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  	raceFiniLock mutex
   119  )
   120  
   121  // This slice records the initializing tasks that need to be
   122  // done to start up the runtime. It is built by the linker.
   123  var runtime_inittasks []*initTask
   124  
   125  // main_init_done is a signal used by cgocallbackg that initialization
   126  // has been completed. It is made before _cgo_notify_runtime_init_done,
   127  // so all cgo calls can rely on it existing. When main_init is complete,
   128  // it is closed, meaning cgocallbackg can reliably receive from it.
   129  var main_init_done chan bool
   130  
   131  //go:linkname main_main main.main
   132  func main_main()
   133  
   134  // mainStarted indicates that the main M has started.
   135  var mainStarted bool
   136  
   137  // runtimeInitTime is the nanotime() at which the runtime started.
   138  var runtimeInitTime int64
   139  
   140  // Value to use for signal mask for newly created M's.
   141  var initSigmask sigset
   142  
   143  // The main goroutine.
   144  func main() {
   145  	mp := getg().m
   146  
   147  	// Racectx of m0->g0 is used only as the parent of the main goroutine.
   148  	// It must not be used for anything else.
   149  	mp.g0.racectx = 0
   150  
   151  	// Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
   152  	// Using decimal instead of binary GB and MB because
   153  	// they look nicer in the stack overflow failure message.
   154  	if goarch.PtrSize == 8 {
   155  		maxstacksize = 1000000000
   156  	} else {
   157  		maxstacksize = 250000000
   158  	}
   159  
   160  	// An upper limit for max stack size. Used to avoid random crashes
   161  	// after calling SetMaxStack and trying to allocate a stack that is too big,
   162  	// since stackalloc works with 32-bit sizes.
   163  	maxstackceiling = 2 * maxstacksize
   164  
   165  	// Allow newproc to start new Ms.
   166  	mainStarted = true
   167  
   168  	if GOARCH != "wasm" { // no threads on wasm yet, so no sysmon
   169  		systemstack(func() {
   170  			newm(sysmon, nil, -1)
   171  		})
   172  	}
   173  
   174  	// Lock the main goroutine onto this, the main OS thread,
   175  	// during initialization. Most programs won't care, but a few
   176  	// do require certain calls to be made by the main thread.
   177  	// Those can arrange for main.main to run in the main thread
   178  	// by calling runtime.LockOSThread during initialization
   179  	// to preserve the lock.
   180  	lockOSThread()
   181  
   182  	if mp != &m0 {
   183  		throw("runtime.main not on m0")
   184  	}
   185  
   186  	// Record when the world started.
   187  	// Must be before doInit for tracing init.
   188  	runtimeInitTime = nanotime()
   189  	if runtimeInitTime == 0 {
   190  		throw("nanotime returning zero")
   191  	}
   192  
   193  	if debug.inittrace != 0 {
   194  		inittrace.id = getg().goid
   195  		inittrace.active = true
   196  	}
   197  
   198  	doInit(runtime_inittasks) // Must be before defer.
   199  
   200  	// Defer unlock so that runtime.Goexit during init does the unlock too.
   201  	needUnlock := true
   202  	defer func() {
   203  		if needUnlock {
   204  			unlockOSThread()
   205  		}
   206  	}()
   207  
   208  	gcenable()
   209  
   210  	main_init_done = make(chan bool)
   211  	if iscgo {
   212  		if _cgo_pthread_key_created == nil {
   213  			throw("_cgo_pthread_key_created missing")
   214  		}
   215  
   216  		if _cgo_thread_start == nil {
   217  			throw("_cgo_thread_start missing")
   218  		}
   219  		if GOOS != "windows" {
   220  			if _cgo_setenv == nil {
   221  				throw("_cgo_setenv missing")
   222  			}
   223  			if _cgo_unsetenv == nil {
   224  				throw("_cgo_unsetenv missing")
   225  			}
   226  		}
   227  		if _cgo_notify_runtime_init_done == nil {
   228  			throw("_cgo_notify_runtime_init_done missing")
   229  		}
   230  
   231  		// Set the x_crosscall2_ptr C function pointer variable point to crosscall2.
   232  		if set_crosscall2 == nil {
   233  			throw("set_crosscall2 missing")
   234  		}
   235  		set_crosscall2()
   236  
   237  		// Start the template thread in case we enter Go from
   238  		// a C-created thread and need to create a new thread.
   239  		startTemplateThread()
   240  		cgocall(_cgo_notify_runtime_init_done, nil)
   241  	}
   242  
   243  	// Run the initializing tasks. Depending on build mode this
   244  	// list can arrive a few different ways, but it will always
   245  	// contain the init tasks computed by the linker for all the
   246  	// packages in the program (excluding those added at runtime
   247  	// by package plugin).
   248  	for _, m := range activeModules() {
   249  		doInit(m.inittasks)
   250  	}
   251  
   252  	// Disable init tracing after main init done to avoid overhead
   253  	// of collecting statistics in malloc and newproc
   254  	inittrace.active = false
   255  
   256  	close(main_init_done)
   257  
   258  	needUnlock = false
   259  	unlockOSThread()
   260  
   261  	if isarchive || islibrary {
   262  		// A program compiled with -buildmode=c-archive or c-shared
   263  		// has a main, but it is not executed.
   264  		return
   265  	}
   266  	fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
   267  	fn()
   268  	if raceenabled {
   269  		runExitHooks(0) // run hooks now, since racefini does not return
   270  		racefini()
   271  	}
   272  
   273  	// Make racy client program work: if panicking on
   274  	// another goroutine at the same time as main returns,
   275  	// let the other goroutine finish printing the panic trace.
   276  	// Once it does, it will exit. See issues 3934 and 20018.
   277  	if runningPanicDefers.Load() != 0 {
   278  		// Running deferred functions should not take long.
   279  		for c := 0; c < 1000; c++ {
   280  			if runningPanicDefers.Load() == 0 {
   281  				break
   282  			}
   283  			Gosched()
   284  		}
   285  	}
   286  	if panicking.Load() != 0 {
   287  		gopark(nil, nil, waitReasonPanicWait, traceBlockForever, 1)
   288  	}
   289  	runExitHooks(0)
   290  
   291  	exit(0)
   292  	for {
   293  		var x *int32
   294  		*x = 0
   295  	}
   296  }
   297  
   298  // os_beforeExit is called from os.Exit(0).
   299  //
   300  //go:linkname os_beforeExit os.runtime_beforeExit
   301  func os_beforeExit(exitCode int) {
   302  	runExitHooks(exitCode)
   303  	if exitCode == 0 && raceenabled {
   304  		racefini()
   305  	}
   306  }
   307  
   308  // start forcegc helper goroutine
   309  func init() {
   310  	go forcegchelper()
   311  }
   312  
   313  func forcegchelper() {
   314  	forcegc.g = getg()
   315  	lockInit(&forcegc.lock, lockRankForcegc)
   316  	for {
   317  		lock(&forcegc.lock)
   318  		if forcegc.idle.Load() {
   319  			throw("forcegc: phase error")
   320  		}
   321  		forcegc.idle.Store(true)
   322  		goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceBlockSystemGoroutine, 1)
   323  		// this goroutine is explicitly resumed by sysmon
   324  		if debug.gctrace > 0 {
   325  			println("GC forced")
   326  		}
   327  		// Time-triggered, fully concurrent.
   328  		gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
   329  	}
   330  }
   331  
   332  // Gosched yields the processor, allowing other goroutines to run. It does not
   333  // suspend the current goroutine, so execution resumes automatically.
   334  //
   335  //go:nosplit
   336  func Gosched() {
   337  	checkTimeouts()
   338  	mcall(gosched_m)
   339  }
   340  
   341  // goschedguarded yields the processor like gosched, but also checks
   342  // for forbidden states and opts out of the yield in those cases.
   343  //
   344  //go:nosplit
   345  func goschedguarded() {
   346  	mcall(goschedguarded_m)
   347  }
   348  
   349  // goschedIfBusy yields the processor like gosched, but only does so if
   350  // there are no idle Ps or if we're on the only P and there's nothing in
   351  // the run queue. In both cases, there is freely available idle time.
   352  //
   353  //go:nosplit
   354  func goschedIfBusy() {
   355  	gp := getg()
   356  	// Call gosched if gp.preempt is set; we may be in a tight loop that
   357  	// doesn't otherwise yield.
   358  	if !gp.preempt && sched.npidle.Load() > 0 {
   359  		return
   360  	}
   361  	mcall(gosched_m)
   362  }
   363  
   364  // Puts the current goroutine into a waiting state and calls unlockf on the
   365  // system stack.
   366  //
   367  // If unlockf returns false, the goroutine is resumed.
   368  //
   369  // unlockf must not access this G's stack, as it may be moved between
   370  // the call to gopark and the call to unlockf.
   371  //
   372  // Note that because unlockf is called after putting the G into a waiting
   373  // state, the G may have already been readied by the time unlockf is called
   374  // unless there is external synchronization preventing the G from being
   375  // readied. If unlockf returns false, it must guarantee that the G cannot be
   376  // externally readied.
   377  //
   378  // Reason explains why the goroutine has been parked. It is displayed in stack
   379  // traces and heap dumps. Reasons should be unique and descriptive. Do not
   380  // re-use reasons, add new ones.
   381  func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceReason traceBlockReason, traceskip int) {
   382  	if reason != waitReasonSleep {
   383  		checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
   384  	}
   385  	mp := acquirem()
   386  	gp := mp.curg
   387  	status := readgstatus(gp)
   388  	if status != _Grunning && status != _Gscanrunning {
   389  		throw("gopark: bad g status")
   390  	}
   391  	mp.waitlock = lock
   392  	mp.waitunlockf = unlockf
   393  	gp.waitreason = reason
   394  	mp.waitTraceBlockReason = traceReason
   395  	mp.waitTraceSkip = traceskip
   396  	releasem(mp)
   397  	// can't do anything that might move the G between Ms here.
   398  	mcall(park_m)
   399  }
   400  
   401  // Puts the current goroutine into a waiting state and unlocks the lock.
   402  // The goroutine can be made runnable again by calling goready(gp).
   403  func goparkunlock(lock *mutex, reason waitReason, traceReason traceBlockReason, traceskip int) {
   404  	gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceReason, traceskip)
   405  }
   406  
   407  func goready(gp *g, traceskip int) {
   408  	systemstack(func() {
   409  		ready(gp, traceskip, true)
   410  	})
   411  }
   412  
   413  //go:nosplit
   414  func acquireSudog() *sudog {
   415  	// Delicate dance: the semaphore implementation calls
   416  	// acquireSudog, acquireSudog calls new(sudog),
   417  	// new calls malloc, malloc can call the garbage collector,
   418  	// and the garbage collector calls the semaphore implementation
   419  	// in stopTheWorld.
   420  	// Break the cycle by doing acquirem/releasem around new(sudog).
   421  	// The acquirem/releasem increments m.locks during new(sudog),
   422  	// which keeps the garbage collector from being invoked.
   423  	mp := acquirem()
   424  	pp := mp.p.ptr()
   425  	if len(pp.sudogcache) == 0 {
   426  		lock(&sched.sudoglock)
   427  		// First, try to grab a batch from central cache.
   428  		for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
   429  			s := sched.sudogcache
   430  			sched.sudogcache = s.next
   431  			s.next = nil
   432  			pp.sudogcache = append(pp.sudogcache, s)
   433  		}
   434  		unlock(&sched.sudoglock)
   435  		// If the central cache is empty, allocate a new one.
   436  		if len(pp.sudogcache) == 0 {
   437  			pp.sudogcache = append(pp.sudogcache, new(sudog))
   438  		}
   439  	}
   440  	n := len(pp.sudogcache)
   441  	s := pp.sudogcache[n-1]
   442  	pp.sudogcache[n-1] = nil
   443  	pp.sudogcache = pp.sudogcache[:n-1]
   444  	if s.elem != nil {
   445  		throw("acquireSudog: found s.elem != nil in cache")
   446  	}
   447  	releasem(mp)
   448  	return s
   449  }
   450  
   451  //go:nosplit
   452  func releaseSudog(s *sudog) {
   453  	if s.elem != nil {
   454  		throw("runtime: sudog with non-nil elem")
   455  	}
   456  	if s.isSelect {
   457  		throw("runtime: sudog with non-false isSelect")
   458  	}
   459  	if s.next != nil {
   460  		throw("runtime: sudog with non-nil next")
   461  	}
   462  	if s.prev != nil {
   463  		throw("runtime: sudog with non-nil prev")
   464  	}
   465  	if s.waitlink != nil {
   466  		throw("runtime: sudog with non-nil waitlink")
   467  	}
   468  	if s.c != nil {
   469  		throw("runtime: sudog with non-nil c")
   470  	}
   471  	gp := getg()
   472  	if gp.param != nil {
   473  		throw("runtime: releaseSudog with non-nil gp.param")
   474  	}
   475  	mp := acquirem() // avoid rescheduling to another P
   476  	pp := mp.p.ptr()
   477  	if len(pp.sudogcache) == cap(pp.sudogcache) {
   478  		// Transfer half of local cache to the central cache.
   479  		var first, last *sudog
   480  		for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
   481  			n := len(pp.sudogcache)
   482  			p := pp.sudogcache[n-1]
   483  			pp.sudogcache[n-1] = nil
   484  			pp.sudogcache = pp.sudogcache[:n-1]
   485  			if first == nil {
   486  				first = p
   487  			} else {
   488  				last.next = p
   489  			}
   490  			last = p
   491  		}
   492  		lock(&sched.sudoglock)
   493  		last.next = sched.sudogcache
   494  		sched.sudogcache = first
   495  		unlock(&sched.sudoglock)
   496  	}
   497  	pp.sudogcache = append(pp.sudogcache, s)
   498  	releasem(mp)
   499  }
   500  
   501  // called from assembly.
   502  func badmcall(fn func(*g)) {
   503  	throw("runtime: mcall called on m->g0 stack")
   504  }
   505  
   506  func badmcall2(fn func(*g)) {
   507  	throw("runtime: mcall function returned")
   508  }
   509  
   510  func badreflectcall() {
   511  	panic(plainError("arg size to reflect.call more than 1GB"))
   512  }
   513  
   514  //go:nosplit
   515  //go:nowritebarrierrec
   516  func badmorestackg0() {
   517  	writeErrStr("fatal: morestack on g0\n")
   518  }
   519  
   520  //go:nosplit
   521  //go:nowritebarrierrec
   522  func badmorestackgsignal() {
   523  	writeErrStr("fatal: morestack on gsignal\n")
   524  }
   525  
   526  //go:nosplit
   527  func badctxt() {
   528  	throw("ctxt != 0")
   529  }
   530  
   531  func lockedOSThread() bool {
   532  	gp := getg()
   533  	return gp.lockedm != 0 && gp.m.lockedg != 0
   534  }
   535  
   536  var (
   537  	// allgs contains all Gs ever created (including dead Gs), and thus
   538  	// never shrinks.
   539  	//
   540  	// Access via the slice is protected by allglock or stop-the-world.
   541  	// Readers that cannot take the lock may (carefully!) use the atomic
   542  	// variables below.
   543  	allglock mutex
   544  	allgs    []*g
   545  
   546  	// allglen and allgptr are atomic variables that contain len(allgs) and
   547  	// &allgs[0] respectively. Proper ordering depends on totally-ordered
   548  	// loads and stores. Writes are protected by allglock.
   549  	//
   550  	// allgptr is updated before allglen. Readers should read allglen
   551  	// before allgptr to ensure that allglen is always <= len(allgptr). New
   552  	// Gs appended during the race can be missed. For a consistent view of
   553  	// all Gs, allglock must be held.
   554  	//
   555  	// allgptr copies should always be stored as a concrete type or
   556  	// unsafe.Pointer, not uintptr, to ensure that GC can still reach it
   557  	// even if it points to a stale array.
   558  	allglen uintptr
   559  	allgptr **g
   560  )
   561  
   562  func allgadd(gp *g) {
   563  	if readgstatus(gp) == _Gidle {
   564  		throw("allgadd: bad status Gidle")
   565  	}
   566  
   567  	lock(&allglock)
   568  	allgs = append(allgs, gp)
   569  	if &allgs[0] != allgptr {
   570  		atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
   571  	}
   572  	atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
   573  	unlock(&allglock)
   574  }
   575  
   576  // allGsSnapshot returns a snapshot of the slice of all Gs.
   577  //
   578  // The world must be stopped or allglock must be held.
   579  func allGsSnapshot() []*g {
   580  	assertWorldStoppedOrLockHeld(&allglock)
   581  
   582  	// Because the world is stopped or allglock is held, allgadd
   583  	// cannot happen concurrently with this. allgs grows
   584  	// monotonically and existing entries never change, so we can
   585  	// simply return a copy of the slice header. For added safety,
   586  	// we trim everything past len because that can still change.
   587  	return allgs[:len(allgs):len(allgs)]
   588  }
   589  
   590  // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
   591  func atomicAllG() (**g, uintptr) {
   592  	length := atomic.Loaduintptr(&allglen)
   593  	ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
   594  	return ptr, length
   595  }
   596  
   597  // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
   598  func atomicAllGIndex(ptr **g, i uintptr) *g {
   599  	return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
   600  }
   601  
   602  // forEachG calls fn on every G from allgs.
   603  //
   604  // forEachG takes a lock to exclude concurrent addition of new Gs.
   605  func forEachG(fn func(gp *g)) {
   606  	lock(&allglock)
   607  	for _, gp := range allgs {
   608  		fn(gp)
   609  	}
   610  	unlock(&allglock)
   611  }
   612  
   613  // forEachGRace calls fn on every G from allgs.
   614  //
   615  // forEachGRace avoids locking, but does not exclude addition of new Gs during
   616  // execution, which may be missed.
   617  func forEachGRace(fn func(gp *g)) {
   618  	ptr, length := atomicAllG()
   619  	for i := uintptr(0); i < length; i++ {
   620  		gp := atomicAllGIndex(ptr, i)
   621  		fn(gp)
   622  	}
   623  	return
   624  }
   625  
   626  const (
   627  	// Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
   628  	// 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
   629  	_GoidCacheBatch = 16
   630  )
   631  
   632  // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
   633  // value of the GODEBUG environment variable.
   634  func cpuinit(env string) {
   635  	switch GOOS {
   636  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   637  		cpu.DebugOptions = true
   638  	}
   639  	cpu.Initialize(env)
   640  
   641  	// Support cpu feature variables are used in code generated by the compiler
   642  	// to guard execution of instructions that can not be assumed to be always supported.
   643  	switch GOARCH {
   644  	case "386", "amd64":
   645  		x86HasPOPCNT = cpu.X86.HasPOPCNT
   646  		x86HasSSE41 = cpu.X86.HasSSE41
   647  		x86HasFMA = cpu.X86.HasFMA
   648  
   649  	case "arm":
   650  		armHasVFPv4 = cpu.ARM.HasVFPv4
   651  
   652  	case "arm64":
   653  		arm64HasATOMICS = cpu.ARM64.HasATOMICS
   654  	}
   655  }
   656  
   657  // getGodebugEarly extracts the environment variable GODEBUG from the environment on
   658  // Unix-like operating systems and returns it. This function exists to extract GODEBUG
   659  // early before much of the runtime is initialized.
   660  func getGodebugEarly() string {
   661  	const prefix = "GODEBUG="
   662  	var env string
   663  	switch GOOS {
   664  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   665  		// Similar to goenv_unix but extracts the environment value for
   666  		// GODEBUG directly.
   667  		// TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
   668  		n := int32(0)
   669  		for argv_index(argv, argc+1+n) != nil {
   670  			n++
   671  		}
   672  
   673  		for i := int32(0); i < n; i++ {
   674  			p := argv_index(argv, argc+1+i)
   675  			s := unsafe.String(p, findnull(p))
   676  
   677  			if hasPrefix(s, prefix) {
   678  				env = gostring(p)[len(prefix):]
   679  				break
   680  			}
   681  		}
   682  	}
   683  	return env
   684  }
   685  
   686  // The bootstrap sequence is:
   687  //
   688  //	call osinit
   689  //	call schedinit
   690  //	make & queue new G
   691  //	call runtime·mstart
   692  //
   693  // The new G calls runtime·main.
   694  func schedinit() {
   695  	lockInit(&sched.lock, lockRankSched)
   696  	lockInit(&sched.sysmonlock, lockRankSysmon)
   697  	lockInit(&sched.deferlock, lockRankDefer)
   698  	lockInit(&sched.sudoglock, lockRankSudog)
   699  	lockInit(&deadlock, lockRankDeadlock)
   700  	lockInit(&paniclk, lockRankPanic)
   701  	lockInit(&allglock, lockRankAllg)
   702  	lockInit(&allpLock, lockRankAllp)
   703  	lockInit(&reflectOffs.lock, lockRankReflectOffs)
   704  	lockInit(&finlock, lockRankFin)
   705  	lockInit(&cpuprof.lock, lockRankCpuprof)
   706  	traceLockInit()
   707  	// Enforce that this lock is always a leaf lock.
   708  	// All of this lock's critical sections should be
   709  	// extremely short.
   710  	lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
   711  
   712  	// raceinit must be the first call to race detector.
   713  	// In particular, it must be done before mallocinit below calls racemapshadow.
   714  	gp := getg()
   715  	if raceenabled {
   716  		gp.racectx, raceprocctx0 = raceinit()
   717  	}
   718  
   719  	sched.maxmcount = 10000
   720  
   721  	// The world starts stopped.
   722  	worldStopped()
   723  
   724  	moduledataverify()
   725  	stackinit()
   726  	mallocinit()
   727  	godebug := getGodebugEarly()
   728  	initPageTrace(godebug) // must run after mallocinit but before anything allocates
   729  	cpuinit(godebug)       // must run before alginit
   730  	alginit()              // maps, hash, fastrand must not be used before this call
   731  	fastrandinit()         // must run before mcommoninit
   732  	mcommoninit(gp.m, -1)
   733  	modulesinit()   // provides activeModules
   734  	typelinksinit() // uses maps, activeModules
   735  	itabsinit()     // uses activeModules
   736  	stkobjinit()    // must run before GC starts
   737  
   738  	sigsave(&gp.m.sigmask)
   739  	initSigmask = gp.m.sigmask
   740  
   741  	goargs()
   742  	goenvs()
   743  	secure()
   744  	parsedebugvars()
   745  	gcinit()
   746  
   747  	// if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
   748  	// Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
   749  	// set to true by the linker, it means that nothing is consuming the profile, it is
   750  	// safe to set MemProfileRate to 0.
   751  	if disableMemoryProfiling {
   752  		MemProfileRate = 0
   753  	}
   754  
   755  	lock(&sched.lock)
   756  	sched.lastpoll.Store(nanotime())
   757  	procs := ncpu
   758  	if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
   759  		procs = n
   760  	}
   761  	if procresize(procs) != nil {
   762  		throw("unknown runnable goroutine during bootstrap")
   763  	}
   764  	unlock(&sched.lock)
   765  
   766  	// World is effectively started now, as P's can run.
   767  	worldStarted()
   768  
   769  	if buildVersion == "" {
   770  		// Condition should never trigger. This code just serves
   771  		// to ensure runtime·buildVersion is kept in the resulting binary.
   772  		buildVersion = "unknown"
   773  	}
   774  	if len(modinfo) == 1 {
   775  		// Condition should never trigger. This code just serves
   776  		// to ensure runtime·modinfo is kept in the resulting binary.
   777  		modinfo = ""
   778  	}
   779  }
   780  
   781  func dumpgstatus(gp *g) {
   782  	thisg := getg()
   783  	print("runtime:   gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   784  	print("runtime: getg:  g=", thisg, ", goid=", thisg.goid, ",  g->atomicstatus=", readgstatus(thisg), "\n")
   785  }
   786  
   787  // sched.lock must be held.
   788  func checkmcount() {
   789  	assertLockHeld(&sched.lock)
   790  
   791  	// Exclude extra M's, which are used for cgocallback from threads
   792  	// created in C.
   793  	//
   794  	// The purpose of the SetMaxThreads limit is to avoid accidental fork
   795  	// bomb from something like millions of goroutines blocking on system
   796  	// calls, causing the runtime to create millions of threads. By
   797  	// definition, this isn't a problem for threads created in C, so we
   798  	// exclude them from the limit. See https://go.dev/issue/60004.
   799  	count := mcount() - int32(extraMInUse.Load()) - int32(extraMLength.Load())
   800  	if count > sched.maxmcount {
   801  		print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
   802  		throw("thread exhaustion")
   803  	}
   804  }
   805  
   806  // mReserveID returns the next ID to use for a new m. This new m is immediately
   807  // considered 'running' by checkdead.
   808  //
   809  // sched.lock must be held.
   810  func mReserveID() int64 {
   811  	assertLockHeld(&sched.lock)
   812  
   813  	if sched.mnext+1 < sched.mnext {
   814  		throw("runtime: thread ID overflow")
   815  	}
   816  	id := sched.mnext
   817  	sched.mnext++
   818  	checkmcount()
   819  	return id
   820  }
   821  
   822  // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
   823  func mcommoninit(mp *m, id int64) {
   824  	gp := getg()
   825  
   826  	// g0 stack won't make sense for user (and is not necessary unwindable).
   827  	if gp != gp.m.g0 {
   828  		callers(1, mp.createstack[:])
   829  	}
   830  
   831  	lock(&sched.lock)
   832  
   833  	if id >= 0 {
   834  		mp.id = id
   835  	} else {
   836  		mp.id = mReserveID()
   837  	}
   838  
   839  	lo := uint32(int64Hash(uint64(mp.id), fastrandseed))
   840  	hi := uint32(int64Hash(uint64(cputicks()), ^fastrandseed))
   841  	if lo|hi == 0 {
   842  		hi = 1
   843  	}
   844  	// Same behavior as for 1.17.
   845  	// TODO: Simplify this.
   846  	if goarch.BigEndian {
   847  		mp.fastrand = uint64(lo)<<32 | uint64(hi)
   848  	} else {
   849  		mp.fastrand = uint64(hi)<<32 | uint64(lo)
   850  	}
   851  
   852  	mpreinit(mp)
   853  	if mp.gsignal != nil {
   854  		mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
   855  	}
   856  
   857  	// Add to allm so garbage collector doesn't free g->m
   858  	// when it is just in a register or thread-local storage.
   859  	mp.alllink = allm
   860  
   861  	// NumCgoCall() iterates over allm w/o schedlock,
   862  	// so we need to publish it safely.
   863  	atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
   864  	unlock(&sched.lock)
   865  
   866  	// Allocate memory to hold a cgo traceback if the cgo call crashes.
   867  	if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
   868  		mp.cgoCallers = new(cgoCallers)
   869  	}
   870  }
   871  
   872  func (mp *m) becomeSpinning() {
   873  	mp.spinning = true
   874  	sched.nmspinning.Add(1)
   875  	sched.needspinning.Store(0)
   876  }
   877  
   878  func (mp *m) hasCgoOnStack() bool {
   879  	return mp.ncgo > 0 || mp.isextra
   880  }
   881  
   882  var fastrandseed uintptr
   883  
   884  func fastrandinit() {
   885  	s := (*[unsafe.Sizeof(fastrandseed)]byte)(unsafe.Pointer(&fastrandseed))[:]
   886  	getRandomData(s)
   887  }
   888  
   889  // Mark gp ready to run.
   890  func ready(gp *g, traceskip int, next bool) {
   891  	if traceEnabled() {
   892  		traceGoUnpark(gp, traceskip)
   893  	}
   894  
   895  	status := readgstatus(gp)
   896  
   897  	// Mark runnable.
   898  	mp := acquirem() // disable preemption because it can be holding p in a local var
   899  	if status&^_Gscan != _Gwaiting {
   900  		dumpgstatus(gp)
   901  		throw("bad g->status in ready")
   902  	}
   903  
   904  	// status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
   905  	casgstatus(gp, _Gwaiting, _Grunnable)
   906  	runqput(mp.p.ptr(), gp, next)
   907  	wakep()
   908  	releasem(mp)
   909  }
   910  
   911  // freezeStopWait is a large value that freezetheworld sets
   912  // sched.stopwait to in order to request that all Gs permanently stop.
   913  const freezeStopWait = 0x7fffffff
   914  
   915  // freezing is set to non-zero if the runtime is trying to freeze the
   916  // world.
   917  var freezing atomic.Bool
   918  
   919  // Similar to stopTheWorld but best-effort and can be called several times.
   920  // There is no reverse operation, used during crashing.
   921  // This function must not lock any mutexes.
   922  func freezetheworld() {
   923  	freezing.Store(true)
   924  	if debug.dontfreezetheworld > 0 {
   925  		// Don't prempt Ps to stop goroutines. That will perturb
   926  		// scheduler state, making debugging more difficult. Instead,
   927  		// allow goroutines to continue execution.
   928  		//
   929  		// fatalpanic will tracebackothers to trace all goroutines. It
   930  		// is unsafe to trace a running goroutine, so tracebackothers
   931  		// will skip running goroutines. That is OK and expected, we
   932  		// expect users of dontfreezetheworld to use core files anyway.
   933  		//
   934  		// However, allowing the scheduler to continue running free
   935  		// introduces a race: a goroutine may be stopped when
   936  		// tracebackothers checks its status, and then start running
   937  		// later when we are in the middle of traceback, potentially
   938  		// causing a crash.
   939  		//
   940  		// To mitigate this, when an M naturally enters the scheduler,
   941  		// schedule checks if freezing is set and if so stops
   942  		// execution. This guarantees that while Gs can transition from
   943  		// running to stopped, they can never transition from stopped
   944  		// to running.
   945  		//
   946  		// The sleep here allows racing Ms that missed freezing and are
   947  		// about to run a G to complete the transition to running
   948  		// before we start traceback.
   949  		usleep(1000)
   950  		return
   951  	}
   952  
   953  	// stopwait and preemption requests can be lost
   954  	// due to races with concurrently executing threads,
   955  	// so try several times
   956  	for i := 0; i < 5; i++ {
   957  		// this should tell the scheduler to not start any new goroutines
   958  		sched.stopwait = freezeStopWait
   959  		sched.gcwaiting.Store(true)
   960  		// this should stop running goroutines
   961  		if !preemptall() {
   962  			break // no running goroutines
   963  		}
   964  		usleep(1000)
   965  	}
   966  	// to be sure
   967  	usleep(1000)
   968  	preemptall()
   969  	usleep(1000)
   970  }
   971  
   972  // All reads and writes of g's status go through readgstatus, casgstatus
   973  // castogscanstatus, casfrom_Gscanstatus.
   974  //
   975  //go:nosplit
   976  func readgstatus(gp *g) uint32 {
   977  	return gp.atomicstatus.Load()
   978  }
   979  
   980  // The Gscanstatuses are acting like locks and this releases them.
   981  // If it proves to be a performance hit we should be able to make these
   982  // simple atomic stores but for now we are going to throw if
   983  // we see an inconsistent state.
