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  	"internal/goos"
    12  	"internal/runtime/atomic"
    13  	"internal/runtime/exithook"
    14  	"internal/runtime/strconv"
    15  	"internal/runtime/sys"
    16  	"internal/stringslite"
    17  	"unsafe"
    18  )
    19  
    20  // set using cmd/go/internal/modload.ModInfoProg
    21  var modinfo string
    22  
    23  // Goroutine scheduler
    24  // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
    25  //
    26  // The main concepts are:
    27  // G - goroutine.
    28  // M - worker thread, or machine.
    29  // P - processor, a resource that is required to execute Go code.
    30  //     M must have an associated P to execute Go code, however it can be
    31  //     blocked or in a syscall w/o an associated P.
    32  //
    33  // Design doc at https://golang.org/s/go11sched.
    34  
    35  // Worker thread parking/unparking.
    36  // We need to balance between keeping enough running worker threads to utilize
    37  // available hardware parallelism and parking excessive running worker threads
    38  // to conserve CPU resources and power. This is not simple for two reasons:
    39  // (1) scheduler state is intentionally distributed (in particular, per-P work
    40  // queues), so it is not possible to compute global predicates on fast paths;
    41  // (2) for optimal thread management we would need to know the future (don't park
    42  // a worker thread when a new goroutine will be readied in near future).
    43  //
    44  // Three rejected approaches that would work badly:
    45  // 1. Centralize all scheduler state (would inhibit scalability).
    46  // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
    47  //    is a spare P, unpark a thread and handoff it the thread and the goroutine.
    48  //    This would lead to thread state thrashing, as the thread that readied the
    49  //    goroutine can be out of work the very next moment, we will need to park it.
    50  //    Also, it would destroy locality of computation as we want to preserve
    51  //    dependent goroutines on the same thread; and introduce additional latency.
    52  // 3. Unpark an additional thread whenever we ready a goroutine and there is an
    53  //    idle P, but don't do handoff. This would lead to excessive thread parking/
    54  //    unparking as the additional threads will instantly park without discovering
    55  //    any work to do.
    56  //
    57  // The current approach:
    58  //
    59  // This approach applies to three primary sources of potential work: readying a
    60  // goroutine, new/modified-earlier timers, and idle-priority GC. See below for
    61  // additional details.
    62  //
    63  // We unpark an additional thread when we submit work if (this is wakep()):
    64  // 1. There is an idle P, and
    65  // 2. There are no "spinning" worker threads.
    66  //
    67  // A worker thread is considered spinning if it is out of local work and did
    68  // not find work in the global run queue or netpoller; the spinning state is
    69  // denoted in m.spinning and in sched.nmspinning. Threads unparked this way are
    70  // also considered spinning; we don't do goroutine handoff so such threads are
    71  // out of work initially. Spinning threads spin on looking for work in per-P
    72  // run queues and timer heaps or from the GC before parking. If a spinning
    73  // thread finds work it takes itself out of the spinning state and proceeds to
    74  // execution. If it does not find work it takes itself out of the spinning
    75  // state and then parks.
    76  //
    77  // If there is at least one spinning thread (sched.nmspinning>1), we don't
    78  // unpark new threads when submitting work. To compensate for that, if the last
    79  // spinning thread finds work and stops spinning, it must unpark a new spinning
    80  // thread. This approach smooths out unjustified spikes of thread unparking,
    81  // but at the same time guarantees eventual maximal CPU parallelism
    82  // utilization.
    83  //
    84  // The main implementation complication is that we need to be very careful
    85  // during spinning->non-spinning thread transition. This transition can race
    86  // with submission of new work, and either one part or another needs to unpark
    87  // another worker thread. If they both fail to do that, we can end up with
    88  // semi-persistent CPU underutilization.
    89  //
    90  // The general pattern for submission is:
    91  // 1. Submit work to the local or global run queue, timer heap, or GC state.
    92  // 2. #StoreLoad-style memory barrier.
    93  // 3. Check sched.nmspinning.
    94  //
    95  // The general pattern for spinning->non-spinning transition is:
    96  // 1. Decrement nmspinning.
    97  // 2. #StoreLoad-style memory barrier.
    98  // 3. Check all per-P work queues and GC for new work.
    99  //
   100  // Note that all this complexity does not apply to global run queue as we are
   101  // not sloppy about thread unparking when submitting to global queue. Also see
   102  // comments for nmspinning manipulation.
   103  //
   104  // How these different sources of work behave varies, though it doesn't affect
   105  // the synchronization approach:
   106  // * Ready goroutine: this is an obvious source of work; the goroutine is
   107  //   immediately ready and must run on some thread eventually.
   108  // * New/modified-earlier timer: The current timer implementation (see time.go)
   109  //   uses netpoll in a thread with no work available to wait for the soonest
   110  //   timer. If there is no thread waiting, we want a new spinning thread to go
   111  //   wait.
   112  // * Idle-priority GC: The GC wakes a stopped idle thread to contribute to
   113  //   background GC work (note: currently disabled per golang.org/issue/19112).
   114  //   Also see golang.org/issue/44313, as this should be extended to all GC
   115  //   workers.
   116  
   117  var (
   118  	m0           m
   119  	g0           g
   120  	mcache0      *mcache
   121  	raceprocctx0 uintptr
   122  	raceFiniLock mutex
   123  )
   124  
   125  // This slice records the initializing tasks that need to be
   126  // done to start up the runtime. It is built by the linker.
   127  var runtime_inittasks []*initTask
   128  
   129  // main_init_done is a signal used by cgocallbackg that initialization
   130  // has been completed. It is made before _cgo_notify_runtime_init_done,
   131  // so all cgo calls can rely on it existing. When main_init is complete,
   132  // it is closed, meaning cgocallbackg can reliably receive from it.
   133  var main_init_done chan bool
   134  
   135  //go:linkname main_main main.main
   136  func main_main()
   137  
   138  // mainStarted indicates that the main M has started.
   139  var mainStarted bool
   140  
   141  // runtimeInitTime is the nanotime() at which the runtime started.
   142  var runtimeInitTime int64
   143  
   144  // Value to use for signal mask for newly created M's.
   145  var initSigmask sigset
   146  
   147  // The main goroutine.
   148  func main() {
   149  	mp := getg().m
   150  
   151  	// Racectx of m0->g0 is used only as the parent of the main goroutine.
   152  	// It must not be used for anything else.
   153  	mp.g0.racectx = 0
   154  
   155  	// Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
   156  	// Using decimal instead of binary GB and MB because
   157  	// they look nicer in the stack overflow failure message.
   158  	if goarch.PtrSize == 8 {
   159  		maxstacksize = 1000000000
   160  	} else {
   161  		maxstacksize = 250000000
   162  	}
   163  
   164  	// An upper limit for max stack size. Used to avoid random crashes
   165  	// after calling SetMaxStack and trying to allocate a stack that is too big,
   166  	// since stackalloc works with 32-bit sizes.
   167  	maxstackceiling = 2 * maxstacksize
   168  
   169  	// Allow newproc to start new Ms.
   170  	mainStarted = true
   171  
   172  	if haveSysmon {
   173  		systemstack(func() {
   174  			newm(sysmon, nil, -1)
   175  		})
   176  	}
   177  
   178  	// Lock the main goroutine onto this, the main OS thread,
   179  	// during initialization. Most programs won't care, but a few
   180  	// do require certain calls to be made by the main thread.
   181  	// Those can arrange for main.main to run in the main thread
   182  	// by calling runtime.LockOSThread during initialization
   183  	// to preserve the lock.
   184  	lockOSThread()
   185  
   186  	if mp != &m0 {
   187  		throw("runtime.main not on m0")
   188  	}
   189  
   190  	// Record when the world started.
   191  	// Must be before doInit for tracing init.
   192  	runtimeInitTime = nanotime()
   193  	if runtimeInitTime == 0 {
   194  		throw("nanotime returning zero")
   195  	}
   196  
   197  	if debug.inittrace != 0 {
   198  		inittrace.id = getg().goid
   199  		inittrace.active = true
   200  	}
   201  
   202  	doInit(runtime_inittasks) // Must be before defer.
   203  
   204  	// Defer unlock so that runtime.Goexit during init does the unlock too.
   205  	needUnlock := true
   206  	defer func() {
   207  		if needUnlock {
   208  			unlockOSThread()
   209  		}
   210  	}()
   211  
   212  	gcenable()
   213  	defaultGOMAXPROCSUpdateEnable() // don't STW before runtime initialized.
   214  
   215  	main_init_done = make(chan bool)
   216  	if iscgo {
   217  		if _cgo_pthread_key_created == nil {
   218  			throw("_cgo_pthread_key_created missing")
   219  		}
   220  
   221  		if _cgo_thread_start == nil {
   222  			throw("_cgo_thread_start missing")
   223  		}
   224  		if GOOS != "windows" {
   225  			if _cgo_setenv == nil {
   226  				throw("_cgo_setenv missing")
   227  			}
   228  			if _cgo_unsetenv == nil {
   229  				throw("_cgo_unsetenv missing")
   230  			}
   231  		}
   232  		if _cgo_notify_runtime_init_done == nil {
   233  			throw("_cgo_notify_runtime_init_done missing")
   234  		}
   235  
   236  		// Set the x_crosscall2_ptr C function pointer variable point to crosscall2.
   237  		if set_crosscall2 == nil {
   238  			throw("set_crosscall2 missing")
   239  		}
   240  		set_crosscall2()
   241  
   242  		// Start the template thread in case we enter Go from
   243  		// a C-created thread and need to create a new thread.
   244  		startTemplateThread()
   245  		cgocall(_cgo_notify_runtime_init_done, nil)
   246  	}
   247  
   248  	// Run the initializing tasks. Depending on build mode this
   249  	// list can arrive a few different ways, but it will always
   250  	// contain the init tasks computed by the linker for all the
   251  	// packages in the program (excluding those added at runtime
   252  	// by package plugin). Run through the modules in dependency
   253  	// order (the order they are initialized by the dynamic
   254  	// loader, i.e. they are added to the moduledata linked list).
