitReasonZero
  3646  	gp.param = nil
  3647  	gp.labels = nil
  3648  	gp.timer = nil
  3649  
  3650  	if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
  3651  		// Flush assist credit to the global pool. This gives
  3652  		// better information to pacing if the application is
  3653  		// rapidly creating an exiting goroutines.
  3654  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
  3655  		scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
  3656  		gcController.bgScanCredit.Add(scanCredit)
  3657  		gp.gcAssistBytes = 0
  3658  	}
  3659  
  3660  	dropg()
  3661  
  3662  	if GOARCH == "wasm" { // no threads yet on wasm
  3663  		gfput(pp, gp)
  3664  		schedule() // never returns
  3665  	}
  3666  
  3667  	if mp.lockedInt != 0 {
  3668  		print("invalid m->lockedInt = ", mp.lockedInt, "\n")
  3669  		throw("internal lockOSThread error")
  3670  	}
  3671  	gfput(pp, gp)
  3672  	if locked {
  3673  		// The goroutine may have locked this thread because
  3674  		// it put it in an unusual kernel state. Kill it
  3675  		// rather than returning it to the thread pool.
  3676  
  3677  		// Return to mstart, which will release the P and exit
  3678  		// the thread.
  3679  		if GOOS != "plan9" { // See golang.org/issue/22227.
  3680  			gogo(&mp.g0.sched)
  3681  		} else {
  3682  			// Clear lockedExt on plan9 since we may end up re-using
  3683  			// this thread.
  3684  			mp.lockedExt = 0
  3685  		}
  3686  	}
  3687  	schedule()
  3688  }
  3689  
  3690  // save updates getg().sched to refer to pc and sp so that a following
  3691  // gogo will restore pc and sp.
  3692  //
  3693  // save must not have write barriers because invoking a write barrier
  3694  // can clobber getg().sched.
  3695  //
  3696  //go:nosplit
  3697  //go:nowritebarrierrec
  3698  func save(pc, sp uintptr) {
  3699  	gp := getg()
  3700  
  3701  	if gp == gp.m.g0 || gp == gp.m.gsignal {
  3702  		// m.g0.sched is special and must describe the context
  3703  		// for exiting the thread. mstart1 writes to it directly.
  3704  		// m.gsignal.sched should not be used at all.
  3705  		// This check makes sure save calls do not accidentally
  3706  		// run in contexts where they'd write to system g's.
  3707  		throw("save on system g not allowed")
  3708  	}
  3709  
  3710  	gp.sched.pc = pc
  3711  	gp.sched.sp = sp
  3712  	gp.sched.lr = 0
  3713  	gp.sched.ret = 0
  3714  	// We need to ensure ctxt is zero, but can't have a write
  3715  	// barrier here. However, it should always already be zero.
  3716  	// Assert that.
  3717  	if gp.sched.ctxt != nil {
  3718  		badctxt()
  3719  	}
  3720  }
  3721  
  3722  // The goroutine g is about to enter a system call.
  3723  // Record that it's not using the cpu anymore.
  3724  // This is called only from the go syscall library and cgocall,
  3725  // not from the low-level system calls used by the runtime.
  3726  //
  3727  // Entersyscall cannot split the stack: the save must
  3728  // make g->sched refer to the caller's stack segment, because
  3729  // entersyscall is going to return immediately after.
  3730  //
  3731  // Nothing entersyscall calls can split the stack either.
  3732  // We cannot safely move the stack during an active call to syscall,
  3733  // because we do not know which of the uintptr arguments are
  3734  // really pointers (back into the stack).
  3735  // In practice, this means that we make the fast path run through
  3736  // entersyscall doing no-split things, and the slow path has to use systemstack
  3737  // to run bigger things on the system stack.
  3738  //
  3739  // reentersyscall is the entry point used by cgo callbacks, where explicitly
  3740  // saved SP and PC are restored. This is needed when exitsyscall will be called
  3741  // from a function further up in the call stack than the parent, as g->syscallsp
  3742  // must always point to a valid stack frame. entersyscall below is the normal
  3743  // entry point for syscalls, which obtains the SP and PC from the caller.
  3744  //
  3745  // Syscall tracing:
  3746  // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
  3747  // If the syscall does not block, that is it, we do not emit any other events.
  3748  // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
  3749  // when syscall returns we emit traceGoSysExit and when the goroutine starts running
  3750  // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
  3751  // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
  3752  // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
  3753  // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
  3754  // and we wait for the increment before emitting traceGoSysExit.
  3755  // Note that the increment is done even if tracing is not enabled,
  3756  // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
  3757  //
  3758  //go:nosplit
  3759  func reentersyscall(pc, sp uintptr) {
  3760  	gp := getg()
  3761  
  3762  	// Disable preemption because during this function g is in Gsyscall status,
  3763  	// but can have inconsistent g->sched, do not let GC observe it.
  3764  	gp.m.locks++
  3765  
  3766  	// Entersyscall must not call any function that might split/grow the stack.
  3767  	// (See details in comment above.)
  3768  	// Catch calls that might, by replacing the stack guard with something that
  3769  	// will trip any stack check and leaving a flag to tell newstack to die.
  3770  	gp.stackguard0 = stackPreempt
  3771  	gp.throwsplit = true
  3772  
  3773  	// Leave SP around for GC and traceback.
  3774  	save(pc, sp)
  3775  	gp.syscallsp = sp
  3776  	gp.syscallpc = pc
  3777  	casgstatus(gp, _Grunning, _Gsyscall)
  3778  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  3779  		systemstack(func() {
  3780  			print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  3781  			throw("entersyscall")
  3782  		})
  3783  	}
  3784  
  3785  	if trace.enabled {
  3786  		systemstack(traceGoSysCall)
  3787  		// systemstack itself clobbers g.sched.{pc,sp} and we might
  3788  		// need them later when the G is genuinely blocked in a
  3789  		// syscall
  3790  		save(pc, sp)
  3791  	}
  3792  
  3793  	if sched.sysmonwait.Load() {
  3794  		systemstack(entersyscall_sysmon)
  3795  		save(pc, sp)
  3796  	}
  3797  
  3798  	if gp.m.p.ptr().runSafePointFn != 0 {
  3799  		// runSafePointFn may stack split if run on this stack
  3800  		systemstack(runSafePointFn)
  3801  		save(pc, sp)
  3802  	}
  3803  
  3804  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  3805  	gp.sysblocktraced = true
  3806  	pp := gp.m.p.ptr()
  3807  	pp.m = 0
  3808  	gp.m.oldp.set(pp)
  3809  	gp.m.p = 0
  3810  	atomic.Store(&pp.status, _Psyscall)
  3811  	if sched.gcwaiting.Load() {
  3812  		systemstack(entersyscall_gcwait)
  3813  		save(pc, sp)
  3814  	}
  3815  
  3816  	gp.m.locks--
  3817  }
  3818  
  3819  // Standard syscall entry used by the go syscall library and normal cgo calls.
  3820  //
  3821  // This is exported via linkname to assembly in the syscall package and x/sys.
  3822  //
  3823  //go:nosplit
  3824  //go:linkname entersyscall
  3825  func entersyscall() {
  3826  	reentersyscall(getcallerpc(), getcallersp())
  3827  }
  3828  
  3829  func entersyscall_sysmon() {
  3830  	lock(&sched.lock)
  3831  	if sched.sysmonwait.Load() {
  3832  		sched.sysmonwait.Store(false)
  3833  		notewakeup(&sched.sysmonnote)
  3834  	}
  3835  	unlock(&sched.lock)
  3836  }
  3837  
  3838  func entersyscall_gcwait() {
  3839  	gp := getg()
  3840  	pp := gp.m.oldp.ptr()
  3841  
  3842  	lock(&sched.lock)
  3843  	if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
  3844  		if trace.enabled {
  3845  			traceGoSysBlock(pp)
  3846  			traceProcStop(pp)
  3847  		}
  3848  		pp.syscalltick++
  3849  		if sched.stopwait--; sched.stopwait == 0 {
  3850  			notewakeup(&sched.stopnote)
  3851  		}
  3852  	}
  3853  	unlock(&sched.lock)
  3854  }
  3855  
  3856  // The same as entersyscall(), but with a hint that the syscall is blocking.
  3857  //
  3858  //go:nosplit
  3859  func entersyscallblock() {
  3860  	gp := getg()
  3861  
  3862  	gp.m.locks++ // see comment in entersyscall
  3863  	gp.throwsplit = true
  3864  	gp.stackguard0 = stackPreempt // see comment in entersyscall
  3865  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  3866  	gp.sysblocktraced = true
  3867  	gp.m.p.ptr().syscalltick++
  3868  
  3869  	// Leave SP around for GC and traceback.
  3870  	pc := getcallerpc()
  3871  	sp := getcallersp()
  3872  	save(pc, sp)
  3873  	gp.syscallsp = gp.sched.sp
  3874  	gp.syscallpc = gp.sched.pc
  3875  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  3876  		sp1 := sp
  3877  		sp2 := gp.sched.sp
  3878  		sp3 := gp.syscallsp
  3879  		systemstack(func() {
  3880  			print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  3881  			throw("entersyscallblock")
  3882  		})
  3883  	}
  3884  	casgstatus(gp, _Grunning, _Gsyscall)
  3885  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  3886  		systemstack(func() {
  3887  			print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  3888  			throw("entersyscallblock")
  3889  		})
  3890  	}
  3891  
  3892  	systemstack(entersyscallblock_handoff)
  3893  
  3894  	// Resave for traceback during blocked call.
  3895  	save(getcallerpc(), getcallersp())
  3896  
  3897  	gp.m.locks--
  3898  }
  3899  
  3900  func entersyscallblock_handoff() {
  3901  	if trace.enabled {
  3902  		traceGoSysCall()
  3903  		traceGoSysBlock(getg().m.p.ptr())
  3904  	}
  3905  	handoffp(releasep())
  3906  }
  3907  
  3908  // The goroutine g exited its system call.
  3909  // Arrange for it to run on a cpu again.
  3910  // This is called only from the go syscall library, not
  3911  // from the low-level system calls used by the runtime.
  3912  //
  3913  // Write barriers are not allowed because our P may have been stolen.
  3914  //
  3915  // This is exported via linkname to assembly in the syscall package.
  3916  //
  3917  //go:nosplit
  3918  //go:nowritebarrierrec
  3919  //go:linkname exitsyscall
  3920  func exitsyscall() {
  3921  	gp := getg()
  3922  
  3923  	gp.m.locks++ // see comment in entersyscall
  3924  	if getcallersp() > gp.syscallsp {
  3925  		throw("exitsyscall: syscall frame is no longer valid")
  3926  	}
  3927  
  3928  	gp.waitsince = 0
  3929  	oldp := gp.m.oldp.ptr()
  3930  	gp.m.oldp = 0
  3931  	if exitsyscallfast(oldp) {
  3932  		// When exitsyscallfast returns success, we have a P so can now use
  3933  		// write barriers
  3934  		if goroutineProfile.active {
  3935  			// Make sure that gp has had its stack written out to the goroutine
  3936  			// profile, exactly as it was when the goroutine profiler first
  3937  			// stopped the world.
  3938  			systemstack(func() {
  3939  				tryRecordGoroutineProfileWB(gp)
  3940  			})
  3941  		}
  3942  		if trace.enabled {
  3943  			if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  3944  				systemstack(traceGoStart)
  3945  			}
  3946  		}
  3947  		// There's a cpu for us, so we can run.
  3948  		gp.m.p.ptr().syscalltick++
  3949  		// We need to cas the status and scan before resuming...
  3950  		casgstatus(gp, _Gsyscall, _Grunning)
  3951  
  3952  		// Garbage collector isn't running (since we are),
  3953  		// so okay to clear syscallsp.
  3954  		gp.syscallsp = 0
  3955  		gp.m.locks--
  3956  		if gp.preempt {
  3957  			// restore the preemption request in case we've cleared it in newstack
  3958  			gp.stackguard0 = stackPreempt
  3959  		} else {
  3960  			// otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
  3961  			gp.stackguard0 = gp.stack.lo + _StackGuard
  3962  		}
  3963  		gp.throwsplit = false
  3964  
  3965  		if sched.disable.user && !schedEnabled(gp) {
  3966  			// Scheduling of this goroutine is disabled.
  3967  			Gosched()
  3968  		}
  3969  
  3970  		return
  3971  	}
  3972  
  3973  	gp.sysexitticks = 0
  3974  	if 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 != nil && oldp.syscalltick == gp.m.syscalltick {
  3978  			osyield()
  3979  		}
  3980  		// We can't trace syscall exit right now because we don't have a P.
  3981  		// Tracing code can invoke write barriers that cannot run without a P.
  3982  		// So instead we remember the syscall exit time and emit the event
  3983  		// in execute when we have a P.
  3984  		gp.sysexitticks = cputicks()
  3985  	}
  3986  
  3987  	gp.m.locks--
  3988  
  3989  	// Call the scheduler.
  3990  	mcall(exitsyscall0)
  3991  
  3992  	// Scheduler returned, so we're allowed to run now.
  3993  	// Delete the syscallsp information that we left for
  3994  	// the garbage collector during the system call.
  3995  	// Must wait until now because until gosched returns
  3996  	// we don't know for sure that the garbage collector
  3997  	// is not running.