   984  func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
   985  	success := false
   986  
   987  	// Check that transition is valid.
   988  	switch oldval {
   989  	default:
   990  		print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
   991  		dumpgstatus(gp)
   992  		throw("casfrom_Gscanstatus:top gp->status is not in scan state")
   993  	case _Gscanrunnable,
   994  		_Gscanwaiting,
   995  		_Gscanrunning,
   996  		_Gscansyscall,
   997  		_Gscanpreempted:
   998  		if newval == oldval&^_Gscan {
   999  			success = gp.atomicstatus.CompareAndSwap(oldval, newval)
  1000  		}
  1001  	}
  1002  	if !success {
  1003  		print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
  1004  		dumpgstatus(gp)
  1005  		throw("casfrom_Gscanstatus: gp->status is not in scan state")
  1006  	}
  1007  	releaseLockRank(lockRankGscan)
  1008  }
  1009  
  1010  // This will return false if the gp is not in the expected status and the cas fails.
  1011  // This acts like a lock acquire while the casfromgstatus acts like a lock release.
  1012  func castogscanstatus(gp *g, oldval, newval uint32) bool {
  1013  	switch oldval {
  1014  	case _Grunnable,
  1015  		_Grunning,
  1016  		_Gwaiting,
  1017  		_Gsyscall:
  1018  		if newval == oldval|_Gscan {
  1019  			r := gp.atomicstatus.CompareAndSwap(oldval, newval)
  1020  			if r {
  1021  				acquireLockRank(lockRankGscan)
  1022  			}
  1023  			return r
  1024  
  1025  		}
  1026  	}
  1027  	print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1028  	throw("castogscanstatus")
  1029  	panic("not reached")
  1030  }
  1031  
  1032  // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
  1033  // various latencies on every transition instead of sampling them.
  1034  var casgstatusAlwaysTrack = false
  1035  
  1036  // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
  1037  // and casfrom_Gscanstatus instead.
  1038  // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
  1039  // put it in the Gscan state is finished.
  1040  //
  1041  //go:nosplit
  1042  func casgstatus(gp *g, oldval, newval uint32) {
  1043  	if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
  1044  		systemstack(func() {
  1045  			print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1046  			throw("casgstatus: bad incoming values")
  1047  		})
  1048  	}
  1049  
  1050  	acquireLockRank(lockRankGscan)
  1051  	releaseLockRank(lockRankGscan)
  1052  
  1053  	// See https://golang.org/cl/21503 for justification of the yield delay.
  1054  	const yieldDelay = 5 * 1000
  1055  	var nextYield int64
  1056  
  1057  	// loop if gp->atomicstatus is in a scan state giving
  1058  	// GC time to finish and change the state to oldval.
  1059  	for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
  1060  		if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
  1061  			throw("casgstatus: waiting for Gwaiting but is Grunnable")
  1062  		}
  1063  		if i == 0 {
  1064  			nextYield = nanotime() + yieldDelay
  1065  		}
  1066  		if nanotime() < nextYield {
  1067  			for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
  1068  				procyield(1)
  1069  			}
  1070  		} else {
  1071  			osyield()
  1072  			nextYield = nanotime() + yieldDelay/2
  1073  		}
  1074  	}
  1075  
  1076  	if oldval == _Grunning {
  1077  		// Track every gTrackingPeriod time a goroutine transitions out of running.
  1078  		if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
  1079  			gp.tracking = true
  1080  		}
  1081  		gp.trackingSeq++
  1082  	}
  1083  	if !gp.tracking {
  1084  		return
  1085  	}
  1086  
  1087  	// Handle various kinds of tracking.
  1088  	//
  1089  	// Currently:
  1090  	// - Time spent in runnable.
  1091  	// - Time spent blocked on a sync.Mutex or sync.RWMutex.
  1092  	switch oldval {
  1093  	case _Grunnable:
  1094  		// We transitioned out of runnable, so measure how much
  1095  		// time we spent in this state and add it to
  1096  		// runnableTime.
  1097  		now := nanotime()
  1098  		gp.runnableTime += now - gp.trackingStamp
  1099  		gp.trackingStamp = 0
  1100  	case _Gwaiting:
  1101  		if !gp.waitreason.isMutexWait() {
  1102  			// Not blocking on a lock.
  1103  			break
  1104  		}
  1105  		// Blocking on a lock, measure it. Note that because we're
  1106  		// sampling, we have to multiply by our sampling period to get
  1107  		// a more representative estimate of the absolute value.
  1108  		// gTrackingPeriod also represents an accurate sampling period
  1109  		// because we can only enter this state from _Grunning.
  1110  		now := nanotime()
  1111  		sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
  1112  		gp.trackingStamp = 0
  1113  	}
  1114  	switch newval {
  1115  	case _Gwaiting:
  1116  		if !gp.waitreason.isMutexWait() {
  1117  			// Not blocking on a lock.
  1118  			break
  1119  		}
  1120  		// Blocking on a lock. Write down the timestamp.
  1121  		now := nanotime()
  1122  		gp.trackingStamp = now
  1123  	case _Grunnable:
  1124  		// We just transitioned into runnable, so record what
  1125  		// time that happened.
  1126  		now := nanotime()
  1127  		gp.trackingStamp = now
  1128  	case _Grunning:
  1129  		// We're transitioning into running, so turn off
  1130  		// tracking and record how much time we spent in
  1131  		// runnable.
  1132  		gp.tracking = false
  1133  		sched.timeToRun.record(gp.runnableTime)
  1134  		gp.runnableTime = 0
  1135  	}
  1136  }
  1137  
  1138  // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
  1139  //
  1140  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1141  func casGToWaiting(gp *g, old uint32, reason waitReason) {
  1142  	// Set the wait reason before calling casgstatus, because casgstatus will use it.
  1143  	gp.waitreason = reason
  1144  	casgstatus(gp, old, _Gwaiting)
  1145  }
  1146  
  1147  // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
  1148  // Returns old status. Cannot call casgstatus directly, because we are racing with an
  1149  // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
  1150  // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
  1151  // it would loop waiting for the status to go back to Gwaiting, which it never will.
  1152  //
  1153  //go:nosplit
  1154  func casgcopystack(gp *g) uint32 {
  1155  	for {
  1156  		oldstatus := readgstatus(gp) &^ _Gscan
  1157  		if oldstatus != _Gwaiting && oldstatus != _Grunnable {
  1158  			throw("copystack: bad status, not Gwaiting or Grunnable")
  1159  		}
  1160  		if gp.atomicstatus.CompareAndSwap(oldstatus, _Gcopystack) {
  1161  			return oldstatus
  1162  		}
  1163  	}
  1164  }
  1165  
  1166  // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
  1167  //
  1168  // TODO(austin): This is the only status operation that both changes
  1169  // the status and locks the _Gscan bit. Rethink this.
  1170  func casGToPreemptScan(gp *g, old, new uint32) {
  1171  	if old != _Grunning || new != _Gscan|_Gpreempted {
  1172  		throw("bad g transition")
  1173  	}
  1174  	acquireLockRank(lockRankGscan)
  1175  	for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
  1176  	}
  1177  }
  1178  
  1179  // casGFromPreempted attempts to transition gp from _Gpreempted to
  1180  // _Gwaiting. If successful, the caller is responsible for
  1181  // re-scheduling gp.
  1182  func casGFromPreempted(gp *g, old, new uint32) bool {
  1183  	if old != _Gpreempted || new != _Gwaiting {
  1184  		throw("bad g transition")
  1185  	}
  1186  	gp.waitreason = waitReasonPreempted
  1187  	return gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting)
  1188  }
  1189  
  1190  // stwReason is an enumeration of reasons the world is stopping.
  1191  type stwReason uint8
  1192  
  1193  // Reasons to stop-the-world.
  1194  //
  1195  // Avoid reusing reasons and add new ones instead.
  1196  const (
  1197  	stwUnknown                     stwReason = iota // "unknown"
  1198  	stwGCMarkTerm                                   // "GC mark termination"
  1199  	stwGCSweepTerm                                  // "GC sweep termination"
  1200  	stwWriteHeapDump                                // "write heap dump"
  1201  	stwGoroutineProfile                             // "goroutine profile"
  1202  	stwGoroutineProfileCleanup                      // "goroutine profile cleanup"
  1203  	stwAllGoroutinesStack                           // "all goroutines stack trace"
  1204  	stwReadMemStats                                 // "read mem stats"
  1205  	stwAllThreadsSyscall                            // "AllThreadsSyscall"
  1206  	stwGOMAXPROCS                                   // "GOMAXPROCS"
  1207  	stwStartTrace                                   // "start trace"
  1208  	stwStopTrace                                    // "stop trace"
  1209  	stwForTestCountPagesInUse                       // "CountPagesInUse (test)"
  1210  	stwForTestReadMetricsSlow                       // "ReadMetricsSlow (test)"
  1211  	stwForTestReadMemStatsSlow                      // "ReadMemStatsSlow (test)"
  1212  	stwForTestPageCachePagesLeaked                  // "PageCachePagesLeaked (test)"
  1213  	stwForTestResetDebugLog                         // "ResetDebugLog (test)"
  1214  )
  1215  
  1216  func (r stwReason) String() string {
  1217  	return stwReasonStrings[r]
  1218  }
  1219  
  1220  // If you add to this list, also add it to src/internal/trace/parser.go.
  1221  // If you change the values of any of the stw* constants, bump the trace
  1222  // version number and make a copy of this.
  1223  var stwReasonStrings = [...]string{
  1224  	stwUnknown:                     "unknown",
  1225  	stwGCMarkTerm:                  "GC mark termination",
  1226  	stwGCSweepTerm:                 "GC sweep termination",
  1227  	stwWriteHeapDump:               "write heap dump",
  1228  	stwGoroutineProfile:            "goroutine profile",
  1229  	stwGoroutineProfileCleanup:     "goroutine profile cleanup",
  1230  	stwAllGoroutinesStack:          "all goroutines stack trace",
  1231  	stwReadMemStats:                "read mem stats",
  1232  	stwAllThreadsSyscall:           "AllThreadsSyscall",
  1233  	stwGOMAXPROCS:                  "GOMAXPROCS",
  1234  	stwStartTrace:                  "start trace",
  1235  	stwStopTrace:                   "stop trace",
  1236  	stwForTestCountPagesInUse:      "CountPagesInUse (test)",
  1237  	stwForTestReadMetricsSlow:      "ReadMetricsSlow (test)",
  1238  	stwForTestReadMemStatsSlow:     "ReadMemStatsSlow (test)",
  1239  	stwForTestPageCachePagesLeaked: "PageCachePagesLeaked (test)",
  1240  	stwForTestResetDebugLog:        "ResetDebugLog (test)",
  1241  }
  1242  
  1243  // stopTheWorld stops all P's from executing goroutines, interrupting
  1244  // all goroutines at GC safe points and records reason as the reason
  1245  // for the stop. On return, only the current goroutine's P is running.
  1246  // stopTheWorld must not be called from a system stack and the caller
  1247  // must not hold worldsema. The caller must call startTheWorld when
  1248  // other P's should resume execution.
  1249  //
  1250  // stopTheWorld is safe for multiple goroutines to call at the
  1251  // same time. Each will execute its own stop, and the stops will
  1252  // be serialized.
  1253  //
  1254  // This is also used by routines that do stack dumps. If the system is
  1255  // in panic or being exited, this may not reliably stop all
  1256  // goroutines.
  1257  func stopTheWorld(reason stwReason) {
  1258  	semacquire(&worldsema)
  1259  	gp := getg()
  1260  	gp.m.preemptoff = reason.String()
  1261  	systemstack(func() {
  1262  		// Mark the goroutine which called stopTheWorld preemptible so its
  1263  		// stack may be scanned.
  1264  		// This lets a mark worker scan us while we try to stop the world
  1265  		// since otherwise we could get in a mutual preemption deadlock.
  1266  		// We must not modify anything on the G stack because a stack shrink
  1267  		// may occur. A stack shrink is otherwise OK though because in order
  1268  		// to return from this function (and to leave the system stack) we
  1269  		// must have preempted all goroutines, including any attempting
  1270  		// to scan our stack, in which case, any stack shrinking will
  1271  		// have already completed by the time we exit.
  1272  		// Don't provide a wait reason because we're still executing.
  1273  		casGToWaiting(gp, _Grunning, waitReasonStoppingTheWorld)
  1274  		stopTheWorldWithSema(reason)
  1275  		casgstatus(gp, _Gwaiting, _Grunning)
  1276  	})
  1277  }
  1278  
  1279  // startTheWorld undoes the effects of stopTheWorld.
  1280  func startTheWorld() {
  1281  	systemstack(func() { startTheWorldWithSema() })
  1282  
  1283  	// worldsema must be held over startTheWorldWithSema to ensure
  1284  	// gomaxprocs cannot change while worldsema is held.
  1285  	//
  1286  	// Release worldsema with direct handoff to the next waiter, but
  1287  	// acquirem so that semrelease1 doesn't try to yield our time.
  1288  	//
  1289  	// Otherwise if e.g. ReadMemStats is being called in a loop,
  1290  	// it might stomp on other attempts to stop the world, such as
  1291  	// for starting or ending GC. The operation this blocks is
  1292  	// so heavy-weight that we should just try to be as fair as
  1293  	// possible here.
  1294  	//
  1295  	// We don't want to just allow us to get preempted between now
  1296  	// and releasing the semaphore because then we keep everyone
  1297  	// (including, for example, GCs) waiting longer.
  1298  	mp := acquirem()
  1299  	mp.preemptoff = ""
  1300  	semrelease1(&worldsema, true, 0)
  1301  	releasem(mp)
  1302  }
  1303  
  1304  // stopTheWorldGC has the same effect as stopTheWorld, but blocks
  1305  // until the GC is not running. It also blocks a GC from starting
  1306  // until startTheWorldGC is called.
  1307  func stopTheWorldGC(reason stwReason) {
  1308  	semacquire(&gcsema)
  1309  	stopTheWorld(reason)
  1310  }
  1311  
  1312  // startTheWorldGC undoes the effects of stopTheWorldGC.
  1313  func startTheWorldGC() {
  1314  	startTheWorld()
  1315  	semrelease(&gcsema)
  1316  }
  1317  
  1318  // Holding worldsema grants an M the right to try to stop the world.
  1319  var worldsema uint32 = 1
  1320  
  1321  // Holding gcsema grants the M the right to block a GC, and blocks
  1322  // until the current GC is done. In particular, it prevents gomaxprocs
  1323  // from changing concurrently.
  1324  //
  1325  // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
  1326  // being changed/enabled during a GC, remove this.
  1327  var gcsema uint32 = 1
  1328  
  1329  // stopTheWorldWithSema is the core implementation of stopTheWorld.
  1330  // The caller is responsible for acquiring worldsema and disabling
  1331  // preemption first and then should stopTheWorldWithSema on the system
  1332  // stack:
  1333  //
  1334  //	semacquire(&worldsema, 0)
  1335  //	m.preemptoff = "reason"
  1336  //	systemstack(stopTheWorldWithSema)
  1337  //
  1338  // When finished, the caller must either call startTheWorld or undo
  1339  // these three operations separately:
  1340  //
  1341  //	m.preemptoff = ""
  1342  //	systemstack(startTheWorldWithSema)
  1343  //	semrelease(&worldsema)
  1344  //
  1345  // It is allowed to acquire worldsema once and then execute multiple
  1346  // startTheWorldWithSema/stopTheWorldWithSema pairs.
  1347  // Other P's are able to execute between successive calls to
  1348  // startTheWorldWithSema and stopTheWorldWithSema.
  1349  // Holding worldsema causes any other goroutines invoking
  1350  // stopTheWorld to block.
  1351  func stopTheWorldWithSema(reason stwReason) {
  1352  	if traceEnabled() {
  1353  		traceSTWStart(reason)
  1354  	}
  1355  	gp := getg()
  1356  
  1357  	// If we hold a lock, then we won't be able to stop another M
  1358  	// that is blocked trying to acquire the lock.
  1359  	if gp.m.locks > 0 {
  1360  		throw("stopTheWorld: holding locks")
  1361  	}
  1362  
  1363  	lock(&sched.lock)
  1364  	sched.stopwait = gomaxprocs
  1365  	sched.gcwaiting.Store(true)
  1366  	preemptall()
  1367  	// stop current P
  1368  	gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
  1369  	sched.stopwait--
  1370  	// try to retake all P's in Psyscall status
  1371  	for _, pp := range allp {
  1372  		s := pp.status
  1373  		if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
  1374  			if traceEnabled() {
  1375  				traceGoSysBlock(pp)
  1376  				traceProcStop(pp)
  1377  			}
  1378  			pp.syscalltick++
  1379  			sched.stopwait--
  1380  		}
  1381  	}
  1382  	// stop idle P's
  1383  	now := nanotime()
  1384  	for {
  1385  		pp, _ := pidleget(now)
  1386  		if pp == nil {
  1387  			break
  1388  		}
  1389  		pp.status = _Pgcstop
  1390  		sched.stopwait--
  1391  	}
  1392  	wait := sched.stopwait > 0
  1393  	unlock(&sched.lock)
  1394  
  1395  	// wait for remaining P's to stop voluntarily
  1396  	if wait {
  1397  		for {
  1398  			// wait for 100us, then try to re-preempt in case of any races
  1399  			if notetsleep(&sched.stopnote, 100*1000) {
  1400  				noteclear(&sched.stopnote)
  1401  				break
  1402  			}
  1403  			preemptall()
  1404  		}
  1405  	}
  1406  
  1407  	// sanity checks
  1408  	bad := ""
  1409  	if sched.stopwait != 0 {
  1410  		bad = "stopTheWorld: not stopped (stopwait != 0)"
  1411  	} else {
  1412  		for _, pp := range allp {
  1413  			if pp.status != _Pgcstop {
  1414  				bad = "stopTheWorld: not stopped (status != _Pgcstop)"
  1415  			}
  1416  		}
  1417  	}
  1418  	if freezing.Load() {
  1419  		// Some other thread is panicking. This can cause the
  1420  		// sanity checks above to fail if the panic happens in
  1421  		// the signal handler on a stopped thread. Either way,
  1422  		// we should halt this thread.
  1423  		lock(&deadlock)
  1424  		lock(&deadlock)
  1425  	}
  1426  	if bad != "" {
  1427  		throw(bad)
  1428  	}
  1429  
  1430  	worldStopped()
  1431  }
  1432  
  1433  func startTheWorldWithSema() int64 {
  1434  	assertWorldStopped()
  1435  
  1436  	mp := acquirem() // disable preemption because it can be holding p in a local var
  1437  	if netpollinited() {
  1438  		list := netpoll(0) // non-blocking
  1439  		injectglist(&list)
  1440  	}
  1441  	lock(&sched.lock)
  1442  
  1443  	procs := gomaxprocs
  1444  	if newprocs != 0 {
  1445  		procs = newprocs
  1446  		newprocs = 0
  1447  	}
  1448  	p1 := procresize(procs)
  1449  	sched.gcwaiting.Store(false)
  1450  	if sched.sysmonwait.Load() {
  1451  		sched.sysmonwait.Store(false)
  1452  		notewakeup(&sched.sysmonnote)
  1453  	}
  1454  	unlock(&sched.lock)
  1455  
  1456  	worldStarted()
  1457  
  1458  	for p1 != nil {
  1459  		p := p1
  1460  		p1 = p1.link.ptr()
  1461  		if p.m != 0 {
  1462  			mp := p.m.ptr()
  1463  			p.m = 0
  1464  			if mp.nextp != 0 {
  1465  				throw("startTheWorld: inconsistent mp->nextp")
  1466  			}
  1467  			mp.nextp.set(p)
  1468  			notewakeup(&mp.park)
  1469  		} else {
  1470  			// Start M to run P.  Do not start another M below.
  1471  			newm(nil, p, -1)
  1472  		}
  1473  	}
  1474  
  1475  	// Capture start-the-world time before doing clean-up tasks.
  1476  	startTime := nanotime()
  1477  	if traceEnabled() {
  1478  		traceSTWDone()
  1479  	}
  1480  
  1481  	// Wakeup an additional proc in case we have excessive runnable goroutines
  1482  	// in local queues or in the global queue. If we don't, the proc will park itself.
  1483  	// If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
  1484  	wakep()
  1485  
  1486  	releasem(mp)
  1487  
  1488  	return startTime
  1489  }
  1490  
  1491  // usesLibcall indicates whether this runtime performs system calls
  1492  // via libcall.
  1493  func usesLibcall() bool {
  1494  	switch GOOS {
  1495  	case "aix", "darwin", "illumos", "ios", "solaris", "windows":
  1496  		return true
  1497  	case "openbsd":
  1498  		return GOARCH == "386" || GOARCH == "amd64" || GOARCH == "arm" || GOARCH == "arm64"
  1499  	}
  1500  	return false
  1501  }
  1502  
  1503  // mStackIsSystemAllocated indicates whether this runtime starts on a
  1504  // system-allocated stack.
  1505  func mStackIsSystemAllocated() bool {
  1506  	switch GOOS {
  1507  	case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
  1508  		return true
  1509  	case "openbsd":
  1510  		switch GOARCH {
  1511  		case "386", "amd64", "arm", "arm64":
  1512  			return true
  1513  		}
  1514  	}
  1515  	return false
  1516  }
  1517  
  1518  // mstart is the entry-point for new Ms.
  1519  // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
  1520  func mstart()
  1521  
  1522  // mstart0 is the Go entry-point for new Ms.
  1523  // This must not split the stack because we may not even have stack
  1524  // bounds set up yet.
  1525  //
  1526  // May run during STW (because it doesn't have a P yet), so write
  1527  // barriers are not allowed.
  1528  //
  1529  //go:nosplit
  1530  //go:nowritebarrierrec
  1531  func mstart0() {
  1532  	gp := getg()
  1533  
  1534  	osStack := gp.stack.lo == 0
  1535  	if osStack {
  1536  		// Initialize stack bounds from system stack.
  1537  		// Cgo may have left stack size in stack.hi.
  1538  		// minit may update the stack bounds.
  1539  		//
  1540  		// Note: these bounds may not be very accurate.
  1541  		// We set hi to &size, but there are things above
  1542  		// it. The 1024 is supposed to compensate this,
  1543  		// but is somewhat arbitrary.
  1544  		size := gp.stack.hi
  1545  		if size == 0 {
  1546  			size = 16384 * sys.StackGuardMultiplier
  1547  		}
  1548  		gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
  1549  		gp.stack.lo = gp.stack.hi - size + 1024
  1550  	}
  1551  	// Initialize stack guard so that we can start calling regular
  1552  	// Go code.
  1553  	gp.stackguard0 = gp.stack.lo + stackGuard
  1554  	// This is the g0, so we can also call go:systemstack
  1555  	// functions, which check stackguard1.
  1556  	gp.stackguard1 = gp.stackguard0
  1557  	mstart1()
  1558  
  1559  	// Exit this thread.
  1560  	if mStackIsSystemAllocated() {
  1561  		// Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
  1562  		// the stack, but put it in gp.stack before mstart,
  1563  		// so the logic above hasn't set osStack yet.
  1564  		osStack = true
  1565  	}
  1566  	mexit(osStack)
  1567  }
  1568  
  1569  // The go:noinline is to guarantee the getcallerpc/getcallersp below are safe,
  1570  // so that we can set up g0.sched to return to the call of mstart1 above.
  1571  //
  1572  //go:noinline
  1573  func mstart1() {
  1574  	gp := getg()
  1575  
  1576  	if gp != gp.m.g0 {
  1577  		throw("bad runtime·mstart")
  1578  	}
  1579  
  1580  	// Set up m.g0.sched as a label returning to just
  1581  	// after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
  1582  	// We're never coming back to mstart1 after we call schedule,
  1583  	// so other calls can reuse the current frame.
  1584  	// And goexit0 does a gogo that needs to return from mstart1
  1585  	// and let mstart0 exit the thread.
  1586  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1587  	gp.sched.pc = getcallerpc()
  1588  	gp.sched.sp = getcallersp()
  1589  
  1590  	asminit()
  1591  	minit()
  1592  
  1593  	// Install signal handlers; after minit so that minit can
  1594  	// prepare the thread to be able to handle the signals.
  1595  	if gp.m == &m0 {
  1596  		mstartm0()
  1597  	}
  1598  
  1599  	if fn := gp.m.mstartfn; fn != nil {
  1600  		fn()
  1601  	}
  1602  
  1603  	if gp.m != &m0 {
  1604  		acquirep(gp.m.nextp.ptr())
  1605  		gp.m.nextp = 0
  1606  	}
  1607  	schedule()
  1608  }
  1609  
  1610  // mstartm0 implements part of mstart1 that only runs on the m0.
  1611  //
  1612  // Write barriers are allowed here because we know the GC can't be
  1613  // running yet, so they'll be no-ops.
  1614  //
  1615  //go:yeswritebarrierrec
  1616  func mstartm0() {
  1617  	// Create an extra M for callbacks on threads not created by Go.
  1618  	// An extra M is also needed on Windows for callbacks created by
  1619  	// syscall.NewCallback. See issue #6751 for details.
  1620  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1621  		cgoHasExtraM = true
  1622  		newextram()
  1623  	}
  1624  	initsig(false)
  1625  }
  1626  
  1627  // mPark causes a thread to park itself, returning once woken.
  1628  //
  1629  //go:nosplit
  1630  func mPark() {
  1631  	gp := getg()
  1632  	notesleep(&gp.m.park)
  1633  	noteclear(&gp.m.park)
  1634  }
  1635  
  1636  // mexit tears down and exits the current thread.
  1637  //
  1638  // Don't call this directly to exit the thread, since it must run at
  1639  // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
  1640  // unwind the stack to the point that exits the thread.
  1641  //
  1642  // It is entered with m.p != nil, so write barriers are allowed. It
  1643  // will release the P before exiting.
  1644  //
  1645  //go:yeswritebarrierrec
  1646  func mexit(osStack bool) {
  1647  	mp := getg().m
  1648  
  1649  	if mp == &m0 {
  1650  		// This is the main thread. Just wedge it.
  1651  		//
  1652  		// On Linux, exiting the main thread puts the process
  1653  		// into a non-waitable zombie state. On Plan 9,
  1654  		// exiting the main thread unblocks wait even though
  1655  		// other threads are still running. On Solaris we can
  1656  		// neither exitThread nor return from mstart. Other
  1657  		// bad things probably happen on other platforms.
  1658  		//
  1659  		// We could try to clean up this M more before wedging
  1660  		// it, but that complicates signal handling.
  1661  		handoffp(releasep())
  1662  		lock(&sched.lock)
  1663  		sched.nmfreed++
  1664  		checkdead()
  1665  		unlock(&sched.lock)
  1666  		mPark()
  1667  		throw("locked m0 woke up")
  1668  	}
  1669  
  1670  	sigblock(true)
  1671  	unminit()
  1672  
  1673  	// Free the gsignal stack.
  1674  	if mp.gsignal != nil {
  1675  		stackfree(mp.gsignal.stack)
  1676  		// On some platforms, when calling into VDSO (e.g. nanotime)
  1677  		// we store our g on the gsignal stack, if there is one.
  1678  		// Now the stack is freed, unlink it from the m, so we
  1679  		// won't write to it when calling VDSO code.
  1680  		mp.gsignal = nil
  1681  	}
  1682  
  1683  	// Remove m from allm.
  1684  	lock(&sched.lock)
  1685  	for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
  1686  		if *pprev == mp {
  1687  			*pprev = mp.alllink
  1688  			goto found
  1689  		}
  1690  	}
  1691  	throw("m not found in allm")
  1692  found:
  1693  	// Delay reaping m until it's done with the stack.
  1694  	//
  1695  	// Put mp on the free list, though it will not be reaped while freeWait
  1696  	// is freeMWait. mp is no longer reachable via allm, so even if it is
  1697  	// on an OS stack, we must keep a reference to mp alive so that the GC
  1698  	// doesn't free mp while we are still using it.
  1699  	//
  1700  	// Note that the free list must not be linked through alllink because
  1701  	// some functions walk allm without locking, so may be using alllink.
  1702  	mp.freeWait.Store(freeMWait)
  1703  	mp.freelink = sched.freem
  1704  	sched.freem = mp
  1705  	unlock(&sched.lock)
  1706  
  1707  	atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
  1708  
  1709  	// Release the P.
  1710  	handoffp(releasep())
  1711  	// After this point we must not have write barriers.
  1712  
  1713  	// Invoke the deadlock detector. This must happen after
  1714  	// handoffp because it may have started a new M to take our
  1715  	// P's work.
  1716  	lock(&sched.lock)
  1717  	sched.nmfreed++
  1718  	checkdead()
  1719  	unlock(&sched.lock)
  1720  
  1721  	if GOOS == "darwin" || GOOS == "ios" {
  1722  		// Make sure pendingPreemptSignals is correct when an M exits.
  1723  		// For #41702.
  1724  		if mp.signalPending.Load() != 0 {
  1725  			pendingPreemptSignals.Add(-1)
  1726  		}
  1727  	}
  1728  
  1729  	// Destroy all allocated resources. After this is called, we may no
  1730  	// longer take any locks.
  1731  	mdestroy(mp)
  1732  
  1733  	if osStack {
  1734  		// No more uses of mp, so it is safe to drop the reference.
  1735  		mp.freeWait.Store(freeMRef)
  1736  
  1737  		// Return from mstart and let the system thread
  1738  		// library free the g0 stack and terminate the thread.
  1739  		return
  1740  	}
  1741  
  1742  	// mstart is the thread's entry point, so there's nothing to
  1743  	// return to. Exit the thread directly. exitThread will clear
  1744  	// m.freeWait when it's done with the stack and the m can be
  1745  	// reaped.
  1746  	exitThread(&mp.freeWait)
  1747  }
  1748  
  1749  // forEachP calls fn(p) for every P p when p reaches a GC safe point.
  1750  // If a P is currently executing code, this will bring the P to a GC
  1751  // safe point and execute fn on that P. If the P is not executing code
  1752  // (it is idle or in a syscall), this will call fn(p) directly while
  1753  // preventing the P from exiting its state. This does not ensure that
  1754  // fn will run on every CPU executing Go code, but it acts as a global
  1755  // memory barrier. GC uses this as a "ragged barrier."
  1756  //
  1757  // The caller must hold worldsema.
  1758  //
  1759  //go:systemstack
  1760  func forEachP(fn func(*p)) {
  1761  	mp := acquirem()
  1762  	pp := getg().m.p.ptr()
  1763  
  1764  	lock(&sched.lock)
  1765  	if sched.safePointWait != 0 {
  1766  		throw("forEachP: sched.safePointWait != 0")
  1767  	}
  1768  	sched.safePointWait = gomaxprocs - 1
  1769  	sched.safePointFn = fn
  1770  
  1771  	// Ask all Ps to run the safe point function.