   255  	last := lastmoduledatap // grab before loop starts. Any added modules after this point will do their own doInit calls.
   256  	for m := &firstmoduledata; true; m = m.next {
   257  		doInit(m.inittasks)
   258  		if m == last {
   259  			break
   260  		}
   261  	}
   262  
   263  	// Disable init tracing after main init done to avoid overhead
   264  	// of collecting statistics in malloc and newproc
   265  	inittrace.active = false
   266  
   267  	close(main_init_done)
   268  
   269  	needUnlock = false
   270  	unlockOSThread()
   271  
   272  	if isarchive || islibrary {
   273  		// A program compiled with -buildmode=c-archive or c-shared
   274  		// has a main, but it is not executed.
   275  		if GOARCH == "wasm" {
   276  			// On Wasm, pause makes it return to the host.
   277  			// Unlike cgo callbacks where Ms are created on demand,
   278  			// on Wasm we have only one M. So we keep this M (and this
   279  			// G) for callbacks.
   280  			// Using the caller's SP unwinds this frame and backs to
   281  			// goexit. The -16 is: 8 for goexit's (fake) return PC,
   282  			// and pause's epilogue pops 8.
   283  			pause(sys.GetCallerSP() - 16) // should not return
   284  			panic("unreachable")
   285  		}
   286  		return
   287  	}
   288  	fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
   289  	fn()
   290  
   291  	exitHooksRun := false
   292  	if raceenabled {
   293  		runExitHooks(0) // run hooks now, since racefini does not return
   294  		exitHooksRun = true
   295  		racefini()
   296  	}
   297  
   298  	// Check for C memory leaks if using ASAN and we've made cgo calls,
   299  	// or if we are running as a library in a C program.
   300  	// We always make one cgo call, above, to notify_runtime_init_done,
   301  	// so we ignore that one.
   302  	// No point in leak checking if no cgo calls, since leak checking
   303  	// just looks for objects allocated using malloc and friends.
   304  	// Just checking iscgo doesn't help because asan implies iscgo.
   305  	if asanenabled && (isarchive || islibrary || NumCgoCall() > 1) {
   306  		runExitHooks(0) // lsandoleakcheck may not return
   307  		exitHooksRun = true
   308  		lsandoleakcheck()
   309  	}
   310  
   311  	// Make racy client program work: if panicking on
   312  	// another goroutine at the same time as main returns,
   313  	// let the other goroutine finish printing the panic trace.
   314  	// Once it does, it will exit. See issues 3934 and 20018.
   315  	if runningPanicDefers.Load() != 0 {
   316  		// Running deferred functions should not take long.
   317  		for c := 0; c < 1000; c++ {
   318  			if runningPanicDefers.Load() == 0 {
   319  				break
   320  			}
   321  			Gosched()
   322  		}
   323  	}
   324  	if panicking.Load() != 0 {
   325  		gopark(nil, nil, waitReasonPanicWait, traceBlockForever, 1)
   326  	}
   327  	if !exitHooksRun {
   328  		runExitHooks(0)
   329  	}
   330  
   331  	exit(0)
   332  	for {
   333  		var x *int32
   334  		*x = 0
   335  	}
   336  }
   337  
   338  // os_beforeExit is called from os.Exit(0).
   339  //
   340  //go:linkname os_beforeExit os.runtime_beforeExit
   341  func os_beforeExit(exitCode int) {
   342  	runExitHooks(exitCode)
   343  	if exitCode == 0 && raceenabled {
   344  		racefini()
   345  	}
   346  
   347  	// See comment in main, above.
   348  	if exitCode == 0 && asanenabled && (isarchive || islibrary || NumCgoCall() > 1) {
   349  		lsandoleakcheck()
   350  	}
   351  }
   352  
   353  func init() {
   354  	exithook.Gosched = Gosched
   355  	exithook.Goid = func() uint64 { return getg().goid }
   356  	exithook.Throw = throw
   357  }
   358  
   359  func runExitHooks(code int) {
   360  	exithook.Run(code)
   361  }
   362  
   363  // start forcegc helper goroutine
   364  func init() {
   365  	go forcegchelper()
   366  }
   367  
   368  func forcegchelper() {
   369  	forcegc.g = getg()
   370  	lockInit(&forcegc.lock, lockRankForcegc)
   371  	for {
   372  		lock(&forcegc.lock)
   373  		if forcegc.idle.Load() {
   374  			throw("forcegc: phase error")
   375  		}
   376  		forcegc.idle.Store(true)
   377  		goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceBlockSystemGoroutine, 1)
   378  		// this goroutine is explicitly resumed by sysmon
   379  		if debug.gctrace > 0 {
   380  			println("GC forced")
   381  		}
   382  		// Time-triggered, fully concurrent.
   383  		gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
   384  	}
   385  }
   386  
   387  // Gosched yields the processor, allowing other goroutines to run. It does not
   388  // suspend the current goroutine, so execution resumes automatically.
   389  //
   390  //go:nosplit
   391  func Gosched() {
   392  	checkTimeouts()
   393  	mcall(gosched_m)
   394  }
   395  
   396  // goschedguarded yields the processor like gosched, but also checks
   397  // for forbidden states and opts out of the yield in those cases.
   398  //
   399  //go:nosplit
   400  func goschedguarded() {
   401  	mcall(goschedguarded_m)
   402  }
   403  
   404  // goschedIfBusy yields the processor like gosched, but only does so if
   405  // there are no idle Ps or if we're on the only P and there's nothing in
   406  // the run queue. In both cases, there is freely available idle time.
   407  //
   408  //go:nosplit
   409  func goschedIfBusy() {
   410  	gp := getg()
   411  	// Call gosched if gp.preempt is set; we may be in a tight loop that
   412  	// doesn't otherwise yield.
   413  	if !gp.preempt && sched.npidle.Load() > 0 {
   414  		return
   415  	}
   416  	mcall(gosched_m)
   417  }
   418  
   419  // Puts the current goroutine into a waiting state and calls unlockf on the
   420  // system stack.
   421  //
   422  // If unlockf returns false, the goroutine is resumed.
   423  //
   424  // unlockf must not access this G's stack, as it may be moved between
   425  // the call to gopark and the call to unlockf.
   426  //
   427  // Note that because unlockf is called after putting the G into a waiting
   428  // state, the G may have already been readied by the time unlockf is called
   429  // unless there is external synchronization preventing the G from being
   430  // readied. If unlockf returns false, it must guarantee that the G cannot be
   431  // externally readied.
   432  //
   433  // Reason explains why the goroutine has been parked. It is displayed in stack
   434  // traces and heap dumps. Reasons should be unique and descriptive. Do not
   435  // re-use reasons, add new ones.
   436  //
   437  // gopark should be an internal detail,
   438  // but widely used packages access it using linkname.
   439  // Notable members of the hall of shame include:
   440  //   - gvisor.dev/gvisor
   441  //   - github.com/sagernet/gvisor
   442  //
   443  // Do not remove or change the type signature.
   444  // See go.dev/issue/67401.
   445  //
   446  //go:linkname gopark
   447  func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceReason traceBlockReason, traceskip int) {
   448  	if reason != waitReasonSleep {
   449  		checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
   450  	}
   451  	mp := acquirem()
   452  	gp := mp.curg
   453  	status := readgstatus(gp)
   454  	if status != _Grunning && status != _Gscanrunning {
   455  		throw("gopark: bad g status")
   456  	}
   457  	mp.waitlock = lock
   458  	mp.waitunlockf = unlockf
   459  	gp.waitreason = reason
   460  	mp.waitTraceBlockReason = traceReason
   461  	mp.waitTraceSkip = traceskip
   462  	releasem(mp)
   463  	// can't do anything that might move the G between Ms here.
   464  	mcall(park_m)
   465  }
   466  
   467  // Puts the current goroutine into a waiting state and unlocks the lock.
   468  // The goroutine can be made runnable again by calling goready(gp).
   469  func goparkunlock(lock *mutex, reason waitReason, traceReason traceBlockReason, traceskip int) {
   470  	gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceReason, traceskip)
   471  }
   472  
   473  // goready should be an internal detail,
   474  // but widely used packages access it using linkname.
   475  // Notable members of the hall of shame include:
   476  //   - gvisor.dev/gvisor
   477  //   - github.com/sagernet/gvisor
   478  //
   479  // Do not remove or change the type signature.
   480  // See go.dev/issue/67401.
   481  //
   482  //go:linkname goready
   483  func goready(gp *g, traceskip int) {
   484  	systemstack(func() {
   485  		ready(gp, traceskip, true)
   486  	})
   487  }
   488  
   489  //go:nosplit
   490  func acquireSudog() *sudog {
   491  	// Delicate dance: the semaphore implementation calls
   492  	// acquireSudog, acquireSudog calls new(sudog),
   493  	// new calls malloc, malloc can call the garbage collector,
   494  	// and the garbage collector calls the semaphore implementation
   495  	// in stopTheWorld.
   496  	// Break the cycle by doing acquirem/releasem around new(sudog).
   497  	// The acquirem/releasem increments m.locks during new(sudog),
   498  	// which keeps the garbage collector from being invoked.
   499  	mp := acquirem()
   500  	pp := mp.p.ptr()
   501  	if len(pp.sudogcache) == 0 {
   502  		lock(&sched.sudoglock)
   503  		// First, try to grab a batch from central cache.
   504  		for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
   505  			s := sched.sudogcache
   506  			sched.sudogcache = s.next
   507  			s.next = nil
   508  			pp.sudogcache = append(pp.sudogcache, s)
   509  		}
   510  		unlock(&sched.sudoglock)
   511  		// If the central cache is empty, allocate a new one.