  3998  	gp.syscallsp = 0
  3999  	gp.m.p.ptr().syscalltick++
  4000  	gp.throwsplit = false
  4001  }
  4002  
  4003  //go:nosplit
  4004  func exitsyscallfast(oldp *p) bool {
  4005  	gp := getg()
  4006  
  4007  	// Freezetheworld sets stopwait but does not retake P's.
  4008  	if sched.stopwait == freezeStopWait {
  4009  		return false
  4010  	}
  4011  
  4012  	// Try to re-acquire the last P.
  4013  	if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
  4014  		// There's a cpu for us, so we can run.
  4015  		wirep(oldp)
  4016  		exitsyscallfast_reacquired()
  4017  		return true
  4018  	}
  4019  
  4020  	// Try to get any other idle P.
  4021  	if sched.pidle != 0 {
  4022  		var ok bool
  4023  		systemstack(func() {
  4024  			ok = exitsyscallfast_pidle()
  4025  			if ok && trace.enabled {
  4026  				if oldp != nil {
  4027  					// Wait till traceGoSysBlock event is emitted.
  4028  					// This ensures consistency of the trace (the goroutine is started after it is blocked).
  4029  					for oldp.syscalltick == gp.m.syscalltick {
  4030  						osyield()
  4031  					}
  4032  				}
  4033  				traceGoSysExit(0)
  4034  			}
  4035  		})
  4036  		if ok {
  4037  			return true
  4038  		}
  4039  	}
  4040  	return false
  4041  }
  4042  
  4043  // exitsyscallfast_reacquired is the exitsyscall path on which this G
  4044  // has successfully reacquired the P it was running on before the
  4045  // syscall.
  4046  //
  4047  //go:nosplit
  4048  func exitsyscallfast_reacquired() {
  4049  	gp := getg()
  4050  	if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  4051  		if trace.enabled {
  4052  			// The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
  4053  			// traceGoSysBlock for this syscall was already emitted,
  4054  			// but here we effectively retake the p from the new syscall running on the same p.
  4055  			systemstack(func() {
  4056  				// Denote blocking of the new syscall.
  4057  				traceGoSysBlock(gp.m.p.ptr())
  4058  				// Denote completion of the current syscall.
  4059  				traceGoSysExit(0)
  4060  			})
  4061  		}
  4062  		gp.m.p.ptr().syscalltick++
  4063  	}
  4064  }
  4065  
  4066  func exitsyscallfast_pidle() bool {
  4067  	lock(&sched.lock)
  4068  	pp, _ := pidleget(0)
  4069  	if pp != nil && sched.sysmonwait.Load() {
  4070  		sched.sysmonwait.Store(false)
  4071  		notewakeup(&sched.sysmonnote)
  4072  	}
  4073  	unlock(&sched.lock)
  4074  	if pp != nil {
  4075  		acquirep(pp)
  4076  		return true
  4077  	}
  4078  	return false
  4079  }
  4080  
  4081  // exitsyscall slow path on g0.
  4082  // Failed to acquire P, enqueue gp as runnable.
  4083  //
  4084  // Called via mcall, so gp is the calling g from this M.
  4085  //
  4086  //go:nowritebarrierrec
  4087  func exitsyscall0(gp *g) {
  4088  	casgstatus(gp, _Gsyscall, _Grunnable)
  4089  	dropg()
  4090  	lock(&sched.lock)
  4091  	var pp *p
  4092  	if schedEnabled(gp) {
  4093  		pp, _ = pidleget(0)
  4094  	}
  4095  	var locked bool
  4096  	if pp == nil {
  4097  		globrunqput(gp)
  4098  
  4099  		// Below, we stoplockedm if gp is locked. globrunqput releases
  4100  		// ownership of gp, so we must check if gp is locked prior to
  4101  		// committing the release by unlocking sched.lock, otherwise we
  4102  		// could race with another M transitioning gp from unlocked to
  4103  		// locked.
  4104  		locked = gp.lockedm != 0
  4105  	} else if sched.sysmonwait.Load() {
  4106  		sched.sysmonwait.Store(false)
  4107  		notewakeup(&sched.sysmonnote)
  4108  	}
  4109  	unlock(&sched.lock)
  4110  	if pp != nil {
  4111  		acquirep(pp)
  4112  		execute(gp, false) // Never returns.
  4113  	}
  4114  	if locked {
  4115  		// Wait until another thread schedules gp and so m again.
  4116  		//
  4117  		// N.B. lockedm must be this M, as this g was running on this M
  4118  		// before entersyscall.
  4119  		stoplockedm()
  4120  		execute(gp, false) // Never returns.
  4121  	}
  4122  	stopm()
  4123  	schedule() // Never returns.
  4124  }
  4125  
  4126  // Called from syscall package before fork.
  4127  //
  4128  //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
  4129  //go:nosplit
  4130  func syscall_runtime_BeforeFork() {
  4131  	gp := getg().m.curg
  4132  
  4133  	// Block signals during a fork, so that the child does not run
  4134  	// a signal handler before exec if a signal is sent to the process
  4135  	// group. See issue #18600.
  4136  	gp.m.locks++
  4137  	sigsave(&gp.m.sigmask)
  4138  	sigblock(false)
  4139  
  4140  	// This function is called before fork in syscall package.
  4141  	// Code between fork and exec must not allocate memory nor even try to grow stack.
  4142  	// Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
  4143  	// runtime_AfterFork will undo this in parent process, but not in child.
  4144  	gp.stackguard0 = stackFork
  4145  }
  4146  
  4147  // Called from syscall package after fork in parent.
  4148  //
  4149  //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
  4150  //go:nosplit
  4151  func syscall_runtime_AfterFork() {
  4152  	gp := getg().m.curg
  4153  
  4154  	// See the comments in beforefork.
  4155  	gp.stackguard0 = gp.stack.lo + _StackGuard
  4156  
  4157  	msigrestore(gp.m.sigmask)
  4158  
  4159  	gp.m.locks--
  4160  }
  4161  
  4162  // inForkedChild is true while manipulating signals in the child process.
  4163  // This is used to avoid calling libc functions in case we are using vfork.
  4164  var inForkedChild bool
  4165  
  4166  // Called from syscall package after fork in child.
  4167  // It resets non-sigignored signals to the default handler, and
  4168  // restores the signal mask in preparation for the exec.
  4169  //
  4170  // Because this might be called during a vfork, and therefore may be
  4171  // temporarily sharing address space with the parent process, this must
  4172  // not change any global variables or calling into C code that may do so.
  4173  //
  4174  //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
  4175  //go:nosplit
  4176  //go:nowritebarrierrec
  4177  func syscall_runtime_AfterForkInChild() {
  4178  	// It's OK to change the global variable inForkedChild here
  4179  	// because we are going to change it back. There is no race here,
  4180  	// because if we are sharing address space with the parent process,
  4181  	// then the parent process can not be running concurrently.
  4182  	inForkedChild = true
  4183  
  4184  	clearSignalHandlers()
  4185  
  4186  	// When we are the child we are the only thread running,
  4187  	// so we know that nothing else has changed gp.m.sigmask.
  4188  	msigrestore(getg().m.sigmask)
  4189  
  4190  	inForkedChild = false
  4191  }
  4192  
  4193  // pendingPreemptSignals is the number of preemption signals
  4194  // that have been sent but not received. This is only used on Darwin.
  4195  // For #41702.
  4196  var pendingPreemptSignals atomic.Int32
  4197  
  4198  // Called from syscall package before Exec.
  4199  //
  4200  //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
  4201  func syscall_runtime_BeforeExec() {
  4202  	// Prevent thread creation during exec.
  4203  	execLock.lock()
  4204  
  4205  	// On Darwin, wait for all pending preemption signals to
  4206  	// be received. See issue #41702.
  4207  	if GOOS == "darwin" || GOOS == "ios" {
  4208  		for pendingPreemptSignals.Load() > 0 {
  4209  			osyield()
  4210  		}
  4211  	}
  4212  }
  4213  
  4214  // Called from syscall package after Exec.
  4215  //
  4216  //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
  4217  func syscall_runtime_AfterExec() {
  4218  	execLock.unlock()
  4219  }
  4220  
  4221  // Allocate a new g, with a stack big enough for stacksize bytes.
  4222  func malg(stacksize int32) *g {
  4223  	newg := new(g)
  4224  	if stacksize >= 0 {
  4225  		stacksize = round2(_StackSystem + stacksize)
  4226  		systemstack(func() {
  4227  			newg.stack = stackalloc(uint32(stacksize))
  4228  		})
  4229  		newg.stackguard0 = newg.stack.lo + _StackGuard
  4230  		newg.stackguard1 = ^uintptr(0)
  4231  		// Clear the bottom word of the stack. We record g
  4232  		// there on gsignal stack during VDSO on ARM and ARM64.
  4233  		*(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
  4234  	}
  4235  	return newg
  4236  }
  4237  
  4238  // Create a new g running fn.
  4239  // Put it on the queue of g's waiting to run.
  4240  // The compiler turns a go statement into a call to this.
  4241  func newproc(fn *funcval) {
  4242  	gp := getg()
  4243  	pc := getcallerpc()
  4244  	systemstack(func() {
  4245  		newg := newproc1(fn, gp, pc)
  4246  
  4247  		pp := getg().m.p.ptr()
  4248  		runqput(pp, newg, true)
  4249  
  4250  		if mainStarted {
  4251  			wakep()
  4252  		}
  4253  	})
  4254  }
  4255  
  4256  // Create a new g in state _Grunnable, starting at fn. callerpc is the
  4257  // address of the go statement that created this. The caller is responsible
  4258  // for adding the new g to the scheduler.
  4259  func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
  4260  	if fn == nil {
  4261  		fatal("go of nil func value")
  4262  	}
  4263  
  4264  	mp := acquirem() // disable preemption because we hold M and P in local vars.
  4265  	pp := mp.p.ptr()
  4266  	newg := gfget(pp)
  4267  	if newg == nil {
  4268  		newg = malg(_StackMin)
  4269  		casgstatus(newg, _Gidle, _Gdead)
  4270  		allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
  4271  	}
  4272  	if newg.stack.hi == 0 {
  4273  		throw("newproc1: newg missing stack")
  4274  	}
  4275  
  4276  	if readgstatus(newg) != _Gdead {
  4277  		throw("newproc1: new g is not Gdead")
  4278  	}
  4279  
  4280  	totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
  4281  	totalSize = alignUp(totalSize, sys.StackAlign)
  4282  	sp := newg.stack.hi - totalSize
  4283  	spArg := sp
  4284  	if usesLR {
  4285  		// caller's LR
  4286  		*(*uintptr)(unsafe.Pointer(sp)) = 0
  4287  		prepGoExitFrame(sp)
  4288  		spArg += sys.MinFrameSize
  4289  	}
  4290  
  4291  	memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
  4292  	newg.sched.sp = sp
  4293  	newg.stktopsp = sp
  4294  	newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
  4295  	newg.sched.g = guintptr(unsafe.Pointer(newg))
  4296  	gostartcallfn(&newg.sched, fn)
  4297  	newg.gopc = callerpc
  4298  	newg.ancestors = saveAncestors(callergp)
  4299  	newg.startpc = fn.fn
  4300  	if isSystemGoroutine(newg, false) {
  4301  		sched.ngsys.Add(1)
  4302  	} else {
  4303  		// Only user goroutines inherit pprof labels.
  4304  		if mp.curg != nil {
  4305  			newg.labels = mp.curg.labels
  4306  		}
  4307  		if goroutineProfile.active {
  4308  			// A concurrent goroutine profile is running. It should include
  4309  			// exactly the set of goroutines that were alive when the goroutine
  4310  			// profiler first stopped the world. That does not include newg, so
  4311  			// mark it as not needing a profile before transitioning it from
  4312  			// _Gdead.
  4313  			newg.goroutineProfiled.Store(goroutineProfileSatisfied)
  4314  		}
  4315  	}
  4316  	// Track initial transition?
  4317  	newg.trackingSeq = uint8(fastrand())
  4318  	if newg.trackingSeq%gTrackingPeriod == 0 {
  4319  		newg.tracking = true
  4320  	}
  4321  	casgstatus(newg, _Gdead, _Grunnable)
  4322  	gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
  4323  
  4324  	if pp.goidcache == pp.goidcacheend {
  4325  		// Sched.goidgen is the last allocated id,
  4326  		// this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
  4327  		// At startup sched.goidgen=0, so main goroutine receives goid=1.
  4328  		pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
  4329  		pp.goidcache -= _GoidCacheBatch - 1
  4330  		pp.goidcacheend = pp.goidcache + _GoidCacheBatch
  4331  	}
  4332  	newg.goid = pp.goidcache
  4333  	pp.goidcache++
  4334  	if raceenabled {
  4335  		newg.racectx = racegostart(callerpc)
  4336  		if newg.labels != nil {
  4337  			// See note in proflabel.go on labelSync's role in synchronizing
  4338  			// with the reads in the signal handler.
  4339  			racereleasemergeg(newg, unsafe.Pointer(&labelSync))
  4340  		}
  4341  	}
  4342  	if trace.enabled {
  4343  		traceGoCreate(newg, newg.startpc)
  4344  	}
  4345  	releasem(mp)
  4346  
  4347  	return newg
  4348  }
  4349  
  4350  // saveAncestors copies previous ancestors of the given caller g and
  4351  // includes infor for the current caller into a new set of tracebacks for
  4352  // a g being created.