  1772  	for _, p2 := range allp {
  1773  		if p2 != pp {
  1774  			atomic.Store(&p2.runSafePointFn, 1)
  1775  		}
  1776  	}
  1777  	preemptall()
  1778  
  1779  	// Any P entering _Pidle or _Psyscall from now on will observe
  1780  	// p.runSafePointFn == 1 and will call runSafePointFn when
  1781  	// changing its status to _Pidle/_Psyscall.
  1782  
  1783  	// Run safe point function for all idle Ps. sched.pidle will
  1784  	// not change because we hold sched.lock.
  1785  	for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
  1786  		if atomic.Cas(&p.runSafePointFn, 1, 0) {
  1787  			fn(p)
  1788  			sched.safePointWait--
  1789  		}
  1790  	}
  1791  
  1792  	wait := sched.safePointWait > 0
  1793  	unlock(&sched.lock)
  1794  
  1795  	// Run fn for the current P.
  1796  	fn(pp)
  1797  
  1798  	// Force Ps currently in _Psyscall into _Pidle and hand them
  1799  	// off to induce safe point function execution.
  1800  	for _, p2 := range allp {
  1801  		s := p2.status
  1802  		if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
  1803  			if traceEnabled() {
  1804  				traceGoSysBlock(p2)
  1805  				traceProcStop(p2)
  1806  			}
  1807  			p2.syscalltick++
  1808  			handoffp(p2)
  1809  		}
  1810  	}
  1811  
  1812  	// Wait for remaining Ps to run fn.
  1813  	if wait {
  1814  		for {
  1815  			// Wait for 100us, then try to re-preempt in
  1816  			// case of any races.
  1817  			//
  1818  			// Requires system stack.
  1819  			if notetsleep(&sched.safePointNote, 100*1000) {
  1820  				noteclear(&sched.safePointNote)
  1821  				break
  1822  			}
  1823  			preemptall()
  1824  		}
  1825  	}
  1826  	if sched.safePointWait != 0 {
  1827  		throw("forEachP: not done")
  1828  	}
  1829  	for _, p2 := range allp {
  1830  		if p2.runSafePointFn != 0 {
  1831  			throw("forEachP: P did not run fn")
  1832  		}
  1833  	}
  1834  
  1835  	lock(&sched.lock)
  1836  	sched.safePointFn = nil
  1837  	unlock(&sched.lock)
  1838  	releasem(mp)
  1839  }
  1840  
  1841  // runSafePointFn runs the safe point function, if any, for this P.
  1842  // This should be called like
  1843  //
  1844  //	if getg().m.p.runSafePointFn != 0 {
  1845  //	    runSafePointFn()
  1846  //	}
  1847  //
  1848  // runSafePointFn must be checked on any transition in to _Pidle or
  1849  // _Psyscall to avoid a race where forEachP sees that the P is running
  1850  // just before the P goes into _Pidle/_Psyscall and neither forEachP
  1851  // nor the P run the safe-point function.
  1852  func runSafePointFn() {
  1853  	p := getg().m.p.ptr()
  1854  	// Resolve the race between forEachP running the safe-point
  1855  	// function on this P's behalf and this P running the
  1856  	// safe-point function directly.
  1857  	if !atomic.Cas(&p.runSafePointFn, 1, 0) {
  1858  		return
  1859  	}
  1860  	sched.safePointFn(p)
  1861  	lock(&sched.lock)
  1862  	sched.safePointWait--
  1863  	if sched.safePointWait == 0 {
  1864  		notewakeup(&sched.safePointNote)
  1865  	}
  1866  	unlock(&sched.lock)
  1867  }
  1868  
  1869  // When running with cgo, we call _cgo_thread_start
  1870  // to start threads for us so that we can play nicely with
  1871  // foreign code.
  1872  var cgoThreadStart unsafe.Pointer
  1873  
  1874  type cgothreadstart struct {
  1875  	g   guintptr
  1876  	tls *uint64
  1877  	fn  unsafe.Pointer
  1878  }
  1879  
  1880  // Allocate a new m unassociated with any thread.
  1881  // Can use p for allocation context if needed.
  1882  // fn is recorded as the new m's m.mstartfn.
  1883  // id is optional pre-allocated m ID. Omit by passing -1.
  1884  //
  1885  // This function is allowed to have write barriers even if the caller
  1886  // isn't because it borrows pp.
  1887  //
  1888  //go:yeswritebarrierrec
  1889  func allocm(pp *p, fn func(), id int64) *m {
  1890  	allocmLock.rlock()
  1891  
  1892  	// The caller owns pp, but we may borrow (i.e., acquirep) it. We must
  1893  	// disable preemption to ensure it is not stolen, which would make the
  1894  	// caller lose ownership.
  1895  	acquirem()
  1896  
  1897  	gp := getg()
  1898  	if gp.m.p == 0 {
  1899  		acquirep(pp) // temporarily borrow p for mallocs in this function
  1900  	}
  1901  
  1902  	// Release the free M list. We need to do this somewhere and
  1903  	// this may free up a stack we can use.
  1904  	if sched.freem != nil {
  1905  		lock(&sched.lock)
  1906  		var newList *m
  1907  		for freem := sched.freem; freem != nil; {
  1908  			wait := freem.freeWait.Load()
  1909  			if wait == freeMWait {
  1910  				next := freem.freelink
  1911  				freem.freelink = newList
  1912  				newList = freem
  1913  				freem = next
  1914  				continue
  1915  			}
  1916  			// Free the stack if needed. For freeMRef, there is
  1917  			// nothing to do except drop freem from the sched.freem
  1918  			// list.
  1919  			if wait == freeMStack {
  1920  				// stackfree must be on the system stack, but allocm is
  1921  				// reachable off the system stack transitively from
  1922  				// startm.
  1923  				systemstack(func() {
  1924  					stackfree(freem.g0.stack)
  1925  				})
  1926  			}
  1927  			freem = freem.freelink
  1928  		}
  1929  		sched.freem = newList
  1930  		unlock(&sched.lock)
  1931  	}
  1932  
  1933  	mp := new(m)
  1934  	mp.mstartfn = fn
  1935  	mcommoninit(mp, id)
  1936  
  1937  	// In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
  1938  	// Windows and Plan 9 will layout sched stack on OS stack.
  1939  	if iscgo || mStackIsSystemAllocated() {
  1940  		mp.g0 = malg(-1)
  1941  	} else {
  1942  		mp.g0 = malg(16384 * sys.StackGuardMultiplier)
  1943  	}
  1944  	mp.g0.m = mp
  1945  
  1946  	if pp == gp.m.p.ptr() {
  1947  		releasep()
  1948  	}
  1949  
  1950  	releasem(gp.m)
  1951  	allocmLock.runlock()
  1952  	return mp
  1953  }
  1954  
  1955  // needm is called when a cgo callback happens on a
  1956  // thread without an m (a thread not created by Go).
  1957  // In this case, needm is expected to find an m to use
  1958  // and return with m, g initialized correctly.
  1959  // Since m and g are not set now (likely nil, but see below)
  1960  // needm is limited in what routines it can call. In particular
  1961  // it can only call nosplit functions (textflag 7) and cannot
  1962  // do any scheduling that requires an m.
  1963  //
  1964  // In order to avoid needing heavy lifting here, we adopt
  1965  // the following strategy: there is a stack of available m's
  1966  // that can be stolen. Using compare-and-swap
  1967  // to pop from the stack has ABA races, so we simulate
  1968  // a lock by doing an exchange (via Casuintptr) to steal the stack
  1969  // head and replace the top pointer with MLOCKED (1).
  1970  // This serves as a simple spin lock that we can use even
  1971  // without an m. The thread that locks the stack in this way
  1972  // unlocks the stack by storing a valid stack head pointer.
  1973  //
  1974  // In order to make sure that there is always an m structure
  1975  // available to be stolen, we maintain the invariant that there
  1976  // is always one more than needed. At the beginning of the
  1977  // program (if cgo is in use) the list is seeded with a single m.
  1978  // If needm finds that it has taken the last m off the list, its job
  1979  // is - once it has installed its own m so that it can do things like
  1980  // allocate memory - to create a spare m and put it on the list.
  1981  //
  1982  // Each of these extra m's also has a g0 and a curg that are
  1983  // pressed into service as the scheduling stack and current
  1984  // goroutine for the duration of the cgo callback.
  1985  //
  1986  // It calls dropm to put the m back on the list,
  1987  // 1. when the callback is done with the m in non-pthread platforms,
  1988  // 2. or when the C thread exiting on pthread platforms.
  1989  //
  1990  // The signal argument indicates whether we're called from a signal
  1991  // handler.
  1992  //
  1993  //go:nosplit
  1994  func needm(signal bool) {
  1995  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1996  		// Can happen if C/C++ code calls Go from a global ctor.
  1997  		// Can also happen on Windows if a global ctor uses a
  1998  		// callback created by syscall.NewCallback. See issue #6751
  1999  		// for details.
  2000  		//
  2001  		// Can not throw, because scheduler is not initialized yet.
  2002  		writeErrStr("fatal error: cgo callback before cgo call\n")
  2003  		exit(1)
  2004  	}
  2005  
  2006  	// Save and block signals before getting an M.
  2007  	// The signal handler may call needm itself,
  2008  	// and we must avoid a deadlock. Also, once g is installed,
  2009  	// any incoming signals will try to execute,
  2010  	// but we won't have the sigaltstack settings and other data
  2011  	// set up appropriately until the end of minit, which will
  2012  	// unblock the signals. This is the same dance as when
  2013  	// starting a new m to run Go code via newosproc.
  2014  	var sigmask sigset
  2015  	sigsave(&sigmask)
  2016  	sigblock(false)
  2017  
  2018  	// getExtraM is safe here because of the invariant above,
  2019  	// that the extra list always contains or will soon contain
  2020  	// at least one m.
  2021  	mp, last := getExtraM()
  2022  
  2023  	// Set needextram when we've just emptied the list,
  2024  	// so that the eventual call into cgocallbackg will
  2025  	// allocate a new m for the extra list. We delay the
  2026  	// allocation until then so that it can be done
  2027  	// after exitsyscall makes sure it is okay to be
  2028  	// running at all (that is, there's no garbage collection
  2029  	// running right now).
  2030  	mp.needextram = last
  2031  
  2032  	// Store the original signal mask for use by minit.
  2033  	mp.sigmask = sigmask
  2034  
  2035  	// Install TLS on some platforms (previously setg
  2036  	// would do this if necessary).
  2037  	osSetupTLS(mp)
  2038  
  2039  	// Install g (= m->g0) and set the stack bounds
  2040  	// to match the current stack. If we don't actually know
  2041  	// how big the stack is, like we don't know how big any
  2042  	// scheduling stack is, but we assume there's at least 32 kB.
  2043  	// If we can get a more accurate stack bound from pthread,
  2044  	// use that.
  2045  	setg(mp.g0)
  2046  	gp := getg()
  2047  	gp.stack.hi = getcallersp() + 1024
  2048  	gp.stack.lo = getcallersp() - 32*1024
  2049  	if !signal && _cgo_getstackbound != nil {
  2050  		// Don't adjust if called from the signal handler.
  2051  		// We are on the signal stack, not the pthread stack.
  2052  		// (We could get the stack bounds from sigaltstack, but
  2053  		// we're getting out of the signal handler very soon
  2054  		// anyway. Not worth it.)
  2055  		var bounds [2]uintptr
  2056  		asmcgocall(_cgo_getstackbound, unsafe.Pointer(&bounds))
  2057  		// getstackbound is an unsupported no-op on Windows.
  2058  		if bounds[0] != 0 {
  2059  			gp.stack.lo = bounds[0]
  2060  			gp.stack.hi = bounds[1]
  2061  		}
  2062  	}
  2063  	gp.stackguard0 = gp.stack.lo + stackGuard
  2064  
  2065  	// Should mark we are already in Go now.
  2066  	// Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
  2067  	// which means the extram list may be empty, that will cause a deadlock.
  2068  	mp.isExtraInC = false
  2069  
  2070  	// Initialize this thread to use the m.
  2071  	asminit()
  2072  	minit()
  2073  
  2074  	// mp.curg is now a real goroutine.
  2075  	casgstatus(mp.curg, _Gdead, _Gsyscall)
  2076  	sched.ngsys.Add(-1)
  2077  }
  2078  
  2079  // Acquire an extra m and bind it to the C thread when a pthread key has been created.
  2080  //
  2081  //go:nosplit
  2082  func needAndBindM() {
  2083  	needm(false)
  2084  
  2085  	if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
  2086  		cgoBindM()
  2087  	}
  2088  }
  2089  
  2090  // newextram allocates m's and puts them on the extra list.
  2091  // It is called with a working local m, so that it can do things
  2092  // like call schedlock and allocate.
  2093  func newextram() {
  2094  	c := extraMWaiters.Swap(0)
  2095  	if c > 0 {
  2096  		for i := uint32(0); i < c; i++ {
  2097  			oneNewExtraM()
  2098  		}
  2099  	} else if extraMLength.Load() == 0 {
  2100  		// Make sure there is at least one extra M.
  2101  		oneNewExtraM()
  2102  	}
  2103  }
  2104  
  2105  // oneNewExtraM allocates an m and puts it on the extra list.
  2106  func oneNewExtraM() {
  2107  	// Create extra goroutine locked to extra m.
  2108  	// The goroutine is the context in which the cgo callback will run.
  2109  	// The sched.pc will never be returned to, but setting it to
  2110  	// goexit makes clear to the traceback routines where
  2111  	// the goroutine stack ends.
  2112  	mp := allocm(nil, nil, -1)
  2113  	gp := malg(4096)
  2114  	gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
  2115  	gp.sched.sp = gp.stack.hi
  2116  	gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
  2117  	gp.sched.lr = 0
  2118  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  2119  	gp.syscallpc = gp.sched.pc
  2120  	gp.syscallsp = gp.sched.sp
  2121  	gp.stktopsp = gp.sched.sp
  2122  	// malg returns status as _Gidle. Change to _Gdead before
  2123  	// adding to allg where GC can see it. We use _Gdead to hide
  2124  	// this from tracebacks and stack scans since it isn't a
  2125  	// "real" goroutine until needm grabs it.
  2126  	casgstatus(gp, _Gidle, _Gdead)
  2127  	gp.m = mp
  2128  	mp.curg = gp
  2129  	mp.isextra = true
  2130  	// mark we are in C by default.
  2131  	mp.isExtraInC = true
  2132  	mp.lockedInt++
  2133  	mp.lockedg.set(gp)
  2134  	gp.lockedm.set(mp)
  2135  	gp.goid = sched.goidgen.Add(1)
  2136  	if raceenabled {
  2137  		gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
  2138  	}
  2139  	if traceEnabled() {
  2140  		traceOneNewExtraM(gp)
  2141  	}
  2142  	// put on allg for garbage collector
  2143  	allgadd(gp)
  2144  
  2145  	// gp is now on the allg list, but we don't want it to be
  2146  	// counted by gcount. It would be more "proper" to increment
  2147  	// sched.ngfree, but that requires locking. Incrementing ngsys
  2148  	// has the same effect.
  2149  	sched.ngsys.Add(1)
  2150  
  2151  	// Add m to the extra list.
  2152  	addExtraM(mp)
  2153  }
  2154  
  2155  // dropm puts the current m back onto the extra list.
  2156  //
  2157  // 1. On systems without pthreads, like Windows
  2158  // dropm is called when a cgo callback has called needm but is now
  2159  // done with the callback and returning back into the non-Go thread.
  2160  //
  2161  // The main expense here is the call to signalstack to release the
  2162  // m's signal stack, and then the call to needm on the next callback
  2163  // from this thread. It is tempting to try to save the m for next time,
  2164  // which would eliminate both these costs, but there might not be
  2165  // a next time: the current thread (which Go does not control) might exit.
  2166  // If we saved the m for that thread, there would be an m leak each time
  2167  // such a thread exited. Instead, we acquire and release an m on each
  2168  // call. These should typically not be scheduling operations, just a few
  2169  // atomics, so the cost should be small.
  2170  //
  2171  // 2. On systems with pthreads
  2172  // dropm is called while a non-Go thread is exiting.
  2173  // We allocate a pthread per-thread variable using pthread_key_create,
  2174  // to register a thread-exit-time destructor.
  2175  // And store the g into a thread-specific value associated with the pthread key,
  2176  // when first return back to C.
  2177  // So that the destructor would invoke dropm while the non-Go thread is exiting.
  2178  // This is much faster since it avoids expensive signal-related syscalls.
  2179  //
  2180  // NOTE: this always runs without a P, so, nowritebarrierrec required.
  2181  //
  2182  //go:nowritebarrierrec
  2183  func dropm() {
  2184  	// Clear m and g, and return m to the extra list.
  2185  	// After the call to setg we can only call nosplit functions
  2186  	// with no pointer manipulation.
  2187  	mp := getg().m
  2188  
  2189  	// Return mp.curg to dead state.
  2190  	casgstatus(mp.curg, _Gsyscall, _Gdead)
  2191  	mp.curg.preemptStop = false
  2192  	sched.ngsys.Add(1)
  2193  
  2194  	// Block signals before unminit.
  2195  	// Unminit unregisters the signal handling stack (but needs g on some systems).
  2196  	// Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
  2197  	// It's important not to try to handle a signal between those two steps.
  2198  	sigmask := mp.sigmask
  2199  	sigblock(false)
  2200  	unminit()
  2201  
  2202  	setg(nil)
  2203  
  2204  	putExtraM(mp)
  2205  
  2206  	msigrestore(sigmask)
  2207  }
  2208  
  2209  // bindm store the g0 of the current m into a thread-specific value.
  2210  //
  2211  // We allocate a pthread per-thread variable using pthread_key_create,
  2212  // to register a thread-exit-time destructor.
  2213  // We are here setting the thread-specific value of the pthread key, to enable the destructor.
  2214  // So that the pthread_key_destructor would dropm while the C thread is exiting.
  2215  //
  2216  // And the saved g will be used in pthread_key_destructor,
  2217  // since the g stored in the TLS by Go might be cleared in some platforms,
  2218  // before the destructor invoked, so, we restore g by the stored g, before dropm.
  2219  //
  2220  // We store g0 instead of m, to make the assembly code simpler,
  2221  // since we need to restore g0 in runtime.cgocallback.
  2222  //
  2223  // On systems without pthreads, like Windows, bindm shouldn't be used.
  2224  //
  2225  // NOTE: this always runs without a P, so, nowritebarrierrec required.
  2226  //
  2227  //go:nosplit
  2228  //go:nowritebarrierrec
  2229  func cgoBindM() {
  2230  	if GOOS == "windows" || GOOS == "plan9" {
  2231  		fatal("bindm in unexpected GOOS")
  2232  	}
  2233  	g := getg()
  2234  	if g.m.g0 != g {
  2235  		fatal("the current g is not g0")
  2236  	}
  2237  	if _cgo_bindm != nil {
  2238  		asmcgocall(_cgo_bindm, unsafe.Pointer(g))
  2239  	}
  2240  }
  2241  
  2242  // A helper function for EnsureDropM.
  2243  func getm() uintptr {
  2244  	return uintptr(unsafe.Pointer(getg().m))
  2245  }
  2246  
  2247  var (
  2248  	// Locking linked list of extra M's, via mp.schedlink. Must be accessed
  2249  	// only via lockextra/unlockextra.
  2250  	//
  2251  	// Can't be atomic.Pointer[m] because we use an invalid pointer as a
  2252  	// "locked" sentinel value. M's on this list remain visible to the GC
  2253  	// because their mp.curg is on allgs.
  2254  	extraM atomic.Uintptr
  2255  	// Number of M's in the extraM list.
  2256  	extraMLength atomic.Uint32
  2257  	// Number of waiters in lockextra.
  2258  	extraMWaiters atomic.Uint32
  2259  
  2260  	// Number of extra M's in use by threads.
  2261  	extraMInUse atomic.Uint32
  2262  )
  2263  
  2264  // lockextra locks the extra list and returns the list head.
  2265  // The caller must unlock the list by storing a new list head
  2266  // to extram. If nilokay is true, then lockextra will
  2267  // return a nil list head if that's what it finds. If nilokay is false,
  2268  // lockextra will keep waiting until the list head is no longer nil.
  2269  //
  2270  //go:nosplit
  2271  func lockextra(nilokay bool) *m {
  2272  	const locked = 1
  2273  
  2274  	incr := false
  2275  	for {
  2276  		old := extraM.Load()
  2277  		if old == locked {
  2278  			osyield_no_g()
  2279  			continue
  2280  		}
  2281  		if old == 0 && !nilokay {
  2282  			if !incr {
  2283  				// Add 1 to the number of threads
  2284  				// waiting for an M.
  2285  				// This is cleared by newextram.
  2286  				extraMWaiters.Add(1)
  2287  				incr = true
  2288  			}
  2289  			usleep_no_g(1)
  2290  			continue
  2291  		}
  2292  		if extraM.CompareAndSwap(old, locked) {
  2293  			return (*m)(unsafe.Pointer(old))
  2294  		}
  2295  		osyield_no_g()
  2296  		continue
  2297  	}
  2298  }
  2299  
  2300  //go:nosplit
  2301  func unlockextra(mp *m, delta int32) {
  2302  	extraMLength.Add(delta)
  2303  	extraM.Store(uintptr(unsafe.Pointer(mp)))
  2304  }
  2305  
  2306  // Return an M from the extra M list. Returns last == true if the list becomes
  2307  // empty because of this call.
  2308  //
  2309  // Spins waiting for an extra M, so caller must ensure that the list always
  2310  // contains or will soon contain at least one M.
  2311  //
  2312  //go:nosplit
  2313  func getExtraM() (mp *m, last bool) {
  2314  	mp = lockextra(false)
  2315  	extraMInUse.Add(1)
  2316  	unlockextra(mp.schedlink.ptr(), -1)
  2317  	return mp, mp.schedlink.ptr() == nil
  2318  }
  2319  
  2320  // Returns an extra M back to the list. mp must be from getExtraM. Newly
  2321  // allocated M's should use addExtraM.
  2322  //
  2323  //go:nosplit
  2324  func putExtraM(mp *m) {
  2325  	extraMInUse.Add(-1)
  2326  	addExtraM(mp)
  2327  }
  2328  
  2329  // Adds a newly allocated M to the extra M list.
  2330  //
  2331  //go:nosplit
  2332  func addExtraM(mp *m) {
  2333  	mnext := lockextra(true)
  2334  	mp.schedlink.set(mnext)
  2335  	unlockextra(mp, 1)
  2336  }
  2337  
  2338  var (
  2339  	// allocmLock is locked for read when creating new Ms in allocm and their
  2340  	// addition to allm. Thus acquiring this lock for write blocks the
  2341  	// creation of new Ms.
  2342  	allocmLock rwmutex
  2343  
  2344  	// execLock serializes exec and clone to avoid bugs or unspecified
  2345  	// behaviour around exec'ing while creating/destroying threads. See
  2346  	// issue #19546.
  2347  	execLock rwmutex
  2348  )
  2349  
  2350  // These errors are reported (via writeErrStr) by some OS-specific
  2351  // versions of newosproc and newosproc0.
  2352  const (
  2353  	failthreadcreate  = "runtime: failed to create new OS thread\n"
  2354  	failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
  2355  )
  2356  
  2357  // newmHandoff contains a list of m structures that need new OS threads.
  2358  // This is used by newm in situations where newm itself can't safely
  2359  // start an OS thread.
  2360  var newmHandoff struct {
  2361  	lock mutex
  2362  
  2363  	// newm points to a list of M structures that need new OS
  2364  	// threads. The list is linked through m.schedlink.
  2365  	newm muintptr
  2366  
  2367  	// waiting indicates that wake needs to be notified when an m
  2368  	// is put on the list.
  2369  	waiting bool
  2370  	wake    note
  2371  
  2372  	// haveTemplateThread indicates that the templateThread has
  2373  	// been started. This is not protected by lock. Use cas to set
  2374  	// to 1.
  2375  	haveTemplateThread uint32
  2376  }
  2377  
  2378  // Create a new m. It will start off with a call to fn, or else the scheduler.
  2379  // fn needs to be static and not a heap allocated closure.
  2380  // May run with m.p==nil, so write barriers are not allowed.
  2381  //
  2382  // id is optional pre-allocated m ID. Omit by passing -1.
  2383  //
  2384  //go:nowritebarrierrec
  2385  func newm(fn func(), pp *p, id int64) {
  2386  	// allocm adds a new M to allm, but they do not start until created by
  2387  	// the OS in newm1 or the template thread.
  2388  	//
  2389  	// doAllThreadsSyscall requires that every M in allm will eventually
  2390  	// start and be signal-able, even with a STW.
  2391  	//
  2392  	// Disable preemption here until we start the thread to ensure that
  2393  	// newm is not preempted between allocm and starting the new thread,
  2394  	// ensuring that anything added to allm is guaranteed to eventually
  2395  	// start.
  2396  	acquirem()
  2397  
  2398  	mp := allocm(pp, fn, id)
  2399  	mp.nextp.set(pp)
  2400  	mp.sigmask = initSigmask
  2401  	if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
  2402  		// We're on a locked M or a thread that may have been
  2403  		// started by C. The kernel state of this thread may
  2404  		// be strange (the user may have locked it for that
  2405  		// purpose). We don't want to clone that into another
  2406  		// thread. Instead, ask a known-good thread to create
  2407  		// the thread for us.
  2408  		//
  2409  		// This is disabled on Plan 9. See golang.org/issue/22227.
  2410  		//
  2411  		// TODO: This may be unnecessary on Windows, which
  2412  		// doesn't model thread creation off fork.
  2413  		lock(&newmHandoff.lock)
  2414  		if newmHandoff.haveTemplateThread == 0 {
  2415  			throw("on a locked thread with no template thread")
  2416  		}
  2417  		mp.schedlink = newmHandoff.newm
  2418  		newmHandoff.newm.set(mp)
  2419  		if newmHandoff.waiting {
  2420  			newmHandoff.waiting = false
  2421  			notewakeup(&newmHandoff.wake)
  2422  		}
  2423  		unlock(&newmHandoff.lock)
  2424  		// The M has not started yet, but the template thread does not
  2425  		// participate in STW, so it will always process queued Ms and
  2426  		// it is safe to releasem.
  2427  		releasem(getg().m)
  2428  		return
  2429  	}
  2430  	newm1(mp)
  2431  	releasem(getg().m)
  2432  }
  2433  
  2434  func newm1(mp *m) {
  2435  	if iscgo {
  2436  		var ts cgothreadstart
  2437  		if _cgo_thread_start == nil {
  2438  			throw("_cgo_thread_start missing")
  2439  		}
  2440  		ts.g.set(mp.g0)
  2441  		ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
  2442  		ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
  2443  		if msanenabled {
  2444  			msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2445  		}
  2446  		if asanenabled {
  2447  			asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2448  		}
  2449  		execLock.rlock() // Prevent process clone.
  2450  		asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
  2451  		execLock.runlock()
  2452  		return
  2453  	}
  2454  	execLock.rlock() // Prevent process clone.
  2455  	newosproc(mp)
  2456  	execLock.runlock()
  2457  }
  2458  
  2459  // startTemplateThread starts the template thread if it is not already
  2460  // running.
  2461  //
  2462  // The calling thread must itself be in a known-good state.
  2463  func startTemplateThread() {
  2464  	if GOARCH == "wasm" { // no threads on wasm yet
  2465  		return
  2466  	}
  2467  
  2468  	// Disable preemption to guarantee that the template thread will be
  2469  	// created before a park once haveTemplateThread is set.
  2470  	mp := acquirem()
  2471  	if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
  2472  		releasem(mp)
  2473  		return
  2474  	}
  2475  	newm(templateThread, nil, -1)
  2476  	releasem(mp)
  2477  }
  2478  
  2479  // templateThread is a thread in a known-good state that exists solely
  2480  // to start new threads in known-good states when the calling thread
  2481  // may not be in a good state.
  2482  //
  2483  // Many programs never need this, so templateThread is started lazily
  2484  // when we first enter a state that might lead to running on a thread
  2485  // in an unknown state.
  2486  //
  2487  // templateThread runs on an M without a P, so it must not have write
  2488  // barriers.
  2489  //
  2490  //go:nowritebarrierrec
  2491  func templateThread() {
  2492  	lock(&sched.lock)
  2493  	sched.nmsys++
  2494  	checkdead()
  2495  	unlock(&sched.lock)
  2496  
  2497  	for {
  2498  		lock(&newmHandoff.lock)
  2499  		for newmHandoff.newm != 0 {
  2500  			newm := newmHandoff.newm.ptr()
  2501  			newmHandoff.newm = 0
  2502  			unlock(&newmHandoff.lock)
  2503  			for newm != nil {
  2504  				next := newm.schedlink.ptr()
  2505  				newm.schedlink = 0
  2506  				newm1(newm)
  2507  				newm = next
  2508  			}
  2509  			lock(&newmHandoff.lock)
  2510  		}
  2511  		newmHandoff.waiting = true
  2512  		noteclear(&newmHandoff.wake)
  2513  		unlock(&newmHandoff.lock)
  2514  		notesleep(&newmHandoff.wake)
  2515  	}
  2516  }
  2517  
  2518  // Stops execution of the current m until new work is available.
  2519  // Returns with acquired P.
  2520  func stopm() {
  2521  	gp := getg()
  2522  
  2523  	if gp.m.locks != 0 {
  2524  		throw("stopm holding locks")
  2525  	}
  2526  	if gp.m.p != 0 {
  2527  		throw("stopm holding p")
  2528  	}
  2529  	if gp.m.spinning {
  2530  		throw("stopm spinning")
  2531  	}
  2532  
  2533  	lock(&sched.lock)
  2534  	mput(gp.m)
  2535  	unlock(&sched.lock)
  2536  	mPark()
  2537  	acquirep(gp.m.nextp.ptr())
  2538  	gp.m.nextp = 0
  2539  }
  2540  
  2541  func mspinning() {
  2542  	// startm's caller incremented nmspinning. Set the new M's spinning.
  2543  	getg().m.spinning = true
  2544  }
  2545  
  2546  // Schedules some M to run the p (creates an M if necessary).
  2547  // If p==nil, tries to get an idle P, if no idle P's does nothing.
  2548  // May run with m.p==nil, so write barriers are not allowed.
  2549  // If spinning is set, the caller has incremented nmspinning and must provide a
  2550  // P. startm will set m.spinning in the newly started M.
  2551  //
  2552  // Callers passing a non-nil P must call from a non-preemptible context. See
  2553  // comment on acquirem below.
  2554  //
  2555  // Argument lockheld indicates whether the caller already acquired the
  2556  // scheduler lock. Callers holding the lock when making the call must pass
  2557  // true. The lock might be temporarily dropped, but will be reacquired before
  2558  // returning.
  2559  //
  2560  // Must not have write barriers because this may be called without a P.
  2561  //
  2562  //go:nowritebarrierrec
  2563  func startm(pp *p, spinning, lockheld bool) {
  2564  	// Disable preemption.