   512  		if len(pp.sudogcache) == 0 {
   513  			pp.sudogcache = append(pp.sudogcache, new(sudog))
   514  		}
   515  	}
   516  	n := len(pp.sudogcache)
   517  	s := pp.sudogcache[n-1]
   518  	pp.sudogcache[n-1] = nil
   519  	pp.sudogcache = pp.sudogcache[:n-1]
   520  	if s.elem.get() != nil {
   521  		throw("acquireSudog: found s.elem != nil in cache")
   522  	}
   523  	releasem(mp)
   524  	return s
   525  }
   526  
   527  //go:nosplit
   528  func releaseSudog(s *sudog) {
   529  	if s.elem.get() != nil {
   530  		throw("runtime: sudog with non-nil elem")
   531  	}
   532  	if s.isSelect {
   533  		throw("runtime: sudog with non-false isSelect")
   534  	}
   535  	if s.next != nil {
   536  		throw("runtime: sudog with non-nil next")
   537  	}
   538  	if s.prev != nil {
   539  		throw("runtime: sudog with non-nil prev")
   540  	}
   541  	if s.waitlink != nil {
   542  		throw("runtime: sudog with non-nil waitlink")
   543  	}
   544  	if s.c.get() != nil {
   545  		throw("runtime: sudog with non-nil c")
   546  	}
   547  	gp := getg()
   548  	if gp.param != nil {
   549  		throw("runtime: releaseSudog with non-nil gp.param")
   550  	}
   551  	mp := acquirem() // avoid rescheduling to another P
   552  	pp := mp.p.ptr()
   553  	if len(pp.sudogcache) == cap(pp.sudogcache) {
   554  		// Transfer half of local cache to the central cache.
   555  		var first, last *sudog
   556  		for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
   557  			n := len(pp.sudogcache)
   558  			p := pp.sudogcache[n-1]
   559  			pp.sudogcache[n-1] = nil
   560  			pp.sudogcache = pp.sudogcache[:n-1]
   561  			if first == nil {
   562  				first = p
   563  			} else {
   564  				last.next = p
   565  			}
   566  			last = p
   567  		}
   568  		lock(&sched.sudoglock)
   569  		last.next = sched.sudogcache
   570  		sched.sudogcache = first
   571  		unlock(&sched.sudoglock)
   572  	}
   573  	pp.sudogcache = append(pp.sudogcache, s)
   574  	releasem(mp)
   575  }
   576  
   577  // called from assembly.
   578  func badmcall(fn func(*g)) {
   579  	throw("runtime: mcall called on m->g0 stack")
   580  }
   581  
   582  func badmcall2(fn func(*g)) {
   583  	throw("runtime: mcall function returned")
   584  }
   585  
   586  func badreflectcall() {
   587  	panic(plainError("arg size to reflect.call more than 1GB"))
   588  }
   589  
   590  //go:nosplit
   591  //go:nowritebarrierrec
   592  func badmorestackg0() {
   593  	if !crashStackImplemented {
   594  		writeErrStr("fatal: morestack on g0\n")
   595  		return
   596  	}
   597  
   598  	g := getg()
   599  	switchToCrashStack(func() {
   600  		print("runtime: morestack on g0, stack [", hex(g.stack.lo), " ", hex(g.stack.hi), "], sp=", hex(g.sched.sp), ", called from\n")
   601  		g.m.traceback = 2 // include pc and sp in stack trace
   602  		traceback1(g.sched.pc, g.sched.sp, g.sched.lr, g, 0)
   603  		print("\n")
   604  
   605  		throw("morestack on g0")
   606  	})
   607  }
   608  
   609  //go:nosplit
   610  //go:nowritebarrierrec
   611  func badmorestackgsignal() {
   612  	writeErrStr("fatal: morestack on gsignal\n")
   613  }
   614  
   615  //go:nosplit
   616  func badctxt() {
   617  	throw("ctxt != 0")
   618  }
   619  
   620  // gcrash is a fake g that can be used when crashing due to bad
   621  // stack conditions.
   622  var gcrash g
   623  
   624  var crashingG atomic.Pointer[g]
   625  
   626  // Switch to crashstack and call fn, with special handling of
   627  // concurrent and recursive cases.
   628  //
   629  // Nosplit as it is called in a bad stack condition (we know
   630  // morestack would fail).
   631  //
   632  //go:nosplit
   633  //go:nowritebarrierrec
   634  func switchToCrashStack(fn func()) {
   635  	me := getg()
   636  	if crashingG.CompareAndSwapNoWB(nil, me) {
   637  		switchToCrashStack0(fn) // should never return
   638  		abort()
   639  	}
   640  	if crashingG.Load() == me {
   641  		// recursive crashing. too bad.
   642  		writeErrStr("fatal: recursive switchToCrashStack\n")
   643  		abort()
   644  	}
   645  	// Another g is crashing. Give it some time, hopefully it will finish traceback.
   646  	usleep_no_g(100)
   647  	writeErrStr("fatal: concurrent switchToCrashStack\n")
   648  	abort()
   649  }
   650  
   651  // Disable crash stack on Windows for now. Apparently, throwing an exception
   652  // on a non-system-allocated crash stack causes EXCEPTION_STACK_OVERFLOW and
   653  // hangs the process (see issue 63938).
   654  const crashStackImplemented = GOOS != "windows"
   655  
   656  //go:noescape
   657  func switchToCrashStack0(fn func()) // in assembly
   658  
   659  func lockedOSThread() bool {
   660  	gp := getg()
   661  	return gp.lockedm != 0 && gp.m.lockedg != 0
   662  }
   663  
   664  var (
   665  	// allgs contains all Gs ever created (including dead Gs), and thus
   666  	// never shrinks.
   667  	//
   668  	// Access via the slice is protected by allglock or stop-the-world.
   669  	// Readers that cannot take the lock may (carefully!) use the atomic
   670  	// variables below.
   671  	allglock mutex
   672  	allgs    []*g
   673  
   674  	// allglen and allgptr are atomic variables that contain len(allgs) and
   675  	// &allgs[0] respectively. Proper ordering depends on totally-ordered
   676  	// loads and stores. Writes are protected by allglock.
   677  	//
   678  	// allgptr is updated before allglen. Readers should read allglen
   679  	// before allgptr to ensure that allglen is always <= len(allgptr). New
   680  	// Gs appended during the race can be missed. For a consistent view of
   681  	// all Gs, allglock must be held.
   682  	//
   683  	// allgptr copies should always be stored as a concrete type or
   684  	// unsafe.Pointer, not uintptr, to ensure that GC can still reach it
   685  	// even if it points to a stale array.
   686  	allglen uintptr
   687  	allgptr **g
   688  )
   689  
   690  func allgadd(gp *g) {
   691  	if readgstatus(gp) == _Gidle {
   692  		throw("allgadd: bad status Gidle")
   693  	}
   694  
   695  	lock(&allglock)
   696  	allgs = append(allgs, gp)
   697  	if &allgs[0] != allgptr {
   698  		atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
   699  	}
   700  	atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
   701  	unlock(&allglock)
   702  }
   703  
   704  // allGsSnapshot returns a snapshot of the slice of all Gs.
   705  //
   706  // The world must be stopped or allglock must be held.
   707  func allGsSnapshot() []*g {
   708  	assertWorldStoppedOrLockHeld(&allglock)
   709  
   710  	// Because the world is stopped or allglock is held, allgadd
   711  	// cannot happen concurrently with this. allgs grows
   712  	// monotonically and existing entries never change, so we can
   713  	// simply return a copy of the slice header. For added safety,
   714  	// we trim everything past len because that can still change.
   715  	return allgs[:len(allgs):len(allgs)]
   716  }
   717  
   718  // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
   719  func atomicAllG() (**g, uintptr) {
   720  	length := atomic.Loaduintptr(&allglen)
   721  	ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
   722  	return ptr, length
   723  }
   724  
   725  // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
   726  func atomicAllGIndex(ptr **g, i uintptr) *g {
   727  	return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
   728  }
   729  
   730  // forEachG calls fn on every G from allgs.
   731  //
   732  // forEachG takes a lock to exclude concurrent addition of new Gs.
   733  func forEachG(fn func(gp *g)) {
   734  	lock(&allglock)
   735  	for _, gp := range allgs {
   736  		fn(gp)
   737  	}
   738  	unlock(&allglock)
   739  }
   740  
   741  // forEachGRace calls fn on every G from allgs.
   742  //
   743  // forEachGRace avoids locking, but does not exclude addition of new Gs during
   744  // execution, which may be missed.
   745  func forEachGRace(fn func(gp *g)) {
   746  	ptr, length := atomicAllG()
   747  	for i := uintptr(0); i < length; i++ {
   748  		gp := atomicAllGIndex(ptr, i)
   749  		fn(gp)
   750  	}
   751  	return
   752  }
   753  
   754  const (
   755  	// Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
   756  	// 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
   757  	_GoidCacheBatch = 16
   758  )
   759  
   760  // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
   761  // value of the GODEBUG environment variable.
   762  func cpuinit(env string) {
   763  	cpu.Initialize(env)
   764  
   765  	// Support cpu feature variables are used in code generated by the compiler
   766  	// to guard execution of instructions that can not be assumed to be always supported.
   767  	switch GOARCH {
   768  	case "386", "amd64":
   769  		x86HasPOPCNT = cpu.X86.HasPOPCNT
   770  		x86HasSSE41 = cpu.X86.HasSSE41
   771  		x86HasFMA = cpu.X86.HasFMA
   772  
   773  	case "arm":
   774  		armHasVFPv4 = cpu.ARM.HasVFPv4
   775  
   776  	case "arm64":
   777  		arm64HasATOMICS = cpu.ARM64.HasATOMICS
   778  
   779  	case "loong64":
   780  		loong64HasLAMCAS = cpu.Loong64.HasLAMCAS
   781  		loong64HasLAM_BH = cpu.Loong64.HasLAM_BH
   782  		loong64HasLSX = cpu.Loong64.HasLSX
   783  
   784  	case "riscv64":
   785  		riscv64HasZbb = cpu.RISCV64.HasZbb
   786  	}
   787  }
   788  
   789  // getGodebugEarly extracts the environment variable GODEBUG from the environment on
   790  // Unix-like operating systems and returns it. This function exists to extract GODEBUG
   791  // early before much of the runtime is initialized.