  4353  func saveAncestors(callergp *g) *[]ancestorInfo {
  4354  	// Copy all prior info, except for the root goroutine (goid 0).
  4355  	if debug.tracebackancestors <= 0 || callergp.goid == 0 {
  4356  		return nil
  4357  	}
  4358  	var callerAncestors []ancestorInfo
  4359  	if callergp.ancestors != nil {
  4360  		callerAncestors = *callergp.ancestors
  4361  	}
  4362  	n := int32(len(callerAncestors)) + 1
  4363  	if n > debug.tracebackancestors {
  4364  		n = debug.tracebackancestors
  4365  	}
  4366  	ancestors := make([]ancestorInfo, n)
  4367  	copy(ancestors[1:], callerAncestors)
  4368  
  4369  	var pcs [_TracebackMaxFrames]uintptr
  4370  	npcs := gcallers(callergp, 0, pcs[:])
  4371  	ipcs := make([]uintptr, npcs)
  4372  	copy(ipcs, pcs[:])
  4373  	ancestors[0] = ancestorInfo{
  4374  		pcs:  ipcs,
  4375  		goid: callergp.goid,
  4376  		gopc: callergp.gopc,
  4377  	}
  4378  
  4379  	ancestorsp := new([]ancestorInfo)
  4380  	*ancestorsp = ancestors
  4381  	return ancestorsp
  4382  }
  4383  
  4384  // Put on gfree list.
  4385  // If local list is too long, transfer a batch to the global list.
  4386  func gfput(pp *p, gp *g) {
  4387  	if readgstatus(gp) != _Gdead {
  4388  		throw("gfput: bad status (not Gdead)")
  4389  	}
  4390  
  4391  	stksize := gp.stack.hi - gp.stack.lo
  4392  
  4393  	if stksize != uintptr(startingStackSize) {
  4394  		// non-standard stack size - free it.
  4395  		stackfree(gp.stack)
  4396  		gp.stack.lo = 0
  4397  		gp.stack.hi = 0
  4398  		gp.stackguard0 = 0
  4399  	}
  4400  
  4401  	pp.gFree.push(gp)
  4402  	pp.gFree.n++
  4403  	if pp.gFree.n >= 64 {
  4404  		var (
  4405  			inc      int32
  4406  			stackQ   gQueue
  4407  			noStackQ gQueue
  4408  		)
  4409  		for pp.gFree.n >= 32 {
  4410  			gp := pp.gFree.pop()
  4411  			pp.gFree.n--
  4412  			if gp.stack.lo == 0 {
  4413  				noStackQ.push(gp)
  4414  			} else {
  4415  				stackQ.push(gp)
  4416  			}
  4417  			inc++
  4418  		}
  4419  		lock(&sched.gFree.lock)
  4420  		sched.gFree.noStack.pushAll(noStackQ)
  4421  		sched.gFree.stack.pushAll(stackQ)
  4422  		sched.gFree.n += inc
  4423  		unlock(&sched.gFree.lock)
  4424  	}
  4425  }
  4426  
  4427  // Get from gfree list.
  4428  // If local list is empty, grab a batch from global list.
  4429  func gfget(pp *p) *g {
  4430  retry:
  4431  	if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
  4432  		lock(&sched.gFree.lock)
  4433  		// Move a batch of free Gs to the P.
  4434  		for pp.gFree.n < 32 {
  4435  			// Prefer Gs with stacks.
  4436  			gp := sched.gFree.stack.pop()
  4437  			if gp == nil {
  4438  				gp = sched.gFree.noStack.pop()
  4439  				if gp == nil {
  4440  					break
  4441  				}
  4442  			}
  4443  			sched.gFree.n--
  4444  			pp.gFree.push(gp)
  4445  			pp.gFree.n++
  4446  		}
  4447  		unlock(&sched.gFree.lock)
  4448  		goto retry
  4449  	}
  4450  	gp := pp.gFree.pop()
  4451  	if gp == nil {
  4452  		return nil
  4453  	}
  4454  	pp.gFree.n--
  4455  	if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
  4456  		// Deallocate old stack. We kept it in gfput because it was the
  4457  		// right size when the goroutine was put on the free list, but
  4458  		// the right size has changed since then.
  4459  		systemstack(func() {
  4460  			stackfree(gp.stack)
  4461  			gp.stack.lo = 0
  4462  			gp.stack.hi = 0
  4463  			gp.stackguard0 = 0
  4464  		})
  4465  	}
  4466  	if gp.stack.lo == 0 {
  4467  		// Stack was deallocated in gfput or just above. Allocate a new one.
  4468  		systemstack(func() {
  4469  			gp.stack = stackalloc(startingStackSize)
  4470  		})
  4471  		gp.stackguard0 = gp.stack.lo + _StackGuard
  4472  	} else {
  4473  		if raceenabled {
  4474  			racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4475  		}
  4476  		if msanenabled {
  4477  			msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4478  		}
  4479  		if asanenabled {
  4480  			asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4481  		}
  4482  	}
  4483  	return gp
  4484  }
  4485  
  4486  // Purge all cached G's from gfree list to the global list.
  4487  func gfpurge(pp *p) {
  4488  	var (
  4489  		inc      int32
  4490  		stackQ   gQueue
  4491  		noStackQ gQueue
  4492  	)
  4493  	for !pp.gFree.empty() {
  4494  		gp := pp.gFree.pop()
  4495  		pp.gFree.n--
  4496  		if gp.stack.lo == 0 {
  4497  			noStackQ.push(gp)
  4498  		} else {
  4499  			stackQ.push(gp)
  4500  		}
  4501  		inc++
  4502  	}
  4503  	lock(&sched.gFree.lock)
  4504  	sched.gFree.noStack.pushAll(noStackQ)
  4505  	sched.gFree.stack.pushAll(stackQ)
  4506  	sched.gFree.n += inc
  4507  	unlock(&sched.gFree.lock)
  4508  }
  4509  
  4510  // Breakpoint executes a breakpoint trap.
  4511  func Breakpoint() {
  4512  	breakpoint()
  4513  }
  4514  
  4515  // dolockOSThread is called by LockOSThread and lockOSThread below
  4516  // after they modify m.locked. Do not allow preemption during this call,
  4517  // or else the m might be different in this function than in the caller.
  4518  //
  4519  //go:nosplit
  4520  func dolockOSThread() {
  4521  	if GOARCH == "wasm" {
  4522  		return // no threads on wasm yet
  4523  	}
  4524  	gp := getg()
  4525  	gp.m.lockedg.set(gp)
  4526  	gp.lockedm.set(gp.m)
  4527  }
  4528  
  4529  //go:nosplit
  4530  
  4531  // LockOSThread wires the calling goroutine to its current operating system thread.
  4532  // The calling goroutine will always execute in that thread,
  4533  // and no other goroutine will execute in it,
  4534  // until the calling goroutine has made as many calls to
  4535  // UnlockOSThread as to LockOSThread.
  4536  // If the calling goroutine exits without unlocking the thread,
  4537  // the thread will be terminated.
  4538  //
  4539  // All init functions are run on the startup thread. Calling LockOSThread
  4540  // from an init function will cause the main function to be invoked on
  4541  // that thread.
  4542  //
  4543  // A goroutine should call LockOSThread before calling OS services or
  4544  // non-Go library functions that depend on per-thread state.
  4545  func LockOSThread() {
  4546  	if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
  4547  		// If we need to start a new thread from the locked
  4548  		// thread, we need the template thread. Start it now
  4549  		// while we're in a known-good state.
  4550  		startTemplateThread()
  4551  	}
  4552  	gp := getg()
  4553  	gp.m.lockedExt++
  4554  	if gp.m.lockedExt == 0 {
  4555  		gp.m.lockedExt--
  4556  		panic("LockOSThread nesting overflow")
  4557  	}
  4558  	dolockOSThread()
  4559  }
  4560  
  4561  //go:nosplit
  4562  func lockOSThread() {
  4563  	getg().m.lockedInt++
  4564  	dolockOSThread()
  4565  }
  4566  
  4567  // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
  4568  // after they update m->locked. Do not allow preemption during this call,
  4569  // or else the m might be in different in this function than in the caller.
  4570  //
  4571  //go:nosplit
  4572  func dounlockOSThread() {
  4573  	if GOARCH == "wasm" {
  4574  		return // no threads on wasm yet
  4575  	}
  4576  	gp := getg()
  4577  	if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
  4578  		return
  4579  	}
  4580  	gp.m.lockedg = 0
  4581  	gp.lockedm = 0
  4582  }
  4583  
  4584  //go:nosplit
  4585  
  4586  // UnlockOSThread undoes an earlier call to LockOSThread.
  4587  // If this drops the number of active LockOSThread calls on the
  4588  // calling goroutine to zero, it unwires the calling goroutine from
  4589  // its fixed operating system thread.
  4590  // If there are no active LockOSThread calls, this is a no-op.
  4591  //
  4592  // Before calling UnlockOSThread, the caller must ensure that the OS
  4593  // thread is suitable for running other goroutines. If the caller made
  4594  // any permanent changes to the state of the thread that would affect
  4595  // other goroutines, it should not call this function and thus leave
  4596  // the goroutine locked to the OS thread until the goroutine (and
  4597  // hence the thread) exits.
  4598  func UnlockOSThread() {
  4599  	gp := getg()
  4600  	if gp.m.lockedExt == 0 {
  4601  		return
  4602  	}
  4603  	gp.m.lockedExt--
  4604  	dounlockOSThread()
  4605  }
  4606  
  4607  //go:nosplit
  4608  func unlockOSThread() {
  4609  	gp := getg()
  4610  	if gp.m.lockedInt == 0 {
  4611  		systemstack(badunlockosthread)
  4612  	}
  4613  	gp.m.lockedInt--
  4614  	dounlockOSThread()
  4615  }
  4616  
  4617  func badunlockosthread() {
  4618  	throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
  4619  }
  4620  
  4621  func gcount() int32 {
  4622  	n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
  4623  	for _, pp := range allp {
  4624  		n -= pp.gFree.n
  4625  	}
  4626  
  4627  	// All these variables can be changed concurrently, so the result can be inconsistent.
  4628  	// But at least the current goroutine is running.
  4629  	if n < 1 {
  4630  		n = 1
  4631  	}
  4632  	return n
  4633  }
  4634  
  4635  func mcount() int32 {
  4636  	return int32(sched.mnext - sched.nmfreed)
  4637  }
  4638  
  4639  var prof struct {
  4640  	signalLock atomic.Uint32
  4641  
  4642  	// Must hold signalLock to write. Reads may be lock-free, but
  4643  	// signalLock should be taken to synchronize with changes.
  4644  	hz atomic.Int32
  4645  }
  4646  
  4647  func _System()                    { _System() }
  4648  func _ExternalCode()              { _ExternalCode() }
  4649  func _LostExternalCode()          { _LostExternalCode() }
  4650  func _GC()                        { _GC() }
  4651  func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
  4652  func _VDSO()                      { _VDSO() }
  4653  
  4654  // Called if we receive a SIGPROF signal.
  4655  // Called by the signal handler, may run during STW.
  4656  //
  4657  //go:nowritebarrierrec
  4658  func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
  4659  	if prof.hz.Load() == 0 {
  4660  		return
  4661  	}
  4662  
  4663  	// If mp.profilehz is 0, then profiling is not enabled for this thread.
  4664  	// We must check this to avoid a deadlock between setcpuprofilerate
  4665  	// and the call to cpuprof.add, below.
  4666  	if mp != nil && mp.profilehz == 0 {
  4667  		return
  4668  	}
  4669  
  4670  	// On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
  4671  	// runtime/internal/atomic. If SIGPROF arrives while the program is inside
  4672  	// the critical section, it creates a deadlock (when writing the sample).
  4673  	// As a workaround, create a counter of SIGPROFs while in critical section
  4674  	// to store the count, and pass it to sigprof.add() later when SIGPROF is
  4675  	// received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
  4676  	if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
  4677  		if f := findfunc(pc); f.valid() {
  4678  			if hasPrefix(funcname(f), "runtime/internal/atomic") {
  4679  				cpuprof.lostAtomic++
  4680  				return
  4681  			}
  4682  		}
  4683  		if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
  4684  			// runtime/internal/atomic functions call into kernel
  4685  			// helpers on arm < 7. See
  4686  			// runtime/internal/atomic/sys_linux_arm.s.
  4687  			cpuprof.lostAtomic++
  4688  			return
  4689  		}
  4690  	}
  4691  
  4692  	// Profiling runs concurrently with GC, so it must not allocate.
  4693  	// Set a trap in case the code does allocate.
  4694  	// Note that on windows, one thread takes profiles of all the
  4695  	// other threads, so mp is usually not getg().m.
  4696  	// In fact mp may not even be stopped.
  4697  	// See golang.org/issue/17165.
  4698  	getg().m.mallocing++
  4699  
  4700  	var stk [maxCPUProfStack]uintptr
  4701  	n := 0
  4702  	if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
  4703  		cgoOff := 0
  4704  		// Check cgoCallersUse to make sure that we are not
  4705  		// interrupting other code that is fiddling with
  4706  		// cgoCallers.  We are running in a signal handler
  4707  		// with all signals blocked, so we don't have to worry
  4708  		// about any other code interrupting us.