  2565  	//
  2566  	// Every owned P must have an owner that will eventually stop it in the
  2567  	// event of a GC stop request. startm takes transient ownership of a P
  2568  	// (either from argument or pidleget below) and transfers ownership to
  2569  	// a started M, which will be responsible for performing the stop.
  2570  	//
  2571  	// Preemption must be disabled during this transient ownership,
  2572  	// otherwise the P this is running on may enter GC stop while still
  2573  	// holding the transient P, leaving that P in limbo and deadlocking the
  2574  	// STW.
  2575  	//
  2576  	// Callers passing a non-nil P must already be in non-preemptible
  2577  	// context, otherwise such preemption could occur on function entry to
  2578  	// startm. Callers passing a nil P may be preemptible, so we must
  2579  	// disable preemption before acquiring a P from pidleget below.
  2580  	mp := acquirem()
  2581  	if !lockheld {
  2582  		lock(&sched.lock)
  2583  	}
  2584  	if pp == nil {
  2585  		if spinning {
  2586  			// TODO(prattmic): All remaining calls to this function
  2587  			// with _p_ == nil could be cleaned up to find a P
  2588  			// before calling startm.
  2589  			throw("startm: P required for spinning=true")
  2590  		}
  2591  		pp, _ = pidleget(0)
  2592  		if pp == nil {
  2593  			if !lockheld {
  2594  				unlock(&sched.lock)
  2595  			}
  2596  			releasem(mp)
  2597  			return
  2598  		}
  2599  	}
  2600  	nmp := mget()
  2601  	if nmp == nil {
  2602  		// No M is available, we must drop sched.lock and call newm.
  2603  		// However, we already own a P to assign to the M.
  2604  		//
  2605  		// Once sched.lock is released, another G (e.g., in a syscall),
  2606  		// could find no idle P while checkdead finds a runnable G but
  2607  		// no running M's because this new M hasn't started yet, thus
  2608  		// throwing in an apparent deadlock.
  2609  		// This apparent deadlock is possible when startm is called
  2610  		// from sysmon, which doesn't count as a running M.
  2611  		//
  2612  		// Avoid this situation by pre-allocating the ID for the new M,
  2613  		// thus marking it as 'running' before we drop sched.lock. This
  2614  		// new M will eventually run the scheduler to execute any
  2615  		// queued G's.
  2616  		id := mReserveID()
  2617  		unlock(&sched.lock)
  2618  
  2619  		var fn func()
  2620  		if spinning {
  2621  			// The caller incremented nmspinning, so set m.spinning in the new M.
  2622  			fn = mspinning
  2623  		}
  2624  		newm(fn, pp, id)
  2625  
  2626  		if lockheld {
  2627  			lock(&sched.lock)
  2628  		}
  2629  		// Ownership transfer of pp committed by start in newm.
  2630  		// Preemption is now safe.
  2631  		releasem(mp)
  2632  		return
  2633  	}
  2634  	if !lockheld {
  2635  		unlock(&sched.lock)
  2636  	}
  2637  	if nmp.spinning {
  2638  		throw("startm: m is spinning")
  2639  	}
  2640  	if nmp.nextp != 0 {
  2641  		throw("startm: m has p")
  2642  	}
  2643  	if spinning && !runqempty(pp) {
  2644  		throw("startm: p has runnable gs")
  2645  	}
  2646  	// The caller incremented nmspinning, so set m.spinning in the new M.
  2647  	nmp.spinning = spinning
  2648  	nmp.nextp.set(pp)
  2649  	notewakeup(&nmp.park)
  2650  	// Ownership transfer of pp committed by wakeup. Preemption is now
  2651  	// safe.
  2652  	releasem(mp)
  2653  }
  2654  
  2655  // Hands off P from syscall or locked M.
  2656  // Always runs without a P, so write barriers are not allowed.
  2657  //
  2658  //go:nowritebarrierrec
  2659  func handoffp(pp *p) {
  2660  	// handoffp must start an M in any situation where
  2661  	// findrunnable would return a G to run on pp.
  2662  
  2663  	// if it has local work, start it straight away
  2664  	if !runqempty(pp) || sched.runqsize != 0 {
  2665  		startm(pp, false, false)
  2666  		return
  2667  	}
  2668  	// if there's trace work to do, start it straight away
  2669  	if (traceEnabled() || traceShuttingDown()) && traceReaderAvailable() != nil {
  2670  		startm(pp, false, false)
  2671  		return
  2672  	}
  2673  	// if it has GC work, start it straight away
  2674  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
  2675  		startm(pp, false, false)
  2676  		return
  2677  	}
  2678  	// no local work, check that there are no spinning/idle M's,
  2679  	// otherwise our help is not required
  2680  	if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
  2681  		sched.needspinning.Store(0)
  2682  		startm(pp, true, false)
  2683  		return
  2684  	}
  2685  	lock(&sched.lock)
  2686  	if sched.gcwaiting.Load() {
  2687  		pp.status = _Pgcstop
  2688  		sched.stopwait--
  2689  		if sched.stopwait == 0 {
  2690  			notewakeup(&sched.stopnote)
  2691  		}
  2692  		unlock(&sched.lock)
  2693  		return
  2694  	}
  2695  	if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
  2696  		sched.safePointFn(pp)
  2697  		sched.safePointWait--
  2698  		if sched.safePointWait == 0 {
  2699  			notewakeup(&sched.safePointNote)
  2700  		}
  2701  	}
  2702  	if sched.runqsize != 0 {
  2703  		unlock(&sched.lock)
  2704  		startm(pp, false, false)
  2705  		return
  2706  	}
  2707  	// If this is the last running P and nobody is polling network,
  2708  	// need to wakeup another M to poll network.
  2709  	if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
  2710  		unlock(&sched.lock)
  2711  		startm(pp, false, false)
  2712  		return
  2713  	}
  2714  
  2715  	// The scheduler lock cannot be held when calling wakeNetPoller below
  2716  	// because wakeNetPoller may call wakep which may call startm.
  2717  	when := nobarrierWakeTime(pp)
  2718  	pidleput(pp, 0)
  2719  	unlock(&sched.lock)
  2720  
  2721  	if when != 0 {
  2722  		wakeNetPoller(when)
  2723  	}
  2724  }
  2725  
  2726  // Tries to add one more P to execute G's.
  2727  // Called when a G is made runnable (newproc, ready).
  2728  // Must be called with a P.
  2729  func wakep() {
  2730  	// Be conservative about spinning threads, only start one if none exist
  2731  	// already.
  2732  	if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
  2733  		return
  2734  	}
  2735  
  2736  	// Disable preemption until ownership of pp transfers to the next M in
  2737  	// startm. Otherwise preemption here would leave pp stuck waiting to
  2738  	// enter _Pgcstop.
  2739  	//
  2740  	// See preemption comment on acquirem in startm for more details.
  2741  	mp := acquirem()
  2742  
  2743  	var pp *p
  2744  	lock(&sched.lock)
  2745  	pp, _ = pidlegetSpinning(0)
  2746  	if pp == nil {
  2747  		if sched.nmspinning.Add(-1) < 0 {
  2748  			throw("wakep: negative nmspinning")
  2749  		}
  2750  		unlock(&sched.lock)
  2751  		releasem(mp)
  2752  		return
  2753  	}
  2754  	// Since we always have a P, the race in the "No M is available"
  2755  	// comment in startm doesn't apply during the small window between the
  2756  	// unlock here and lock in startm. A checkdead in between will always
  2757  	// see at least one running M (ours).
  2758  	unlock(&sched.lock)
  2759  
  2760  	startm(pp, true, false)
  2761  
  2762  	releasem(mp)
  2763  }
  2764  
  2765  // Stops execution of the current m that is locked to a g until the g is runnable again.
  2766  // Returns with acquired P.
  2767  func stoplockedm() {
  2768  	gp := getg()
  2769  
  2770  	if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
  2771  		throw("stoplockedm: inconsistent locking")
  2772  	}
  2773  	if gp.m.p != 0 {
  2774  		// Schedule another M to run this p.
  2775  		pp := releasep()
  2776  		handoffp(pp)
  2777  	}
  2778  	incidlelocked(1)
  2779  	// Wait until another thread schedules lockedg again.
  2780  	mPark()
  2781  	status := readgstatus(gp.m.lockedg.ptr())
  2782  	if status&^_Gscan != _Grunnable {
  2783  		print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
  2784  		dumpgstatus(gp.m.lockedg.ptr())
  2785  		throw("stoplockedm: not runnable")
  2786  	}
  2787  	acquirep(gp.m.nextp.ptr())
  2788  	gp.m.nextp = 0
  2789  }
  2790  
  2791  // Schedules the locked m to run the locked gp.
  2792  // May run during STW, so write barriers are not allowed.
  2793  //
  2794  //go:nowritebarrierrec
  2795  func startlockedm(gp *g) {
  2796  	mp := gp.lockedm.ptr()
  2797  	if mp == getg().m {
  2798  		throw("startlockedm: locked to me")
  2799  	}
  2800  	if mp.nextp != 0 {
  2801  		throw("startlockedm: m has p")
  2802  	}
  2803  	// directly handoff current P to the locked m
  2804  	incidlelocked(-1)
  2805  	pp := releasep()
  2806  	mp.nextp.set(pp)
  2807  	notewakeup(&mp.park)
  2808  	stopm()
  2809  }
  2810  
  2811  // Stops the current m for stopTheWorld.
  2812  // Returns when the world is restarted.
  2813  func gcstopm() {
  2814  	gp := getg()
  2815  
  2816  	if !sched.gcwaiting.Load() {
  2817  		throw("gcstopm: not waiting for gc")
  2818  	}
  2819  	if gp.m.spinning {
  2820  		gp.m.spinning = false
  2821  		// OK to just drop nmspinning here,
  2822  		// startTheWorld will unpark threads as necessary.
  2823  		if sched.nmspinning.Add(-1) < 0 {
  2824  			throw("gcstopm: negative nmspinning")
  2825  		}
  2826  	}
  2827  	pp := releasep()
  2828  	lock(&sched.lock)
  2829  	pp.status = _Pgcstop
  2830  	sched.stopwait--
  2831  	if sched.stopwait == 0 {
  2832  		notewakeup(&sched.stopnote)
  2833  	}
  2834  	unlock(&sched.lock)
  2835  	stopm()
  2836  }
  2837  
  2838  // Schedules gp to run on the current M.
  2839  // If inheritTime is true, gp inherits the remaining time in the
  2840  // current time slice. Otherwise, it starts a new time slice.
  2841  // Never returns.
  2842  //
  2843  // Write barriers are allowed because this is called immediately after
  2844  // acquiring a P in several places.
  2845  //
  2846  //go:yeswritebarrierrec
  2847  func execute(gp *g, inheritTime bool) {
  2848  	mp := getg().m
  2849  
  2850  	if goroutineProfile.active {
  2851  		// Make sure that gp has had its stack written out to the goroutine
  2852  		// profile, exactly as it was when the goroutine profiler first stopped
  2853  		// the world.
  2854  		tryRecordGoroutineProfile(gp, osyield)
  2855  	}
  2856  
  2857  	// Assign gp.m before entering _Grunning so running Gs have an
  2858  	// M.
  2859  	mp.curg = gp
  2860  	gp.m = mp
  2861  	casgstatus(gp, _Grunnable, _Grunning)
  2862  	gp.waitsince = 0
  2863  	gp.preempt = false
  2864  	gp.stackguard0 = gp.stack.lo + stackGuard
  2865  	if !inheritTime {
  2866  		mp.p.ptr().schedtick++
  2867  	}
  2868  
  2869  	// Check whether the profiler needs to be turned on or off.
  2870  	hz := sched.profilehz
  2871  	if mp.profilehz != hz {
  2872  		setThreadCPUProfiler(hz)
  2873  	}
  2874  
  2875  	if traceEnabled() {
  2876  		// GoSysExit has to happen when we have a P, but before GoStart.
  2877  		// So we emit it here.
  2878  		if gp.syscallsp != 0 {
  2879  			traceGoSysExit()
  2880  		}
  2881  		traceGoStart()
  2882  	}
  2883  
  2884  	gogo(&gp.sched)
  2885  }
  2886  
  2887  // Finds a runnable goroutine to execute.
  2888  // Tries to steal from other P's, get g from local or global queue, poll network.
  2889  // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
  2890  // reader) so the caller should try to wake a P.
  2891  func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
  2892  	mp := getg().m
  2893  
  2894  	// The conditions here and in handoffp must agree: if
  2895  	// findrunnable would return a G to run, handoffp must start
  2896  	// an M.
  2897  
  2898  top:
  2899  	pp := mp.p.ptr()
  2900  	if sched.gcwaiting.Load() {
  2901  		gcstopm()
  2902  		goto top
  2903  	}
  2904  	if pp.runSafePointFn != 0 {
  2905  		runSafePointFn()
  2906  	}
  2907  
  2908  	// now and pollUntil are saved for work stealing later,
  2909  	// which may steal timers. It's important that between now
  2910  	// and then, nothing blocks, so these numbers remain mostly
  2911  	// relevant.
  2912  	now, pollUntil, _ := checkTimers(pp, 0)
  2913  
  2914  	// Try to schedule the trace reader.
  2915  	if traceEnabled() || traceShuttingDown() {
  2916  		gp := traceReader()
  2917  		if gp != nil {
  2918  			casgstatus(gp, _Gwaiting, _Grunnable)
  2919  			traceGoUnpark(gp, 0)
  2920  			return gp, false, true
  2921  		}
  2922  	}
  2923  
  2924  	// Try to schedule a GC worker.
  2925  	if gcBlackenEnabled != 0 {
  2926  		gp, tnow := gcController.findRunnableGCWorker(pp, now)
  2927  		if gp != nil {
  2928  			return gp, false, true
  2929  		}
  2930  		now = tnow
  2931  	}
  2932  
  2933  	// Check the global runnable queue once in a while to ensure fairness.
  2934  	// Otherwise two goroutines can completely occupy the local runqueue
  2935  	// by constantly respawning each other.
  2936  	if pp.schedtick%61 == 0 && sched.runqsize > 0 {
  2937  		lock(&sched.lock)
  2938  		gp := globrunqget(pp, 1)
  2939  		unlock(&sched.lock)
  2940  		if gp != nil {
  2941  			return gp, false, false
  2942  		}
  2943  	}
  2944  
  2945  	// Wake up the finalizer G.
  2946  	if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
  2947  		if gp := wakefing(); gp != nil {
  2948  			ready(gp, 0, true)
  2949  		}
  2950  	}
  2951  	if *cgo_yield != nil {
  2952  		asmcgocall(*cgo_yield, nil)
  2953  	}
  2954  
  2955  	// local runq
  2956  	if gp, inheritTime := runqget(pp); gp != nil {
  2957  		return gp, inheritTime, false
  2958  	}
  2959  
  2960  	// global runq
  2961  	if sched.runqsize != 0 {
  2962  		lock(&sched.lock)
  2963  		gp := globrunqget(pp, 0)
  2964  		unlock(&sched.lock)
  2965  		if gp != nil {
  2966  			return gp, false, false
  2967  		}
  2968  	}
  2969  
  2970  	// Poll network.
  2971  	// This netpoll is only an optimization before we resort to stealing.
  2972  	// We can safely skip it if there are no waiters or a thread is blocked
  2973  	// in netpoll already. If there is any kind of logical race with that
  2974  	// blocked thread (e.g. it has already returned from netpoll, but does
  2975  	// not set lastpoll yet), this thread will do blocking netpoll below
  2976  	// anyway.
  2977  	if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
  2978  		if list := netpoll(0); !list.empty() { // non-blocking
  2979  			gp := list.pop()
  2980  			injectglist(&list)
  2981  			casgstatus(gp, _Gwaiting, _Grunnable)
  2982  			if traceEnabled() {
  2983  				traceGoUnpark(gp, 0)
  2984  			}
  2985  			return gp, false, false
  2986  		}
  2987  	}
  2988  
  2989  	// Spinning Ms: steal work from other Ps.
  2990  	//
  2991  	// Limit the number of spinning Ms to half the number of busy Ps.
  2992  	// This is necessary to prevent excessive CPU consumption when
  2993  	// GOMAXPROCS>>1 but the program parallelism is low.
  2994  	if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
  2995  		if !mp.spinning {
  2996  			mp.becomeSpinning()
  2997  		}
  2998  
  2999  		gp, inheritTime, tnow, w, newWork := stealWork(now)
  3000  		if gp != nil {
  3001  			// Successfully stole.
  3002  			return gp, inheritTime, false
  3003  		}
  3004  		if newWork {
  3005  			// There may be new timer or GC work; restart to
  3006  			// discover.
  3007  			goto top
  3008  		}
  3009  
  3010  		now = tnow
  3011  		if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3012  			// Earlier timer to wait for.
  3013  			pollUntil = w
  3014  		}
  3015  	}
  3016  
  3017  	// We have nothing to do.
  3018  	//
  3019  	// If we're in the GC mark phase, can safely scan and blacken objects,
  3020  	// and have work to do, run idle-time marking rather than give up the P.
  3021  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
  3022  		node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3023  		if node != nil {
  3024  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3025  			gp := node.gp.ptr()
  3026  			casgstatus(gp, _Gwaiting, _Grunnable)
  3027  			if traceEnabled() {
  3028  				traceGoUnpark(gp, 0)
  3029  			}
  3030  			return gp, false, false
  3031  		}
  3032  		gcController.removeIdleMarkWorker()
  3033  	}
  3034  
  3035  	// wasm only:
  3036  	// If a callback returned and no other goroutine is awake,
  3037  	// then wake event handler goroutine which pauses execution
  3038  	// until a callback was triggered.
  3039  	gp, otherReady := beforeIdle(now, pollUntil)
  3040  	if gp != nil {
  3041  		casgstatus(gp, _Gwaiting, _Grunnable)
  3042  		if traceEnabled() {
  3043  			traceGoUnpark(gp, 0)
  3044  		}
  3045  		return gp, false, false
  3046  	}
  3047  	if otherReady {
  3048  		goto top
  3049  	}
  3050  
  3051  	// Before we drop our P, make a snapshot of the allp slice,
  3052  	// which can change underfoot once we no longer block
  3053  	// safe-points. We don't need to snapshot the contents because
  3054  	// everything up to cap(allp) is immutable.
  3055  	allpSnapshot := allp
  3056  	// Also snapshot masks. Value changes are OK, but we can't allow
  3057  	// len to change out from under us.
  3058  	idlepMaskSnapshot := idlepMask
  3059  	timerpMaskSnapshot := timerpMask
  3060  
  3061  	// return P and block
  3062  	lock(&sched.lock)
  3063  	if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
  3064  		unlock(&sched.lock)
  3065  		goto top
  3066  	}
  3067  	if sched.runqsize != 0 {
  3068  		gp := globrunqget(pp, 0)
  3069  		unlock(&sched.lock)
  3070  		return gp, false, false
  3071  	}
  3072  	if !mp.spinning && sched.needspinning.Load() == 1 {
  3073  		// See "Delicate dance" comment below.
  3074  		mp.becomeSpinning()
  3075  		unlock(&sched.lock)
  3076  		goto top
  3077  	}
  3078  	if releasep() != pp {
  3079  		throw("findrunnable: wrong p")
  3080  	}
  3081  	now = pidleput(pp, now)
  3082  	unlock(&sched.lock)
  3083  
  3084  	// Delicate dance: thread transitions from spinning to non-spinning
  3085  	// state, potentially concurrently with submission of new work. We must
  3086  	// drop nmspinning first and then check all sources again (with
  3087  	// #StoreLoad memory barrier in between). If we do it the other way
  3088  	// around, another thread can submit work after we've checked all
  3089  	// sources but before we drop nmspinning; as a result nobody will
  3090  	// unpark a thread to run the work.
  3091  	//
  3092  	// This applies to the following sources of work:
  3093  	//
  3094  	// * Goroutines added to a per-P run queue.
  3095  	// * New/modified-earlier timers on a per-P timer heap.
  3096  	// * Idle-priority GC work (barring golang.org/issue/19112).
  3097  	//
  3098  	// If we discover new work below, we need to restore m.spinning as a
  3099  	// signal for resetspinning to unpark a new worker thread (because
  3100  	// there can be more than one starving goroutine).
  3101  	//
  3102  	// However, if after discovering new work we also observe no idle Ps
  3103  	// (either here or in resetspinning), we have a problem. We may be
  3104  	// racing with a non-spinning M in the block above, having found no
  3105  	// work and preparing to release its P and park. Allowing that P to go
  3106  	// idle will result in loss of work conservation (idle P while there is
  3107  	// runnable work). This could result in complete deadlock in the
  3108  	// unlikely event that we discover new work (from netpoll) right as we
  3109  	// are racing with _all_ other Ps going idle.
  3110  	//
  3111  	// We use sched.needspinning to synchronize with non-spinning Ms going
  3112  	// idle. If needspinning is set when they are about to drop their P,
  3113  	// they abort the drop and instead become a new spinning M on our
  3114  	// behalf. If we are not racing and the system is truly fully loaded
  3115  	// then no spinning threads are required, and the next thread to
  3116  	// naturally become spinning will clear the flag.
  3117  	//
  3118  	// Also see "Worker thread parking/unparking" comment at the top of the
  3119  	// file.
  3120  	wasSpinning := mp.spinning
  3121  	if mp.spinning {
  3122  		mp.spinning = false
  3123  		if sched.nmspinning.Add(-1) < 0 {
  3124  			throw("findrunnable: negative nmspinning")
  3125  		}
  3126  
  3127  		// Note the for correctness, only the last M transitioning from
  3128  		// spinning to non-spinning must perform these rechecks to
  3129  		// ensure no missed work. However, the runtime has some cases
  3130  		// of transient increments of nmspinning that are decremented
  3131  		// without going through this path, so we must be conservative
  3132  		// and perform the check on all spinning Ms.
  3133  		//
  3134  		// See https://go.dev/issue/43997.
  3135  
  3136  		// Check all runqueues once again.
  3137  		pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
  3138  		if pp != nil {
  3139  			acquirep(pp)
  3140  			mp.becomeSpinning()
  3141  			goto top
  3142  		}
  3143  
  3144  		// Check for idle-priority GC work again.
  3145  		pp, gp := checkIdleGCNoP()
  3146  		if pp != nil {
  3147  			acquirep(pp)
  3148  			mp.becomeSpinning()
  3149  
  3150  			// Run the idle worker.
  3151  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3152  			casgstatus(gp, _Gwaiting, _Grunnable)
  3153  			if traceEnabled() {
  3154  				traceGoUnpark(gp, 0)
  3155  			}
  3156  			return gp, false, false
  3157  		}
  3158  
  3159  		// Finally, check for timer creation or expiry concurrently with
  3160  		// transitioning from spinning to non-spinning.
  3161  		//
  3162  		// Note that we cannot use checkTimers here because it calls
  3163  		// adjusttimers which may need to allocate memory, and that isn't
  3164  		// allowed when we don't have an active P.
  3165  		pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
  3166  	}
  3167  
  3168  	// Poll network until next timer.
  3169  	if netpollinited() && (netpollWaiters.Load() > 0 || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
  3170  		sched.pollUntil.Store(pollUntil)
  3171  		if mp.p != 0 {
  3172  			throw("findrunnable: netpoll with p")
  3173  		}
  3174  		if mp.spinning {
  3175  			throw("findrunnable: netpoll with spinning")
  3176  		}
  3177  		delay := int64(-1)
  3178  		if pollUntil != 0 {
  3179  			if now == 0 {
  3180  				now = nanotime()
  3181  			}
  3182  			delay = pollUntil - now
  3183  			if delay < 0 {
  3184  				delay = 0
  3185  			}
  3186  		}
  3187  		if faketime != 0 {
  3188  			// When using fake time, just poll.
  3189  			delay = 0
  3190  		}
  3191  		list := netpoll(delay) // block until new work is available
  3192  		// Refresh now again, after potentially blocking.
  3193  		now = nanotime()
  3194  		sched.pollUntil.Store(0)
  3195  		sched.lastpoll.Store(now)
  3196  		if faketime != 0 && list.empty() {
  3197  			// Using fake time and nothing is ready; stop M.
  3198  			// When all M's stop, checkdead will call timejump.
  3199  			stopm()
  3200  			goto top
  3201  		}
  3202  		lock(&sched.lock)
  3203  		pp, _ := pidleget(now)
  3204  		unlock(&sched.lock)
  3205  		if pp == nil {
  3206  			injectglist(&list)
  3207  		} else {
  3208  			acquirep(pp)
  3209  			if !list.empty() {
  3210  				gp := list.pop()
  3211  				injectglist(&list)
  3212  				casgstatus(gp, _Gwaiting, _Grunnable)
  3213  				if traceEnabled() {
  3214  					traceGoUnpark(gp, 0)
  3215  				}
  3216  				return gp, false, false
  3217  			}
  3218  			if wasSpinning {
  3219  				mp.becomeSpinning()
  3220  			}
  3221  			goto top
  3222  		}
  3223  	} else if pollUntil != 0 && netpollinited() {
  3224  		pollerPollUntil := sched.pollUntil.Load()
  3225  		if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
  3226  			netpollBreak()
  3227  		}
  3228  	}
  3229  	stopm()
  3230  	goto top
  3231  }
  3232  
  3233  // pollWork reports whether there is non-background work this P could
  3234  // be doing. This is a fairly lightweight check to be used for
  3235  // background work loops, like idle GC. It checks a subset of the
  3236  // conditions checked by the actual scheduler.
  3237  func pollWork() bool {
  3238  	if sched.runqsize != 0 {
  3239  		return true
  3240  	}
  3241  	p := getg().m.p.ptr()
  3242  	if !runqempty(p) {
  3243  		return true
  3244  	}
  3245  	if netpollinited() && netpollWaiters.Load() > 0 && sched.lastpoll.Load() != 0 {
  3246  		if list := netpoll(0); !list.empty() {
  3247  			injectglist(&list)
  3248  			return true
  3249  		}
  3250  	}
  3251  	return false
  3252  }
  3253  
  3254  // stealWork attempts to steal a runnable goroutine or timer from any P.
  3255  //
  3256  // If newWork is true, new work may have been readied.
  3257  //
  3258  // If now is not 0 it is the current time. stealWork returns the passed time or
  3259  // the current time if now was passed as 0.
  3260  func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
  3261  	pp := getg().m.p.ptr()
  3262  
  3263  	ranTimer := false
  3264  
  3265  	const stealTries = 4
  3266  	for i := 0; i < stealTries; i++ {
  3267  		stealTimersOrRunNextG := i == stealTries-1
  3268  
  3269  		for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
  3270  			if sched.gcwaiting.Load() {
  3271  				// GC work may be available.
  3272  				return nil, false, now, pollUntil, true
  3273  			}
  3274  			p2 := allp[enum.position()]
  3275  			if pp == p2 {
  3276  				continue
  3277  			}
  3278  
  3279  			// Steal timers from p2. This call to checkTimers is the only place
  3280  			// where we might hold a lock on a different P's timers. We do this
  3281  			// once on the last pass before checking runnext because stealing
  3282  			// from the other P's runnext should be the last resort, so if there
  3283  			// are timers to steal do that first.
  3284  			//
  3285  			// We only check timers on one of the stealing iterations because
  3286  			// the time stored in now doesn't change in this loop and checking
  3287  			// the timers for each P more than once with the same value of now
  3288  			// is probably a waste of time.
  3289  			//
  3290  			// timerpMask tells us whether the P may have timers at all. If it
  3291  			// can't, no need to check at all.
  3292  			if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
  3293  				tnow, w, ran := checkTimers(p2, now)
  3294  				now = tnow
  3295  				if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3296  					pollUntil = w
  3297  				}
  3298  				if ran {
  3299  					// Running the timers may have
  3300  					// made an arbitrary number of G's
  3301  					// ready and added them to this P's
  3302  					// local run queue. That invalidates
  3303  					// the assumption of runqsteal
  3304  					// that it always has room to add
  3305  					// stolen G's. So check now if there
  3306  					// is a local G to run.
  3307  					if gp, inheritTime := runqget(pp); gp != nil {
  3308  						return gp, inheritTime, now, pollUntil, ranTimer
  3309  					}
  3310  					ranTimer = true
  3311  				}
  3312  			}
  3313  
  3314  			// Don't bother to attempt to steal if p2 is idle.
  3315  			if !idlepMask.read(enum.position()) {
  3316  				if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
  3317  					return gp, false, now, pollUntil, ranTimer
  3318  				}
  3319  			}
  3320  		}
  3321  	}
  3322  
  3323  	// No goroutines found to steal. Regardless, running a timer may have
  3324  	// made some goroutine ready that we missed. Indicate the next timer to
  3325  	// wait for.
  3326  	return nil, false, now, pollUntil, ranTimer
  3327  }
  3328  
  3329  // Check all Ps for a runnable G to steal.
  3330  //
  3331  // On entry we have no P. If a G is available to steal and a P is available,
  3332  // the P is returned which the caller should acquire and attempt to steal the
  3333  // work to.
  3334  func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
  3335  	for id, p2 := range allpSnapshot {
  3336  		if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
  3337  			lock(&sched.lock)
  3338  			pp, _ := pidlegetSpinning(0)
  3339  			if pp == nil {
  3340  				// Can't get a P, don't bother checking remaining Ps.
  3341  				unlock(&sched.lock)
  3342  				return nil
  3343  			}
  3344  			unlock(&sched.lock)
  3345  			return pp
  3346  		}
  3347  	}
  3348  
  3349  	// No work available.
  3350  	return nil
  3351  }
  3352  
  3353  // Check all Ps for a timer expiring sooner than pollUntil.
  3354  //
  3355  // Returns updated pollUntil value.
  3356  func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
  3357  	for id, p2 := range allpSnapshot {
  3358  		if timerpMaskSnapshot.read(uint32(id)) {
  3359  			w := nobarrierWakeTime(p2)
  3360  			if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3361  				pollUntil = w
  3362  			}
  3363  		}
  3364  	}
  3365  
  3366  	return pollUntil
  3367  }
  3368  
  3369  // Check for idle-priority GC, without a P on entry.
  3370  //
  3371  // If some GC work, a P, and a worker G are all available, the P and G will be
  3372  // returned. The returned P has not been wired yet.
  3373  func checkIdleGCNoP() (*p, *g) {
  3374  	// N.B. Since we have no P, gcBlackenEnabled may change at any time; we
  3375  	// must check again after acquiring a P. As an optimization, we also check
  3376  	// if an idle mark worker is needed at all. This is OK here, because if we
  3377  	// observe that one isn't needed, at least one is currently running. Even if
  3378  	// it stops running, its own journey into the scheduler should schedule it
  3379  	// again, if need be (at which point, this check will pass, if relevant).
  3380  	if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
  3381  		return nil, nil
  3382  	}
  3383  	if !gcMarkWorkAvailable(nil) {
  3384  		return nil, nil
  3385  	}
  3386  
  3387  	// Work is available; we can start an idle GC worker only if there is
  3388  	// an available P and available worker G.