   792  //
   793  // Returns nil, false if OS doesn't provide env vars early in the init sequence.
   794  func getGodebugEarly() (string, bool) {
   795  	const prefix = "GODEBUG="
   796  	var env string
   797  	switch GOOS {
   798  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   799  		// Similar to goenv_unix but extracts the environment value for
   800  		// GODEBUG directly.
   801  		// TODO(moehrmann): remove when general goenvs() can be called before cpuinit()
   802  		n := int32(0)
   803  		for argv_index(argv, argc+1+n) != nil {
   804  			n++
   805  		}
   806  
   807  		for i := int32(0); i < n; i++ {
   808  			p := argv_index(argv, argc+1+i)
   809  			s := unsafe.String(p, findnull(p))
   810  
   811  			if stringslite.HasPrefix(s, prefix) {
   812  				env = gostringnocopy(p)[len(prefix):]
   813  				break
   814  			}
   815  		}
   816  		break
   817  
   818  	default:
   819  		return "", false
   820  	}
   821  	return env, true
   822  }
   823  
   824  // The bootstrap sequence is:
   825  //
   826  //	call osinit
   827  //	call schedinit
   828  //	make & queue new G
   829  //	call runtime·mstart
   830  //
   831  // The new G calls runtime·main.
   832  func schedinit() {
   833  	lockInit(&sched.lock, lockRankSched)
   834  	lockInit(&sched.sysmonlock, lockRankSysmon)
   835  	lockInit(&sched.deferlock, lockRankDefer)
   836  	lockInit(&sched.sudoglock, lockRankSudog)
   837  	lockInit(&deadlock, lockRankDeadlock)
   838  	lockInit(&paniclk, lockRankPanic)
   839  	lockInit(&allglock, lockRankAllg)
   840  	lockInit(&allpLock, lockRankAllp)
   841  	lockInit(&reflectOffs.lock, lockRankReflectOffs)
   842  	lockInit(&finlock, lockRankFin)
   843  	lockInit(&cpuprof.lock, lockRankCpuprof)
   844  	lockInit(&computeMaxProcsLock, lockRankComputeMaxProcs)
   845  	allocmLock.init(lockRankAllocmR, lockRankAllocmRInternal, lockRankAllocmW)
   846  	execLock.init(lockRankExecR, lockRankExecRInternal, lockRankExecW)
   847  	traceLockInit()
   848  	// Enforce that this lock is always a leaf lock.
   849  	// All of this lock's critical sections should be
   850  	// extremely short.
   851  	lockInit(&memstats.heapStats.noPLock, lockRankLeafRank)
   852  
   853  	lockVerifyMSize()
   854  
   855  	// raceinit must be the first call to race detector.
   856  	// In particular, it must be done before mallocinit below calls racemapshadow.
   857  	gp := getg()
   858  	if raceenabled {
   859  		gp.racectx, raceprocctx0 = raceinit()
   860  	}
   861  
   862  	sched.maxmcount = 10000
   863  	crashFD.Store(^uintptr(0))
   864  
   865  	// The world starts stopped.
   866  	worldStopped()
   867  
   868  	godebug, parsedGodebug := getGodebugEarly()
   869  	if parsedGodebug {
   870  		parseRuntimeDebugVars(godebug)
   871  	}
   872  	ticks.init() // run as early as possible
   873  	moduledataverify()
   874  	stackinit()
   875  	randinit() // must run before mallocinit, alginit, mcommoninit
   876  	mallocinit()
   877  	cpuinit(godebug) // must run before alginit
   878  	alginit()        // maps, hash, rand must not be used before this call
   879  	mcommoninit(gp.m, -1)
   880  	modulesinit()   // provides activeModules
   881  	typelinksinit() // uses maps, activeModules
   882  	itabsinit()     // uses activeModules
   883  	stkobjinit()    // must run before GC starts
   884  
   885  	sigsave(&gp.m.sigmask)
   886  	initSigmask = gp.m.sigmask
   887  
   888  	goargs()
   889  	goenvs()
   890  	secure()
   891  	checkfds()
   892  	if !parsedGodebug {
   893  		// Some platforms, e.g., Windows, didn't make env vars available "early",
   894  		// so try again now.
   895  		parseRuntimeDebugVars(gogetenv("GODEBUG"))
   896  	}
   897  	finishDebugVarsSetup()
   898  	gcinit()
   899  
   900  	// Allocate stack space that can be used when crashing due to bad stack
   901  	// conditions, e.g. morestack on g0.
   902  	gcrash.stack = stackalloc(16384)
   903  	gcrash.stackguard0 = gcrash.stack.lo + 1000
   904  	gcrash.stackguard1 = gcrash.stack.lo + 1000
   905  
   906  	// if disableMemoryProfiling is set, update MemProfileRate to 0 to turn off memprofile.
   907  	// Note: parsedebugvars may update MemProfileRate, but when disableMemoryProfiling is
   908  	// set to true by the linker, it means that nothing is consuming the profile, it is
   909  	// safe to set MemProfileRate to 0.
   910  	if disableMemoryProfiling {
   911  		MemProfileRate = 0
   912  	}
   913  
   914  	// mcommoninit runs before parsedebugvars, so init profstacks again.
   915  	mProfStackInit(gp.m)
   916  	defaultGOMAXPROCSInit()
   917  
   918  	lock(&sched.lock)
   919  	sched.lastpoll.Store(nanotime())
   920  	var procs int32
   921  	if n, ok := strconv.Atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
   922  		procs = n
   923  		sched.customGOMAXPROCS = true
   924  	} else {
   925  		// Use numCPUStartup for initial GOMAXPROCS for two reasons:
   926  		//
   927  		// 1. We just computed it in osinit, recomputing is (minorly) wasteful.
   928  		//
   929  		// 2. More importantly, if debug.containermaxprocs == 0 &&
   930  		//    debug.updatemaxprocs == 0, we want to guarantee that
   931  		//    runtime.GOMAXPROCS(0) always equals runtime.NumCPU (which is
   932  		//    just numCPUStartup).
   933  		procs = defaultGOMAXPROCS(numCPUStartup)
   934  	}
   935  	if procresize(procs) != nil {
   936  		throw("unknown runnable goroutine during bootstrap")
   937  	}
   938  	unlock(&sched.lock)
   939  
   940  	// World is effectively started now, as P's can run.
   941  	worldStarted()
   942  
   943  	if buildVersion == "" {
   944  		// Condition should never trigger. This code just serves
   945  		// to ensure runtime·buildVersion is kept in the resulting binary.
   946  		buildVersion = "unknown"
   947  	}
   948  	if len(modinfo) == 1 {
   949  		// Condition should never trigger. This code just serves
   950  		// to ensure runtime·modinfo is kept in the resulting binary.
   951  		modinfo = ""
   952  	}
   953  }
   954  
   955  func dumpgstatus(gp *g) {
   956  	thisg := getg()
   957  	print("runtime:   gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   958  	print("runtime: getg:  g=", thisg, ", goid=", thisg.goid, ",  g->atomicstatus=", readgstatus(thisg), "\n")
   959  }
   960  
   961  // sched.lock must be held.
   962  func checkmcount() {
   963  	assertLockHeld(&sched.lock)
   964  
   965  	// Exclude extra M's, which are used for cgocallback from threads
   966  	// created in C.
   967  	//
   968  	// The purpose of the SetMaxThreads limit is to avoid accidental fork
   969  	// bomb from something like millions of goroutines blocking on system
   970  	// calls, causing the runtime to create millions of threads. By
   971  	// definition, this isn't a problem for threads created in C, so we
   972  	// exclude them from the limit. See https://go.dev/issue/60004.
   973  	count := mcount() - int32(extraMInUse.Load()) - int32(extraMLength.Load())
   974  	if count > sched.maxmcount {
   975  		print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
   976  		throw("thread exhaustion")
   977  	}
   978  }
   979  
   980  // mReserveID returns the next ID to use for a new m. This new m is immediately
   981  // considered 'running' by checkdead.
   982  //
   983  // sched.lock must be held.
   984  func mReserveID() int64 {
   985  	assertLockHeld(&sched.lock)
   986  
   987  	if sched.mnext+1 < sched.mnext {
   988  		throw("runtime: thread ID overflow")
   989  	}
   990  	id := sched.mnext
   991  	sched.mnext++
   992  	checkmcount()
   993  	return id
   994  }
   995  
   996  // Pre-allocated ID may be passed as 'id', or omitted by passing -1.
   997  func mcommoninit(mp *m, id int64) {
   998  	gp := getg()
   999  
  1000  	// g0 stack won't make sense for user (and is not necessary unwindable).
  1001  	if gp != gp.m.g0 {
  1002  		callers(1, mp.createstack[:])
  1003  	}
  1004  
  1005  	lock(&sched.lock)
  1006  
  1007  	if id >= 0 {
  1008  		mp.id = id
  1009  	} else {
  1010  		mp.id = mReserveID()
  1011  	}
  1012  
  1013  	mrandinit(mp)
  1014  
  1015  	mpreinit(mp)
  1016  	if mp.gsignal != nil {
  1017  		mp.gsignal.stackguard1 = mp.gsignal.stack.lo + stackGuard
  1018  	}
  1019  
  1020  	// Add to allm so garbage collector doesn't free g->m
  1021  	// when it is just in a register or thread-local storage.
  1022  	mp.alllink = allm
  1023  
  1024  	// NumCgoCall and others iterate over allm w/o schedlock,
  1025  	// so we need to publish it safely.
  1026  	atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
  1027  	unlock(&sched.lock)
  1028  
  1029  	// Allocate memory to hold a cgo traceback if the cgo call crashes.
  1030  	if iscgo || GOOS == "solaris" || GOOS == "illumos" || GOOS == "windows" {
  1031  		mp.cgoCallers = new(cgoCallers)
  1032  	}
  1033  	mProfStackInit(mp)
  1034  }
  1035  
  1036  // mProfStackInit is used to eagerly initialize stack trace buffers for
  1037  // profiling. Lazy allocation would have to deal with reentrancy issues in
  1038  // malloc and runtime locks for mLockProfile.