  4709  		if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
  4710  			for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
  4711  				cgoOff++
  4712  			}
  4713  			copy(stk[:], mp.cgoCallers[:cgoOff])
  4714  			mp.cgoCallers[0] = 0
  4715  		}
  4716  
  4717  		// Collect Go stack that leads to the cgo call.
  4718  		n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
  4719  		if n > 0 {
  4720  			n += cgoOff
  4721  		}
  4722  	} else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
  4723  		// Libcall, i.e. runtime syscall on windows.
  4724  		// Collect Go stack that leads to the call.
  4725  		n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[n], len(stk[n:]), nil, nil, 0)
  4726  	} else if mp != nil && mp.vdsoSP != 0 {
  4727  		// VDSO call, e.g. nanotime1 on Linux.
  4728  		// Collect Go stack that leads to the call.
  4729  		n = gentraceback(mp.vdsoPC, mp.vdsoSP, 0, gp, 0, &stk[n], len(stk[n:]), nil, nil, _TraceJumpStack)
  4730  	} else {
  4731  		n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
  4732  	}
  4733  
  4734  	if n <= 0 {
  4735  		// Normal traceback is impossible or has failed.
  4736  		// Account it against abstract "System" or "GC".
  4737  		n = 2
  4738  		if inVDSOPage(pc) {
  4739  			pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
  4740  		} else if pc > firstmoduledata.etext {
  4741  			// "ExternalCode" is better than "etext".
  4742  			pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
  4743  		}
  4744  		stk[0] = pc
  4745  		if mp.preemptoff != "" {
  4746  			stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
  4747  		} else {
  4748  			stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
  4749  		}
  4750  	}
  4751  
  4752  	if prof.hz.Load() != 0 {
  4753  		// Note: it can happen on Windows that we interrupted a system thread
  4754  		// with no g, so gp could nil. The other nil checks are done out of
  4755  		// caution, but not expected to be nil in practice.
  4756  		var tagPtr *unsafe.Pointer
  4757  		if gp != nil && gp.m != nil && gp.m.curg != nil {
  4758  			tagPtr = &gp.m.curg.labels
  4759  		}
  4760  		cpuprof.add(tagPtr, stk[:n])
  4761  
  4762  		gprof := gp
  4763  		var pp *p
  4764  		if gp != nil && gp.m != nil {
  4765  			if gp.m.curg != nil {
  4766  				gprof = gp.m.curg
  4767  			}
  4768  			pp = gp.m.p.ptr()
  4769  		}
  4770  		traceCPUSample(gprof, pp, stk[:n])
  4771  	}
  4772  	getg().m.mallocing--
  4773  }
  4774  
  4775  // setcpuprofilerate sets the CPU profiling rate to hz times per second.
  4776  // If hz <= 0, setcpuprofilerate turns off CPU profiling.
  4777  func setcpuprofilerate(hz int32) {
  4778  	// Force sane arguments.
  4779  	if hz < 0 {
  4780  		hz = 0
  4781  	}
  4782  
  4783  	// Disable preemption, otherwise we can be rescheduled to another thread
  4784  	// that has profiling enabled.
  4785  	gp := getg()
  4786  	gp.m.locks++
  4787  
  4788  	// Stop profiler on this thread so that it is safe to lock prof.
  4789  	// if a profiling signal came in while we had prof locked,
  4790  	// it would deadlock.
  4791  	setThreadCPUProfiler(0)
  4792  
  4793  	for !prof.signalLock.CompareAndSwap(0, 1) {
  4794  		osyield()
  4795  	}
  4796  	if prof.hz.Load() != hz {
  4797  		setProcessCPUProfiler(hz)
  4798  		prof.hz.Store(hz)
  4799  	}
  4800  	prof.signalLock.Store(0)
  4801  
  4802  	lock(&sched.lock)
  4803  	sched.profilehz = hz
  4804  	unlock(&sched.lock)
  4805  
  4806  	if hz != 0 {
  4807  		setThreadCPUProfiler(hz)
  4808  	}
  4809  
  4810  	gp.m.locks--
  4811  }
  4812  
  4813  // init initializes pp, which may be a freshly allocated p or a
  4814  // previously destroyed p, and transitions it to status _Pgcstop.
  4815  func (pp *p) init(id int32) {
  4816  	pp.id = id
  4817  	pp.status = _Pgcstop
  4818  	pp.sudogcache = pp.sudogbuf[:0]
  4819  	pp.deferpool = pp.deferpoolbuf[:0]
  4820  	pp.wbBuf.reset()
  4821  	if pp.mcache == nil {
  4822  		if id == 0 {
  4823  			if mcache0 == nil {
  4824  				throw("missing mcache?")
  4825  			}
  4826  			// Use the bootstrap mcache0. Only one P will get
  4827  			// mcache0: the one with ID 0.
  4828  			pp.mcache = mcache0
  4829  		} else {
  4830  			pp.mcache = allocmcache()
  4831  		}
  4832  	}
  4833  	if raceenabled && pp.raceprocctx == 0 {
  4834  		if id == 0 {
  4835  			pp.raceprocctx = raceprocctx0
  4836  			raceprocctx0 = 0 // bootstrap
  4837  		} else {
  4838  			pp.raceprocctx = raceproccreate()
  4839  		}
  4840  	}
  4841  	lockInit(&pp.timersLock, lockRankTimers)
  4842  
  4843  	// This P may get timers when it starts running. Set the mask here
  4844  	// since the P may not go through pidleget (notably P 0 on startup).
  4845  	timerpMask.set(id)
  4846  	// Similarly, we may not go through pidleget before this P starts
  4847  	// running if it is P 0 on startup.
  4848  	idlepMask.clear(id)
  4849  }
  4850  
  4851  // destroy releases all of the resources associated with pp and
  4852  // transitions it to status _Pdead.
  4853  //
  4854  // sched.lock must be held and the world must be stopped.
  4855  func (pp *p) destroy() {
  4856  	assertLockHeld(&sched.lock)
  4857  	assertWorldStopped()
  4858  
  4859  	// Move all runnable goroutines to the global queue
  4860  	for pp.runqhead != pp.runqtail {
  4861  		// Pop from tail of local queue
  4862  		pp.runqtail--
  4863  		gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
  4864  		// Push onto head of global queue
  4865  		globrunqputhead(gp)
  4866  	}
  4867  	if pp.runnext != 0 {
  4868  		globrunqputhead(pp.runnext.ptr())
  4869  		pp.runnext = 0
  4870  	}
  4871  	if len(pp.timers) > 0 {
  4872  		plocal := getg().m.p.ptr()
  4873  		// The world is stopped, but we acquire timersLock to
  4874  		// protect against sysmon calling timeSleepUntil.
  4875  		// This is the only case where we hold the timersLock of
  4876  		// more than one P, so there are no deadlock concerns.
  4877  		lock(&plocal.timersLock)
  4878  		lock(&pp.timersLock)
  4879  		moveTimers(plocal, pp.timers)
  4880  		pp.timers = nil
  4881  		pp.numTimers.Store(0)
  4882  		pp.deletedTimers.Store(0)
  4883  		pp.timer0When.Store(0)
  4884  		unlock(&pp.timersLock)
  4885  		unlock(&plocal.timersLock)
  4886  	}
  4887  	// Flush p's write barrier buffer.
  4888  	if gcphase != _GCoff {
  4889  		wbBufFlush1(pp)
  4890  		pp.gcw.dispose()
  4891  	}
  4892  	for i := range pp.sudogbuf {
  4893  		pp.sudogbuf[i] = nil
  4894  	}
  4895  	pp.sudogcache = pp.sudogbuf[:0]
  4896  	for j := range pp.deferpoolbuf {
  4897  		pp.deferpoolbuf[j] = nil
  4898  	}
  4899  	pp.deferpool = pp.deferpoolbuf[:0]
  4900  	systemstack(func() {
  4901  		for i := 0; i < pp.mspancache.len; i++ {
  4902  			// Safe to call since the world is stopped.
  4903  			mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
  4904  		}
  4905  		pp.mspancache.len = 0
  4906  		lock(&mheap_.lock)
  4907  		pp.pcache.flush(&mheap_.pages)
  4908  		unlock(&mheap_.lock)
  4909  	})
  4910  	freemcache(pp.mcache)
  4911  	pp.mcache = nil
  4912  	gfpurge(pp)
  4913  	traceProcFree(pp)
  4914  	if raceenabled {
  4915  		if pp.timerRaceCtx != 0 {
  4916  			// The race detector code uses a callback to fetch
  4917  			// the proc context, so arrange for that callback
  4918  			// to see the right thing.
  4919  			// This hack only works because we are the only
  4920  			// thread running.
  4921  			mp := getg().m
  4922  			phold := mp.p.ptr()
  4923  			mp.p.set(pp)
  4924  
  4925  			racectxend(pp.timerRaceCtx)
  4926  			pp.timerRaceCtx = 0
  4927  
  4928  			mp.p.set(phold)
  4929  		}
  4930  		raceprocdestroy(pp.raceprocctx)
  4931  		pp.raceprocctx = 0
  4932  	}
  4933  	pp.gcAssistTime = 0
  4934  	pp.status = _Pdead
  4935  }
  4936  
  4937  // Change number of processors.
  4938  //
  4939  // sched.lock must be held, and the world must be stopped.
  4940  //
  4941  // gcworkbufs must not be being modified by either the GC or the write barrier
  4942  // code, so the GC must not be running if the number of Ps actually changes.
  4943  //
  4944  // Returns list of Ps with local work, they need to be scheduled by the caller.
  4945  func procresize(nprocs int32) *p {
  4946  	assertLockHeld(&sched.lock)
  4947  	assertWorldStopped()
  4948  
  4949  	old := gomaxprocs
  4950  	if old < 0 || nprocs <= 0 {
  4951  		throw("procresize: invalid arg")
  4952  	}
  4953  	if trace.enabled {
  4954  		traceGomaxprocs(nprocs)
  4955  	}
  4956  
  4957  	// update statistics
  4958  	now := nanotime()
  4959  	if sched.procresizetime != 0 {
  4960  		sched.totaltime += int64(old) * (now - sched.procresizetime)
  4961  	}
  4962  	sched.procresizetime = now
  4963  
  4964  	maskWords := (nprocs + 31) / 32
  4965  
  4966  	// Grow allp if necessary.
  4967  	if nprocs > int32(len(allp)) {
  4968  		// Synchronize with retake, which could be running
  4969  		// concurrently since it doesn't run on a P.
  4970  		lock(&allpLock)
  4971  		if nprocs <= int32(cap(allp)) {
  4972  			allp = allp[:nprocs]
  4973  		} else {
  4974  			nallp := make([]*p, nprocs)
  4975  			// Copy everything up to allp's cap so we
  4976  			// never lose old allocated Ps.
  4977  			copy(nallp, allp[:cap(allp)])
  4978  			allp = nallp
  4979  		}
  4980  
  4981  		if maskWords <= int32(cap(idlepMask)) {
  4982  			idlepMask = idlepMask[:maskWords]
  4983  			timerpMask = timerpMask[:maskWords]
  4984  		} else {
  4985  			nidlepMask := make([]uint32, maskWords)
  4986  			// No need to copy beyond len, old Ps are irrelevant.
  4987  			copy(nidlepMask, idlepMask)
  4988  			idlepMask = nidlepMask
  4989  
  4990  			ntimerpMask := make([]uint32, maskWords)
  4991  			copy(ntimerpMask, timerpMask)
  4992  			timerpMask = ntimerpMask
  4993  		}
  4994  		unlock(&allpLock)
  4995  	}
  4996  
  4997  	// initialize new P's
  4998  	for i := old; i < nprocs; i++ {
  4999  		pp := allp[i]
  5000  		if pp == nil {
  5001  			pp = new(p)
  5002  		}
  5003  		pp.init(i)
  5004  		atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
  5005  	}
  5006  
  5007  	gp := getg()
  5008  	if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
  5009  		// continue to use the current P
  5010  		gp.m.p.ptr().status = _Prunning
  5011  		gp.m.p.ptr().mcache.prepareForSweep()
  5012  	} else {
  5013  		// release the current P and acquire allp[0].
  5014  		//
  5015  		// We must do this before destroying our current P
  5016  		// because p.destroy itself has write barriers, so we
  5017  		// need to do that from a valid P.
  5018  		if gp.m.p != 0 {
  5019  			if trace.enabled {
  5020  				// Pretend that we were descheduled
  5021  				// and then scheduled again to keep
  5022  				// the trace sane.
  5023  				traceGoSched()
  5024  				traceProcStop(gp.m.p.ptr())
  5025  			}
  5026  			gp.m.p.ptr().m = 0
  5027  		}
  5028  		gp.m.p = 0
  5029  		pp := allp[0]
  5030  		pp.m = 0
  5031  		pp.status = _Pidle
  5032  		acquirep(pp)
  5033  		if trace.enabled {
  5034  			traceGoStart()
  5035  		}
  5036  	}
  5037  
  5038  	// g.m.p is now set, so we no longer need mcache0 for bootstrapping.
  5039  	mcache0 = nil
  5040  
  5041  	// release resources from unused P's
  5042  	for i := nprocs; i < old; i++ {
  5043  		pp := allp[i]
  5044  		pp.destroy()
  5045  		// can't free P itself because it can be referenced by an M in syscall
  5046  	}
  5047  
  5048  	// Trim allp.