  3389  	//
  3390  	// We can attempt to acquire these in either order, though both have
  3391  	// synchronization concerns (see below). Workers are almost always
  3392  	// available (see comment in findRunnableGCWorker for the one case
  3393  	// there may be none). Since we're slightly less likely to find a P,
  3394  	// check for that first.
  3395  	//
  3396  	// Synchronization: note that we must hold sched.lock until we are
  3397  	// committed to keeping it. Otherwise we cannot put the unnecessary P
  3398  	// back in sched.pidle without performing the full set of idle
  3399  	// transition checks.
  3400  	//
  3401  	// If we were to check gcBgMarkWorkerPool first, we must somehow handle
  3402  	// the assumption in gcControllerState.findRunnableGCWorker that an
  3403  	// empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
  3404  	lock(&sched.lock)
  3405  	pp, now := pidlegetSpinning(0)
  3406  	if pp == nil {
  3407  		unlock(&sched.lock)
  3408  		return nil, nil
  3409  	}
  3410  
  3411  	// Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
  3412  	if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
  3413  		pidleput(pp, now)
  3414  		unlock(&sched.lock)
  3415  		return nil, nil
  3416  	}
  3417  
  3418  	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3419  	if node == nil {
  3420  		pidleput(pp, now)
  3421  		unlock(&sched.lock)
  3422  		gcController.removeIdleMarkWorker()
  3423  		return nil, nil
  3424  	}
  3425  
  3426  	unlock(&sched.lock)
  3427  
  3428  	return pp, node.gp.ptr()
  3429  }
  3430  
  3431  // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
  3432  // going to wake up before the when argument; or it wakes an idle P to service
  3433  // timers and the network poller if there isn't one already.
  3434  func wakeNetPoller(when int64) {
  3435  	if sched.lastpoll.Load() == 0 {
  3436  		// In findrunnable we ensure that when polling the pollUntil
  3437  		// field is either zero or the time to which the current
  3438  		// poll is expected to run. This can have a spurious wakeup
  3439  		// but should never miss a wakeup.
  3440  		pollerPollUntil := sched.pollUntil.Load()
  3441  		if pollerPollUntil == 0 || pollerPollUntil > when {
  3442  			netpollBreak()
  3443  		}
  3444  	} else {
  3445  		// There are no threads in the network poller, try to get
  3446  		// one there so it can handle new timers.
  3447  		if GOOS != "plan9" { // Temporary workaround - see issue #42303.
  3448  			wakep()
  3449  		}
  3450  	}
  3451  }
  3452  
  3453  func resetspinning() {
  3454  	gp := getg()
  3455  	if !gp.m.spinning {
  3456  		throw("resetspinning: not a spinning m")
  3457  	}
  3458  	gp.m.spinning = false
  3459  	nmspinning := sched.nmspinning.Add(-1)
  3460  	if nmspinning < 0 {
  3461  		throw("findrunnable: negative nmspinning")
  3462  	}
  3463  	// M wakeup policy is deliberately somewhat conservative, so check if we
  3464  	// need to wakeup another P here. See "Worker thread parking/unparking"
  3465  	// comment at the top of the file for details.
  3466  	wakep()
  3467  }
  3468  
  3469  // injectglist adds each runnable G on the list to some run queue,
  3470  // and clears glist. If there is no current P, they are added to the
  3471  // global queue, and up to npidle M's are started to run them.
  3472  // Otherwise, for each idle P, this adds a G to the global queue
  3473  // and starts an M. Any remaining G's are added to the current P's
  3474  // local run queue.
  3475  // This may temporarily acquire sched.lock.
  3476  // Can run concurrently with GC.
  3477  func injectglist(glist *gList) {
  3478  	if glist.empty() {
  3479  		return
  3480  	}
  3481  	if traceEnabled() {
  3482  		for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
  3483  			traceGoUnpark(gp, 0)
  3484  		}
  3485  	}
  3486  
  3487  	// Mark all the goroutines as runnable before we put them
  3488  	// on the run queues.
  3489  	head := glist.head.ptr()
  3490  	var tail *g
  3491  	qsize := 0
  3492  	for gp := head; gp != nil; gp = gp.schedlink.ptr() {
  3493  		tail = gp
  3494  		qsize++
  3495  		casgstatus(gp, _Gwaiting, _Grunnable)
  3496  	}
  3497  
  3498  	// Turn the gList into a gQueue.
  3499  	var q gQueue
  3500  	q.head.set(head)
  3501  	q.tail.set(tail)
  3502  	*glist = gList{}
  3503  
  3504  	startIdle := func(n int) {
  3505  		for i := 0; i < n; i++ {
  3506  			mp := acquirem() // See comment in startm.
  3507  			lock(&sched.lock)
  3508  
  3509  			pp, _ := pidlegetSpinning(0)
  3510  			if pp == nil {
  3511  				unlock(&sched.lock)
  3512  				releasem(mp)
  3513  				break
  3514  			}
  3515  
  3516  			startm(pp, false, true)
  3517  			unlock(&sched.lock)
  3518  			releasem(mp)
  3519  		}
  3520  	}
  3521  
  3522  	pp := getg().m.p.ptr()
  3523  	if pp == nil {
  3524  		lock(&sched.lock)
  3525  		globrunqputbatch(&q, int32(qsize))
  3526  		unlock(&sched.lock)
  3527  		startIdle(qsize)
  3528  		return
  3529  	}
  3530  
  3531  	npidle := int(sched.npidle.Load())
  3532  	var globq gQueue
  3533  	var n int
  3534  	for n = 0; n < npidle && !q.empty(); n++ {
  3535  		g := q.pop()
  3536  		globq.pushBack(g)
  3537  	}
  3538  	if n > 0 {
  3539  		lock(&sched.lock)
  3540  		globrunqputbatch(&globq, int32(n))
  3541  		unlock(&sched.lock)
  3542  		startIdle(n)
  3543  		qsize -= n
  3544  	}
  3545  
  3546  	if !q.empty() {
  3547  		runqputbatch(pp, &q, qsize)
  3548  	}
  3549  }
  3550  
  3551  // One round of scheduler: find a runnable goroutine and execute it.
  3552  // Never returns.
  3553  func schedule() {
  3554  	mp := getg().m
  3555  
  3556  	if mp.locks != 0 {
  3557  		throw("schedule: holding locks")
  3558  	}
  3559  
  3560  	if mp.lockedg != 0 {
  3561  		stoplockedm()
  3562  		execute(mp.lockedg.ptr(), false) // Never returns.
  3563  	}
  3564  
  3565  	// We should not schedule away from a g that is executing a cgo call,
  3566  	// since the cgo call is using the m's g0 stack.
  3567  	if mp.incgo {
  3568  		throw("schedule: in cgo")
  3569  	}
  3570  
  3571  top:
  3572  	pp := mp.p.ptr()
  3573  	pp.preempt = false
  3574  
  3575  	// Safety check: if we are spinning, the run queue should be empty.
  3576  	// Check this before calling checkTimers, as that might call
  3577  	// goready to put a ready goroutine on the local run queue.
  3578  	if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
  3579  		throw("schedule: spinning with local work")
  3580  	}
  3581  
  3582  	gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
  3583  
  3584  	if debug.dontfreezetheworld > 0 && freezing.Load() {
  3585  		// See comment in freezetheworld. We don't want to perturb
  3586  		// scheduler state, so we didn't gcstopm in findRunnable, but
  3587  		// also don't want to allow new goroutines to run.
  3588  		//
  3589  		// Deadlock here rather than in the findRunnable loop so if
  3590  		// findRunnable is stuck in a loop we don't perturb that
  3591  		// either.
  3592  		lock(&deadlock)
  3593  		lock(&deadlock)
  3594  	}
  3595  
  3596  	// This thread is going to run a goroutine and is not spinning anymore,
  3597  	// so if it was marked as spinning we need to reset it now and potentially
  3598  	// start a new spinning M.
  3599  	if mp.spinning {
  3600  		resetspinning()
  3601  	}
  3602  
  3603  	if sched.disable.user && !schedEnabled(gp) {
  3604  		// Scheduling of this goroutine is disabled. Put it on
  3605  		// the list of pending runnable goroutines for when we
  3606  		// re-enable user scheduling and look again.
  3607  		lock(&sched.lock)
  3608  		if schedEnabled(gp) {
  3609  			// Something re-enabled scheduling while we
  3610  			// were acquiring the lock.
  3611  			unlock(&sched.lock)
  3612  		} else {
  3613  			sched.disable.runnable.pushBack(gp)
  3614  			sched.disable.n++
  3615  			unlock(&sched.lock)
  3616  			goto top
  3617  		}
  3618  	}
  3619  
  3620  	// If about to schedule a not-normal goroutine (a GCworker or tracereader),
  3621  	// wake a P if there is one.
  3622  	if tryWakeP {
  3623  		wakep()
  3624  	}
  3625  	if gp.lockedm != 0 {
  3626  		// Hands off own p to the locked m,
  3627  		// then blocks waiting for a new p.
  3628  		startlockedm(gp)
  3629  		goto top
  3630  	}
  3631  
  3632  	execute(gp, inheritTime)
  3633  }
  3634  
  3635  // dropg removes the association between m and the current goroutine m->curg (gp for short).
  3636  // Typically a caller sets gp's status away from Grunning and then
  3637  // immediately calls dropg to finish the job. The caller is also responsible
  3638  // for arranging that gp will be restarted using ready at an
  3639  // appropriate time. After calling dropg and arranging for gp to be
  3640  // readied later, the caller can do other work but eventually should
  3641  // call schedule to restart the scheduling of goroutines on this m.
  3642  func dropg() {
  3643  	gp := getg()
  3644  
  3645  	setMNoWB(&gp.m.curg.m, nil)
  3646  	setGNoWB(&gp.m.curg, nil)
  3647  }
  3648  
  3649  // checkTimers runs any timers for the P that are ready.
  3650  // If now is not 0 it is the current time.
  3651  // It returns the passed time or the current time if now was passed as 0.
  3652  // and the time when the next timer should run or 0 if there is no next timer,
  3653  // and reports whether it ran any timers.
  3654  // If the time when the next timer should run is not 0,
  3655  // it is always larger than the returned time.
  3656  // We pass now in and out to avoid extra calls of nanotime.
  3657  //
  3658  //go:yeswritebarrierrec
  3659  func checkTimers(pp *p, now int64) (rnow, pollUntil int64, ran bool) {
  3660  	// If it's not yet time for the first timer, or the first adjusted
  3661  	// timer, then there is nothing to do.
  3662  	next := pp.timer0When.Load()
  3663  	nextAdj := pp.timerModifiedEarliest.Load()
  3664  	if next == 0 || (nextAdj != 0 && nextAdj < next) {
  3665  		next = nextAdj
  3666  	}
  3667  
  3668  	if next == 0 {
  3669  		// No timers to run or adjust.
  3670  		return now, 0, false
  3671  	}
  3672  
  3673  	if now == 0 {
  3674  		now = nanotime()
  3675  	}
  3676  	if now < next {
  3677  		// Next timer is not ready to run, but keep going
  3678  		// if we would clear deleted timers.
  3679  		// This corresponds to the condition below where
  3680  		// we decide whether to call clearDeletedTimers.
  3681  		if pp != getg().m.p.ptr() || int(pp.deletedTimers.Load()) <= int(pp.numTimers.Load()/4) {
  3682  			return now, next, false
  3683  		}
  3684  	}
  3685  
  3686  	lock(&pp.timersLock)
  3687  
  3688  	if len(pp.timers) > 0 {
  3689  		adjusttimers(pp, now)
  3690  		for len(pp.timers) > 0 {
  3691  			// Note that runtimer may temporarily unlock
  3692  			// pp.timersLock.
  3693  			if tw := runtimer(pp, now); tw != 0 {
  3694  				if tw > 0 {
  3695  					pollUntil = tw
  3696  				}
  3697  				break
  3698  			}
  3699  			ran = true
  3700  		}
  3701  	}
  3702  
  3703  	// If this is the local P, and there are a lot of deleted timers,
  3704  	// clear them out. We only do this for the local P to reduce
  3705  	// lock contention on timersLock.
  3706  	if pp == getg().m.p.ptr() && int(pp.deletedTimers.Load()) > len(pp.timers)/4 {
  3707  		clearDeletedTimers(pp)
  3708  	}
  3709  
  3710  	unlock(&pp.timersLock)
  3711  
  3712  	return now, pollUntil, ran
  3713  }
  3714  
  3715  func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
  3716  	unlock((*mutex)(lock))
  3717  	return true
  3718  }
  3719  
  3720  // park continuation on g0.
  3721  func park_m(gp *g) {
  3722  	mp := getg().m
  3723  
  3724  	if traceEnabled() {
  3725  		traceGoPark(mp.waitTraceBlockReason, mp.waitTraceSkip)
  3726  	}
  3727  
  3728  	// N.B. Not using casGToWaiting here because the waitreason is
  3729  	// set by park_m's caller.
  3730  	casgstatus(gp, _Grunning, _Gwaiting)
  3731  	dropg()
  3732  
  3733  	if fn := mp.waitunlockf; fn != nil {
  3734  		ok := fn(gp, mp.waitlock)
  3735  		mp.waitunlockf = nil
  3736  		mp.waitlock = nil
  3737  		if !ok {
  3738  			if traceEnabled() {
  3739  				traceGoUnpark(gp, 2)
  3740  			}
  3741  			casgstatus(gp, _Gwaiting, _Grunnable)
  3742  			execute(gp, true) // Schedule it back, never returns.
  3743  		}
  3744  	}
  3745  	schedule()
  3746  }
  3747  
  3748  func goschedImpl(gp *g) {
  3749  	status := readgstatus(gp)
  3750  	if status&^_Gscan != _Grunning {
  3751  		dumpgstatus(gp)
  3752  		throw("bad g status")
  3753  	}
  3754  	casgstatus(gp, _Grunning, _Grunnable)
  3755  	dropg()
  3756  	lock(&sched.lock)
  3757  	globrunqput(gp)
  3758  	unlock(&sched.lock)
  3759  
  3760  	schedule()
  3761  }
  3762  
  3763  // Gosched continuation on g0.
  3764  func gosched_m(gp *g) {
  3765  	if traceEnabled() {
  3766  		traceGoSched()
  3767  	}
  3768  	goschedImpl(gp)
  3769  }
  3770  
  3771  // goschedguarded is a forbidden-states-avoided version of gosched_m.
  3772  func goschedguarded_m(gp *g) {
  3773  
  3774  	if !canPreemptM(gp.m) {
  3775  		gogo(&gp.sched) // never return
  3776  	}
  3777  
  3778  	if traceEnabled() {
  3779  		traceGoSched()
  3780  	}
  3781  	goschedImpl(gp)
  3782  }
  3783  
  3784  func gopreempt_m(gp *g) {
  3785  	if traceEnabled() {
  3786  		traceGoPreempt()
  3787  	}
  3788  	goschedImpl(gp)
  3789  }
  3790  
  3791  // preemptPark parks gp and puts it in _Gpreempted.
  3792  //
  3793  //go:systemstack
  3794  func preemptPark(gp *g) {
  3795  	if traceEnabled() {
  3796  		traceGoPark(traceBlockPreempted, 0)
  3797  	}
  3798  	status := readgstatus(gp)
  3799  	if status&^_Gscan != _Grunning {
  3800  		dumpgstatus(gp)
  3801  		throw("bad g status")
  3802  	}
  3803  
  3804  	if gp.asyncSafePoint {
  3805  		// Double-check that async preemption does not
  3806  		// happen in SPWRITE assembly functions.
  3807  		// isAsyncSafePoint must exclude this case.
  3808  		f := findfunc(gp.sched.pc)
  3809  		if !f.valid() {
  3810  			throw("preempt at unknown pc")
  3811  		}
  3812  		if f.flag&abi.FuncFlagSPWrite != 0 {
  3813  			println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
  3814  			throw("preempt SPWRITE")
  3815  		}
  3816  	}
  3817  
  3818  	// Transition from _Grunning to _Gscan|_Gpreempted. We can't
  3819  	// be in _Grunning when we dropg because then we'd be running
  3820  	// without an M, but the moment we're in _Gpreempted,
  3821  	// something could claim this G before we've fully cleaned it
  3822  	// up. Hence, we set the scan bit to lock down further
  3823  	// transitions until we can dropg.
  3824  	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
  3825  	dropg()
  3826  	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
  3827  	schedule()
  3828  }
  3829  
  3830  // goyield is like Gosched, but it:
  3831  // - emits a GoPreempt trace event instead of a GoSched trace event
  3832  // - puts the current G on the runq of the current P instead of the globrunq
  3833  func goyield() {
  3834  	checkTimeouts()
  3835  	mcall(goyield_m)
  3836  }
  3837  
  3838  func goyield_m(gp *g) {
  3839  	if traceEnabled() {
  3840  		traceGoPreempt()
  3841  	}
  3842  	pp := gp.m.p.ptr()
  3843  	casgstatus(gp, _Grunning, _Grunnable)
  3844  	dropg()
  3845  	runqput(pp, gp, false)
  3846  	schedule()
  3847  }
  3848  
  3849  // Finishes execution of the current goroutine.
  3850  func goexit1() {
  3851  	if raceenabled {
  3852  		racegoend()
  3853  	}
  3854  	if traceEnabled() {
  3855  		traceGoEnd()
  3856  	}
  3857  	mcall(goexit0)
  3858  }
  3859  
  3860  // goexit continuation on g0.
  3861  func goexit0(gp *g) {
  3862  	mp := getg().m
  3863  	pp := mp.p.ptr()
  3864  
  3865  	casgstatus(gp, _Grunning, _Gdead)
  3866  	gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
  3867  	if isSystemGoroutine(gp, false) {
  3868  		sched.ngsys.Add(-1)
  3869  	}
  3870  	gp.m = nil
  3871  	locked := gp.lockedm != 0
  3872  	gp.lockedm = 0
  3873  	mp.lockedg = 0
  3874  	gp.preemptStop = false
  3875  	gp.paniconfault = false
  3876  	gp._defer = nil // should be true already but just in case.
  3877  	gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
  3878  	gp.writebuf = nil
  3879  	gp.waitreason = waitReasonZero
  3880  	gp.param = nil
  3881  	gp.labels = nil
  3882  	gp.timer = nil
  3883  
  3884  	if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
  3885  		// Flush assist credit to the global pool. This gives
  3886  		// better information to pacing if the application is
  3887  		// rapidly creating an exiting goroutines.
  3888  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
  3889  		scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
  3890  		gcController.bgScanCredit.Add(scanCredit)
  3891  		gp.gcAssistBytes = 0
  3892  	}
  3893  
  3894  	dropg()
  3895  
  3896  	if GOARCH == "wasm" { // no threads yet on wasm
  3897  		gfput(pp, gp)
  3898  		schedule() // never returns
  3899  	}
  3900  
  3901  	if mp.lockedInt != 0 {
  3902  		print("invalid m->lockedInt = ", mp.lockedInt, "\n")
  3903  		throw("internal lockOSThread error")
  3904  	}
  3905  	gfput(pp, gp)
  3906  	if locked {
  3907  		// The goroutine may have locked this thread because
  3908  		// it put it in an unusual kernel state. Kill it
  3909  		// rather than returning it to the thread pool.
  3910  
  3911  		// Return to mstart, which will release the P and exit
  3912  		// the thread.
  3913  		if GOOS != "plan9" { // See golang.org/issue/22227.
  3914  			gogo(&mp.g0.sched)
  3915  		} else {
  3916  			// Clear lockedExt on plan9 since we may end up re-using
  3917  			// this thread.
  3918  			mp.lockedExt = 0
  3919  		}
  3920  	}
  3921  	schedule()
  3922  }
  3923  
  3924  // save updates getg().sched to refer to pc and sp so that a following
  3925  // gogo will restore pc and sp.
  3926  //
  3927  // save must not have write barriers because invoking a write barrier
  3928  // can clobber getg().sched.
  3929  //
  3930  //go:nosplit
  3931  //go:nowritebarrierrec
  3932  func save(pc, sp uintptr) {
  3933  	gp := getg()
  3934  
  3935  	if gp == gp.m.g0 || gp == gp.m.gsignal {
  3936  		// m.g0.sched is special and must describe the context
  3937  		// for exiting the thread. mstart1 writes to it directly.
  3938  		// m.gsignal.sched should not be used at all.
  3939  		// This check makes sure save calls do not accidentally
  3940  		// run in contexts where they'd write to system g's.
  3941  		throw("save on system g not allowed")
  3942  	}
  3943  
  3944  	gp.sched.pc = pc
  3945  	gp.sched.sp = sp
  3946  	gp.sched.lr = 0
  3947  	gp.sched.ret = 0
  3948  	// We need to ensure ctxt is zero, but can't have a write
  3949  	// barrier here. However, it should always already be zero.
  3950  	// Assert that.
  3951  	if gp.sched.ctxt != nil {
  3952  		badctxt()
  3953  	}
  3954  }
  3955  
  3956  // The goroutine g is about to enter a system call.
  3957  // Record that it's not using the cpu anymore.
  3958  // This is called only from the go syscall library and cgocall,
  3959  // not from the low-level system calls used by the runtime.
  3960  //
  3961  // Entersyscall cannot split the stack: the save must
  3962  // make g->sched refer to the caller's stack segment, because
  3963  // entersyscall is going to return immediately after.
  3964  //
  3965  // Nothing entersyscall calls can split the stack either.
  3966  // We cannot safely move the stack during an active call to syscall,
  3967  // because we do not know which of the uintptr arguments are
  3968  // really pointers (back into the stack).
  3969  // In practice, this means that we make the fast path run through
  3970  // entersyscall doing no-split things, and the slow path has to use systemstack
  3971  // to run bigger things on the system stack.
  3972  //
  3973  // reentersyscall is the entry point used by cgo callbacks, where explicitly
  3974  // saved SP and PC are restored. This is needed when exitsyscall will be called
  3975  // from a function further up in the call stack than the parent, as g->syscallsp
  3976  // must always point to a valid stack frame. entersyscall below is the normal
  3977  // entry point for syscalls, which obtains the SP and PC from the caller.
  3978  //
  3979  // Syscall tracing:
  3980  // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
  3981  // If the syscall does not block, that is it, we do not emit any other events.
  3982  // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
  3983  // when syscall returns we emit traceGoSysExit and when the goroutine starts running
  3984  // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
  3985  // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
  3986  // we remember current value of syscalltick in m (gp.m.syscalltick = gp.m.p.ptr().syscalltick),
  3987  // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
  3988  // and we wait for the increment before emitting traceGoSysExit.
  3989  // Note that the increment is done even if tracing is not enabled,
  3990  // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
  3991  //
  3992  //go:nosplit
  3993  func reentersyscall(pc, sp uintptr) {
  3994  	gp := getg()
  3995  
  3996  	// Disable preemption because during this function g is in Gsyscall status,
  3997  	// but can have inconsistent g->sched, do not let GC observe it.
  3998  	gp.m.locks++
  3999  
  4000  	// Entersyscall must not call any function that might split/grow the stack.
  4001  	// (See details in comment above.)
  4002  	// Catch calls that might, by replacing the stack guard with something that
  4003  	// will trip any stack check and leaving a flag to tell newstack to die.
  4004  	gp.stackguard0 = stackPreempt
  4005  	gp.throwsplit = true
  4006  
  4007  	// Leave SP around for GC and traceback.
  4008  	save(pc, sp)
  4009  	gp.syscallsp = sp
  4010  	gp.syscallpc = pc
  4011  	casgstatus(gp, _Grunning, _Gsyscall)
  4012  	if staticLockRanking {
  4013  		// When doing static lock ranking casgstatus can call
  4014  		// systemstack which clobbers g.sched.
  4015  		save(pc, sp)
  4016  	}
  4017  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4018  		systemstack(func() {
  4019  			print("entersyscall inconsistent ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4020  			throw("entersyscall")
  4021  		})
  4022  	}
  4023  
  4024  	if traceEnabled() {
  4025  		systemstack(traceGoSysCall)
  4026  		// systemstack itself clobbers g.sched.{pc,sp} and we might
  4027  		// need them later when the G is genuinely blocked in a
  4028  		// syscall
  4029  		save(pc, sp)
  4030  	}
  4031  
  4032  	if sched.sysmonwait.Load() {
  4033  		systemstack(entersyscall_sysmon)
  4034  		save(pc, sp)
  4035  	}
  4036  
  4037  	if gp.m.p.ptr().runSafePointFn != 0 {
  4038  		// runSafePointFn may stack split if run on this stack
  4039  		systemstack(runSafePointFn)
  4040  		save(pc, sp)
  4041  	}
  4042  
  4043  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4044  	pp := gp.m.p.ptr()
  4045  	pp.m = 0
  4046  	gp.m.oldp.set(pp)
  4047  	gp.m.p = 0
  4048  	atomic.Store(&pp.status, _Psyscall)
  4049  	if sched.gcwaiting.Load() {
  4050  		systemstack(entersyscall_gcwait)
  4051  		save(pc, sp)
  4052  	}
  4053  
  4054  	gp.m.locks--
  4055  }
  4056  
  4057  // Standard syscall entry used by the go syscall library and normal cgo calls.
  4058  //
  4059  // This is exported via linkname to assembly in the syscall package and x/sys.
  4060  //
  4061  //go:nosplit
  4062  //go:linkname entersyscall
  4063  func entersyscall() {
  4064  	reentersyscall(getcallerpc(), getcallersp())
  4065  }
  4066  
  4067  func entersyscall_sysmon() {
  4068  	lock(&sched.lock)
  4069  	if sched.sysmonwait.Load() {
  4070  		sched.sysmonwait.Store(false)
  4071  		notewakeup(&sched.sysmonnote)
  4072  	}
  4073  	unlock(&sched.lock)
  4074  }
  4075  
  4076  func entersyscall_gcwait() {
  4077  	gp := getg()
  4078  	pp := gp.m.oldp.ptr()
  4079  
  4080  	lock(&sched.lock)
  4081  	if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
  4082  		if traceEnabled() {
  4083  			traceGoSysBlock(pp)
  4084  			traceProcStop(pp)
  4085  		}
  4086  		pp.syscalltick++
  4087  		if sched.stopwait--; sched.stopwait == 0 {
  4088  			notewakeup(&sched.stopnote)
  4089  		}
  4090  	}
  4091  	unlock(&sched.lock)
  4092  }
  4093  
  4094  // The same as entersyscall(), but with a hint that the syscall is blocking.
  4095  //
  4096  //go:nosplit
  4097  func entersyscallblock() {
  4098  	gp := getg()
  4099  
  4100  	gp.m.locks++ // see comment in entersyscall
  4101  	gp.throwsplit = true
  4102  	gp.stackguard0 = stackPreempt // see comment in entersyscall
  4103  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4104  	gp.m.p.ptr().syscalltick++
  4105  
  4106  	// Leave SP around for GC and traceback.
  4107  	pc := getcallerpc()
  4108  	sp := getcallersp()
  4109  	save(pc, sp)
  4110  	gp.syscallsp = gp.sched.sp
  4111  	gp.syscallpc = gp.sched.pc
  4112  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4113  		sp1 := sp
  4114  		sp2 := gp.sched.sp
  4115  		sp3 := gp.syscallsp
  4116  		systemstack(func() {
  4117  			print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4118  			throw("entersyscallblock")
  4119  		})
  4120  	}
  4121  	casgstatus(gp, _Grunning, _Gsyscall)
  4122  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4123  		systemstack(func() {
  4124  			print("entersyscallblock inconsistent ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4125  			throw("entersyscallblock")
  4126  		})
  4127  	}
  4128  
  4129  	systemstack(entersyscallblock_handoff)
  4130  
  4131  	// Resave for traceback during blocked call.
  4132  	save(getcallerpc(), getcallersp())
  4133  
  4134  	gp.m.locks--
  4135  }
  4136  
  4137  func entersyscallblock_handoff() {
  4138  	if traceEnabled() {
  4139  		traceGoSysCall()
  4140  		traceGoSysBlock(getg().m.p.ptr())
  4141  	}
  4142  	handoffp(releasep())
  4143  }
  4144  
  4145  // The goroutine g exited its system call.
  4146  // Arrange for it to run on a cpu again.
  4147  // This is called only from the go syscall library, not
  4148  // from the low-level system calls used by the runtime.
  4149  //
  4150  // Write barriers are not allowed because our P may have been stolen.
  4151  //
  4152  // This is exported via linkname to assembly in the syscall package.
  4153  //
  4154  //go:nosplit
  4155  //go:nowritebarrierrec
  4156  //go:linkname exitsyscall
  4157  func exitsyscall() {
  4158  	gp := getg()
  4159  
  4160  	gp.m.locks++ // see comment in entersyscall
  4161  	if getcallersp() > gp.syscallsp {
  4162  		throw("exitsyscall: syscall frame is no longer valid")
  4163  	}
  4164  
  4165  	gp.waitsince = 0
  4166  	oldp := gp.m.oldp.ptr()
  4167  	gp.m.oldp = 0
  4168  	if exitsyscallfast(oldp) {
  4169  		// When exitsyscallfast returns success, we have a P so can now use
  4170  		// write barriers
  4171  		if goroutineProfile.active {
  4172  			// Make sure that gp has had its stack written out to the goroutine
  4173  			// profile, exactly as it was when the goroutine profiler first
  4174  			// stopped the world.
  4175  			systemstack(func() {
  4176  				tryRecordGoroutineProfileWB(gp)
  4177  			})
  4178  		}
  4179  		if traceEnabled() {
  4180  			if oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  4181  				systemstack(traceGoStart)
  4182  			}
  4183  		}
  4184  		// There's a cpu for us, so we can run.
  4185  		gp.m.p.ptr().syscalltick++
  4186  		// We need to cas the status and scan before resuming...
  4187  		casgstatus(gp, _Gsyscall, _Grunning)
  4188  
  4189  		// Garbage collector isn't running (since we are),
  4190  		// so okay to clear syscallsp.
  4191  		gp.syscallsp = 0
  4192  		gp.m.locks--
  4193  		if gp.preempt {
  4194  			// restore the preemption request in case we've cleared it in newstack
  4195  			gp.stackguard0 = stackPreempt
  4196  		} else {
  4197  			// otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
  4198  			gp.stackguard0 = gp.stack.lo + stackGuard
  4199  		}
  4200  		gp.throwsplit = false
  4201  
  4202  		if sched.disable.user && !schedEnabled(gp) {
  4203  			// Scheduling of this goroutine is disabled.
  4204  			Gosched()
  4205  		}
  4206  
  4207  		return
  4208  	}
  4209  
  4210  	if traceEnabled() {
  4211  		// Wait till traceGoSysBlock event is emitted.
  4212  		// This ensures consistency of the trace (the goroutine is started after it is blocked).