  1039  // TODO(mknyszek): Implement lazy allocation if this becomes a problem.
  1040  func mProfStackInit(mp *m) {
  1041  	if debug.profstackdepth == 0 {
  1042  		// debug.profstack is set to 0 by the user, or we're being called from
  1043  		// schedinit before parsedebugvars.
  1044  		return
  1045  	}
  1046  	mp.profStack = makeProfStackFP()
  1047  	mp.mLockProfile.stack = makeProfStackFP()
  1048  }
  1049  
  1050  // makeProfStackFP creates a buffer large enough to hold a maximum-sized stack
  1051  // trace as well as any additional frames needed for frame pointer unwinding
  1052  // with delayed inline expansion.
  1053  func makeProfStackFP() []uintptr {
  1054  	// The "1" term is to account for the first stack entry being
  1055  	// taken up by a "skip" sentinel value for profilers which
  1056  	// defer inline frame expansion until the profile is reported.
  1057  	// The "maxSkip" term is for frame pointer unwinding, where we
  1058  	// want to end up with debug.profstackdebth frames but will discard
  1059  	// some "physical" frames to account for skipping.
  1060  	return make([]uintptr, 1+maxSkip+debug.profstackdepth)
  1061  }
  1062  
  1063  // makeProfStack returns a buffer large enough to hold a maximum-sized stack
  1064  // trace.
  1065  func makeProfStack() []uintptr { return make([]uintptr, debug.profstackdepth) }
  1066  
  1067  //go:linkname pprof_makeProfStack
  1068  func pprof_makeProfStack() []uintptr { return makeProfStack() }
  1069  
  1070  func (mp *m) becomeSpinning() {
  1071  	mp.spinning = true
  1072  	sched.nmspinning.Add(1)
  1073  	sched.needspinning.Store(0)
  1074  }
  1075  
  1076  // Take a snapshot of allp, for use after dropping the P.
  1077  //
  1078  // Must be called with a P, but the returned slice may be used after dropping
  1079  // the P. The M holds a reference on the snapshot to keep the backing array
  1080  // alive.
  1081  //
  1082  //go:yeswritebarrierrec
  1083  func (mp *m) snapshotAllp() []*p {
  1084  	mp.allpSnapshot = allp
  1085  	return mp.allpSnapshot
  1086  }
  1087  
  1088  // Clear the saved allp snapshot. Should be called as soon as the snapshot is
  1089  // no longer required.
  1090  //
  1091  // Must be called after reacquiring a P, as it requires a write barrier.
  1092  //
  1093  //go:yeswritebarrierrec
  1094  func (mp *m) clearAllpSnapshot() {
  1095  	mp.allpSnapshot = nil
  1096  }
  1097  
  1098  func (mp *m) hasCgoOnStack() bool {
  1099  	return mp.ncgo > 0 || mp.isextra
  1100  }
  1101  
  1102  const (
  1103  	// osHasLowResTimer indicates that the platform's internal timer system has a low resolution,
  1104  	// typically on the order of 1 ms or more.
  1105  	osHasLowResTimer = GOOS == "windows" || GOOS == "openbsd" || GOOS == "netbsd"
  1106  
  1107  	// osHasLowResClockInt is osHasLowResClock but in integer form, so it can be used to create
  1108  	// constants conditionally.
  1109  	osHasLowResClockInt = goos.IsWindows
  1110  
  1111  	// osHasLowResClock indicates that timestamps produced by nanotime on the platform have a
  1112  	// low resolution, typically on the order of 1 ms or more.
  1113  	osHasLowResClock = osHasLowResClockInt > 0
  1114  )
  1115  
  1116  // Mark gp ready to run.
  1117  func ready(gp *g, traceskip int, next bool) {
  1118  	status := readgstatus(gp)
  1119  
  1120  	// Mark runnable.
  1121  	mp := acquirem() // disable preemption because it can be holding p in a local var
  1122  	if status&^_Gscan != _Gwaiting {
  1123  		dumpgstatus(gp)
  1124  		throw("bad g->status in ready")
  1125  	}
  1126  
  1127  	// status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
  1128  	trace := traceAcquire()
  1129  	casgstatus(gp, _Gwaiting, _Grunnable)
  1130  	if trace.ok() {
  1131  		trace.GoUnpark(gp, traceskip)
  1132  		traceRelease(trace)
  1133  	}
  1134  	runqput(mp.p.ptr(), gp, next)
  1135  	wakep()
  1136  	releasem(mp)
  1137  }
  1138  
  1139  // freezeStopWait is a large value that freezetheworld sets
  1140  // sched.stopwait to in order to request that all Gs permanently stop.
  1141  const freezeStopWait = 0x7fffffff
  1142  
  1143  // freezing is set to non-zero if the runtime is trying to freeze the
  1144  // world.
  1145  var freezing atomic.Bool
  1146  
  1147  // Similar to stopTheWorld but best-effort and can be called several times.
  1148  // There is no reverse operation, used during crashing.
  1149  // This function must not lock any mutexes.
  1150  func freezetheworld() {
  1151  	freezing.Store(true)
  1152  	if debug.dontfreezetheworld > 0 {
  1153  		// Don't prempt Ps to stop goroutines. That will perturb
  1154  		// scheduler state, making debugging more difficult. Instead,
  1155  		// allow goroutines to continue execution.
  1156  		//
  1157  		// fatalpanic will tracebackothers to trace all goroutines. It
  1158  		// is unsafe to trace a running goroutine, so tracebackothers
  1159  		// will skip running goroutines. That is OK and expected, we
  1160  		// expect users of dontfreezetheworld to use core files anyway.
  1161  		//
  1162  		// However, allowing the scheduler to continue running free
  1163  		// introduces a race: a goroutine may be stopped when
  1164  		// tracebackothers checks its status, and then start running
  1165  		// later when we are in the middle of traceback, potentially
  1166  		// causing a crash.
  1167  		//
  1168  		// To mitigate this, when an M naturally enters the scheduler,
  1169  		// schedule checks if freezing is set and if so stops
  1170  		// execution. This guarantees that while Gs can transition from
  1171  		// running to stopped, they can never transition from stopped
  1172  		// to running.
  1173  		//
  1174  		// The sleep here allows racing Ms that missed freezing and are
  1175  		// about to run a G to complete the transition to running
  1176  		// before we start traceback.
  1177  		usleep(1000)
  1178  		return
  1179  	}
  1180  
  1181  	// stopwait and preemption requests can be lost
  1182  	// due to races with concurrently executing threads,
  1183  	// so try several times
  1184  	for i := 0; i < 5; i++ {
  1185  		// this should tell the scheduler to not start any new goroutines
  1186  		sched.stopwait = freezeStopWait
  1187  		sched.gcwaiting.Store(true)
  1188  		// this should stop running goroutines
  1189  		if !preemptall() {
  1190  			break // no running goroutines
  1191  		}
  1192  		usleep(1000)
  1193  	}
  1194  	// to be sure
  1195  	usleep(1000)
  1196  	preemptall()
  1197  	usleep(1000)
  1198  }
  1199  
  1200  // All reads and writes of g's status go through readgstatus, casgstatus
  1201  // castogscanstatus, casfrom_Gscanstatus.
  1202  //
  1203  //go:nosplit
  1204  func readgstatus(gp *g) uint32 {
  1205  	return gp.atomicstatus.Load()
  1206  }
  1207  
  1208  // The Gscanstatuses are acting like locks and this releases them.
  1209  // If it proves to be a performance hit we should be able to make these
  1210  // simple atomic stores but for now we are going to throw if
  1211  // we see an inconsistent state.
  1212  func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
  1213  	success := false
  1214  
  1215  	// Check that transition is valid.
  1216  	switch oldval {
  1217  	default:
  1218  		print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
  1219  		dumpgstatus(gp)
  1220  		throw("casfrom_Gscanstatus:top gp->status is not in scan state")
  1221  	case _Gscanrunnable,
  1222  		_Gscanwaiting,
  1223  		_Gscanrunning,
  1224  		_Gscansyscall,
  1225  		_Gscanleaked,
  1226  		_Gscanpreempted:
  1227  		if newval == oldval&^_Gscan {
  1228  			success = gp.atomicstatus.CompareAndSwap(oldval, newval)
  1229  		}
  1230  	}
  1231  	if !success {
  1232  		print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
  1233  		dumpgstatus(gp)
  1234  		throw("casfrom_Gscanstatus: gp->status is not in scan state")
  1235  	}
  1236  	releaseLockRankAndM(lockRankGscan)
  1237  }
  1238  
  1239  // This will return false if the gp is not in the expected status and the cas fails.
  1240  // This acts like a lock acquire while the casfromgstatus acts like a lock release.
  1241  func castogscanstatus(gp *g, oldval, newval uint32) bool {
  1242  	switch oldval {
  1243  	case _Grunnable,
  1244  		_Grunning,
  1245  		_Gwaiting,
  1246  		_Gleaked,
  1247  		_Gsyscall:
  1248  		if newval == oldval|_Gscan {
  1249  			r := gp.atomicstatus.CompareAndSwap(oldval, newval)
  1250  			if r {
  1251  				acquireLockRankAndM(lockRankGscan)
  1252  			}
  1253  			return r
  1254  
  1255  		}
  1256  	}
  1257  	print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1258  	throw("castogscanstatus")
  1259  	panic("not reached")
  1260  }
  1261  
  1262  // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
  1263  // various latencies on every transition instead of sampling them.
  1264  var casgstatusAlwaysTrack = false
  1265  
  1266  // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
  1267  // and casfrom_Gscanstatus instead.
  1268  // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
  1269  // put it in the Gscan state is finished.
  1270  //
  1271  //go:nosplit
  1272  func casgstatus(gp *g, oldval, newval uint32) {
  1273  	if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
  1274  		systemstack(func() {
  1275  			// Call on the systemstack to prevent print and throw from counting
  1276  			// against the nosplit stack reservation.