  5049  	if int32(len(allp)) != nprocs {
  5050  		lock(&allpLock)
  5051  		allp = allp[:nprocs]
  5052  		idlepMask = idlepMask[:maskWords]
  5053  		timerpMask = timerpMask[:maskWords]
  5054  		unlock(&allpLock)
  5055  	}
  5056  
  5057  	var runnablePs *p
  5058  	for i := nprocs - 1; i >= 0; i-- {
  5059  		pp := allp[i]
  5060  		if gp.m.p.ptr() == pp {
  5061  			continue
  5062  		}
  5063  		pp.status = _Pidle
  5064  		if runqempty(pp) {
  5065  			pidleput(pp, now)
  5066  		} else {
  5067  			pp.m.set(mget())
  5068  			pp.link.set(runnablePs)
  5069  			runnablePs = pp
  5070  		}
  5071  	}
  5072  	stealOrder.reset(uint32(nprocs))
  5073  	var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
  5074  	atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
  5075  	if old != nprocs {
  5076  		// Notify the limiter that the amount of procs has changed.
  5077  		gcCPULimiter.resetCapacity(now, nprocs)
  5078  	}
  5079  	return runnablePs
  5080  }
  5081  
  5082  // Associate p and the current m.
  5083  //
  5084  // This function is allowed to have write barriers even if the caller
  5085  // isn't because it immediately acquires pp.
  5086  //
  5087  //go:yeswritebarrierrec
  5088  func acquirep(pp *p) {
  5089  	// Do the part that isn't allowed to have write barriers.
  5090  	wirep(pp)
  5091  
  5092  	// Have p; write barriers now allowed.
  5093  
  5094  	// Perform deferred mcache flush before this P can allocate
  5095  	// from a potentially stale mcache.
  5096  	pp.mcache.prepareForSweep()
  5097  
  5098  	if trace.enabled {
  5099  		traceProcStart()
  5100  	}
  5101  }
  5102  
  5103  // wirep is the first step of acquirep, which actually associates the
  5104  // current M to pp. This is broken out so we can disallow write
  5105  // barriers for this part, since we don't yet have a P.
  5106  //
  5107  //go:nowritebarrierrec
  5108  //go:nosplit
  5109  func wirep(pp *p) {
  5110  	gp := getg()
  5111  
  5112  	if gp.m.p != 0 {
  5113  		throw("wirep: already in go")
  5114  	}
  5115  	if pp.m != 0 || pp.status != _Pidle {
  5116  		id := int64(0)
  5117  		if pp.m != 0 {
  5118  			id = pp.m.ptr().id
  5119  		}
  5120  		print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
  5121  		throw("wirep: invalid p state")
  5122  	}
  5123  	gp.m.p.set(pp)
  5124  	pp.m.set(gp.m)
  5125  	pp.status = _Prunning
  5126  }
  5127  
  5128  // Disassociate p and the current m.
  5129  func releasep() *p {
  5130  	gp := getg()
  5131  
  5132  	if gp.m.p == 0 {
  5133  		throw("releasep: invalid arg")
  5134  	}
  5135  	pp := gp.m.p.ptr()
  5136  	if pp.m.ptr() != gp.m || pp.status != _Prunning {
  5137  		print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
  5138  		throw("releasep: invalid p state")
  5139  	}
  5140  	if trace.enabled {
  5141  		traceProcStop(gp.m.p.ptr())
  5142  	}
  5143  	gp.m.p = 0
  5144  	pp.m = 0
  5145  	pp.status = _Pidle
  5146  	return pp
  5147  }
  5148  
  5149  func incidlelocked(v int32) {
  5150  	lock(&sched.lock)
  5151  	sched.nmidlelocked += v
  5152  	if v > 0 {
  5153  		checkdead()
  5154  	}
  5155  	unlock(&sched.lock)
  5156  }
  5157  
  5158  // Check for deadlock situation.
  5159  // The check is based on number of running M's, if 0 -> deadlock.
  5160  // sched.lock must be held.
  5161  func checkdead() {
  5162  	assertLockHeld(&sched.lock)
  5163  
  5164  	// For -buildmode=c-shared or -buildmode=c-archive it's OK if
  5165  	// there are no running goroutines. The calling program is
  5166  	// assumed to be running.
  5167  	if islibrary || isarchive {
  5168  		return
  5169  	}
  5170  
  5171  	// If we are dying because of a signal caught on an already idle thread,
  5172  	// freezetheworld will cause all running threads to block.
  5173  	// And runtime will essentially enter into deadlock state,
  5174  	// except that there is a thread that will call exit soon.
  5175  	if panicking.Load() > 0 {
  5176  		return
  5177  	}
  5178  
  5179  	// If we are not running under cgo, but we have an extra M then account
  5180  	// for it. (It is possible to have an extra M on Windows without cgo to
  5181  	// accommodate callbacks created by syscall.NewCallback. See issue #6751
  5182  	// for details.)
  5183  	var run0 int32
  5184  	if !iscgo && cgoHasExtraM {
  5185  		mp := lockextra(true)
  5186  		haveExtraM := extraMCount > 0
  5187  		unlockextra(mp)
  5188  		if haveExtraM {
  5189  			run0 = 1
  5190  		}
  5191  	}
  5192  
  5193  	run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
  5194  	if run > run0 {
  5195  		return
  5196  	}
  5197  	if run < 0 {
  5198  		print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
  5199  		throw("checkdead: inconsistent counts")
  5200  	}
  5201  
  5202  	grunning := 0
  5203  	forEachG(func(gp *g) {
  5204  		if isSystemGoroutine(gp, false) {
  5205  			return
  5206  		}
  5207  		s := readgstatus(gp)
  5208  		switch s &^ _Gscan {
  5209  		case _Gwaiting,
  5210  			_Gpreempted:
  5211  			grunning++
  5212  		case _Grunnable,
  5213  			_Grunning,
  5214  			_Gsyscall:
  5215  			print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
  5216  			throw("checkdead: runnable g")
  5217  		}
  5218  	})
  5219  	if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
  5220  		unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  5221  		fatal("no goroutines (main called runtime.Goexit) - deadlock!")
  5222  	}
  5223  
  5224  	// Maybe jump time forward for playground.
  5225  	if faketime != 0 {
  5226  		if when := timeSleepUntil(); when < maxWhen {
  5227  			faketime = when
  5228  
  5229  			// Start an M to steal the timer.
  5230  			pp, _ := pidleget(faketime)
  5231  			if pp == nil {
  5232  				// There should always be a free P since
  5233  				// nothing is running.
  5234  				throw("checkdead: no p for timer")
  5235  			}
  5236  			mp := mget()
  5237  			if mp == nil {
  5238  				// There should always be a free M since
  5239  				// nothing is running.
  5240  				throw("checkdead: no m for timer")
  5241  			}
  5242  			// M must be spinning to steal. We set this to be
  5243  			// explicit, but since this is the only M it would
  5244  			// become spinning on its own anyways.
  5245  			sched.nmspinning.Add(1)
  5246  			mp.spinning = true
  5247  			mp.nextp.set(pp)
  5248  			notewakeup(&mp.park)
  5249  			return
  5250  		}
  5251  	}
  5252  
  5253  	// There are no goroutines running, so we can look at the P's.
  5254  	for _, pp := range allp {
  5255  		if len(pp.timers) > 0 {
  5256  			return
  5257  		}
  5258  	}
  5259  
  5260  	unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  5261  	fatal("all goroutines are asleep - deadlock!")
  5262  }
  5263  
  5264  // forcegcperiod is the maximum time in nanoseconds between garbage
  5265  // collections. If we go this long without a garbage collection, one
  5266  // is forced to run.
  5267  //
  5268  // This is a variable for testing purposes. It normally doesn't change.
  5269  var forcegcperiod int64 = 2 * 60 * 1e9
  5270  
  5271  // needSysmonWorkaround is true if the workaround for
  5272  // golang.org/issue/42515 is needed on NetBSD.
  5273  var needSysmonWorkaround bool = false
  5274  
  5275  // Always runs without a P, so write barriers are not allowed.
  5276  //
  5277  //go:nowritebarrierrec
  5278  func sysmon() {
  5279  	lock(&sched.lock)
  5280  	sched.nmsys++
  5281  	checkdead()
  5282  	unlock(&sched.lock)
  5283  
  5284  	lasttrace := int64(0)
  5285  	idle := 0 // how many cycles in succession we had not wokeup somebody
  5286  	delay := uint32(0)
  5287  
  5288  	for {
  5289  		if idle == 0 { // start with 20us sleep...
  5290  			delay = 20
  5291  		} else if idle > 50 { // start doubling the sleep after 1ms...
  5292  			delay *= 2
  5293  		}
  5294  		if delay > 10*1000 { // up to 10ms
  5295  			delay = 10 * 1000
  5296  		}
  5297  		usleep(delay)
  5298  
  5299  		// sysmon should not enter deep sleep if schedtrace is enabled so that
  5300  		// it can print that information at the right time.
  5301  		//
  5302  		// It should also not enter deep sleep if there are any active P's so
  5303  		// that it can retake P's from syscalls, preempt long running G's, and
  5304  		// poll the network if all P's are busy for long stretches.
  5305  		//
  5306  		// It should wakeup from deep sleep if any P's become active either due
  5307  		// to exiting a syscall or waking up due to a timer expiring so that it
  5308  		// can resume performing those duties. If it wakes from a syscall it
  5309  		// resets idle and delay as a bet that since it had retaken a P from a
  5310  		// syscall before, it may need to do it again shortly after the
  5311  		// application starts work again. It does not reset idle when waking
  5312  		// from a timer to avoid adding system load to applications that spend
  5313  		// most of their time sleeping.
  5314  		now := nanotime()
  5315  		if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
  5316  			lock(&sched.lock)
  5317  			if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
  5318  				syscallWake := false
  5319  				next := timeSleepUntil()
  5320  				if next > now {
  5321  					sched.sysmonwait.Store(true)
  5322  					unlock(&sched.lock)
  5323  					// Make wake-up period small enough
  5324  					// for the sampling to be correct.
  5325  					sleep := forcegcperiod / 2
  5326  					if next-now < sleep {
  5327  						sleep = next - now
  5328  					}
  5329  					shouldRelax := sleep >= osRelaxMinNS
  5330  					if shouldRelax {
  5331  						osRelax(true)
  5332  					}
  5333  					syscallWake = notetsleep(&sched.sysmonnote, sleep)
  5334  					if shouldRelax {
  5335  						osRelax(false)
  5336  					}
  5337  					lock(&sched.lock)
  5338  					sched.sysmonwait.Store(false)
  5339  					noteclear(&sched.sysmonnote)
  5340  				}
  5341  				if syscallWake {
  5342  					idle = 0
  5343  					delay = 20
  5344  				}
  5345  			}
  5346  			unlock(&sched.lock)
  5347  		}
  5348  
  5349  		lock(&sched.sysmonlock)
  5350  		// Update now in case we blocked on sysmonnote or spent a long time
  5351  		// blocked on schedlock or sysmonlock above.
  5352  		now = nanotime()
  5353  
  5354  		// trigger libc interceptors if needed
  5355  		if *cgo_yield != nil {
  5356  			asmcgocall(*cgo_yield, nil)
  5357  		}
  5358  		// poll network if not polled for more than 10ms
  5359  		lastpoll := sched.lastpoll.Load()
  5360  		if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
  5361  			sched.lastpoll.CompareAndSwap(lastpoll, now)
  5362  			list := netpoll(0) // non-blocking - returns list of goroutines
  5363  			if !list.empty() {
  5364  				// Need to decrement number of idle locked M's
  5365  				// (pretending that one more is running) before injectglist.
  5366  				// Otherwise it can lead to the following situation:
  5367  				// injectglist grabs all P's but before it starts M's to run the P's,
  5368  				// another M returns from syscall, finishes running its G,
  5369  				// observes that there is no work to do and no other running M's
  5370  				// and reports deadlock.
  5371  				incidlelocked(-1)
  5372  				injectglist(&list)
  5373  				incidlelocked(1)
  5374  			}
  5375  		}
  5376  		if GOOS == "netbsd" && needSysmonWorkaround {
  5377  			// netpoll is responsible for waiting for timer
  5378  			// expiration, so we typically don't have to worry
  5379  			// about starting an M to service timers. (Note that
  5380  			// sleep for timeSleepUntil above simply ensures sysmon
  5381  			// starts running again when that timer expiration may
  5382  			// cause Go code to run again).
  5383  			//
  5384  			// However, netbsd has a kernel bug that sometimes
  5385  			// misses netpollBreak wake-ups, which can lead to
  5386  			// unbounded delays servicing timers. If we detect this
  5387  			// overrun, then startm to get something to handle the
  5388  			// timer.
  5389  			//
  5390  			// See issue 42515 and
  5391  			// https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
  5392  			if next := timeSleepUntil(); next < now {
  5393  				startm(nil, false)
  5394  			}
  5395  		}
  5396  		if scavenger.sysmonWake.Load() != 0 {
  5397  			// Kick the scavenger awake if someone requested it.