  4213  		for oldp != nil && oldp.syscalltick == gp.m.syscalltick {
  4214  			osyield()
  4215  		}
  4216  		// We can't trace syscall exit right now because we don't have a P.
  4217  		// Tracing code can invoke write barriers that cannot run without a P.
  4218  		// So instead we remember the syscall exit time and emit the event
  4219  		// in execute when we have a P.
  4220  		gp.trace.sysExitTime = traceClockNow()
  4221  	}
  4222  
  4223  	gp.m.locks--
  4224  
  4225  	// Call the scheduler.
  4226  	mcall(exitsyscall0)
  4227  
  4228  	// Scheduler returned, so we're allowed to run now.
  4229  	// Delete the syscallsp information that we left for
  4230  	// the garbage collector during the system call.
  4231  	// Must wait until now because until gosched returns
  4232  	// we don't know for sure that the garbage collector
  4233  	// is not running.
  4234  	gp.syscallsp = 0
  4235  	gp.m.p.ptr().syscalltick++
  4236  	gp.throwsplit = false
  4237  }
  4238  
  4239  //go:nosplit
  4240  func exitsyscallfast(oldp *p) bool {
  4241  	gp := getg()
  4242  
  4243  	// Freezetheworld sets stopwait but does not retake P's.
  4244  	if sched.stopwait == freezeStopWait {
  4245  		return false
  4246  	}
  4247  
  4248  	// Try to re-acquire the last P.
  4249  	if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
  4250  		// There's a cpu for us, so we can run.
  4251  		wirep(oldp)
  4252  		exitsyscallfast_reacquired()
  4253  		return true
  4254  	}
  4255  
  4256  	// Try to get any other idle P.
  4257  	if sched.pidle != 0 {
  4258  		var ok bool
  4259  		systemstack(func() {
  4260  			ok = exitsyscallfast_pidle()
  4261  			if ok && traceEnabled() {
  4262  				if oldp != nil {
  4263  					// Wait till traceGoSysBlock event is emitted.
  4264  					// This ensures consistency of the trace (the goroutine is started after it is blocked).
  4265  					for oldp.syscalltick == gp.m.syscalltick {
  4266  						osyield()
  4267  					}
  4268  				}
  4269  				traceGoSysExit()
  4270  			}
  4271  		})
  4272  		if ok {
  4273  			return true
  4274  		}
  4275  	}
  4276  	return false
  4277  }
  4278  
  4279  // exitsyscallfast_reacquired is the exitsyscall path on which this G
  4280  // has successfully reacquired the P it was running on before the
  4281  // syscall.
  4282  //
  4283  //go:nosplit
  4284  func exitsyscallfast_reacquired() {
  4285  	gp := getg()
  4286  	if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  4287  		if traceEnabled() {
  4288  			// The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
  4289  			// traceGoSysBlock for this syscall was already emitted,
  4290  			// but here we effectively retake the p from the new syscall running on the same p.
  4291  			systemstack(func() {
  4292  				// Denote blocking of the new syscall.
  4293  				traceGoSysBlock(gp.m.p.ptr())
  4294  				// Denote completion of the current syscall.
  4295  				traceGoSysExit()
  4296  			})
  4297  		}
  4298  		gp.m.p.ptr().syscalltick++
  4299  	}
  4300  }
  4301  
  4302  func exitsyscallfast_pidle() bool {
  4303  	lock(&sched.lock)
  4304  	pp, _ := pidleget(0)
  4305  	if pp != nil && sched.sysmonwait.Load() {
  4306  		sched.sysmonwait.Store(false)
  4307  		notewakeup(&sched.sysmonnote)
  4308  	}
  4309  	unlock(&sched.lock)
  4310  	if pp != nil {
  4311  		acquirep(pp)
  4312  		return true
  4313  	}
  4314  	return false
  4315  }
  4316  
  4317  // exitsyscall slow path on g0.
  4318  // Failed to acquire P, enqueue gp as runnable.
  4319  //
  4320  // Called via mcall, so gp is the calling g from this M.
  4321  //
  4322  //go:nowritebarrierrec
  4323  func exitsyscall0(gp *g) {
  4324  	casgstatus(gp, _Gsyscall, _Grunnable)
  4325  	dropg()
  4326  	lock(&sched.lock)
  4327  	var pp *p
  4328  	if schedEnabled(gp) {
  4329  		pp, _ = pidleget(0)
  4330  	}
  4331  	var locked bool
  4332  	if pp == nil {
  4333  		globrunqput(gp)
  4334  
  4335  		// Below, we stoplockedm if gp is locked. globrunqput releases
  4336  		// ownership of gp, so we must check if gp is locked prior to
  4337  		// committing the release by unlocking sched.lock, otherwise we
  4338  		// could race with another M transitioning gp from unlocked to
  4339  		// locked.
  4340  		locked = gp.lockedm != 0
  4341  	} else if sched.sysmonwait.Load() {
  4342  		sched.sysmonwait.Store(false)
  4343  		notewakeup(&sched.sysmonnote)
  4344  	}
  4345  	unlock(&sched.lock)
  4346  	if pp != nil {
  4347  		acquirep(pp)
  4348  		execute(gp, false) // Never returns.
  4349  	}
  4350  	if locked {
  4351  		// Wait until another thread schedules gp and so m again.
  4352  		//
  4353  		// N.B. lockedm must be this M, as this g was running on this M
  4354  		// before entersyscall.
  4355  		stoplockedm()
  4356  		execute(gp, false) // Never returns.
  4357  	}
  4358  	stopm()
  4359  	schedule() // Never returns.
  4360  }
  4361  
  4362  // Called from syscall package before fork.
  4363  //
  4364  //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
  4365  //go:nosplit
  4366  func syscall_runtime_BeforeFork() {
  4367  	gp := getg().m.curg
  4368  
  4369  	// Block signals during a fork, so that the child does not run
  4370  	// a signal handler before exec if a signal is sent to the process
  4371  	// group. See issue #18600.
  4372  	gp.m.locks++
  4373  	sigsave(&gp.m.sigmask)
  4374  	sigblock(false)
  4375  
  4376  	// This function is called before fork in syscall package.
  4377  	// Code between fork and exec must not allocate memory nor even try to grow stack.
  4378  	// Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
  4379  	// runtime_AfterFork will undo this in parent process, but not in child.
  4380  	gp.stackguard0 = stackFork
  4381  }
  4382  
  4383  // Called from syscall package after fork in parent.
  4384  //
  4385  //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
  4386  //go:nosplit
  4387  func syscall_runtime_AfterFork() {
  4388  	gp := getg().m.curg
  4389  
  4390  	// See the comments in beforefork.
  4391  	gp.stackguard0 = gp.stack.lo + stackGuard
  4392  
  4393  	msigrestore(gp.m.sigmask)
  4394  
  4395  	gp.m.locks--
  4396  }
  4397  
  4398  // inForkedChild is true while manipulating signals in the child process.
  4399  // This is used to avoid calling libc functions in case we are using vfork.
  4400  var inForkedChild bool
  4401  
  4402  // Called from syscall package after fork in child.
  4403  // It resets non-sigignored signals to the default handler, and
  4404  // restores the signal mask in preparation for the exec.
  4405  //
  4406  // Because this might be called during a vfork, and therefore may be
  4407  // temporarily sharing address space with the parent process, this must
  4408  // not change any global variables or calling into C code that may do so.
  4409  //
  4410  //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
  4411  //go:nosplit
  4412  //go:nowritebarrierrec
  4413  func syscall_runtime_AfterForkInChild() {
  4414  	// It's OK to change the global variable inForkedChild here
  4415  	// because we are going to change it back. There is no race here,
  4416  	// because if we are sharing address space with the parent process,
  4417  	// then the parent process can not be running concurrently.
  4418  	inForkedChild = true
  4419  
  4420  	clearSignalHandlers()
  4421  
  4422  	// When we are the child we are the only thread running,
  4423  	// so we know that nothing else has changed gp.m.sigmask.
  4424  	msigrestore(getg().m.sigmask)
  4425  
  4426  	inForkedChild = false
  4427  }
  4428  
  4429  // pendingPreemptSignals is the number of preemption signals
  4430  // that have been sent but not received. This is only used on Darwin.
  4431  // For #41702.
  4432  var pendingPreemptSignals atomic.Int32
  4433  
  4434  // Called from syscall package before Exec.
  4435  //
  4436  //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
  4437  func syscall_runtime_BeforeExec() {
  4438  	// Prevent thread creation during exec.
  4439  	execLock.lock()
  4440  
  4441  	// On Darwin, wait for all pending preemption signals to
  4442  	// be received. See issue #41702.
  4443  	if GOOS == "darwin" || GOOS == "ios" {
  4444  		for pendingPreemptSignals.Load() > 0 {
  4445  			osyield()
  4446  		}
  4447  	}
  4448  }
  4449  
  4450  // Called from syscall package after Exec.
  4451  //
  4452  //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
  4453  func syscall_runtime_AfterExec() {
  4454  	execLock.unlock()
  4455  }
  4456  
  4457  // Allocate a new g, with a stack big enough for stacksize bytes.
  4458  func malg(stacksize int32) *g {
  4459  	newg := new(g)
  4460  	if stacksize >= 0 {
  4461  		stacksize = round2(stackSystem + stacksize)
  4462  		systemstack(func() {
  4463  			newg.stack = stackalloc(uint32(stacksize))
  4464  		})
  4465  		newg.stackguard0 = newg.stack.lo + stackGuard
  4466  		newg.stackguard1 = ^uintptr(0)
  4467  		// Clear the bottom word of the stack. We record g
  4468  		// there on gsignal stack during VDSO on ARM and ARM64.
  4469  		*(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
  4470  	}
  4471  	return newg
  4472  }
  4473  
  4474  // Create a new g running fn.
  4475  // Put it on the queue of g's waiting to run.
  4476  // The compiler turns a go statement into a call to this.
  4477  func newproc(fn *funcval) {
  4478  	gp := getg()
  4479  	pc := getcallerpc()
  4480  	systemstack(func() {
  4481  		newg := newproc1(fn, gp, pc)
  4482  
  4483  		pp := getg().m.p.ptr()
  4484  		runqput(pp, newg, true)
  4485  
  4486  		if mainStarted {
  4487  			wakep()
  4488  		}
  4489  	})
  4490  }
  4491  
  4492  // Create a new g in state _Grunnable, starting at fn. callerpc is the
  4493  // address of the go statement that created this. The caller is responsible
  4494  // for adding the new g to the scheduler.
  4495  func newproc1(fn *funcval, callergp *g, callerpc uintptr) *g {
  4496  	if fn == nil {
  4497  		fatal("go of nil func value")
  4498  	}
  4499  
  4500  	mp := acquirem() // disable preemption because we hold M and P in local vars.
  4501  	pp := mp.p.ptr()
  4502  	newg := gfget(pp)
  4503  	if newg == nil {
  4504  		newg = malg(stackMin)
  4505  		casgstatus(newg, _Gidle, _Gdead)
  4506  		allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
  4507  	}
  4508  	if newg.stack.hi == 0 {
  4509  		throw("newproc1: newg missing stack")
  4510  	}
  4511  
  4512  	if readgstatus(newg) != _Gdead {
  4513  		throw("newproc1: new g is not Gdead")
  4514  	}
  4515  
  4516  	totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
  4517  	totalSize = alignUp(totalSize, sys.StackAlign)
  4518  	sp := newg.stack.hi - totalSize
  4519  	spArg := sp
  4520  	if usesLR {
  4521  		// caller's LR
  4522  		*(*uintptr)(unsafe.Pointer(sp)) = 0
  4523  		prepGoExitFrame(sp)
  4524  		spArg += sys.MinFrameSize
  4525  	}
  4526  
  4527  	memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
  4528  	newg.sched.sp = sp
  4529  	newg.stktopsp = sp
  4530  	newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
  4531  	newg.sched.g = guintptr(unsafe.Pointer(newg))
  4532  	gostartcallfn(&newg.sched, fn)
  4533  	newg.parentGoid = callergp.goid
  4534  	newg.gopc = callerpc
  4535  	newg.ancestors = saveAncestors(callergp)
  4536  	newg.startpc = fn.fn
  4537  	if isSystemGoroutine(newg, false) {
  4538  		sched.ngsys.Add(1)
  4539  	} else {
  4540  		// Only user goroutines inherit pprof labels.
  4541  		if mp.curg != nil {
  4542  			newg.labels = mp.curg.labels
  4543  		}
  4544  		if goroutineProfile.active {
  4545  			// A concurrent goroutine profile is running. It should include
  4546  			// exactly the set of goroutines that were alive when the goroutine
  4547  			// profiler first stopped the world. That does not include newg, so
  4548  			// mark it as not needing a profile before transitioning it from
  4549  			// _Gdead.
  4550  			newg.goroutineProfiled.Store(goroutineProfileSatisfied)
  4551  		}
  4552  	}
  4553  	// Track initial transition?
  4554  	newg.trackingSeq = uint8(fastrand())
  4555  	if newg.trackingSeq%gTrackingPeriod == 0 {
  4556  		newg.tracking = true
  4557  	}
  4558  	casgstatus(newg, _Gdead, _Grunnable)
  4559  	gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
  4560  
  4561  	if pp.goidcache == pp.goidcacheend {
  4562  		// Sched.goidgen is the last allocated id,
  4563  		// this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
  4564  		// At startup sched.goidgen=0, so main goroutine receives goid=1.
  4565  		pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
  4566  		pp.goidcache -= _GoidCacheBatch - 1
  4567  		pp.goidcacheend = pp.goidcache + _GoidCacheBatch
  4568  	}
  4569  	newg.goid = pp.goidcache
  4570  	pp.goidcache++
  4571  	if raceenabled {
  4572  		newg.racectx = racegostart(callerpc)
  4573  		newg.raceignore = 0
  4574  		if newg.labels != nil {
  4575  			// See note in proflabel.go on labelSync's role in synchronizing
  4576  			// with the reads in the signal handler.
  4577  			racereleasemergeg(newg, unsafe.Pointer(&labelSync))
  4578  		}
  4579  	}
  4580  	if traceEnabled() {
  4581  		traceGoCreate(newg, newg.startpc)
  4582  	}
  4583  	releasem(mp)
  4584  
  4585  	return newg
  4586  }
  4587  
  4588  // saveAncestors copies previous ancestors of the given caller g and
  4589  // includes info for the current caller into a new set of tracebacks for
  4590  // a g being created.
  4591  func saveAncestors(callergp *g) *[]ancestorInfo {
  4592  	// Copy all prior info, except for the root goroutine (goid 0).
  4593  	if debug.tracebackancestors <= 0 || callergp.goid == 0 {
  4594  		return nil
  4595  	}
  4596  	var callerAncestors []ancestorInfo
  4597  	if callergp.ancestors != nil {
  4598  		callerAncestors = *callergp.ancestors
  4599  	}
  4600  	n := int32(len(callerAncestors)) + 1
  4601  	if n > debug.tracebackancestors {
  4602  		n = debug.tracebackancestors
  4603  	}
  4604  	ancestors := make([]ancestorInfo, n)
  4605  	copy(ancestors[1:], callerAncestors)
  4606  
  4607  	var pcs [tracebackInnerFrames]uintptr
  4608  	npcs := gcallers(callergp, 0, pcs[:])
  4609  	ipcs := make([]uintptr, npcs)
  4610  	copy(ipcs, pcs[:])
  4611  	ancestors[0] = ancestorInfo{
  4612  		pcs:  ipcs,
  4613  		goid: callergp.goid,
  4614  		gopc: callergp.gopc,
  4615  	}
  4616  
  4617  	ancestorsp := new([]ancestorInfo)
  4618  	*ancestorsp = ancestors
  4619  	return ancestorsp
  4620  }
  4621  
  4622  // Put on gfree list.
  4623  // If local list is too long, transfer a batch to the global list.
  4624  func gfput(pp *p, gp *g) {
  4625  	if readgstatus(gp) != _Gdead {
  4626  		throw("gfput: bad status (not Gdead)")
  4627  	}
  4628  
  4629  	stksize := gp.stack.hi - gp.stack.lo
  4630  
  4631  	if stksize != uintptr(startingStackSize) {
  4632  		// non-standard stack size - free it.
  4633  		stackfree(gp.stack)
  4634  		gp.stack.lo = 0
  4635  		gp.stack.hi = 0
  4636  		gp.stackguard0 = 0
  4637  	}
  4638  
  4639  	pp.gFree.push(gp)
  4640  	pp.gFree.n++
  4641  	if pp.gFree.n >= 64 {
  4642  		var (
  4643  			inc      int32
  4644  			stackQ   gQueue
  4645  			noStackQ gQueue
  4646  		)
  4647  		for pp.gFree.n >= 32 {
  4648  			gp := pp.gFree.pop()
  4649  			pp.gFree.n--
  4650  			if gp.stack.lo == 0 {
  4651  				noStackQ.push(gp)
  4652  			} else {
  4653  				stackQ.push(gp)
  4654  			}
  4655  			inc++
  4656  		}
  4657  		lock(&sched.gFree.lock)
  4658  		sched.gFree.noStack.pushAll(noStackQ)
  4659  		sched.gFree.stack.pushAll(stackQ)
  4660  		sched.gFree.n += inc
  4661  		unlock(&sched.gFree.lock)
  4662  	}
  4663  }
  4664  
  4665  // Get from gfree list.
  4666  // If local list is empty, grab a batch from global list.
  4667  func gfget(pp *p) *g {
  4668  retry:
  4669  	if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
  4670  		lock(&sched.gFree.lock)
  4671  		// Move a batch of free Gs to the P.
  4672  		for pp.gFree.n < 32 {
  4673  			// Prefer Gs with stacks.
  4674  			gp := sched.gFree.stack.pop()
  4675  			if gp == nil {
  4676  				gp = sched.gFree.noStack.pop()
  4677  				if gp == nil {
  4678  					break
  4679  				}
  4680  			}
  4681  			sched.gFree.n--
  4682  			pp.gFree.push(gp)
  4683  			pp.gFree.n++
  4684  		}
  4685  		unlock(&sched.gFree.lock)
  4686  		goto retry
  4687  	}
  4688  	gp := pp.gFree.pop()
  4689  	if gp == nil {
  4690  		return nil
  4691  	}
  4692  	pp.gFree.n--
  4693  	if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
  4694  		// Deallocate old stack. We kept it in gfput because it was the
  4695  		// right size when the goroutine was put on the free list, but
  4696  		// the right size has changed since then.
  4697  		systemstack(func() {
  4698  			stackfree(gp.stack)
  4699  			gp.stack.lo = 0
  4700  			gp.stack.hi = 0
  4701  			gp.stackguard0 = 0
  4702  		})
  4703  	}
  4704  	if gp.stack.lo == 0 {
  4705  		// Stack was deallocated in gfput or just above. Allocate a new one.
  4706  		systemstack(func() {
  4707  			gp.stack = stackalloc(startingStackSize)
  4708  		})
  4709  		gp.stackguard0 = gp.stack.lo + stackGuard
  4710  	} else {
  4711  		if raceenabled {
  4712  			racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4713  		}
  4714  		if msanenabled {
  4715  			msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4716  		}
  4717  		if asanenabled {
  4718  			asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  4719  		}
  4720  	}
  4721  	return gp
  4722  }
  4723  
  4724  // Purge all cached G's from gfree list to the global list.
  4725  func gfpurge(pp *p) {
  4726  	var (
  4727  		inc      int32
  4728  		stackQ   gQueue
  4729  		noStackQ gQueue
  4730  	)
  4731  	for !pp.gFree.empty() {
  4732  		gp := pp.gFree.pop()
  4733  		pp.gFree.n--
  4734  		if gp.stack.lo == 0 {
  4735  			noStackQ.push(gp)
  4736  		} else {
  4737  			stackQ.push(gp)
  4738  		}
  4739  		inc++
  4740  	}
  4741  	lock(&sched.gFree.lock)
  4742  	sched.gFree.noStack.pushAll(noStackQ)
  4743  	sched.gFree.stack.pushAll(stackQ)
  4744  	sched.gFree.n += inc
  4745  	unlock(&sched.gFree.lock)
  4746  }
  4747  
  4748  // Breakpoint executes a breakpoint trap.
  4749  func Breakpoint() {
  4750  	breakpoint()
  4751  }
  4752  
  4753  // dolockOSThread is called by LockOSThread and lockOSThread below
  4754  // after they modify m.locked. Do not allow preemption during this call,
  4755  // or else the m might be different in this function than in the caller.
  4756  //
  4757  //go:nosplit
  4758  func dolockOSThread() {
  4759  	if GOARCH == "wasm" {
  4760  		return // no threads on wasm yet
  4761  	}
  4762  	gp := getg()
  4763  	gp.m.lockedg.set(gp)
  4764  	gp.lockedm.set(gp.m)
  4765  }
  4766  
  4767  // LockOSThread wires the calling goroutine to its current operating system thread.
  4768  // The calling goroutine will always execute in that thread,
  4769  // and no other goroutine will execute in it,
  4770  // until the calling goroutine has made as many calls to
  4771  // UnlockOSThread as to LockOSThread.
  4772  // If the calling goroutine exits without unlocking the thread,
  4773  // the thread will be terminated.
  4774  //
  4775  // All init functions are run on the startup thread. Calling LockOSThread
  4776  // from an init function will cause the main function to be invoked on
  4777  // that thread.
  4778  //
  4779  // A goroutine should call LockOSThread before calling OS services or
  4780  // non-Go library functions that depend on per-thread state.
  4781  //
  4782  //go:nosplit
  4783  func LockOSThread() {
  4784  	if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
  4785  		// If we need to start a new thread from the locked
  4786  		// thread, we need the template thread. Start it now
  4787  		// while we're in a known-good state.
  4788  		startTemplateThread()
  4789  	}
  4790  	gp := getg()
  4791  	gp.m.lockedExt++
  4792  	if gp.m.lockedExt == 0 {
  4793  		gp.m.lockedExt--
  4794  		panic("LockOSThread nesting overflow")
  4795  	}
  4796  	dolockOSThread()
  4797  }
  4798  
  4799  //go:nosplit
  4800  func lockOSThread() {
  4801  	getg().m.lockedInt++
  4802  	dolockOSThread()
  4803  }
  4804  
  4805  // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
  4806  // after they update m->locked. Do not allow preemption during this call,
  4807  // or else the m might be in different in this function than in the caller.
  4808  //
  4809  //go:nosplit
  4810  func dounlockOSThread() {
  4811  	if GOARCH == "wasm" {
  4812  		return // no threads on wasm yet
  4813  	}
  4814  	gp := getg()
  4815  	if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
  4816  		return
  4817  	}
  4818  	gp.m.lockedg = 0
  4819  	gp.lockedm = 0
  4820  }
  4821  
  4822  // UnlockOSThread undoes an earlier call to LockOSThread.
  4823  // If this drops the number of active LockOSThread calls on the
  4824  // calling goroutine to zero, it unwires the calling goroutine from
  4825  // its fixed operating system thread.
  4826  // If there are no active LockOSThread calls, this is a no-op.
  4827  //
  4828  // Before calling UnlockOSThread, the caller must ensure that the OS
  4829  // thread is suitable for running other goroutines. If the caller made
  4830  // any permanent changes to the state of the thread that would affect
  4831  // other goroutines, it should not call this function and thus leave
  4832  // the goroutine locked to the OS thread until the goroutine (and
  4833  // hence the thread) exits.
  4834  //
  4835  //go:nosplit
  4836  func UnlockOSThread() {
  4837  	gp := getg()
  4838  	if gp.m.lockedExt == 0 {
  4839  		return
  4840  	}
  4841  	gp.m.lockedExt--
  4842  	dounlockOSThread()
  4843  }
  4844  
  4845  //go:nosplit
  4846  func unlockOSThread() {
  4847  	gp := getg()
  4848  	if gp.m.lockedInt == 0 {
  4849  		systemstack(badunlockosthread)
  4850  	}
  4851  	gp.m.lockedInt--
  4852  	dounlockOSThread()
  4853  }
  4854  
  4855  func badunlockosthread() {
  4856  	throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
  4857  }
  4858  
  4859  func gcount() int32 {
  4860  	n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.n - sched.ngsys.Load()
  4861  	for _, pp := range allp {
  4862  		n -= pp.gFree.n
  4863  	}
  4864  
  4865  	// All these variables can be changed concurrently, so the result can be inconsistent.
  4866  	// But at least the current goroutine is running.
  4867  	if n < 1 {
  4868  		n = 1
  4869  	}
  4870  	return n
  4871  }
  4872  
  4873  func mcount() int32 {
  4874  	return int32(sched.mnext - sched.nmfreed)
  4875  }
  4876  
  4877  var prof struct {
  4878  	signalLock atomic.Uint32
  4879  
  4880  	// Must hold signalLock to write. Reads may be lock-free, but
  4881  	// signalLock should be taken to synchronize with changes.
  4882  	hz atomic.Int32
  4883  }
  4884  
  4885  func _System()                    { _System() }
  4886  func _ExternalCode()              { _ExternalCode() }
  4887  func _LostExternalCode()          { _LostExternalCode() }
  4888  func _GC()                        { _GC() }
  4889  func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
  4890  func _VDSO()                      { _VDSO() }
  4891  
  4892  // Called if we receive a SIGPROF signal.
  4893  // Called by the signal handler, may run during STW.
  4894  //
  4895  //go:nowritebarrierrec
  4896  func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
  4897  	if prof.hz.Load() == 0 {
  4898  		return
  4899  	}
  4900  
  4901  	// If mp.profilehz is 0, then profiling is not enabled for this thread.
  4902  	// We must check this to avoid a deadlock between setcpuprofilerate
  4903  	// and the call to cpuprof.add, below.
  4904  	if mp != nil && mp.profilehz == 0 {
  4905  		return
  4906  	}
  4907  
  4908  	// On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
  4909  	// runtime/internal/atomic. If SIGPROF arrives while the program is inside
  4910  	// the critical section, it creates a deadlock (when writing the sample).
  4911  	// As a workaround, create a counter of SIGPROFs while in critical section
  4912  	// to store the count, and pass it to sigprof.add() later when SIGPROF is
  4913  	// received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
  4914  	if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
  4915  		if f := findfunc(pc); f.valid() {
  4916  			if hasPrefix(funcname(f), "runtime/internal/atomic") {
  4917  				cpuprof.lostAtomic++
  4918  				return
  4919  			}
  4920  		}
  4921  		if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
  4922  			// runtime/internal/atomic functions call into kernel
  4923  			// helpers on arm < 7. See
  4924  			// runtime/internal/atomic/sys_linux_arm.s.
  4925  			cpuprof.lostAtomic++
  4926  			return
  4927  		}
  4928  	}
  4929  
  4930  	// Profiling runs concurrently with GC, so it must not allocate.
  4931  	// Set a trap in case the code does allocate.
  4932  	// Note that on windows, one thread takes profiles of all the
  4933  	// other threads, so mp is usually not getg().m.
  4934  	// In fact mp may not even be stopped.
  4935  	// See golang.org/issue/17165.
  4936  	getg().m.mallocing++
  4937  
  4938  	var u unwinder
  4939  	var stk [maxCPUProfStack]uintptr
  4940  	n := 0
  4941  	if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
  4942  		cgoOff := 0
  4943  		// Check cgoCallersUse to make sure that we are not
  4944  		// interrupting other code that is fiddling with
  4945  		// cgoCallers.  We are running in a signal handler
  4946  		// with all signals blocked, so we don't have to worry
  4947  		// about any other code interrupting us.
  4948  		if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
  4949  			for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
  4950  				cgoOff++
  4951  			}
  4952  			n += copy(stk[:], mp.cgoCallers[:cgoOff])
  4953  			mp.cgoCallers[0] = 0
  4954  		}
  4955  
  4956  		// Collect Go stack that leads to the cgo call.
  4957  		u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
  4958  	} else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
  4959  		// Libcall, i.e. runtime syscall on windows.
  4960  		// Collect Go stack that leads to the call.
  4961  		u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
  4962  	} else if mp != nil && mp.vdsoSP != 0 {
  4963  		// VDSO call, e.g. nanotime1 on Linux.
  4964  		// Collect Go stack that leads to the call.
  4965  		u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
  4966  	} else {
  4967  		u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
  4968  	}
  4969  	n += tracebackPCs(&u, 0, stk[n:])
  4970  
  4971  	if n <= 0 {
  4972  		// Normal traceback is impossible or has failed.
  4973  		// Account it against abstract "System" or "GC".
  4974  		n = 2
  4975  		if inVDSOPage(pc) {
  4976  			pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
  4977  		} else if pc > firstmoduledata.etext {
  4978  			// "ExternalCode" is better than "etext".
  4979  			pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
  4980  		}
  4981  		stk[0] = pc
  4982  		if mp.preemptoff != "" {
  4983  			stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
  4984  		} else {
  4985  			stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
  4986  		}
  4987  	}
  4988  
  4989  	if prof.hz.Load() != 0 {
  4990  		// Note: it can happen on Windows that we interrupted a system thread
  4991  		// with no g, so gp could nil. The other nil checks are done out of
  4992  		// caution, but not expected to be nil in practice.
  4993  		var tagPtr *unsafe.Pointer
  4994  		if gp != nil && gp.m != nil && gp.m.curg != nil {
  4995  			tagPtr = &gp.m.curg.labels
  4996  		}
  4997  		cpuprof.add(tagPtr, stk[:n])
  4998  
  4999  		gprof := gp
  5000  		var pp *p
  5001  		if gp != nil && gp.m != nil {
  5002  			if gp.m.curg != nil {
  5003  				gprof = gp.m.curg
  5004  			}
  5005  			pp = gp.m.p.ptr()
  5006  		}
  5007  		traceCPUSample(gprof, pp, stk[:n])
  5008  	}
  5009  	getg().m.mallocing--
  5010  }
  5011  
  5012  // setcpuprofilerate sets the CPU profiling rate to hz times per second.
  5013  // If hz <= 0, setcpuprofilerate turns off CPU profiling.
  5014  func setcpuprofilerate(hz int32) {
  5015  	// Force sane arguments.
  5016  	if hz < 0 {
  5017  		hz = 0
  5018  	}
  5019  
  5020  	// Disable preemption, otherwise we can be rescheduled to another thread
  5021  	// that has profiling enabled.
  5022  	gp := getg()
  5023  	gp.m.locks++
  5024  
  5025  	// Stop profiler on this thread so that it is safe to lock prof.
  5026  	// if a profiling signal came in while we had prof locked,
  5027  	// it would deadlock.
  5028  	setThreadCPUProfiler(0)
  5029  
  5030  	for !prof.signalLock.CompareAndSwap(0, 1) {
  5031  		osyield()
  5032  	}
  5033  	if prof.hz.Load() != hz {
  5034  		setProcessCPUProfiler(hz)
  5035  		prof.hz.Store(hz)
  5036  	}
  5037  	prof.signalLock.Store(0)
  5038  
  5039  	lock(&sched.lock)
  5040  	sched.profilehz = hz
  5041  	unlock(&sched.lock)
  5042  
  5043  	if hz != 0 {
  5044  		setThreadCPUProfiler(hz)
  5045  	}
  5046  
  5047  	gp.m.locks--
  5048  }
  5049  
  5050  // init initializes pp, which may be a freshly allocated p or a
  5051  // previously destroyed p, and transitions it to status _Pgcstop.