  1277  			print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1278  			throw("casgstatus: bad incoming values")
  1279  		})
  1280  	}
  1281  
  1282  	lockWithRankMayAcquire(nil, lockRankGscan)
  1283  
  1284  	// See https://golang.org/cl/21503 for justification of the yield delay.
  1285  	const yieldDelay = 5 * 1000
  1286  	var nextYield int64
  1287  
  1288  	// loop if gp->atomicstatus is in a scan state giving
  1289  	// GC time to finish and change the state to oldval.
  1290  	for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
  1291  		if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
  1292  			systemstack(func() {
  1293  				// Call on the systemstack to prevent throw from counting
  1294  				// against the nosplit stack reservation.
  1295  				throw("casgstatus: waiting for Gwaiting but is Grunnable")
  1296  			})
  1297  		}
  1298  		if i == 0 {
  1299  			nextYield = nanotime() + yieldDelay
  1300  		}
  1301  		if nanotime() < nextYield {
  1302  			for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
  1303  				procyield(1)
  1304  			}
  1305  		} else {
  1306  			osyield()
  1307  			nextYield = nanotime() + yieldDelay/2
  1308  		}
  1309  	}
  1310  
  1311  	if gp.bubble != nil {
  1312  		systemstack(func() {
  1313  			gp.bubble.changegstatus(gp, oldval, newval)
  1314  		})
  1315  	}
  1316  
  1317  	if oldval == _Grunning {
  1318  		// Track every gTrackingPeriod time a goroutine transitions out of running.
  1319  		if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
  1320  			gp.tracking = true
  1321  		}
  1322  		gp.trackingSeq++
  1323  	}
  1324  	if !gp.tracking {
  1325  		return
  1326  	}
  1327  
  1328  	// Handle various kinds of tracking.
  1329  	//
  1330  	// Currently:
  1331  	// - Time spent in runnable.
  1332  	// - Time spent blocked on a sync.Mutex or sync.RWMutex.
  1333  	switch oldval {
  1334  	case _Grunnable:
  1335  		// We transitioned out of runnable, so measure how much
  1336  		// time we spent in this state and add it to
  1337  		// runnableTime.
  1338  		now := nanotime()
  1339  		gp.runnableTime += now - gp.trackingStamp
  1340  		gp.trackingStamp = 0
  1341  	case _Gwaiting:
  1342  		if !gp.waitreason.isMutexWait() {
  1343  			// Not blocking on a lock.
  1344  			break
  1345  		}
  1346  		// Blocking on a lock, measure it. Note that because we're
  1347  		// sampling, we have to multiply by our sampling period to get
  1348  		// a more representative estimate of the absolute value.
  1349  		// gTrackingPeriod also represents an accurate sampling period
  1350  		// because we can only enter this state from _Grunning.
  1351  		now := nanotime()
  1352  		sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
  1353  		gp.trackingStamp = 0
  1354  	}
  1355  	switch newval {
  1356  	case _Gwaiting:
  1357  		if !gp.waitreason.isMutexWait() {
  1358  			// Not blocking on a lock.
  1359  			break
  1360  		}
  1361  		// Blocking on a lock. Write down the timestamp.
  1362  		now := nanotime()
  1363  		gp.trackingStamp = now
  1364  	case _Grunnable:
  1365  		// We just transitioned into runnable, so record what
  1366  		// time that happened.
  1367  		now := nanotime()
  1368  		gp.trackingStamp = now
  1369  	case _Grunning:
  1370  		// We're transitioning into running, so turn off
  1371  		// tracking and record how much time we spent in
  1372  		// runnable.
  1373  		gp.tracking = false
  1374  		sched.timeToRun.record(gp.runnableTime)
  1375  		gp.runnableTime = 0
  1376  	}
  1377  }
  1378  
  1379  // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
  1380  //
  1381  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1382  func casGToWaiting(gp *g, old uint32, reason waitReason) {
  1383  	// Set the wait reason before calling casgstatus, because casgstatus will use it.
  1384  	gp.waitreason = reason
  1385  	casgstatus(gp, old, _Gwaiting)
  1386  }
  1387  
  1388  // casGToWaitingForSuspendG transitions gp from old to _Gwaiting, and sets the wait reason.
  1389  // The wait reason must be a valid isWaitingForSuspendG wait reason.
  1390  //
  1391  // While a goroutine is in this state, it's stack is effectively pinned.
  1392  // The garbage collector must not shrink or otherwise mutate the goroutine's stack.
  1393  //
  1394  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1395  func casGToWaitingForSuspendG(gp *g, old uint32, reason waitReason) {
  1396  	if !reason.isWaitingForSuspendG() {
  1397  		throw("casGToWaitingForSuspendG with non-isWaitingForSuspendG wait reason")
  1398  	}
  1399  	casGToWaiting(gp, old, reason)
  1400  }
  1401  
  1402  // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
  1403  //
  1404  // TODO(austin): This is the only status operation that both changes
  1405  // the status and locks the _Gscan bit. Rethink this.
  1406  func casGToPreemptScan(gp *g, old, new uint32) {
  1407  	if old != _Grunning || new != _Gscan|_Gpreempted {
  1408  		throw("bad g transition")
  1409  	}
  1410  	acquireLockRankAndM(lockRankGscan)
  1411  	for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
  1412  	}
  1413  	// We never notify gp.bubble that the goroutine state has moved
  1414  	// from _Grunning to _Gpreempted. We call bubble.changegstatus
  1415  	// after status changes happen, but doing so here would violate the
  1416  	// ordering between the gscan and synctest locks. The bubble doesn't
  1417  	// distinguish between _Grunning and _Gpreempted anyway, so not
  1418  	// notifying it is fine.
  1419  }
  1420  
  1421  // casGFromPreempted attempts to transition gp from _Gpreempted to
  1422  // _Gwaiting. If successful, the caller is responsible for
  1423  // re-scheduling gp.
  1424  func casGFromPreempted(gp *g, old, new uint32) bool {
  1425  	if old != _Gpreempted || new != _Gwaiting {
  1426  		throw("bad g transition")
  1427  	}
  1428  	gp.waitreason = waitReasonPreempted
  1429  	if !gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting) {
  1430  		return false
  1431  	}
  1432  	if bubble := gp.bubble; bubble != nil {
  1433  		bubble.changegstatus(gp, _Gpreempted, _Gwaiting)
  1434  	}
  1435  	return true
  1436  }
  1437  
  1438  // stwReason is an enumeration of reasons the world is stopping.
  1439  type stwReason uint8
  1440  
  1441  // Reasons to stop-the-world.
  1442  //
  1443  // Avoid reusing reasons and add new ones instead.
  1444  const (
  1445  	stwUnknown                     stwReason = iota // "unknown"
  1446  	stwGCMarkTerm                                   // "GC mark termination"
  1447  	stwGCSweepTerm                                  // "GC sweep termination"
  1448  	stwWriteHeapDump                                // "write heap dump"
  1449  	stwGoroutineProfile                             // "goroutine profile"
  1450  	stwGoroutineProfileCleanup                      // "goroutine profile cleanup"
  1451  	stwAllGoroutinesStack                           // "all goroutines stack trace"
  1452  	stwReadMemStats                                 // "read mem stats"
  1453  	stwAllThreadsSyscall                            // "AllThreadsSyscall"
  1454  	stwGOMAXPROCS                                   // "GOMAXPROCS"
  1455  	stwStartTrace                                   // "start trace"
  1456  	stwStopTrace                                    // "stop trace"
  1457  	stwForTestCountPagesInUse                       // "CountPagesInUse (test)"
  1458  	stwForTestReadMetricsSlow                       // "ReadMetricsSlow (test)"
  1459  	stwForTestReadMemStatsSlow                      // "ReadMemStatsSlow (test)"
  1460  	stwForTestPageCachePagesLeaked                  // "PageCachePagesLeaked (test)"
  1461  	stwForTestResetDebugLog                         // "ResetDebugLog (test)"
  1462  )
  1463  
  1464  func (r stwReason) String() string {
  1465  	return stwReasonStrings[r]
  1466  }
  1467  
  1468  func (r stwReason) isGC() bool {
  1469  	return r == stwGCMarkTerm || r == stwGCSweepTerm
  1470  }
  1471  
  1472  // If you add to this list, also add it to src/internal/trace/parser.go.
  1473  // If you change the values of any of the stw* constants, bump the trace
  1474  // version number and make a copy of this.
  1475  var stwReasonStrings = [...]string{
  1476  	stwUnknown:                     "unknown",
  1477  	stwGCMarkTerm:                  "GC mark termination",
  1478  	stwGCSweepTerm:                 "GC sweep termination",
  1479  	stwWriteHeapDump:               "write heap dump",
  1480  	stwGoroutineProfile:            "goroutine profile",
  1481  	stwGoroutineProfileCleanup:     "goroutine profile cleanup",
  1482  	stwAllGoroutinesStack:          "all goroutines stack trace",
  1483  	stwReadMemStats:                "read mem stats",
  1484  	stwAllThreadsSyscall:           "AllThreadsSyscall",
  1485  	stwGOMAXPROCS:                  "GOMAXPROCS",
  1486  	stwStartTrace:                  "start trace",
  1487  	stwStopTrace:                   "stop trace",
  1488  	stwForTestCountPagesInUse:      "CountPagesInUse (test)",
  1489  	stwForTestReadMetricsSlow:      "ReadMetricsSlow (test)",
  1490  	stwForTestReadMemStatsSlow:     "ReadMemStatsSlow (test)",
  1491  	stwForTestPageCachePagesLeaked: "PageCachePagesLeaked (test)",
  1492  	stwForTestResetDebugLog:        "ResetDebugLog (test)",
  1493  }
  1494  
  1495  // worldStop provides context from the stop-the-world required by the
  1496  // start-the-world.