  5398  			scavenger.wake()
  5399  		}
  5400  		// retake P's blocked in syscalls
  5401  		// and preempt long running G's
  5402  		if retake(now) != 0 {
  5403  			idle = 0
  5404  		} else {
  5405  			idle++
  5406  		}
  5407  		// check if we need to force a GC
  5408  		if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
  5409  			lock(&forcegc.lock)
  5410  			forcegc.idle.Store(false)
  5411  			var list gList
  5412  			list.push(forcegc.g)
  5413  			injectglist(&list)
  5414  			unlock(&forcegc.lock)
  5415  		}
  5416  		if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
  5417  			lasttrace = now
  5418  			schedtrace(debug.scheddetail > 0)
  5419  		}
  5420  		unlock(&sched.sysmonlock)
  5421  	}
  5422  }
  5423  
  5424  type sysmontick struct {
  5425  	schedtick   uint32
  5426  	schedwhen   int64
  5427  	syscalltick uint32
  5428  	syscallwhen int64
  5429  }
  5430  
  5431  // forcePreemptNS is the time slice given to a G before it is
  5432  // preempted.
  5433  const forcePreemptNS = 10 * 1000 * 1000 // 10ms
  5434  
  5435  func retake(now int64) uint32 {
  5436  	n := 0
  5437  	// Prevent allp slice changes. This lock will be completely
  5438  	// uncontended unless we're already stopping the world.
  5439  	lock(&allpLock)
  5440  	// We can't use a range loop over allp because we may
  5441  	// temporarily drop the allpLock. Hence, we need to re-fetch
  5442  	// allp each time around the loop.
  5443  	for i := 0; i < len(allp); i++ {
  5444  		pp := allp[i]
  5445  		if pp == nil {
  5446  			// This can happen if procresize has grown
  5447  			// allp but not yet created new Ps.
  5448  			continue
  5449  		}
  5450  		pd := &pp.sysmontick
  5451  		s := pp.status
  5452  		sysretake := false
  5453  		if s == _Prunning || s == _Psyscall {
  5454  			// Preempt G if it's running for too long.
  5455  			t := int64(pp.schedtick)
  5456  			if int64(pd.schedtick) != t {
  5457  				pd.schedtick = uint32(t)
  5458  				pd.schedwhen = now
  5459  			} else if pd.schedwhen+forcePreemptNS <= now {
  5460  				preemptone(pp)
  5461  				// In case of syscall, preemptone() doesn't
  5462  				// work, because there is no M wired to P.
  5463  				sysretake = true
  5464  			}
  5465  		}
  5466  		if s == _Psyscall {
  5467  			// Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
  5468  			t := int64(pp.syscalltick)
  5469  			if !sysretake && int64(pd.syscalltick) != t {
  5470  				pd.syscalltick = uint32(t)
  5471  				pd.syscallwhen = now
  5472  				continue
  5473  			}
  5474  			// On the one hand we don't want to retake Ps if there is no other work to do,
  5475  			// but on the other hand we want to retake them eventually
  5476  			// because they can prevent the sysmon thread from deep sleep.
  5477  			if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
  5478  				continue
  5479  			}
  5480  			// Drop allpLock so we can take sched.lock.
  5481  			unlock(&allpLock)
  5482  			// Need to decrement number of idle locked M's
  5483  			// (pretending that one more is running) before the CAS.
  5484  			// Otherwise the M from which we retake can exit the syscall,
  5485  			// increment nmidle and report deadlock.
  5486  			incidlelocked(-1)
  5487  			if atomic.Cas(&pp.status, s, _Pidle) {
  5488  				if trace.enabled {
  5489  					traceGoSysBlock(pp)
  5490  					traceProcStop(pp)
  5491  				}
  5492  				n++
  5493  				pp.syscalltick++
  5494  				handoffp(pp)
  5495  			}
  5496  			incidlelocked(1)
  5497  			lock(&allpLock)
  5498  		}
  5499  	}
  5500  	unlock(&allpLock)
  5501  	return uint32(n)
  5502  }
  5503  
  5504  // Tell all goroutines that they have been preempted and they should stop.
  5505  // This function is purely best-effort. It can fail to inform a goroutine if a
  5506  // processor just started running it.
  5507  // No locks need to be held.
  5508  // Returns true if preemption request was issued to at least one goroutine.
  5509  func preemptall() bool {
  5510  	res := false
  5511  	for _, pp := range allp {
  5512  		if pp.status != _Prunning {
  5513  			continue
  5514  		}
  5515  		if preemptone(pp) {
  5516  			res = true
  5517  		}
  5518  	}
  5519  	return res
  5520  }
  5521  
  5522  // Tell the goroutine running on processor P to stop.
  5523  // This function is purely best-effort. It can incorrectly fail to inform the
  5524  // goroutine. It can inform the wrong goroutine. Even if it informs the
  5525  // correct goroutine, that goroutine might ignore the request if it is
  5526  // simultaneously executing newstack.
  5527  // No lock needs to be held.
  5528  // Returns true if preemption request was issued.
  5529  // The actual preemption will happen at some point in the future
  5530  // and will be indicated by the gp->status no longer being
  5531  // Grunning
  5532  func preemptone(pp *p) bool {
  5533  	mp := pp.m.ptr()
  5534  	if mp == nil || mp == getg().m {
  5535  		return false
  5536  	}
  5537  	gp := mp.curg
  5538  	if gp == nil || gp == mp.g0 {
  5539  		return false
  5540  	}
  5541  
  5542  	gp.preempt = true
  5543  
  5544  	// Every call in a goroutine checks for stack overflow by
  5545  	// comparing the current stack pointer to gp->stackguard0.
  5546  	// Setting gp->stackguard0 to StackPreempt folds
  5547  	// preemption into the normal stack overflow check.
  5548  	gp.stackguard0 = stackPreempt
  5549  
  5550  	// Request an async preemption of this P.
  5551  	if preemptMSupported && debug.asyncpreemptoff == 0 {
  5552  		pp.preempt = true
  5553  		preemptM(mp)
  5554  	}
  5555  
  5556  	return true
  5557  }
  5558  
  5559  var starttime int64
  5560  
  5561  func schedtrace(detailed bool) {
  5562  	now := nanotime()
  5563  	if starttime == 0 {
  5564  		starttime = now
  5565  	}
  5566  
  5567  	lock(&sched.lock)
  5568  	print("SCHED ", (now-starttime)/1e6, "ms: gomaxprocs=", gomaxprocs, " idleprocs=", sched.npidle.Load(), " threads=", mcount(), " spinningthreads=", sched.nmspinning.Load(), " needspinning=", sched.needspinning.Load(), " idlethreads=", sched.nmidle, " runqueue=", sched.runqsize)
  5569  	if detailed {
  5570  		print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
  5571  	}
  5572  	// We must be careful while reading data from P's, M's and G's.
  5573  	// Even if we hold schedlock, most data can be changed concurrently.
  5574  	// E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
  5575  	for i, pp := range allp {
  5576  		mp := pp.m.ptr()
  5577  		h := atomic.Load(&pp.runqhead)
  5578  		t := atomic.Load(&pp.runqtail)
  5579  		if detailed {
  5580  			print("  P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
  5581  			if mp != nil {
  5582  				print(mp.id)
  5583  			} else {
  5584  				print("nil")
  5585  			}
  5586  			print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
  5587  		} else {
  5588  			// In non-detailed mode format lengths of per-P run queues as:
  5589  			// [len1 len2 len3 len4]
  5590  			print(" ")
  5591  			if i == 0 {
  5592  				print("[")
  5593  			}
  5594  			print(t - h)
  5595  			if i == len(allp)-1 {
  5596  				print("]\n")
  5597  			}
  5598  		}
  5599  	}
  5600  
  5601  	if !detailed {
  5602  		unlock(&sched.lock)
  5603  		return
  5604  	}
  5605  
  5606  	for mp := allm; mp != nil; mp = mp.alllink {
  5607  		pp := mp.p.ptr()
  5608  		print("  M", mp.id, ": p=")
  5609  		if pp != nil {
  5610  			print(pp.id)
  5611  		} else {
  5612  			print("nil")
  5613  		}
  5614  		print(" curg=")
  5615  		if mp.curg != nil {
  5616  			print(mp.curg.goid)
  5617  		} else {
  5618  			print("nil")
  5619  		}
  5620  		print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
  5621  		if lockedg := mp.lockedg.ptr(); lockedg != nil {
  5622  			print(lockedg.goid)
  5623  		} else {
  5624  			print("nil")
  5625  		}
  5626  		print("\n")
  5627  	}
  5628  
  5629  	forEachG(func(gp *g) {
  5630  		print("  G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
  5631  		if gp.m != nil {
  5632  			print(gp.m.id)
  5633  		} else {
  5634  			print("nil")
  5635  		}
  5636  		print(" lockedm=")
  5637  		if lockedm := gp.lockedm.ptr(); lockedm != nil {
  5638  			print(lockedm.id)
  5639  		} else {
  5640  			print("nil")
  5641  		}
  5642  		print("\n")
  5643  	})
  5644  	unlock(&sched.lock)
  5645  }
  5646  
  5647  // schedEnableUser enables or disables the scheduling of user
  5648  // goroutines.
  5649  //
  5650  // This does not stop already running user goroutines, so the caller
  5651  // should first stop the world when disabling user goroutines.
  5652  func schedEnableUser(enable bool) {
  5653  	lock(&sched.lock)
  5654  	if sched.disable.user == !enable {
  5655  		unlock(&sched.lock)
  5656  		return
  5657  	}
  5658  	sched.disable.user = !enable
  5659  	if enable {
  5660  		n := sched.disable.n
  5661  		sched.disable.n = 0
  5662  		globrunqputbatch(&sched.disable.runnable, n)
  5663  		unlock(&sched.lock)
  5664  		for ; n != 0 && sched.npidle.Load() != 0; n-- {
  5665  			startm(nil, false)
  5666  		}
  5667  	} else {
  5668  		unlock(&sched.lock)
  5669  	}
  5670  }
  5671  
  5672  // schedEnabled reports whether gp should be scheduled. It returns
  5673  // false is scheduling of gp is disabled.
  5674  //
  5675  // sched.lock must be held.
  5676  func schedEnabled(gp *g) bool {
  5677  	assertLockHeld(&sched.lock)
  5678  
  5679  	if sched.disable.user {
  5680  		return isSystemGoroutine(gp, true)
  5681  	}
  5682  	return true
  5683  }
  5684  
  5685  // Put mp on midle list.
  5686  // sched.lock must be held.
  5687  // May run during STW, so write barriers are not allowed.
  5688  //
  5689  //go:nowritebarrierrec
  5690  func mput(mp *m) {
  5691  	assertLockHeld(&sched.lock)
  5692  
  5693  	mp.schedlink = sched.midle
  5694  	sched.midle.set(mp)
  5695  	sched.nmidle++
  5696  	checkdead()
  5697  }
  5698  
  5699  // Try to get an m from midle list.
  5700  // sched.lock must be held.
  5701  // May run during STW, so write barriers are not allowed.
  5702  //
  5703  //go:nowritebarrierrec
  5704  func mget() *m {
  5705  	assertLockHeld(&sched.lock)
  5706  
  5707  	mp := sched.midle.ptr()
  5708  	if mp != nil {
  5709  		sched.midle = mp.schedlink
  5710  		sched.nmidle--
  5711  	}
  5712  	return mp
  5713  }
  5714  
  5715  // Put gp on the global runnable queue.
  5716  // sched.lock must be held.
  5717  // May run during STW, so write barriers are not allowed.
  5718  //
  5719  //go:nowritebarrierrec
  5720  func globrunqput(gp *g) {
  5721  	assertLockHeld(&sched.lock)
  5722  
  5723  	sched.runq.pushBack(gp)
  5724  	sched.runqsize++
  5725  }
  5726  
  5727  // Put gp at the head of the global runnable queue.
  5728  // sched.lock must be held.
  5729  // May run during STW, so write barriers are not allowed.
  5730  //
  5731  //go:nowritebarrierrec
  5732  func globrunqputhead(gp *g) {
  5733  	assertLockHeld(&sched.lock)
  5734  
  5735  	sched.runq.push(gp)
  5736  	sched.runqsize++
  5737  }
  5738  
  5739  // Put a batch of runnable goroutines on the global runnable queue.
  5740  // This clears *batch.
  5741  // sched.lock must be held.
  5742  // May run during STW, so write barriers are not allowed.
  5743  //
  5744  //go:nowritebarrierrec
  5745  func globrunqputbatch(batch *gQueue, n int32) {
  5746  	assertLockHeld(&sched.lock)
  5747  
  5748  	sched.runq.pushBackAll(*batch)
  5749  	sched.runqsize += n
  5750  	*batch = gQueue{}
  5751  }
  5752  
  5753  // Try get a batch of G's from the global runnable queue.
  5754  // sched.lock must be held.
  5755  func globrunqget(pp *p, max int32) *g {
  5756  	assertLockHeld(&sched.lock)
  5757  
  5758  	if sched.runqsize == 0 {
  5759  		return nil
  5760  	}
  5761  
  5762  	n := sched.runqsize/gomaxprocs + 1
  5763  	if n > sched.runqsize {
  5764  		n = sched.runqsize
  5765  	}
  5766  	if max > 0 && n > max {
  5767  		n = max
  5768  	}
  5769  	if n > int32(len(pp.runq))/2 {
  5770  		n = int32(len(pp.runq)) / 2
  5771  	}
  5772  
  5773  	sched.runqsize -= n
  5774  
  5775  	gp := sched.runq.pop()
  5776  	n--
  5777  	for ; n > 0; n-- {
  5778  		gp1 := sched.runq.pop()
  5779  		runqput(pp, gp1, false)
  5780  	}
  5781  	return gp
  5782  }
  5783  
  5784  // pMask is an atomic bitstring with one bit per P.