  5052  func (pp *p) init(id int32) {
  5053  	pp.id = id
  5054  	pp.status = _Pgcstop
  5055  	pp.sudogcache = pp.sudogbuf[:0]
  5056  	pp.deferpool = pp.deferpoolbuf[:0]
  5057  	pp.wbBuf.reset()
  5058  	if pp.mcache == nil {
  5059  		if id == 0 {
  5060  			if mcache0 == nil {
  5061  				throw("missing mcache?")
  5062  			}
  5063  			// Use the bootstrap mcache0. Only one P will get
  5064  			// mcache0: the one with ID 0.
  5065  			pp.mcache = mcache0
  5066  		} else {
  5067  			pp.mcache = allocmcache()
  5068  		}
  5069  	}
  5070  	if raceenabled && pp.raceprocctx == 0 {
  5071  		if id == 0 {
  5072  			pp.raceprocctx = raceprocctx0
  5073  			raceprocctx0 = 0 // bootstrap
  5074  		} else {
  5075  			pp.raceprocctx = raceproccreate()
  5076  		}
  5077  	}
  5078  	lockInit(&pp.timersLock, lockRankTimers)
  5079  
  5080  	// This P may get timers when it starts running. Set the mask here
  5081  	// since the P may not go through pidleget (notably P 0 on startup).
  5082  	timerpMask.set(id)
  5083  	// Similarly, we may not go through pidleget before this P starts
  5084  	// running if it is P 0 on startup.
  5085  	idlepMask.clear(id)
  5086  }
  5087  
  5088  // destroy releases all of the resources associated with pp and
  5089  // transitions it to status _Pdead.
  5090  //
  5091  // sched.lock must be held and the world must be stopped.
  5092  func (pp *p) destroy() {
  5093  	assertLockHeld(&sched.lock)
  5094  	assertWorldStopped()
  5095  
  5096  	// Move all runnable goroutines to the global queue
  5097  	for pp.runqhead != pp.runqtail {
  5098  		// Pop from tail of local queue
  5099  		pp.runqtail--
  5100  		gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
  5101  		// Push onto head of global queue
  5102  		globrunqputhead(gp)
  5103  	}
  5104  	if pp.runnext != 0 {
  5105  		globrunqputhead(pp.runnext.ptr())
  5106  		pp.runnext = 0
  5107  	}
  5108  	if len(pp.timers) > 0 {
  5109  		plocal := getg().m.p.ptr()
  5110  		// The world is stopped, but we acquire timersLock to
  5111  		// protect against sysmon calling timeSleepUntil.
  5112  		// This is the only case where we hold the timersLock of
  5113  		// more than one P, so there are no deadlock concerns.
  5114  		lock(&plocal.timersLock)
  5115  		lock(&pp.timersLock)
  5116  		moveTimers(plocal, pp.timers)
  5117  		pp.timers = nil
  5118  		pp.numTimers.Store(0)
  5119  		pp.deletedTimers.Store(0)
  5120  		pp.timer0When.Store(0)
  5121  		unlock(&pp.timersLock)
  5122  		unlock(&plocal.timersLock)
  5123  	}
  5124  	// Flush p's write barrier buffer.
  5125  	if gcphase != _GCoff {
  5126  		wbBufFlush1(pp)
  5127  		pp.gcw.dispose()
  5128  	}
  5129  	for i := range pp.sudogbuf {
  5130  		pp.sudogbuf[i] = nil
  5131  	}
  5132  	pp.sudogcache = pp.sudogbuf[:0]
  5133  	pp.pinnerCache = nil
  5134  	for j := range pp.deferpoolbuf {
  5135  		pp.deferpoolbuf[j] = nil
  5136  	}
  5137  	pp.deferpool = pp.deferpoolbuf[:0]
  5138  	systemstack(func() {
  5139  		for i := 0; i < pp.mspancache.len; i++ {
  5140  			// Safe to call since the world is stopped.
  5141  			mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
  5142  		}
  5143  		pp.mspancache.len = 0
  5144  		lock(&mheap_.lock)
  5145  		pp.pcache.flush(&mheap_.pages)
  5146  		unlock(&mheap_.lock)
  5147  	})
  5148  	freemcache(pp.mcache)
  5149  	pp.mcache = nil
  5150  	gfpurge(pp)
  5151  	traceProcFree(pp)
  5152  	if raceenabled {
  5153  		if pp.timerRaceCtx != 0 {
  5154  			// The race detector code uses a callback to fetch
  5155  			// the proc context, so arrange for that callback
  5156  			// to see the right thing.
  5157  			// This hack only works because we are the only
  5158  			// thread running.
  5159  			mp := getg().m
  5160  			phold := mp.p.ptr()
  5161  			mp.p.set(pp)
  5162  
  5163  			racectxend(pp.timerRaceCtx)
  5164  			pp.timerRaceCtx = 0
  5165  
  5166  			mp.p.set(phold)
  5167  		}
  5168  		raceprocdestroy(pp.raceprocctx)
  5169  		pp.raceprocctx = 0
  5170  	}
  5171  	pp.gcAssistTime = 0
  5172  	pp.status = _Pdead
  5173  }
  5174  
  5175  // Change number of processors.
  5176  //
  5177  // sched.lock must be held, and the world must be stopped.
  5178  //
  5179  // gcworkbufs must not be being modified by either the GC or the write barrier
  5180  // code, so the GC must not be running if the number of Ps actually changes.
  5181  //
  5182  // Returns list of Ps with local work, they need to be scheduled by the caller.
  5183  func procresize(nprocs int32) *p {
  5184  	assertLockHeld(&sched.lock)
  5185  	assertWorldStopped()
  5186  
  5187  	old := gomaxprocs
  5188  	if old < 0 || nprocs <= 0 {
  5189  		throw("procresize: invalid arg")
  5190  	}
  5191  	if traceEnabled() {
  5192  		traceGomaxprocs(nprocs)
  5193  	}
  5194  
  5195  	// update statistics
  5196  	now := nanotime()
  5197  	if sched.procresizetime != 0 {
  5198  		sched.totaltime += int64(old) * (now - sched.procresizetime)
  5199  	}
  5200  	sched.procresizetime = now
  5201  
  5202  	maskWords := (nprocs + 31) / 32
  5203  
  5204  	// Grow allp if necessary.
  5205  	if nprocs > int32(len(allp)) {
  5206  		// Synchronize with retake, which could be running
  5207  		// concurrently since it doesn't run on a P.
  5208  		lock(&allpLock)
  5209  		if nprocs <= int32(cap(allp)) {
  5210  			allp = allp[:nprocs]
  5211  		} else {
  5212  			nallp := make([]*p, nprocs)
  5213  			// Copy everything up to allp's cap so we
  5214  			// never lose old allocated Ps.
  5215  			copy(nallp, allp[:cap(allp)])
  5216  			allp = nallp
  5217  		}
  5218  
  5219  		if maskWords <= int32(cap(idlepMask)) {
  5220  			idlepMask = idlepMask[:maskWords]
  5221  			timerpMask = timerpMask[:maskWords]
  5222  		} else {
  5223  			nidlepMask := make([]uint32, maskWords)
  5224  			// No need to copy beyond len, old Ps are irrelevant.
  5225  			copy(nidlepMask, idlepMask)
  5226  			idlepMask = nidlepMask
  5227  
  5228  			ntimerpMask := make([]uint32, maskWords)
  5229  			copy(ntimerpMask, timerpMask)
  5230  			timerpMask = ntimerpMask
  5231  		}
  5232  		unlock(&allpLock)
  5233  	}
  5234  
  5235  	// initialize new P's
  5236  	for i := old; i < nprocs; i++ {
  5237  		pp := allp[i]
  5238  		if pp == nil {
  5239  			pp = new(p)
  5240  		}
  5241  		pp.init(i)
  5242  		atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
  5243  	}
  5244  
  5245  	gp := getg()
  5246  	if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
  5247  		// continue to use the current P
  5248  		gp.m.p.ptr().status = _Prunning
  5249  		gp.m.p.ptr().mcache.prepareForSweep()
  5250  	} else {
  5251  		// release the current P and acquire allp[0].
  5252  		//
  5253  		// We must do this before destroying our current P
  5254  		// because p.destroy itself has write barriers, so we
  5255  		// need to do that from a valid P.
  5256  		if gp.m.p != 0 {
  5257  			if traceEnabled() {
  5258  				// Pretend that we were descheduled
  5259  				// and then scheduled again to keep
  5260  				// the trace sane.
  5261  				traceGoSched()
  5262  				traceProcStop(gp.m.p.ptr())
  5263  			}
  5264  			gp.m.p.ptr().m = 0
  5265  		}
  5266  		gp.m.p = 0
  5267  		pp := allp[0]
  5268  		pp.m = 0
  5269  		pp.status = _Pidle
  5270  		acquirep(pp)
  5271  		if traceEnabled() {
  5272  			traceGoStart()
  5273  		}
  5274  	}
  5275  
  5276  	// g.m.p is now set, so we no longer need mcache0 for bootstrapping.
  5277  	mcache0 = nil
  5278  
  5279  	// release resources from unused P's
  5280  	for i := nprocs; i < old; i++ {
  5281  		pp := allp[i]
  5282  		pp.destroy()
  5283  		// can't free P itself because it can be referenced by an M in syscall
  5284  	}
  5285  
  5286  	// Trim allp.
  5287  	if int32(len(allp)) != nprocs {
  5288  		lock(&allpLock)
  5289  		allp = allp[:nprocs]
  5290  		idlepMask = idlepMask[:maskWords]
  5291  		timerpMask = timerpMask[:maskWords]
  5292  		unlock(&allpLock)
  5293  	}
  5294  
  5295  	var runnablePs *p
  5296  	for i := nprocs - 1; i >= 0; i-- {
  5297  		pp := allp[i]
  5298  		if gp.m.p.ptr() == pp {
  5299  			continue
  5300  		}
  5301  		pp.status = _Pidle
  5302  		if runqempty(pp) {
  5303  			pidleput(pp, now)
  5304  		} else {
  5305  			pp.m.set(mget())
  5306  			pp.link.set(runnablePs)
  5307  			runnablePs = pp
  5308  		}
  5309  	}
  5310  	stealOrder.reset(uint32(nprocs))
  5311  	var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
  5312  	atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
  5313  	if old != nprocs {
  5314  		// Notify the limiter that the amount of procs has changed.
  5315  		gcCPULimiter.resetCapacity(now, nprocs)
  5316  	}
  5317  	return runnablePs
  5318  }
  5319  
  5320  // Associate p and the current m.
  5321  //
  5322  // This function is allowed to have write barriers even if the caller
  5323  // isn't because it immediately acquires pp.
  5324  //
  5325  //go:yeswritebarrierrec
  5326  func acquirep(pp *p) {
  5327  	// Do the part that isn't allowed to have write barriers.
  5328  	wirep(pp)
  5329  
  5330  	// Have p; write barriers now allowed.
  5331  
  5332  	// Perform deferred mcache flush before this P can allocate
  5333  	// from a potentially stale mcache.
  5334  	pp.mcache.prepareForSweep()
  5335  
  5336  	if traceEnabled() {
  5337  		traceProcStart()
  5338  	}
  5339  }
  5340  
  5341  // wirep is the first step of acquirep, which actually associates the
  5342  // current M to pp. This is broken out so we can disallow write
  5343  // barriers for this part, since we don't yet have a P.
  5344  //
  5345  //go:nowritebarrierrec
  5346  //go:nosplit
  5347  func wirep(pp *p) {
  5348  	gp := getg()
  5349  
  5350  	if gp.m.p != 0 {
  5351  		throw("wirep: already in go")
  5352  	}
  5353  	if pp.m != 0 || pp.status != _Pidle {
  5354  		id := int64(0)
  5355  		if pp.m != 0 {
  5356  			id = pp.m.ptr().id
  5357  		}
  5358  		print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
  5359  		throw("wirep: invalid p state")
  5360  	}
  5361  	gp.m.p.set(pp)
  5362  	pp.m.set(gp.m)
  5363  	pp.status = _Prunning
  5364  }
  5365  
  5366  // Disassociate p and the current m.
  5367  func releasep() *p {
  5368  	gp := getg()
  5369  
  5370  	if gp.m.p == 0 {
  5371  		throw("releasep: invalid arg")
  5372  	}
  5373  	pp := gp.m.p.ptr()
  5374  	if pp.m.ptr() != gp.m || pp.status != _Prunning {
  5375  		print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
  5376  		throw("releasep: invalid p state")
  5377  	}
  5378  	if traceEnabled() {
  5379  		traceProcStop(gp.m.p.ptr())
  5380  	}
  5381  	gp.m.p = 0
  5382  	pp.m = 0
  5383  	pp.status = _Pidle
  5384  	return pp
  5385  }
  5386  
  5387  func incidlelocked(v int32) {
  5388  	lock(&sched.lock)
  5389  	sched.nmidlelocked += v
  5390  	if v > 0 {
  5391  		checkdead()
  5392  	}
  5393  	unlock(&sched.lock)
  5394  }
  5395  
  5396  // Check for deadlock situation.
  5397  // The check is based on number of running M's, if 0 -> deadlock.
  5398  // sched.lock must be held.
  5399  func checkdead() {
  5400  	assertLockHeld(&sched.lock)
  5401  
  5402  	// For -buildmode=c-shared or -buildmode=c-archive it's OK if
  5403  	// there are no running goroutines. The calling program is
  5404  	// assumed to be running.
  5405  	if islibrary || isarchive {
  5406  		return
  5407  	}
  5408  
  5409  	// If we are dying because of a signal caught on an already idle thread,
  5410  	// freezetheworld will cause all running threads to block.
  5411  	// And runtime will essentially enter into deadlock state,
  5412  	// except that there is a thread that will call exit soon.
  5413  	if panicking.Load() > 0 {
  5414  		return
  5415  	}
  5416  
  5417  	// If we are not running under cgo, but we have an extra M then account
  5418  	// for it. (It is possible to have an extra M on Windows without cgo to
  5419  	// accommodate callbacks created by syscall.NewCallback. See issue #6751
  5420  	// for details.)
  5421  	var run0 int32
  5422  	if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
  5423  		run0 = 1
  5424  	}
  5425  
  5426  	run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
  5427  	if run > run0 {
  5428  		return
  5429  	}
  5430  	if run < 0 {
  5431  		print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
  5432  		unlock(&sched.lock)
  5433  		throw("checkdead: inconsistent counts")
  5434  	}
  5435  
  5436  	grunning := 0
  5437  	forEachG(func(gp *g) {
  5438  		if isSystemGoroutine(gp, false) {
  5439  			return
  5440  		}
  5441  		s := readgstatus(gp)
  5442  		switch s &^ _Gscan {
  5443  		case _Gwaiting,
  5444  			_Gpreempted:
  5445  			grunning++
  5446  		case _Grunnable,
  5447  			_Grunning,
  5448  			_Gsyscall:
  5449  			print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
  5450  			unlock(&sched.lock)
  5451  			throw("checkdead: runnable g")
  5452  		}
  5453  	})
  5454  	if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
  5455  		unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  5456  		fatal("no goroutines (main called runtime.Goexit) - deadlock!")
  5457  	}
  5458  
  5459  	// Maybe jump time forward for playground.
  5460  	if faketime != 0 {
  5461  		if when := timeSleepUntil(); when < maxWhen {
  5462  			faketime = when
  5463  
  5464  			// Start an M to steal the timer.
  5465  			pp, _ := pidleget(faketime)
  5466  			if pp == nil {
  5467  				// There should always be a free P since
  5468  				// nothing is running.
  5469  				unlock(&sched.lock)
  5470  				throw("checkdead: no p for timer")
  5471  			}
  5472  			mp := mget()
  5473  			if mp == nil {
  5474  				// There should always be a free M since
  5475  				// nothing is running.
  5476  				unlock(&sched.lock)
  5477  				throw("checkdead: no m for timer")
  5478  			}
  5479  			// M must be spinning to steal. We set this to be
  5480  			// explicit, but since this is the only M it would
  5481  			// become spinning on its own anyways.
  5482  			sched.nmspinning.Add(1)
  5483  			mp.spinning = true
  5484  			mp.nextp.set(pp)
  5485  			notewakeup(&mp.park)
  5486  			return
  5487  		}
  5488  	}
  5489  
  5490  	// There are no goroutines running, so we can look at the P's.
  5491  	for _, pp := range allp {
  5492  		if len(pp.timers) > 0 {
  5493  			return
  5494  		}
  5495  	}
  5496  
  5497  	unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  5498  	fatal("all goroutines are asleep - deadlock!")
  5499  }
  5500  
  5501  // forcegcperiod is the maximum time in nanoseconds between garbage
  5502  // collections. If we go this long without a garbage collection, one
  5503  // is forced to run.
  5504  //
  5505  // This is a variable for testing purposes. It normally doesn't change.
  5506  var forcegcperiod int64 = 2 * 60 * 1e9
  5507  
  5508  // needSysmonWorkaround is true if the workaround for
  5509  // golang.org/issue/42515 is needed on NetBSD.
  5510  var needSysmonWorkaround bool = false
  5511  
  5512  // Always runs without a P, so write barriers are not allowed.
  5513  //
  5514  //go:nowritebarrierrec
  5515  func sysmon() {
  5516  	lock(&sched.lock)
  5517  	sched.nmsys++
  5518  	checkdead()
  5519  	unlock(&sched.lock)
  5520  
  5521  	lasttrace := int64(0)
  5522  	idle := 0 // how many cycles in succession we had not wokeup somebody
  5523  	delay := uint32(0)
  5524  
  5525  	for {
  5526  		if idle == 0 { // start with 20us sleep...
  5527  			delay = 20
  5528  		} else if idle > 50 { // start doubling the sleep after 1ms...
  5529  			delay *= 2
  5530  		}
  5531  		if delay > 10*1000 { // up to 10ms
  5532  			delay = 10 * 1000
  5533  		}
  5534  		usleep(delay)
  5535  
  5536  		// sysmon should not enter deep sleep if schedtrace is enabled so that
  5537  		// it can print that information at the right time.
  5538  		//
  5539  		// It should also not enter deep sleep if there are any active P's so
  5540  		// that it can retake P's from syscalls, preempt long running G's, and
  5541  		// poll the network if all P's are busy for long stretches.
  5542  		//
  5543  		// It should wakeup from deep sleep if any P's become active either due
  5544  		// to exiting a syscall or waking up due to a timer expiring so that it
  5545  		// can resume performing those duties. If it wakes from a syscall it
  5546  		// resets idle and delay as a bet that since it had retaken a P from a
  5547  		// syscall before, it may need to do it again shortly after the
  5548  		// application starts work again. It does not reset idle when waking
  5549  		// from a timer to avoid adding system load to applications that spend
  5550  		// most of their time sleeping.
  5551  		now := nanotime()
  5552  		if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
  5553  			lock(&sched.lock)
  5554  			if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
  5555  				syscallWake := false
  5556  				next := timeSleepUntil()
  5557  				if next > now {
  5558  					sched.sysmonwait.Store(true)
  5559  					unlock(&sched.lock)
  5560  					// Make wake-up period small enough
  5561  					// for the sampling to be correct.
  5562  					sleep := forcegcperiod / 2
  5563  					if next-now < sleep {
  5564  						sleep = next - now
  5565  					}
  5566  					shouldRelax := sleep >= osRelaxMinNS
  5567  					if shouldRelax {
  5568  						osRelax(true)
  5569  					}
  5570  					syscallWake = notetsleep(&sched.sysmonnote, sleep)
  5571  					if shouldRelax {
  5572  						osRelax(false)
  5573  					}
  5574  					lock(&sched.lock)
  5575  					sched.sysmonwait.Store(false)
  5576  					noteclear(&sched.sysmonnote)
  5577  				}
  5578  				if syscallWake {
  5579  					idle = 0
  5580  					delay = 20
  5581  				}
  5582  			}
  5583  			unlock(&sched.lock)
  5584  		}
  5585  
  5586  		lock(&sched.sysmonlock)
  5587  		// Update now in case we blocked on sysmonnote or spent a long time
  5588  		// blocked on schedlock or sysmonlock above.
  5589  		now = nanotime()
  5590  
  5591  		// trigger libc interceptors if needed
  5592  		if *cgo_yield != nil {
  5593  			asmcgocall(*cgo_yield, nil)
  5594  		}
  5595  		// poll network if not polled for more than 10ms
  5596  		lastpoll := sched.lastpoll.Load()
  5597  		if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
  5598  			sched.lastpoll.CompareAndSwap(lastpoll, now)
  5599  			list := netpoll(0) // non-blocking - returns list of goroutines
  5600  			if !list.empty() {
  5601  				// Need to decrement number of idle locked M's
  5602  				// (pretending that one more is running) before injectglist.
  5603  				// Otherwise it can lead to the following situation:
  5604  				// injectglist grabs all P's but before it starts M's to run the P's,
  5605  				// another M returns from syscall, finishes running its G,
  5606  				// observes that there is no work to do and no other running M's
  5607  				// and reports deadlock.
  5608  				incidlelocked(-1)
  5609  				injectglist(&list)
  5610  				incidlelocked(1)
  5611  			}
  5612  		}
  5613  		if GOOS == "netbsd" && needSysmonWorkaround {
  5614  			// netpoll is responsible for waiting for timer
  5615  			// expiration, so we typically don't have to worry
  5616  			// about starting an M to service timers. (Note that
  5617  			// sleep for timeSleepUntil above simply ensures sysmon
  5618  			// starts running again when that timer expiration may
  5619  			// cause Go code to run again).
  5620  			//
  5621  			// However, netbsd has a kernel bug that sometimes
  5622  			// misses netpollBreak wake-ups, which can lead to
  5623  			// unbounded delays servicing timers. If we detect this
  5624  			// overrun, then startm to get something to handle the
  5625  			// timer.
  5626  			//
  5627  			// See issue 42515 and
  5628  			// https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
  5629  			if next := timeSleepUntil(); next < now {
  5630  				startm(nil, false, false)
  5631  			}
  5632  		}
  5633  		if scavenger.sysmonWake.Load() != 0 {
  5634  			// Kick the scavenger awake if someone requested it.
  5635  			scavenger.wake()
  5636  		}
  5637  		// retake P's blocked in syscalls
  5638  		// and preempt long running G's
  5639  		if retake(now) != 0 {
  5640  			idle = 0
  5641  		} else {
  5642  			idle++
  5643  		}
  5644  		// check if we need to force a GC
  5645  		if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
  5646  			lock(&forcegc.lock)
  5647  			forcegc.idle.Store(false)
  5648  			var list gList
  5649  			list.push(forcegc.g)
  5650  			injectglist(&list)
  5651  			unlock(&forcegc.lock)
  5652  		}
  5653  		if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
  5654  			lasttrace = now
  5655  			schedtrace(debug.scheddetail > 0)
  5656  		}
  5657  		unlock(&sched.sysmonlock)
  5658  	}
  5659  }
  5660  
  5661  type sysmontick struct {
  5662  	schedtick   uint32
  5663  	schedwhen   int64
  5664  	syscalltick uint32
  5665  	syscallwhen int64
  5666  }
  5667  
  5668  // forcePreemptNS is the time slice given to a G before it is
  5669  // preempted.
  5670  const forcePreemptNS = 10 * 1000 * 1000 // 10ms
  5671  
  5672  func retake(now int64) uint32 {
  5673  	n := 0
  5674  	// Prevent allp slice changes. This lock will be completely
  5675  	// uncontended unless we're already stopping the world.
  5676  	lock(&allpLock)
  5677  	// We can't use a range loop over allp because we may
  5678  	// temporarily drop the allpLock. Hence, we need to re-fetch
  5679  	// allp each time around the loop.
  5680  	for i := 0; i < len(allp); i++ {
  5681  		pp := allp[i]
  5682  		if pp == nil {
  5683  			// This can happen if procresize has grown
  5684  			// allp but not yet created new Ps.
  5685  			continue
  5686  		}
  5687  		pd := &pp.sysmontick
  5688  		s := pp.status
  5689  		sysretake := false
  5690  		if s == _Prunning || s == _Psyscall {
  5691  			// Preempt G if it's running for too long.
  5692  			t := int64(pp.schedtick)
  5693  			if int64(pd.schedtick) != t {
  5694  				pd.schedtick = uint32(t)
  5695  				pd.schedwhen = now
  5696  			} else if pd.schedwhen+forcePreemptNS <= now {
  5697  				preemptone(pp)
  5698  				// In case of syscall, preemptone() doesn't
  5699  				// work, because there is no M wired to P.
  5700  				sysretake = true
  5701  			}
  5702  		}
  5703  		if s == _Psyscall {
  5704  			// Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
  5705  			t := int64(pp.syscalltick)
  5706  			if !sysretake && int64(pd.syscalltick) != t {
  5707  				pd.syscalltick = uint32(t)
  5708  				pd.syscallwhen = now
  5709  				continue
  5710  			}
  5711  			// On the one hand we don't want to retake Ps if there is no other work to do,
  5712  			// but on the other hand we want to retake them eventually
  5713  			// because they can prevent the sysmon thread from deep sleep.
  5714  			if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
  5715  				continue
  5716  			}
  5717  			// Drop allpLock so we can take sched.lock.
  5718  			unlock(&allpLock)
  5719  			// Need to decrement number of idle locked M's
  5720  			// (pretending that one more is running) before the CAS.
  5721  			// Otherwise the M from which we retake can exit the syscall,
  5722  			// increment nmidle and report deadlock.
  5723  			incidlelocked(-1)
  5724  			if atomic.Cas(&pp.status, s, _Pidle) {
  5725  				if traceEnabled() {
  5726  					traceGoSysBlock(pp)
  5727  					traceProcStop(pp)
  5728  				}
  5729  				n++
  5730  				pp.syscalltick++
  5731  				handoffp(pp)
  5732  			}
  5733  			incidlelocked(1)
  5734  			lock(&allpLock)
  5735  		}
  5736  	}
  5737  	unlock(&allpLock)
  5738  	return uint32(n)
  5739  }
  5740  
  5741  // Tell all goroutines that they have been preempted and they should stop.
  5742  // This function is purely best-effort. It can fail to inform a goroutine if a
  5743  // processor just started running it.
  5744  // No locks need to be held.
  5745  // Returns true if preemption request was issued to at least one goroutine.
  5746  func preemptall() bool {
  5747  	res := false
  5748  	for _, pp := range allp {
  5749  		if pp.status != _Prunning {
  5750  			continue
  5751  		}
  5752  		if preemptone(pp) {
  5753  			res = true
  5754  		}
  5755  	}
  5756  	return res
  5757  }
  5758  
  5759  // Tell the goroutine running on processor P to stop.
  5760  // This function is purely best-effort. It can incorrectly fail to inform the
  5761  // goroutine. It can inform the wrong goroutine. Even if it informs the
  5762  // correct goroutine, that goroutine might ignore the request if it is
  5763  // simultaneously executing newstack.
  5764  // No lock needs to be held.
  5765  // Returns true if preemption request was issued.
  5766  // The actual preemption will happen at some point in the future
  5767  // and will be indicated by the gp->status no longer being
  5768  // Grunning
  5769  func preemptone(pp *p) bool {
  5770  	mp := pp.m.ptr()
  5771  	if mp == nil || mp == getg().m {
  5772  		return false
  5773  	}
  5774  	gp := mp.curg
  5775  	if gp == nil || gp == mp.g0 {
  5776  		return false
  5777  	}
  5778  
  5779  	gp.preempt = true
  5780  
  5781  	// Every call in a goroutine checks for stack overflow by
  5782  	// comparing the current stack pointer to gp->stackguard0.
  5783  	// Setting gp->stackguard0 to StackPreempt folds
  5784  	// preemption into the normal stack overflow check.
  5785  	gp.stackguard0 = stackPreempt
  5786  
  5787  	// Request an async preemption of this P.
  5788  	if preemptMSupported && debug.asyncpreemptoff == 0 {
  5789  		pp.preempt = true
  5790  		preemptM(mp)
  5791  	}
  5792  
  5793  	return true
  5794  }
  5795  
  5796  var starttime int64
  5797  
  5798  func schedtrace(detailed bool) {
  5799  	now := nanotime()
  5800  	if starttime == 0 {
  5801  		starttime = now
  5802  	}
  5803  
  5804  	lock(&sched.lock)
  5805  	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)
  5806  	if detailed {
  5807  		print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
  5808  	}
  5809  	// We must be careful while reading data from P's, M's and G's.
  5810  	// Even if we hold schedlock, most data can be changed concurrently.
  5811  	// E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
  5812  	for i, pp := range allp {
  5813  		mp := pp.m.ptr()
  5814  		h := atomic.Load(&pp.runqhead)
  5815  		t := atomic.Load(&pp.runqtail)
  5816  		if detailed {
  5817  			print("  P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
  5818  			if mp != nil {
  5819  				print(mp.id)
  5820  			} else {
  5821  				print("nil")
  5822  			}
  5823  			print(" runqsize=", t-h, " gfreecnt=", pp.gFree.n, " timerslen=", len(pp.timers), "\n")
  5824  		} else {
  5825  			// In non-detailed mode format lengths of per-P run queues as:
  5826  			// [len1 len2 len3 len4]
  5827  			print(" ")
  5828  			if i == 0 {
  5829  				print("[")
  5830  			}
  5831  			print(t - h)
  5832  			if i == len(allp)-1 {
  5833  				print("]\n")
  5834  			}
  5835  		}
  5836  	}
  5837  
  5838  	if !detailed {
  5839  		unlock(&sched.lock)
  5840  		return
  5841  	}
  5842  
  5843  	for mp := allm; mp != nil; mp = mp.alllink {
  5844  		pp := mp.p.ptr()
  5845  		print("  M", mp.id, ": p=")
  5846  		if pp != nil {
  5847  			print(pp.id)
  5848  		} else {
  5849  			print("nil")
  5850  		}
  5851  		print(" curg=")
  5852  		if mp.curg != nil {
  5853  			print(mp.curg.goid)
  5854  		} else {
  5855  			print("nil")
  5856  		}
  5857  		print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
  5858  		if lockedg := mp.lockedg.ptr(); lockedg != nil {
  5859  			print(lockedg.goid)
  5860  		} else {
  5861  			print("nil")
  5862  		}
  5863  		print("\n")
  5864  	}
  5865  
  5866  	forEachG(func(gp *g) {
  5867  		print("  G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
  5868  		if gp.m != nil {
  5869  			print(gp.m.id)
  5870  		} else {
  5871  			print("nil")
  5872  		}
  5873  		print(" lockedm=")
  5874  		if lockedm := gp.lockedm.ptr(); lockedm != nil {
  5875  			print(lockedm.id)
  5876  		} else {
  5877  			print("nil")
  5878  		}
  5879  		print("\n")
  5880  	})
  5881  	unlock(&sched.lock)
  5882  }
  5883  
  5884  // schedEnableUser enables or disables the scheduling of user
  5885  // goroutines.