  1497  type worldStop struct {
  1498  	reason           stwReason
  1499  	startedStopping  int64
  1500  	finishedStopping int64
  1501  	stoppingCPUTime  int64
  1502  }
  1503  
  1504  // Temporary variable for stopTheWorld, when it can't write to the stack.
  1505  //
  1506  // Protected by worldsema.
  1507  var stopTheWorldContext worldStop
  1508  
  1509  // stopTheWorld stops all P's from executing goroutines, interrupting
  1510  // all goroutines at GC safe points and records reason as the reason
  1511  // for the stop. On return, only the current goroutine's P is running.
  1512  // stopTheWorld must not be called from a system stack and the caller
  1513  // must not hold worldsema. The caller must call startTheWorld when
  1514  // other P's should resume execution.
  1515  //
  1516  // stopTheWorld is safe for multiple goroutines to call at the
  1517  // same time. Each will execute its own stop, and the stops will
  1518  // be serialized.
  1519  //
  1520  // This is also used by routines that do stack dumps. If the system is
  1521  // in panic or being exited, this may not reliably stop all
  1522  // goroutines.
  1523  //
  1524  // Returns the STW context. When starting the world, this context must be
  1525  // passed to startTheWorld.
  1526  func stopTheWorld(reason stwReason) worldStop {
  1527  	semacquire(&worldsema)
  1528  	gp := getg()
  1529  	gp.m.preemptoff = reason.String()
  1530  	systemstack(func() {
  1531  		stopTheWorldContext = stopTheWorldWithSema(reason) // avoid write to stack
  1532  	})
  1533  	return stopTheWorldContext
  1534  }
  1535  
  1536  // startTheWorld undoes the effects of stopTheWorld.
  1537  //
  1538  // w must be the worldStop returned by stopTheWorld.
  1539  func startTheWorld(w worldStop) {
  1540  	systemstack(func() { startTheWorldWithSema(0, w) })
  1541  
  1542  	// worldsema must be held over startTheWorldWithSema to ensure
  1543  	// gomaxprocs cannot change while worldsema is held.
  1544  	//
  1545  	// Release worldsema with direct handoff to the next waiter, but
  1546  	// acquirem so that semrelease1 doesn't try to yield our time.
  1547  	//
  1548  	// Otherwise if e.g. ReadMemStats is being called in a loop,
  1549  	// it might stomp on other attempts to stop the world, such as
  1550  	// for starting or ending GC. The operation this blocks is
  1551  	// so heavy-weight that we should just try to be as fair as
  1552  	// possible here.
  1553  	//
  1554  	// We don't want to just allow us to get preempted between now
  1555  	// and releasing the semaphore because then we keep everyone
  1556  	// (including, for example, GCs) waiting longer.
  1557  	mp := acquirem()
  1558  	mp.preemptoff = ""
  1559  	semrelease1(&worldsema, true, 0)
  1560  	releasem(mp)
  1561  }
  1562  
  1563  // stopTheWorldGC has the same effect as stopTheWorld, but blocks
  1564  // until the GC is not running. It also blocks a GC from starting
  1565  // until startTheWorldGC is called.
  1566  func stopTheWorldGC(reason stwReason) worldStop {
  1567  	semacquire(&gcsema)
  1568  	return stopTheWorld(reason)
  1569  }
  1570  
  1571  // startTheWorldGC undoes the effects of stopTheWorldGC.
  1572  //
  1573  // w must be the worldStop returned by stopTheWorld.
  1574  func startTheWorldGC(w worldStop) {
  1575  	startTheWorld(w)
  1576  	semrelease(&gcsema)
  1577  }
  1578  
  1579  // Holding worldsema grants an M the right to try to stop the world.
  1580  var worldsema uint32 = 1
  1581  
  1582  // Holding gcsema grants the M the right to block a GC, and blocks
  1583  // until the current GC is done. In particular, it prevents gomaxprocs
  1584  // from changing concurrently.
  1585  //
  1586  // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
  1587  // being changed/enabled during a GC, remove this.
  1588  var gcsema uint32 = 1
  1589  
  1590  // stopTheWorldWithSema is the core implementation of stopTheWorld.
  1591  // The caller is responsible for acquiring worldsema and disabling
  1592  // preemption first and then should stopTheWorldWithSema on the system
  1593  // stack:
  1594  //
  1595  //	semacquire(&worldsema, 0)
  1596  //	m.preemptoff = "reason"
  1597  //	var stw worldStop
  1598  //	systemstack(func() {
  1599  //		stw = stopTheWorldWithSema(reason)
  1600  //	})
  1601  //
  1602  // When finished, the caller must either call startTheWorld or undo
  1603  // these three operations separately:
  1604  //
  1605  //	m.preemptoff = ""
  1606  //	systemstack(func() {
  1607  //		now = startTheWorldWithSema(stw)
  1608  //	})
  1609  //	semrelease(&worldsema)
  1610  //
  1611  // It is allowed to acquire worldsema once and then execute multiple
  1612  // startTheWorldWithSema/stopTheWorldWithSema pairs.
  1613  // Other P's are able to execute between successive calls to
  1614  // startTheWorldWithSema and stopTheWorldWithSema.
  1615  // Holding worldsema causes any other goroutines invoking
  1616  // stopTheWorld to block.
  1617  //
  1618  // Returns the STW context. When starting the world, this context must be
  1619  // passed to startTheWorldWithSema.
  1620  //
  1621  //go:systemstack
  1622  func stopTheWorldWithSema(reason stwReason) worldStop {
  1623  	// Mark the goroutine which called stopTheWorld preemptible so its
  1624  	// stack may be scanned by the GC or observed by the execution tracer.
  1625  	//
  1626  	// This lets a mark worker scan us or the execution tracer take our
  1627  	// stack while we try to stop the world since otherwise we could get
  1628  	// in a mutual preemption deadlock.
  1629  	//
  1630  	// casGToWaitingForSuspendG marks the goroutine as ineligible for a
  1631  	// stack shrink, effectively pinning the stack in memory for the duration.
  1632  	//
  1633  	// N.B. The execution tracer is not aware of this status transition and
  1634  	// handles it specially based on the wait reason.
  1635  	casGToWaitingForSuspendG(getg().m.curg, _Grunning, waitReasonStoppingTheWorld)
  1636  
  1637  	trace := traceAcquire()
  1638  	if trace.ok() {
  1639  		trace.STWStart(reason)
  1640  		traceRelease(trace)
  1641  	}
  1642  	gp := getg()
  1643  
  1644  	// If we hold a lock, then we won't be able to stop another M
  1645  	// that is blocked trying to acquire the lock.
  1646  	if gp.m.locks > 0 {
  1647  		throw("stopTheWorld: holding locks")
  1648  	}
  1649  
  1650  	lock(&sched.lock)
  1651  	start := nanotime() // exclude time waiting for sched.lock from start and total time metrics.
  1652  	sched.stopwait = gomaxprocs
  1653  	sched.gcwaiting.Store(true)
  1654  	preemptall()
  1655  	// stop current P
  1656  	gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
  1657  	gp.m.p.ptr().gcStopTime = start
  1658  	sched.stopwait--
  1659  	// try to retake all P's in Psyscall status
  1660  	trace = traceAcquire()
  1661  	for _, pp := range allp {
  1662  		s := pp.status
  1663  		if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
  1664  			if trace.ok() {
  1665  				trace.ProcSteal(pp, false)
  1666  			}
  1667  			sched.nGsyscallNoP.Add(1)
  1668  			pp.syscalltick++
  1669  			pp.gcStopTime = nanotime()
  1670  			sched.stopwait--
  1671  		}
  1672  	}
  1673  	if trace.ok() {
  1674  		traceRelease(trace)
  1675  	}
  1676  
  1677  	// stop idle P's
  1678  	now := nanotime()
  1679  	for {
  1680  		pp, _ := pidleget(now)
  1681  		if pp == nil {
  1682  			break
  1683  		}
  1684  		pp.status = _Pgcstop
  1685  		pp.gcStopTime = nanotime()
  1686  		sched.stopwait--
  1687  	}
  1688  	wait := sched.stopwait > 0
  1689  	unlock(&sched.lock)
  1690  
  1691  	// wait for remaining P's to stop voluntarily
  1692  	if wait {
  1693  		for {
  1694  			// wait for 100us, then try to re-preempt in case of any races
  1695  			if notetsleep(&sched.stopnote, 100*1000) {
  1696  				noteclear(&sched.stopnote)
  1697  				break
  1698  			}
  1699  			preemptall()
  1700  		}
  1701  	}
  1702  
  1703  	finish := nanotime()
  1704  	startTime := finish - start
  1705  	if reason.isGC() {
  1706  		sched.stwStoppingTimeGC.record(startTime)
  1707  	} else {
  1708  		sched.stwStoppingTimeOther.record(startTime)
  1709  	}
  1710  
  1711  	// Double-check we actually stopped everything, and all the invariants hold.
  1712  	// Also accumulate all the time spent by each P in _Pgcstop up to the point
  1713  	// where everything was stopped. This will be accumulated into the total pause
  1714  	// CPU time by the caller.
  1715  	stoppingCPUTime := int64(0)
  1716  	bad := ""
  1717  	if sched.stopwait != 0 {
  1718  		bad = "stopTheWorld: not stopped (stopwait != 0)"
  1719  	} else {
  1720  		for _, pp := range allp {
  1721  			if pp.status != _Pgcstop {
  1722  				bad = "stopTheWorld: not stopped (status != _Pgcstop)"
  1723  			}
  1724  			if pp.gcStopTime == 0 && bad == "" {
  1725  				bad = "stopTheWorld: broken CPU time accounting"
  1726  			}
  1727  			stoppingCPUTime += finish - pp.gcStopTime
  1728  			pp.gcStopTime = 0
  1729  		}
  1730  	}
  1731  	if freezing.Load() {
  1732  		// Some other thread is panicking. This can cause the
  1733  		// sanity checks above to fail if the panic happens in
  1734  		// the signal handler on a stopped thread. Either way,
  1735  		// we should halt this thread.