  5785  type pMask []uint32
  5786  
  5787  // read returns true if P id's bit is set.
  5788  func (p pMask) read(id uint32) bool {
  5789  	word := id / 32
  5790  	mask := uint32(1) << (id % 32)
  5791  	return (atomic.Load(&p[word]) & mask) != 0
  5792  }
  5793  
  5794  // set sets P id's bit.
  5795  func (p pMask) set(id int32) {
  5796  	word := id / 32
  5797  	mask := uint32(1) << (id % 32)
  5798  	atomic.Or(&p[word], mask)
  5799  }
  5800  
  5801  // clear clears P id's bit.
  5802  func (p pMask) clear(id int32) {
  5803  	word := id / 32
  5804  	mask := uint32(1) << (id % 32)
  5805  	atomic.And(&p[word], ^mask)
  5806  }
  5807  
  5808  // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
  5809  //
  5810  // Ideally, the timer mask would be kept immediately consistent on any timer
  5811  // operations. Unfortunately, updating a shared global data structure in the
  5812  // timer hot path adds too much overhead in applications frequently switching
  5813  // between no timers and some timers.
  5814  //
  5815  // As a compromise, the timer mask is updated only on pidleget / pidleput. A
  5816  // running P (returned by pidleget) may add a timer at any time, so its mask
  5817  // must be set. An idle P (passed to pidleput) cannot add new timers while
  5818  // idle, so if it has no timers at that time, its mask may be cleared.
  5819  //
  5820  // Thus, we get the following effects on timer-stealing in findrunnable:
  5821  //
  5822  //   - Idle Ps with no timers when they go idle are never checked in findrunnable
  5823  //     (for work- or timer-stealing; this is the ideal case).
  5824  //   - Running Ps must always be checked.
  5825  //   - Idle Ps whose timers are stolen must continue to be checked until they run
  5826  //     again, even after timer expiration.
  5827  //
  5828  // When the P starts running again, the mask should be set, as a timer may be
  5829  // added at any time.
  5830  //
  5831  // TODO(prattmic): Additional targeted updates may improve the above cases.
  5832  // e.g., updating the mask when stealing a timer.
  5833  func updateTimerPMask(pp *p) {
  5834  	if pp.numTimers.Load() > 0 {
  5835  		return
  5836  	}
  5837  
  5838  	// Looks like there are no timers, however another P may transiently
  5839  	// decrement numTimers when handling a timerModified timer in
  5840  	// checkTimers. We must take timersLock to serialize with these changes.
  5841  	lock(&pp.timersLock)
  5842  	if pp.numTimers.Load() == 0 {
  5843  		timerpMask.clear(pp.id)
  5844  	}
  5845  	unlock(&pp.timersLock)
  5846  }
  5847  
  5848  // pidleput puts p on the _Pidle list. now must be a relatively recent call
  5849  // to nanotime or zero. Returns now or the current time if now was zero.
  5850  //
  5851  // This releases ownership of p. Once sched.lock is released it is no longer
  5852  // safe to use p.
  5853  //
  5854  // sched.lock must be held.
  5855  //
  5856  // May run during STW, so write barriers are not allowed.
  5857  //
  5858  //go:nowritebarrierrec
  5859  func pidleput(pp *p, now int64) int64 {
  5860  	assertLockHeld(&sched.lock)
  5861  
  5862  	if !runqempty(pp) {
  5863  		throw("pidleput: P has non-empty run queue")
  5864  	}
  5865  	if now == 0 {
  5866  		now = nanotime()
  5867  	}
  5868  	updateTimerPMask(pp) // clear if there are no timers.
  5869  	idlepMask.set(pp.id)
  5870  	pp.link = sched.pidle
  5871  	sched.pidle.set(pp)
  5872  	sched.npidle.Add(1)
  5873  	if !pp.limiterEvent.start(limiterEventIdle, now) {
  5874  		throw("must be able to track idle limiter event")
  5875  	}
  5876  	return now
  5877  }
  5878  
  5879  // pidleget tries to get a p from the _Pidle list, acquiring ownership.
  5880  //
  5881  // sched.lock must be held.
  5882  //
  5883  // May run during STW, so write barriers are not allowed.
  5884  //
  5885  //go:nowritebarrierrec
  5886  func pidleget(now int64) (*p, int64) {
  5887  	assertLockHeld(&sched.lock)
  5888  
  5889  	pp := sched.pidle.ptr()
  5890  	if pp != nil {
  5891  		// Timer may get added at any time now.
  5892  		if now == 0 {
  5893  			now = nanotime()
  5894  		}
  5895  		timerpMask.set(pp.id)
  5896  		idlepMask.clear(pp.id)
  5897  		sched.pidle = pp.link
  5898  		sched.npidle.Add(-1)
  5899  		pp.limiterEvent.stop(limiterEventIdle, now)
  5900  	}
  5901  	return pp, now
  5902  }
  5903  
  5904  // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
  5905  // This is called by spinning Ms (or callers than need a spinning M) that have
  5906  // found work. If no P is available, this must synchronized with non-spinning
  5907  // Ms that may be preparing to drop their P without discovering this work.
  5908  //
  5909  // sched.lock must be held.
  5910  //
  5911  // May run during STW, so write barriers are not allowed.
  5912  //
  5913  //go:nowritebarrierrec
  5914  func pidlegetSpinning(now int64) (*p, int64) {
  5915  	assertLockHeld(&sched.lock)
  5916  
  5917  	pp, now := pidleget(now)
  5918  	if pp == nil {
  5919  		// See "Delicate dance" comment in findrunnable. We found work
  5920  		// that we cannot take, we must synchronize with non-spinning
  5921  		// Ms that may be preparing to drop their P.
  5922  		sched.needspinning.Store(1)
  5923  		return nil, now
  5924  	}
  5925  
  5926  	return pp, now
  5927  }
  5928  
  5929  // runqempty reports whether pp has no Gs on its local run queue.
  5930  // It never returns true spuriously.
  5931  func runqempty(pp *p) bool {
  5932  	// Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
  5933  	// 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
  5934  	// Simply observing that runqhead == runqtail and then observing that runqnext == nil
  5935  	// does not mean the queue is empty.
  5936  	for {
  5937  		head := atomic.Load(&pp.runqhead)
  5938  		tail := atomic.Load(&pp.runqtail)
  5939  		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
  5940  		if tail == atomic.Load(&pp.runqtail) {
  5941  			return head == tail && runnext == 0
  5942  		}
  5943  	}
  5944  }
  5945  
  5946  // To shake out latent assumptions about scheduling order,
  5947  // we introduce some randomness into scheduling decisions
  5948  // when running with the race detector.
  5949  // The need for this was made obvious by changing the
  5950  // (deterministic) scheduling order in Go 1.5 and breaking
  5951  // many poorly-written tests.
  5952  // With the randomness here, as long as the tests pass
  5953  // consistently with -race, they shouldn't have latent scheduling
  5954  // assumptions.
  5955  const randomizeScheduler = raceenabled
  5956  
  5957  // runqput tries to put g on the local runnable queue.
  5958  // If next is false, runqput adds g to the tail of the runnable queue.
  5959  // If next is true, runqput puts g in the pp.runnext slot.
  5960  // If the run queue is full, runnext puts g on the global queue.
  5961  // Executed only by the owner P.
  5962  func runqput(pp *p, gp *g, next bool) {
  5963  	if randomizeScheduler && next && fastrandn(2) == 0 {
  5964  		next = false
  5965  	}
  5966  
  5967  	if next {
  5968  	retryNext:
  5969  		oldnext := pp.runnext
  5970  		if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
  5971  			goto retryNext
  5972  		}
  5973  		if oldnext == 0 {
  5974  			return
  5975  		}
  5976  		// Kick the old runnext out to the regular run queue.
  5977  		gp = oldnext.ptr()
  5978  	}
  5979  
  5980  retry:
  5981  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  5982  	t := pp.runqtail
  5983  	if t-h < uint32(len(pp.runq)) {
  5984  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  5985  		atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
  5986  		return
  5987  	}
  5988  	if runqputslow(pp, gp, h, t) {
  5989  		return
  5990  	}
  5991  	// the queue is not full, now the put above must succeed
  5992  	goto retry
  5993  }
  5994  
  5995  // Put g and a batch of work from local runnable queue on global queue.
  5996  // Executed only by the owner P.
  5997  func runqputslow(pp *p, gp *g, h, t uint32) bool {
  5998  	var batch [len(pp.runq)/2 + 1]*g
  5999  
  6000  	// First, grab a batch from local queue.
  6001  	n := t - h
  6002  	n = n / 2
  6003  	if n != uint32(len(pp.runq)/2) {
  6004  		throw("runqputslow: queue is not full")
  6005  	}
  6006  	for i := uint32(0); i < n; i++ {
  6007  		batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  6008  	}
  6009  	if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  6010  		return false
  6011  	}
  6012  	batch[n] = gp
  6013  
  6014  	if randomizeScheduler {
  6015  		for i := uint32(1); i <= n; i++ {
  6016  			j := fastrandn(i + 1)
  6017  			batch[i], batch[j] = batch[j], batch[i]
  6018  		}
  6019  	}
  6020  
  6021  	// Link the goroutines.
  6022  	for i := uint32(0); i < n; i++ {
  6023  		batch[i].schedlink.set(batch[i+1])
  6024  	}
  6025  	var q gQueue
  6026  	q.head.set(batch[0])
  6027  	q.tail.set(batch[n])
  6028  
  6029  	// Now put the batch on global queue.
  6030  	lock(&sched.lock)
  6031  	globrunqputbatch(&q, int32(n+1))
  6032  	unlock(&sched.lock)
  6033  	return true
  6034  }
  6035  
  6036  // runqputbatch tries to put all the G's on q on the local runnable queue.
  6037  // If the queue is full, they are put on the global queue; in that case
  6038  // this will temporarily acquire the scheduler lock.
  6039  // Executed only by the owner P.
  6040  func runqputbatch(pp *p, q *gQueue, qsize int) {
  6041  	h := atomic.LoadAcq(&pp.runqhead)
  6042  	t := pp.runqtail
  6043  	n := uint32(0)
  6044  	for !q.empty() && t-h < uint32(len(pp.runq)) {
  6045  		gp := q.pop()
  6046  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  6047  		t++
  6048  		n++
  6049  	}
  6050  	qsize -= int(n)
  6051  
  6052  	if randomizeScheduler {
  6053  		off := func(o uint32) uint32 {
  6054  			return (pp.runqtail + o) % uint32(len(pp.runq))
  6055  		}
  6056  		for i := uint32(1); i < n; i++ {
  6057  			j := fastrandn(i + 1)
  6058  			pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
  6059  		}
  6060  	}
  6061  
  6062  	atomic.StoreRel(&pp.runqtail, t)
  6063  	if !q.empty() {
  6064  		lock(&sched.lock)
  6065  		globrunqputbatch(q, int32(qsize))
  6066  		unlock(&sched.lock)
  6067  	}
  6068  }
  6069  
  6070  // Get g from local runnable queue.
  6071  // If inheritTime is true, gp should inherit the remaining time in the
  6072  // current time slice. Otherwise, it should start a new time slice.
  6073  // Executed only by the owner P.
  6074  func runqget(pp *p) (gp *g, inheritTime bool) {
  6075  	// If there's a runnext, it's the next G to run.
  6076  	next := pp.runnext
  6077  	// If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
  6078  	// because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
  6079  	// Hence, there's no need to retry this CAS if it fails.
  6080  	if next != 0 && pp.runnext.cas(next, 0) {
  6081  		return next.ptr(), true
  6082  	}
  6083  
  6084  	for {
  6085  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6086  		t := pp.runqtail
  6087  		if t == h {
  6088  			return nil, false
  6089  		}
  6090  		gp := pp.runq[h%uint32(len(pp.runq))].ptr()
  6091  		if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
  6092  			return gp, false
  6093  		}
  6094  	}
  6095  }
  6096  
  6097  // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
  6098  // Executed only by the owner P.
  6099  func runqdrain(pp *p) (drainQ gQueue, n uint32) {
  6100  	oldNext := pp.runnext
  6101  	if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
  6102  		drainQ.pushBack(oldNext.ptr())
  6103  		n++
  6104  	}
  6105  
  6106  retry:
  6107  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6108  	t := pp.runqtail
  6109  	qn := t - h
  6110  	if qn == 0 {
  6111  		return
  6112  	}
  6113  	if qn > uint32(len(pp.runq)) { // read inconsistent h and t
  6114  		goto retry
  6115  	}
  6116  
  6117  	if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
  6118  		goto retry
  6119  	}
  6120  
  6121  	// We've inverted the order in which it gets G's from the local P's runnable queue
  6122  	// and then advances the head pointer because we don't want to mess up the statuses of G's
  6123  	// while runqdrain() and runqsteal() are running in parallel.