  5886  //
  5887  // This does not stop already running user goroutines, so the caller
  5888  // should first stop the world when disabling user goroutines.
  5889  func schedEnableUser(enable bool) {
  5890  	lock(&sched.lock)
  5891  	if sched.disable.user == !enable {
  5892  		unlock(&sched.lock)
  5893  		return
  5894  	}
  5895  	sched.disable.user = !enable
  5896  	if enable {
  5897  		n := sched.disable.n
  5898  		sched.disable.n = 0
  5899  		globrunqputbatch(&sched.disable.runnable, n)
  5900  		unlock(&sched.lock)
  5901  		for ; n != 0 && sched.npidle.Load() != 0; n-- {
  5902  			startm(nil, false, false)
  5903  		}
  5904  	} else {
  5905  		unlock(&sched.lock)
  5906  	}
  5907  }
  5908  
  5909  // schedEnabled reports whether gp should be scheduled. It returns
  5910  // false is scheduling of gp is disabled.
  5911  //
  5912  // sched.lock must be held.
  5913  func schedEnabled(gp *g) bool {
  5914  	assertLockHeld(&sched.lock)
  5915  
  5916  	if sched.disable.user {
  5917  		return isSystemGoroutine(gp, true)
  5918  	}
  5919  	return true
  5920  }
  5921  
  5922  // Put mp on midle list.
  5923  // sched.lock must be held.
  5924  // May run during STW, so write barriers are not allowed.
  5925  //
  5926  //go:nowritebarrierrec
  5927  func mput(mp *m) {
  5928  	assertLockHeld(&sched.lock)
  5929  
  5930  	mp.schedlink = sched.midle
  5931  	sched.midle.set(mp)
  5932  	sched.nmidle++
  5933  	checkdead()
  5934  }
  5935  
  5936  // Try to get an m from midle list.
  5937  // sched.lock must be held.
  5938  // May run during STW, so write barriers are not allowed.
  5939  //
  5940  //go:nowritebarrierrec
  5941  func mget() *m {
  5942  	assertLockHeld(&sched.lock)
  5943  
  5944  	mp := sched.midle.ptr()
  5945  	if mp != nil {
  5946  		sched.midle = mp.schedlink
  5947  		sched.nmidle--
  5948  	}
  5949  	return mp
  5950  }
  5951  
  5952  // Put gp on the global runnable queue.
  5953  // sched.lock must be held.
  5954  // May run during STW, so write barriers are not allowed.
  5955  //
  5956  //go:nowritebarrierrec
  5957  func globrunqput(gp *g) {
  5958  	assertLockHeld(&sched.lock)
  5959  
  5960  	sched.runq.pushBack(gp)
  5961  	sched.runqsize++
  5962  }
  5963  
  5964  // Put gp at the head of the global runnable queue.
  5965  // sched.lock must be held.
  5966  // May run during STW, so write barriers are not allowed.
  5967  //
  5968  //go:nowritebarrierrec
  5969  func globrunqputhead(gp *g) {
  5970  	assertLockHeld(&sched.lock)
  5971  
  5972  	sched.runq.push(gp)
  5973  	sched.runqsize++
  5974  }
  5975  
  5976  // Put a batch of runnable goroutines on the global runnable queue.
  5977  // This clears *batch.
  5978  // sched.lock must be held.
  5979  // May run during STW, so write barriers are not allowed.
  5980  //
  5981  //go:nowritebarrierrec
  5982  func globrunqputbatch(batch *gQueue, n int32) {
  5983  	assertLockHeld(&sched.lock)
  5984  
  5985  	sched.runq.pushBackAll(*batch)
  5986  	sched.runqsize += n
  5987  	*batch = gQueue{}
  5988  }
  5989  
  5990  // Try get a batch of G's from the global runnable queue.
  5991  // sched.lock must be held.
  5992  func globrunqget(pp *p, max int32) *g {
  5993  	assertLockHeld(&sched.lock)
  5994  
  5995  	if sched.runqsize == 0 {
  5996  		return nil
  5997  	}
  5998  
  5999  	n := sched.runqsize/gomaxprocs + 1
  6000  	if n > sched.runqsize {
  6001  		n = sched.runqsize
  6002  	}
  6003  	if max > 0 && n > max {
  6004  		n = max
  6005  	}
  6006  	if n > int32(len(pp.runq))/2 {
  6007  		n = int32(len(pp.runq)) / 2
  6008  	}
  6009  
  6010  	sched.runqsize -= n
  6011  
  6012  	gp := sched.runq.pop()
  6013  	n--
  6014  	for ; n > 0; n-- {
  6015  		gp1 := sched.runq.pop()
  6016  		runqput(pp, gp1, false)
  6017  	}
  6018  	return gp
  6019  }
  6020  
  6021  // pMask is an atomic bitstring with one bit per P.
  6022  type pMask []uint32
  6023  
  6024  // read returns true if P id's bit is set.
  6025  func (p pMask) read(id uint32) bool {
  6026  	word := id / 32
  6027  	mask := uint32(1) << (id % 32)
  6028  	return (atomic.Load(&p[word]) & mask) != 0
  6029  }
  6030  
  6031  // set sets P id's bit.
  6032  func (p pMask) set(id int32) {
  6033  	word := id / 32
  6034  	mask := uint32(1) << (id % 32)
  6035  	atomic.Or(&p[word], mask)
  6036  }
  6037  
  6038  // clear clears P id's bit.
  6039  func (p pMask) clear(id int32) {
  6040  	word := id / 32
  6041  	mask := uint32(1) << (id % 32)
  6042  	atomic.And(&p[word], ^mask)
  6043  }
  6044  
  6045  // updateTimerPMask clears pp's timer mask if it has no timers on its heap.
  6046  //
  6047  // Ideally, the timer mask would be kept immediately consistent on any timer
  6048  // operations. Unfortunately, updating a shared global data structure in the
  6049  // timer hot path adds too much overhead in applications frequently switching
  6050  // between no timers and some timers.
  6051  //
  6052  // As a compromise, the timer mask is updated only on pidleget / pidleput. A
  6053  // running P (returned by pidleget) may add a timer at any time, so its mask
  6054  // must be set. An idle P (passed to pidleput) cannot add new timers while
  6055  // idle, so if it has no timers at that time, its mask may be cleared.
  6056  //
  6057  // Thus, we get the following effects on timer-stealing in findrunnable:
  6058  //
  6059  //   - Idle Ps with no timers when they go idle are never checked in findrunnable
  6060  //     (for work- or timer-stealing; this is the ideal case).
  6061  //   - Running Ps must always be checked.
  6062  //   - Idle Ps whose timers are stolen must continue to be checked until they run
  6063  //     again, even after timer expiration.
  6064  //
  6065  // When the P starts running again, the mask should be set, as a timer may be
  6066  // added at any time.
  6067  //
  6068  // TODO(prattmic): Additional targeted updates may improve the above cases.
  6069  // e.g., updating the mask when stealing a timer.
  6070  func updateTimerPMask(pp *p) {
  6071  	if pp.numTimers.Load() > 0 {
  6072  		return
  6073  	}
  6074  
  6075  	// Looks like there are no timers, however another P may transiently
  6076  	// decrement numTimers when handling a timerModified timer in
  6077  	// checkTimers. We must take timersLock to serialize with these changes.
  6078  	lock(&pp.timersLock)
  6079  	if pp.numTimers.Load() == 0 {
  6080  		timerpMask.clear(pp.id)
  6081  	}
  6082  	unlock(&pp.timersLock)
  6083  }
  6084  
  6085  // pidleput puts p on the _Pidle list. now must be a relatively recent call
  6086  // to nanotime or zero. Returns now or the current time if now was zero.
  6087  //
  6088  // This releases ownership of p. Once sched.lock is released it is no longer
  6089  // safe to use p.
  6090  //
  6091  // sched.lock must be held.
  6092  //
  6093  // May run during STW, so write barriers are not allowed.
  6094  //
  6095  //go:nowritebarrierrec
  6096  func pidleput(pp *p, now int64) int64 {
  6097  	assertLockHeld(&sched.lock)
  6098  
  6099  	if !runqempty(pp) {
  6100  		throw("pidleput: P has non-empty run queue")
  6101  	}
  6102  	if now == 0 {
  6103  		now = nanotime()
  6104  	}
  6105  	updateTimerPMask(pp) // clear if there are no timers.
  6106  	idlepMask.set(pp.id)
  6107  	pp.link = sched.pidle
  6108  	sched.pidle.set(pp)
  6109  	sched.npidle.Add(1)
  6110  	if !pp.limiterEvent.start(limiterEventIdle, now) {
  6111  		throw("must be able to track idle limiter event")
  6112  	}
  6113  	return now
  6114  }
  6115  
  6116  // pidleget tries to get a p from the _Pidle list, acquiring ownership.
  6117  //
  6118  // sched.lock must be held.
  6119  //
  6120  // May run during STW, so write barriers are not allowed.
  6121  //
  6122  //go:nowritebarrierrec
  6123  func pidleget(now int64) (*p, int64) {
  6124  	assertLockHeld(&sched.lock)
  6125  
  6126  	pp := sched.pidle.ptr()
  6127  	if pp != nil {
  6128  		// Timer may get added at any time now.
  6129  		if now == 0 {
  6130  			now = nanotime()
  6131  		}
  6132  		timerpMask.set(pp.id)
  6133  		idlepMask.clear(pp.id)
  6134  		sched.pidle = pp.link
  6135  		sched.npidle.Add(-1)
  6136  		pp.limiterEvent.stop(limiterEventIdle, now)
  6137  	}
  6138  	return pp, now
  6139  }
  6140  
  6141  // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
  6142  // This is called by spinning Ms (or callers than need a spinning M) that have
  6143  // found work. If no P is available, this must synchronized with non-spinning
  6144  // Ms that may be preparing to drop their P without discovering this work.
  6145  //
  6146  // sched.lock must be held.
  6147  //
  6148  // May run during STW, so write barriers are not allowed.
  6149  //
  6150  //go:nowritebarrierrec
  6151  func pidlegetSpinning(now int64) (*p, int64) {
  6152  	assertLockHeld(&sched.lock)
  6153  
  6154  	pp, now := pidleget(now)
  6155  	if pp == nil {
  6156  		// See "Delicate dance" comment in findrunnable. We found work
  6157  		// that we cannot take, we must synchronize with non-spinning
  6158  		// Ms that may be preparing to drop their P.
  6159  		sched.needspinning.Store(1)
  6160  		return nil, now
  6161  	}
  6162  
  6163  	return pp, now
  6164  }
  6165  
  6166  // runqempty reports whether pp has no Gs on its local run queue.
  6167  // It never returns true spuriously.
  6168  func runqempty(pp *p) bool {
  6169  	// Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
  6170  	// 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
  6171  	// Simply observing that runqhead == runqtail and then observing that runqnext == nil
  6172  	// does not mean the queue is empty.
  6173  	for {
  6174  		head := atomic.Load(&pp.runqhead)
  6175  		tail := atomic.Load(&pp.runqtail)
  6176  		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
  6177  		if tail == atomic.Load(&pp.runqtail) {
  6178  			return head == tail && runnext == 0
  6179  		}
  6180  	}
  6181  }
  6182  
  6183  // To shake out latent assumptions about scheduling order,
  6184  // we introduce some randomness into scheduling decisions
  6185  // when running with the race detector.
  6186  // The need for this was made obvious by changing the
  6187  // (deterministic) scheduling order in Go 1.5 and breaking
  6188  // many poorly-written tests.
  6189  // With the randomness here, as long as the tests pass
  6190  // consistently with -race, they shouldn't have latent scheduling
  6191  // assumptions.
  6192  const randomizeScheduler = raceenabled
  6193  
  6194  // runqput tries to put g on the local runnable queue.
  6195  // If next is false, runqput adds g to the tail of the runnable queue.
  6196  // If next is true, runqput puts g in the pp.runnext slot.
  6197  // If the run queue is full, runnext puts g on the global queue.
  6198  // Executed only by the owner P.
  6199  func runqput(pp *p, gp *g, next bool) {
  6200  	if randomizeScheduler && next && fastrandn(2) == 0 {
  6201  		next = false
  6202  	}
  6203  
  6204  	if next {
  6205  	retryNext:
  6206  		oldnext := pp.runnext
  6207  		if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
  6208  			goto retryNext
  6209  		}
  6210  		if oldnext == 0 {
  6211  			return
  6212  		}
  6213  		// Kick the old runnext out to the regular run queue.
  6214  		gp = oldnext.ptr()
  6215  	}
  6216  
  6217  retry:
  6218  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  6219  	t := pp.runqtail
  6220  	if t-h < uint32(len(pp.runq)) {
  6221  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  6222  		atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
  6223  		return
  6224  	}
  6225  	if runqputslow(pp, gp, h, t) {
  6226  		return
  6227  	}
  6228  	// the queue is not full, now the put above must succeed
  6229  	goto retry
  6230  }
  6231  
  6232  // Put g and a batch of work from local runnable queue on global queue.
  6233  // Executed only by the owner P.
  6234  func runqputslow(pp *p, gp *g, h, t uint32) bool {
  6235  	var batch [len(pp.runq)/2 + 1]*g
  6236  
  6237  	// First, grab a batch from local queue.
  6238  	n := t - h
  6239  	n = n / 2
  6240  	if n != uint32(len(pp.runq)/2) {
  6241  		throw("runqputslow: queue is not full")
  6242  	}
  6243  	for i := uint32(0); i < n; i++ {
  6244  		batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  6245  	}
  6246  	if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  6247  		return false
  6248  	}
  6249  	batch[n] = gp
  6250  
  6251  	if randomizeScheduler {
  6252  		for i := uint32(1); i <= n; i++ {
  6253  			j := fastrandn(i + 1)
  6254  			batch[i], batch[j] = batch[j], batch[i]
  6255  		}
  6256  	}
  6257  
  6258  	// Link the goroutines.
  6259  	for i := uint32(0); i < n; i++ {
  6260  		batch[i].schedlink.set(batch[i+1])
  6261  	}
  6262  	var q gQueue
  6263  	q.head.set(batch[0])
  6264  	q.tail.set(batch[n])
  6265  
  6266  	// Now put the batch on global queue.
  6267  	lock(&sched.lock)
  6268  	globrunqputbatch(&q, int32(n+1))
  6269  	unlock(&sched.lock)
  6270  	return true
  6271  }
  6272  
  6273  // runqputbatch tries to put all the G's on q on the local runnable queue.
  6274  // If the queue is full, they are put on the global queue; in that case
  6275  // this will temporarily acquire the scheduler lock.
  6276  // Executed only by the owner P.
  6277  func runqputbatch(pp *p, q *gQueue, qsize int) {
  6278  	h := atomic.LoadAcq(&pp.runqhead)
  6279  	t := pp.runqtail
  6280  	n := uint32(0)
  6281  	for !q.empty() && t-h < uint32(len(pp.runq)) {
  6282  		gp := q.pop()
  6283  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  6284  		t++
  6285  		n++
  6286  	}
  6287  	qsize -= int(n)
  6288  
  6289  	if randomizeScheduler {
  6290  		off := func(o uint32) uint32 {
  6291  			return (pp.runqtail + o) % uint32(len(pp.runq))
  6292  		}
  6293  		for i := uint32(1); i < n; i++ {
  6294  			j := fastrandn(i + 1)
  6295  			pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
  6296  		}
  6297  	}
  6298  
  6299  	atomic.StoreRel(&pp.runqtail, t)
  6300  	if !q.empty() {
  6301  		lock(&sched.lock)
  6302  		globrunqputbatch(q, int32(qsize))
  6303  		unlock(&sched.lock)
  6304  	}
  6305  }
  6306  
  6307  // Get g from local runnable queue.
  6308  // If inheritTime is true, gp should inherit the remaining time in the
  6309  // current time slice. Otherwise, it should start a new time slice.
  6310  // Executed only by the owner P.
  6311  func runqget(pp *p) (gp *g, inheritTime bool) {
  6312  	// If there's a runnext, it's the next G to run.
  6313  	next := pp.runnext
  6314  	// If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
  6315  	// because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
  6316  	// Hence, there's no need to retry this CAS if it fails.
  6317  	if next != 0 && pp.runnext.cas(next, 0) {
  6318  		return next.ptr(), true
  6319  	}
  6320  
  6321  	for {
  6322  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6323  		t := pp.runqtail
  6324  		if t == h {
  6325  			return nil, false
  6326  		}
  6327  		gp := pp.runq[h%uint32(len(pp.runq))].ptr()
  6328  		if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
  6329  			return gp, false
  6330  		}
  6331  	}
  6332  }
  6333  
  6334  // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
  6335  // Executed only by the owner P.
  6336  func runqdrain(pp *p) (drainQ gQueue, n uint32) {
  6337  	oldNext := pp.runnext
  6338  	if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
  6339  		drainQ.pushBack(oldNext.ptr())
  6340  		n++
  6341  	}
  6342  
  6343  retry:
  6344  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6345  	t := pp.runqtail
  6346  	qn := t - h
  6347  	if qn == 0 {
  6348  		return
  6349  	}
  6350  	if qn > uint32(len(pp.runq)) { // read inconsistent h and t
  6351  		goto retry
  6352  	}
  6353  
  6354  	if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
  6355  		goto retry
  6356  	}
  6357  
  6358  	// We've inverted the order in which it gets G's from the local P's runnable queue
  6359  	// and then advances the head pointer because we don't want to mess up the statuses of G's
  6360  	// while runqdrain() and runqsteal() are running in parallel.
  6361  	// Thus we should advance the head pointer before draining the local P into a gQueue,
  6362  	// so that we can update any gp.schedlink only after we take the full ownership of G,
  6363  	// meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
  6364  	// See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
  6365  	for i := uint32(0); i < qn; i++ {
  6366  		gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  6367  		drainQ.pushBack(gp)
  6368  		n++
  6369  	}
  6370  	return
  6371  }
  6372  
  6373  // Grabs a batch of goroutines from pp's runnable queue into batch.
  6374  // Batch is a ring buffer starting at batchHead.
  6375  // Returns number of grabbed goroutines.
  6376  // Can be executed by any P.
  6377  func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
  6378  	for {
  6379  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  6380  		t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
  6381  		n := t - h
  6382  		n = n - n/2
  6383  		if n == 0 {
  6384  			if stealRunNextG {
  6385  				// Try to steal from pp.runnext.
  6386  				if next := pp.runnext; next != 0 {
  6387  					if pp.status == _Prunning {
  6388  						// Sleep to ensure that pp isn't about to run the g
  6389  						// we are about to steal.
  6390  						// The important use case here is when the g running
  6391  						// on pp ready()s another g and then almost
  6392  						// immediately blocks. Instead of stealing runnext
  6393  						// in this window, back off to give pp a chance to
  6394  						// schedule runnext. This will avoid thrashing gs
  6395  						// between different Ps.
  6396  						// A sync chan send/recv takes ~50ns as of time of
  6397  						// writing, so 3us gives ~50x overshoot.
  6398  						if GOOS != "windows" && GOOS != "openbsd" && GOOS != "netbsd" {
  6399  							usleep(3)
  6400  						} else {
  6401  							// On some platforms system timer granularity is
  6402  							// 1-15ms, which is way too much for this
  6403  							// optimization. So just yield.
  6404  							osyield()
  6405  						}
  6406  					}
  6407  					if !pp.runnext.cas(next, 0) {
  6408  						continue
  6409  					}
  6410  					batch[batchHead%uint32(len(batch))] = next
  6411  					return 1
  6412  				}
  6413  			}
  6414  			return 0
  6415  		}
  6416  		if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
  6417  			continue
  6418  		}
  6419  		for i := uint32(0); i < n; i++ {
  6420  			g := pp.runq[(h+i)%uint32(len(pp.runq))]
  6421  			batch[(batchHead+i)%uint32(len(batch))] = g
  6422  		}
  6423  		if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  6424  			return n
  6425  		}
  6426  	}
  6427  }
  6428  
  6429  // Steal half of elements from local runnable queue of p2
  6430  // and put onto local runnable queue of p.
  6431  // Returns one of the stolen elements (or nil if failed).
  6432  func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
  6433  	t := pp.runqtail
  6434  	n := runqgrab(p2, &pp.runq, t, stealRunNextG)
  6435  	if n == 0 {
  6436  		return nil
  6437  	}
  6438  	n--
  6439  	gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
  6440  	if n == 0 {
  6441  		return gp
  6442  	}
  6443  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  6444  	if t-h+n >= uint32(len(pp.runq)) {
  6445  		throw("runqsteal: runq overflow")
  6446  	}
  6447  	atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
  6448  	return gp
  6449  }
  6450  
  6451  // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
  6452  // be on one gQueue or gList at a time.
  6453  type gQueue struct {
  6454  	head guintptr
  6455  	tail guintptr
  6456  }
  6457  
  6458  // empty reports whether q is empty.
  6459  func (q *gQueue) empty() bool {
  6460  	return q.head == 0
  6461  }
  6462  
  6463  // push adds gp to the head of q.
  6464  func (q *gQueue) push(gp *g) {
  6465  	gp.schedlink = q.head
  6466  	q.head.set(gp)
  6467  	if q.tail == 0 {
  6468  		q.tail.set(gp)
  6469  	}
  6470  }
  6471  
  6472  // pushBack adds gp to the tail of q.
  6473  func (q *gQueue) pushBack(gp *g) {
  6474  	gp.schedlink = 0
  6475  	if q.tail != 0 {
  6476  		q.tail.ptr().schedlink.set(gp)
  6477  	} else {
  6478  		q.head.set(gp)
  6479  	}
  6480  	q.tail.set(gp)
  6481  }
  6482  
  6483  // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
  6484  // not be used.
  6485  func (q *gQueue) pushBackAll(q2 gQueue) {
  6486  	if q2.tail == 0 {
  6487  		return
  6488  	}
  6489  	q2.tail.ptr().schedlink = 0
  6490  	if q.tail != 0 {
  6491  		q.tail.ptr().schedlink = q2.head
  6492  	} else {
  6493  		q.head = q2.head
  6494  	}
  6495  	q.tail = q2.tail
  6496  }
  6497  
  6498  // pop removes and returns the head of queue q. It returns nil if
  6499  // q is empty.
  6500  func (q *gQueue) pop() *g {
  6501  	gp := q.head.ptr()
  6502  	if gp != nil {
  6503  		q.head = gp.schedlink
  6504  		if q.head == 0 {
  6505  			q.tail = 0
  6506  		}
  6507  	}
  6508  	return gp
  6509  }
  6510  
  6511  // popList takes all Gs in q and returns them as a gList.
  6512  func (q *gQueue) popList() gList {
  6513  	stack := gList{q.head}
  6514  	*q = gQueue{}
  6515  	return stack
  6516  }
  6517  
  6518  // A gList is a list of Gs linked through g.schedlink. A G can only be
  6519  // on one gQueue or gList at a time.
  6520  type gList struct {
  6521  	head guintptr
  6522  }
  6523  
  6524  // empty reports whether l is empty.
  6525  func (l *gList) empty() bool {
  6526  	return l.head == 0
  6527  }
  6528  
  6529  // push adds gp to the head of l.
  6530  func (l *gList) push(gp *g) {
  6531  	gp.schedlink = l.head
  6532  	l.head.set(gp)
  6533  }
  6534  
  6535  // pushAll prepends all Gs in q to l.
  6536  func (l *gList) pushAll(q gQueue) {
  6537  	if !q.empty() {
  6538  		q.tail.ptr().schedlink = l.head
  6539  		l.head = q.head
  6540  	}
  6541  }
  6542  
  6543  // pop removes and returns the head of l. If l is empty, it returns nil.
  6544  func (l *gList) pop() *g {
  6545  	gp := l.head.ptr()
  6546  	if gp != nil {
  6547  		l.head = gp.schedlink
  6548  	}
  6549  	return gp
  6550  }
  6551  
  6552  //go:linkname setMaxThreads runtime/debug.setMaxThreads
  6553  func setMaxThreads(in int) (out int) {
  6554  	lock(&sched.lock)
  6555  	out = int(sched.maxmcount)
  6556  	if in > 0x7fffffff { // MaxInt32
  6557  		sched.maxmcount = 0x7fffffff
  6558  	} else {
  6559  		sched.maxmcount = int32(in)
  6560  	}
  6561  	checkmcount()
  6562  	unlock(&sched.lock)
  6563  	return
  6564  }
  6565  
  6566  //go:nosplit
  6567  func procPin() int {
  6568  	gp := getg()
  6569  	mp := gp.m
  6570  
  6571  	mp.locks++
  6572  	return int(mp.p.ptr().id)
  6573  }
  6574  
  6575  //go:nosplit
  6576  func procUnpin() {
  6577  	gp := getg()
  6578  	gp.m.locks--
  6579  }
  6580  
  6581  //go:linkname sync_runtime_procPin sync.runtime_procPin
  6582  //go:nosplit
  6583  func sync_runtime_procPin() int {
  6584  	return procPin()
  6585  }
  6586  
  6587  //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
  6588  //go:nosplit
  6589  func sync_runtime_procUnpin() {
  6590  	procUnpin()
  6591  }
  6592  
  6593  //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
  6594  //go:nosplit
  6595  func sync_atomic_runtime_procPin() int {
  6596  	return procPin()
  6597  }
  6598  
  6599  //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
  6600  //go:nosplit
  6601  func sync_atomic_runtime_procUnpin() {
  6602  	procUnpin()
  6603  }
  6604  
  6605  // Active spinning for sync.Mutex.
  6606  //
  6607  //go:linkname sync_runtime_canSpin sync.runtime_canSpin
  6608  //go:nosplit
  6609  func sync_runtime_canSpin(i int) bool {
  6610  	// sync.Mutex is cooperative, so we are conservative with spinning.
  6611  	// Spin only few times and only if running on a multicore machine and
  6612  	// GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
  6613  	// As opposed to runtime mutex we don't do passive spinning here,
  6614  	// because there can be work on global runq or on other Ps.
  6615  	if i >= active_spin || ncpu <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
  6616  		return false
  6617  	}
  6618  	if p := getg().m.p.ptr(); !runqempty(p) {
  6619  		return false
  6620  	}
  6621  	return true
  6622  }
  6623  
  6624  //go:linkname sync_runtime_doSpin sync.runtime_doSpin
  6625  //go:nosplit
  6626  func sync_runtime_doSpin() {
  6627  	procyield(active_spin_cnt)
  6628  }
  6629  
  6630  var stealOrder randomOrder
  6631  
  6632  // randomOrder/randomEnum are helper types for randomized work stealing.
  6633  // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
  6634  // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
  6635  // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
  6636  type randomOrder struct {
  6637  	count    uint32
  6638  	coprimes []uint32
  6639  }
  6640  
  6641  type randomEnum struct {
  6642  	i     uint32
  6643  	count uint32
  6644  	pos   uint32
  6645  	inc   uint32
  6646  }
  6647  
  6648  func (ord *randomOrder) reset(count uint32) {
  6649  	ord.count = count
  6650  	ord.coprimes = ord.coprimes[:0]
  6651  	for i := uint32(1); i <= count; i++ {
  6652  		if gcd(i, count) == 1 {
  6653  			ord.coprimes = append(ord.coprimes, i)
  6654  		}
  6655  	}
  6656  }
  6657  
  6658  func (ord *randomOrder) start(i uint32) randomEnum {
  6659  	return randomEnum{
  6660  		count: ord.count,
  6661  		pos:   i % ord.count,
  6662  		inc:   ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
  6663  	}
  6664  }
  6665  
  6666  func (enum *randomEnum) done() bool {
  6667  	return enum.i == enum.count
  6668  }
  6669  
  6670  func (enum *randomEnum) next() {
  6671  	enum.i++
  6672  	enum.pos = (enum.pos + enum.inc) % enum.count
  6673  }
  6674  
  6675  func (enum *randomEnum) position() uint32 {
  6676  	return enum.pos
  6677  }
  6678  
  6679  func gcd(a, b uint32) uint32 {
  6680  	for b != 0 {
  6681  		a, b = b, a%b
  6682  	}
  6683  	return a
  6684  }
  6685  
  6686  // An initTask represents the set of initializations that need to be done for a package.
  6687  // Keep in sync with ../../test/noinit.go:initTask
  6688  type initTask struct {
  6689  	state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
  6690  	nfns  uint32
  6691  	// followed by nfns pcs, uintptr sized, one per init function to run
  6692  }
  6693  
  6694  // inittrace stores statistics for init functions which are
  6695  // updated by malloc and newproc when active is true.
  6696  var inittrace tracestat
  6697  
  6698  type tracestat struct {
  6699  	active bool   // init tracing activation status
  6700  	id     uint64 // init goroutine id
  6701  	allocs uint64 // heap allocations
  6702  	bytes  uint64 // heap allocated bytes
  6703  }
  6704  
  6705  func doInit(ts []*initTask) {
  6706  	for _, t := range ts {
  6707  		doInit1(t)
  6708  	}
  6709  }
  6710  
  6711  func doInit1(t *initTask) {
  6712  	switch t.state {
  6713  	case 2: // fully initialized
  6714  		return
  6715  	case 1: // initialization in progress
  6716  		throw("recursive call during initialization - linker skew")
  6717  	default: // not initialized yet
  6718  		t.state = 1 // initialization in progress
  6719  
  6720  		var (
  6721  			start  int64
  6722  			before tracestat
  6723  		)
  6724  
  6725  		if inittrace.active {
  6726  			start = nanotime()
  6727  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  6728  			before = inittrace
  6729  		}
  6730  
  6731  		if t.nfns == 0 {
  6732  			// We should have pruned all of these in the linker.
  6733  			throw("inittask with no functions")
  6734  		}
  6735  
  6736  		firstFunc := add(unsafe.Pointer(t), 8)
  6737  		for i := uint32(0); i < t.nfns; i++ {
  6738  			p := add(firstFunc, uintptr(i)*goarch.PtrSize)
  6739  			f := *(*func())(unsafe.Pointer(&p))
  6740  			f()
  6741  		}
  6742  
  6743  		if inittrace.active {
  6744  			end := nanotime()
  6745  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  6746  			after := inittrace
  6747  
  6748  			f := *(*func())(unsafe.Pointer(&firstFunc))
  6749  			pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
  6750  
  6751  			var sbuf [24]byte
  6752  			print("init ", pkg, " @")
  6753  			print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
  6754  			print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
  6755  			print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
  6756  			print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
  6757  			print("\n")
  6758  		}
  6759  
  6760  		t.state = 2 // initialization done
  6761  	}
  6762  }
  6763