  6124  	// Thus we should advance the head pointer before draining the local P into a gQueue,
  6125  	// so that we can update any gp.schedlink only after we take the full ownership of G,
  6126  	// meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
  6127  	// See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
  6128  	for i := uint32(0); i < qn; i++ {
  6129  		gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  6130  		drainQ.pushBack(gp)
  6131  		n++
  6132  	}
  6133  	return
  6134  }
  6135  
  6136  // Grabs a batch of goroutines from pp's runnable queue into batch.
  6137  // Batch is a ring buffer starting at batchHead.
  6138  // Returns number of grabbed goroutines.
  6139  // Can be executed by any P.
  6140  func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
  6141  	for {
  6142  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6143  		t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
  6144  		n := t - h
  6145  		n = n - n/2
  6146  		if n == 0 {
  6147  			if stealRunNextG {
  6148  				// Try to steal from pp.runnext.
  6149  				if next := pp.runnext; next != 0 {
  6150  					if pp.status == _Prunning {
  6151  						// Sleep to ensure that pp isn't about to run the g
  6152  						// we are about to steal.
  6153  						// The important use case here is when the g running
  6154  						// on pp ready()s another g and then almost
  6155  						// immediately blocks. Instead of stealing runnext
  6156  						// in this window, back off to give pp a chance to
  6157  						// schedule runnext. This will avoid thrashing gs
  6158  						// between different Ps.
  6159  						// A sync chan send/recv takes ~50ns as of time of
  6160  						// writing, so 3us gives ~50x overshoot.
  6161  						if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
  6162  							usleep(3)
  6163  						} else {
  6164  							// On some platforms system timer granularity is
  6165  							// 1-15ms, which is way too much for this
  6166  							// optimization. So just yield.
  6167  							osyield()
  6168  						}
  6169  					}
  6170  					if !pp.runnext.cas(next, 0) {
  6171  						continue
  6172  					}
  6173  					batch[batchHead%uint32(len(batch))] = next
  6174  					return 1
  6175  				}
  6176  			}
  6177  			return 0
  6178  		}
  6179  		if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
  6180  			continue
  6181  		}
  6182  		for i := uint32(0); i < n; i++ {
  6183  			g := pp.runq[(h+i)%uint32(len(pp.runq))]
  6184  			batch[(batchHead+i)%uint32(len(batch))] = g
  6185  		}
  6186  		if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  6187  			return n
  6188  		}
  6189  	}
  6190  }
  6191  
  6192  // Steal half of elements from local runnable queue of p2
  6193  // and put onto local runnable queue of p.
  6194  // Returns one of the stolen elements (or nil if failed).
  6195  func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
  6196  	t := pp.runqtail
  6197  	n := runqgrab(p2, &pp.runq, t, stealRunNextG)
  6198  	if n == 0 {
  6199  		return nil
  6200  	}
  6201  	n--
  6202  	gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
  6203  	if n == 0 {
  6204  		return gp
  6205  	}
  6206  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  6207  	if t-h+n >= uint32(len(pp.runq)) {
  6208  		throw("runqsteal: runq overflow")
  6209  	}
  6210  	atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
  6211  	return gp
  6212  }
  6213  
  6214  // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
  6215  // be on one gQueue or gList at a time.
  6216  type gQueue struct {
  6217  	head guintptr
  6218  	tail guintptr
  6219  }
  6220  
  6221  // empty reports whether q is empty.
  6222  func (q *gQueue) empty() bool {
  6223  	return q.head == 0
  6224  }
  6225  
  6226  // push adds gp to the head of q.
  6227  func (q *gQueue) push(gp *g) {
  6228  	gp.schedlink = q.head
  6229  	q.head.set(gp)
  6230  	if q.tail == 0 {
  6231  		q.tail.set(gp)
  6232  	}
  6233  }
  6234  
  6235  // pushBack adds gp to the tail of q.
  6236  func (q *gQueue) pushBack(gp *g) {
  6237  	gp.schedlink = 0
  6238  	if q.tail != 0 {
  6239  		q.tail.ptr().schedlink.set(gp)
  6240  	} else {
  6241  		q.head.set(gp)
  6242  	}
  6243  	q.tail.set(gp)
  6244  }
  6245  
  6246  // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
  6247  // not be used.
  6248  func (q *gQueue) pushBackAll(q2 gQueue) {
  6249  	if q2.tail == 0 {
  6250  		return
  6251  	}
  6252  	q2.tail.ptr().schedlink = 0
  6253  	if q.tail != 0 {
  6254  		q.tail.ptr().schedlink = q2.head
  6255  	} else {
  6256  		q.head = q2.head
  6257  	}
  6258  	q.tail = q2.tail
  6259  }
  6260  
  6261  // pop removes and returns the head of queue q. It returns nil if
  6262  // q is empty.
  6263  func (q *gQueue) pop() *g {
  6264  	gp := q.head.ptr()
  6265  	if gp != nil {
  6266  		q.head = gp.schedlink
  6267  		if q.head == 0 {
  6268  			q.tail = 0
  6269  		}
  6270  	}
  6271  	return gp
  6272  }
  6273  
  6274  // popList takes all Gs in q and returns them as a gList.
  6275  func (q *gQueue) popList() gList {
  6276  	stack := gList{q.head}
  6277  	*q = gQueue{}
  6278  	return stack
  6279  }
  6280  
  6281  // A gList is a list of Gs linked through g.schedlink. A G can only be
  6282  // on one gQueue or gList at a time.
  6283  type gList struct {
  6284  	head guintptr
  6285  }
  6286  
  6287  // empty reports whether l is empty.
  6288  func (l *gList) empty() bool {
  6289  	return l.head == 0
  6290  }
  6291  
  6292  // push adds gp to the head of l.
  6293  func (l *gList) push(gp *g) {
  6294  	gp.schedlink = l.head
  6295  	l.head.set(gp)
  6296  }
  6297  
  6298  // pushAll prepends all Gs in q to l.
  6299  func (l *gList) pushAll(q gQueue) {
  6300  	if !q.empty() {
  6301  		q.tail.ptr().schedlink = l.head
  6302  		l.head = q.head
  6303  	}
  6304  }
  6305  
  6306  // pop removes and returns the head of l. If l is empty, it returns nil.
  6307  func (l *gList) pop() *g {
  6308  	gp := l.head.ptr()
  6309  	if gp != nil {
  6310  		l.head = gp.schedlink
  6311  	}
  6312  	return gp
  6313  }
  6314  
  6315  //go:linkname setMaxThreads runtime/debug.setMaxThreads
  6316  func setMaxThreads(in int) (out int) {
  6317  	lock(&sched.lock)
  6318  	out = int(sched.maxmcount)
  6319  	if in > 0x7fffffff { // MaxInt32
  6320  		sched.maxmcount = 0x7fffffff
  6321  	} else {
  6322  		sched.maxmcount = int32(in)
  6323  	}
  6324  	checkmcount()
  6325  	unlock(&sched.lock)
  6326  	return
  6327  }
  6328  
  6329  //go:nosplit
  6330  func procPin() int {
  6331  	gp := getg()
  6332  	mp := gp.m
  6333  
  6334  	mp.locks++
  6335  	return int(mp.p.ptr().id)
  6336  }
  6337  
  6338  //go:nosplit
  6339  func procUnpin() {
  6340  	gp := getg()
  6341  	gp.m.locks--
  6342  }
  6343  
  6344  //go:linkname sync_runtime_procPin sync.runtime_procPin
  6345  //go:nosplit
  6346  func sync_runtime_procPin() int {
  6347  	return procPin()
  6348  }
  6349  
  6350  //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
  6351  //go:nosplit
  6352  func sync_runtime_procUnpin() {
  6353  	procUnpin()
  6354  }
  6355  
  6356  //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
  6357  //go:nosplit
  6358  func sync_atomic_runtime_procPin() int {
  6359  	return procPin()
  6360  }
  6361  
  6362  //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
  6363  //go:nosplit
  6364  func sync_atomic_runtime_procUnpin() {
  6365  	procUnpin()
  6366  }
  6367  
  6368  // Active spinning for sync.Mutex.
  6369  //
  6370  //go:linkname sync_runtime_canSpin sync.runtime_canSpin
  6371  //go:nosplit
  6372  func sync_runtime_canSpin(i int) bool {
  6373  	// sync.Mutex is cooperative, so we are conservative with spinning.
  6374  	// Spin only few times and only if running on a multicore machine and
  6375  	// GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
  6376  	// As opposed to runtime mutex we don't do passive spinning here,
  6377  	// because there can be work on global runq or on other Ps.
  6378  	if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
  6379  		return false
  6380  	}
  6381  	if p := getg().m.p.ptr(); !runqempty(p) {
  6382  		return false
  6383  	}
  6384  	return true
  6385  }
  6386  
  6387  //go:linkname sync_runtime_doSpin sync.runtime_doSpin
  6388  //go:nosplit
  6389  func sync_runtime_doSpin() {
  6390  	procyield(active_spin_cnt)
  6391  }
  6392  
  6393  var stealOrder randomOrder
  6394  
  6395  // randomOrder/randomEnum are helper types for randomized work stealing.
  6396  // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
  6397  // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
  6398  // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
  6399  type randomOrder struct {
  6400  	count    uint32
  6401  	coprimes []uint32
  6402  }
  6403  
  6404  type randomEnum struct {
  6405  	i     uint32
  6406  	count uint32
  6407  	pos   uint32
  6408  	inc   uint32
  6409  }
  6410  
  6411  func (ord *randomOrder) reset(count uint32) {
  6412  	ord.count = count
  6413  	ord.coprimes = ord.coprimes[:0]
  6414  	for i := uint32(1); i <= count; i++ {
  6415  		if gcd(i, count) == 1 {
  6416  			ord.coprimes = append(ord.coprimes, i)
  6417  		}
  6418  	}
  6419  }
  6420  
  6421  func (ord *randomOrder) start(i uint32) randomEnum {
  6422  	return randomEnum{
  6423  		count: ord.count,
  6424  		pos:   i % ord.count,
  6425  		inc:   ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
  6426  	}
  6427  }
  6428  
  6429  func (enum *randomEnum) done() bool {
  6430  	return enum.i == enum.count
  6431  }
  6432  
  6433  func (enum *randomEnum) next() {
  6434  	enum.i++
  6435  	enum.pos = (enum.pos + enum.inc) % enum.count
  6436  }
  6437  
  6438  func (enum *randomEnum) position() uint32 {
  6439  	return enum.pos
  6440  }
  6441  
  6442  func gcd(a, b uint32) uint32 {
  6443  	for b != 0 {
  6444  		a, b = b, a%b
  6445  	}
  6446  	return a
  6447  }
  6448  
  6449  // An initTask represents the set of initializations that need to be done for a package.
  6450  // Keep in sync with ../../test/initempty.go:initTask
  6451  type initTask struct {
  6452  	// TODO: pack the first 3 fields more tightly?
  6453  	state uintptr // 0 = uninitialized, 1 = in progress, 2 = done
  6454  	ndeps uintptr
  6455  	nfns  uintptr
  6456  	// followed by ndeps instances of an *initTask, one per package depended on
  6457  	// followed by nfns pcs, one per init function to run
  6458  }
  6459  
  6460  // inittrace stores statistics for init functions which are
  6461  // updated by malloc and newproc when active is true.
  6462  var inittrace tracestat
  6463  
  6464  type tracestat struct {
  6465  	active bool   // init tracing activation status
  6466  	id     uint64 // init goroutine id
  6467  	allocs uint64 // heap allocations
  6468  	bytes  uint64 // heap allocated bytes
  6469  }
  6470  
  6471  func doInit(t *initTask) {
  6472  	switch t.state {
  6473  	case 2: // fully initialized
  6474  		return
  6475  	case 1: // initialization in progress
  6476  		throw("recursive call during initialization - linker skew")
  6477  	default: // not initialized yet
  6478  		t.state = 1 // initialization in progress
  6479  
  6480  		for i := uintptr(0); i < t.ndeps; i++ {
  6481  			p := add(unsafe.Pointer(t), (3+i)*goarch.PtrSize)
  6482  			t2 := *(**initTask)(p)
  6483  			doInit(t2)
  6484  		}
  6485  
  6486  		if t.nfns == 0 {
  6487  			t.state = 2 // initialization done
  6488  			return
  6489  		}
  6490  
  6491  		var (
  6492  			start  int64
  6493  			before tracestat
  6494  		)
  6495  
  6496  		if inittrace.active {
  6497  			start = nanotime()
  6498  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  6499  			before = inittrace
  6500  		}
  6501  
  6502  		firstFunc := add(unsafe.Pointer(t), (3+t.ndeps)*goarch.PtrSize)
  6503  		for i := uintptr(0); i < t.nfns; i++ {
  6504  			p := add(firstFunc, i*goarch.PtrSize)
  6505  			f := *(*func())(unsafe.Pointer(&p))
  6506  			f()
  6507  		}
  6508  
  6509  		if inittrace.active {
  6510  			end := nanotime()
  6511  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  6512  			after := inittrace
  6513  
  6514  			f := *(*func())(unsafe.Pointer(&firstFunc))
  6515  			pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
  6516  
  6517  			var sbuf [24]byte
  6518  			print("init ", pkg, " @")
  6519  			print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
  6520  			print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
  6521  			print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
  6522  			print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
  6523  			print("\n")
  6524  		}
  6525  
  6526  		t.state = 2 // initialization done
  6527  	}
  6528  }
  6529