  1736  		lock(&deadlock)
  1737  		lock(&deadlock)
  1738  	}
  1739  	if bad != "" {
  1740  		throw(bad)
  1741  	}
  1742  
  1743  	worldStopped()
  1744  
  1745  	// Switch back to _Grunning, now that the world is stopped.
  1746  	casgstatus(getg().m.curg, _Gwaiting, _Grunning)
  1747  
  1748  	return worldStop{
  1749  		reason:           reason,
  1750  		startedStopping:  start,
  1751  		finishedStopping: finish,
  1752  		stoppingCPUTime:  stoppingCPUTime,
  1753  	}
  1754  }
  1755  
  1756  // reason is the same STW reason passed to stopTheWorld. start is the start
  1757  // time returned by stopTheWorld.
  1758  //
  1759  // now is the current time; prefer to pass 0 to capture a fresh timestamp.
  1760  //
  1761  // stattTheWorldWithSema returns now.
  1762  func startTheWorldWithSema(now int64, w worldStop) int64 {
  1763  	assertWorldStopped()
  1764  
  1765  	mp := acquirem() // disable preemption because it can be holding p in a local var
  1766  	if netpollinited() {
  1767  		list, delta := netpoll(0) // non-blocking
  1768  		injectglist(&list)
  1769  		netpollAdjustWaiters(delta)
  1770  	}
  1771  	lock(&sched.lock)
  1772  
  1773  	procs := gomaxprocs
  1774  	if newprocs != 0 {
  1775  		procs = newprocs
  1776  		newprocs = 0
  1777  	}
  1778  	p1 := procresize(procs)
  1779  	sched.gcwaiting.Store(false)
  1780  	if sched.sysmonwait.Load() {
  1781  		sched.sysmonwait.Store(false)
  1782  		notewakeup(&sched.sysmonnote)
  1783  	}
  1784  	unlock(&sched.lock)
  1785  
  1786  	worldStarted()
  1787  
  1788  	for p1 != nil {
  1789  		p := p1
  1790  		p1 = p1.link.ptr()
  1791  		if p.m != 0 {
  1792  			mp := p.m.ptr()
  1793  			p.m = 0
  1794  			if mp.nextp != 0 {
  1795  				throw("startTheWorld: inconsistent mp->nextp")
  1796  			}
  1797  			mp.nextp.set(p)
  1798  			notewakeup(&mp.park)
  1799  		} else {
  1800  			// Start M to run P.  Do not start another M below.
  1801  			newm(nil, p, -1)
  1802  		}
  1803  	}
  1804  
  1805  	// Capture start-the-world time before doing clean-up tasks.
  1806  	if now == 0 {
  1807  		now = nanotime()
  1808  	}
  1809  	totalTime := now - w.startedStopping
  1810  	if w.reason.isGC() {
  1811  		sched.stwTotalTimeGC.record(totalTime)
  1812  	} else {
  1813  		sched.stwTotalTimeOther.record(totalTime)
  1814  	}
  1815  	trace := traceAcquire()
  1816  	if trace.ok() {
  1817  		trace.STWDone()
  1818  		traceRelease(trace)
  1819  	}
  1820  
  1821  	// Wakeup an additional proc in case we have excessive runnable goroutines
  1822  	// in local queues or in the global queue. If we don't, the proc will park itself.
  1823  	// If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
  1824  	wakep()
  1825  
  1826  	releasem(mp)
  1827  
  1828  	return now
  1829  }
  1830  
  1831  // usesLibcall indicates whether this runtime performs system calls
  1832  // via libcall.
  1833  func usesLibcall() bool {
  1834  	switch GOOS {
  1835  	case "aix", "darwin", "illumos", "ios", "openbsd", "solaris", "windows":
  1836  		return true
  1837  	}
  1838  	return false
  1839  }
  1840  
  1841  // mStackIsSystemAllocated indicates whether this runtime starts on a
  1842  // system-allocated stack.
  1843  func mStackIsSystemAllocated() bool {
  1844  	switch GOOS {
  1845  	case "aix", "darwin", "plan9", "illumos", "ios", "openbsd", "solaris", "windows":
  1846  		return true
  1847  	}
  1848  	return false
  1849  }
  1850  
  1851  // mstart is the entry-point for new Ms.
  1852  // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
  1853  func mstart()
  1854  
  1855  // mstart0 is the Go entry-point for new Ms.
  1856  // This must not split the stack because we may not even have stack
  1857  // bounds set up yet.
  1858  //
  1859  // May run during STW (because it doesn't have a P yet), so write
  1860  // barriers are not allowed.
  1861  //
  1862  //go:nosplit
  1863  //go:nowritebarrierrec
  1864  func mstart0() {
  1865  	gp := getg()
  1866  
  1867  	osStack := gp.stack.lo == 0
  1868  	if osStack {
  1869  		// Initialize stack bounds from system stack.
  1870  		// Cgo may have left stack size in stack.hi.
  1871  		// minit may update the stack bounds.
  1872  		//
  1873  		// Note: these bounds may not be very accurate.
  1874  		// We set hi to &size, but there are things above
  1875  		// it. The 1024 is supposed to compensate this,
  1876  		// but is somewhat arbitrary.
  1877  		size := gp.stack.hi
  1878  		if size == 0 {
  1879  			size = 16384 * sys.StackGuardMultiplier
  1880  		}
  1881  		gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
  1882  		gp.stack.lo = gp.stack.hi - size + 1024
  1883  	}
  1884  	// Initialize stack guard so that we can start calling regular
  1885  	// Go code.
  1886  	gp.stackguard0 = gp.stack.lo + stackGuard
  1887  	// This is the g0, so we can also call go:systemstack
  1888  	// functions, which check stackguard1.
  1889  	gp.stackguard1 = gp.stackguard0
  1890  	mstart1()
  1891  
  1892  	// Exit this thread.
  1893  	if mStackIsSystemAllocated() {
  1894  		// Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
  1895  		// the stack, but put it in gp.stack before mstart,
  1896  		// so the logic above hasn't set osStack yet.
  1897  		osStack = true
  1898  	}
  1899  	mexit(osStack)
  1900  }
  1901  
  1902  // The go:noinline is to guarantee the sys.GetCallerPC/sys.GetCallerSP below are safe,
  1903  // so that we can set up g0.sched to return to the call of mstart1 above.
  1904  //
  1905  //go:noinline
  1906  func mstart1() {
  1907  	gp := getg()
  1908  
  1909  	if gp != gp.m.g0 {
  1910  		throw("bad runtime·mstart")
  1911  	}
  1912  
  1913  	// Set up m.g0.sched as a label returning to just
  1914  	// after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
  1915  	// We're never coming back to mstart1 after we call schedule,
  1916  	// so other calls can reuse the current frame.
  1917  	// And goexit0 does a gogo that needs to return from mstart1
  1918  	// and let mstart0 exit the thread.
  1919  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1920  	gp.sched.pc = sys.GetCallerPC()
  1921  	gp.sched.sp = sys.GetCallerSP()
  1922  
  1923  	asminit()
  1924  	minit()
  1925  
  1926  	// Install signal handlers; after minit so that minit can
  1927  	// prepare the thread to be able to handle the signals.
  1928  	if gp.m == &m0 {
  1929  		mstartm0()
  1930  	}
  1931  
  1932  	if debug.dataindependenttiming == 1 {
  1933  		sys.EnableDIT()
  1934  	}
  1935  
  1936  	if fn := gp.m.mstartfn; fn != nil {
  1937  		fn()
  1938  	}
  1939  
  1940  	if gp.m != &m0 {
  1941  		acquirep(gp.m.nextp.ptr())
  1942  		gp.m.nextp = 0
  1943  	}
  1944  	schedule()
  1945  }
  1946  
  1947  // mstartm0 implements part of mstart1 that only runs on the m0.
  1948  //
  1949  // Write barriers are allowed here because we know the GC can't be
  1950  // running yet, so they'll be no-ops.
  1951  //
  1952  //go:yeswritebarrierrec
  1953  func mstartm0() {
  1954  	// Create an extra M for callbacks on threads not created by Go.
  1955  	// An extra M is also needed on Windows for callbacks created by
  1956  	// syscall.NewCallback. See issue #6751 for details.
  1957  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1958  		cgoHasExtraM = true
  1959  		newextram()
  1960  	}
  1961  	initsig(false)
  1962  }
  1963  
  1964  // mPark causes a thread to park itself, returning once woken.
  1965  //
  1966  //go:nosplit
  1967  func mPark() {
  1968  	gp := getg()
  1969  	notesleep(&gp.m.park)
  1970  	noteclear(&gp.m.park)
  1971  }
  1972  
  1973  // mexit tears down and exits the current thread.
  1974  //
  1975  // Don't call this directly to exit the thread, since it must run at
  1976  // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
  1977  // unwind the stack to the point that exits the thread.
  1978  //
  1979  // It is entered with m.p != nil, so write barriers are allowed. It
  1980  // will release the P before exiting.
  1981  //
  1982  //go:yeswritebarrierrec
  1983  func mexit(osStack bool) {
  1984  	mp := getg().m
  1985  
  1986  	if mp == &m0 {
  1987  		// This is the main thread. Just wedge it.
  1988  		//
  1989  		// On Linux, exiting the main thread puts the process
  1990  		// into a non-waitable zombie state. On Plan 9,
  1991  		// exiting the main thread unblocks wait even though
  1992  		// other threads are still running. On Solaris we can
  1993  		// neither exitThread nor return from mstart. Other
  1994  		// bad things probably happen on other platforms.
  1995  		//
  1996  		// We could try to clean up this M more before wedging
  1997  		// it, but that complicates signal handling.
  1998  		handoffp(releasep())
  1999  		lock(&sched.lock)
  2000  		sched.nmfreed++
  2001  		checkdead()
  2002  		unlock(&sched.lock)
  2003  		mPark()
  2004  		throw("locked m0 woke up")
  2005  	}
  2006  
  2007  	sigblock(true)
  2008  	unminit()
  2009  
  2010  	// Free the gsignal stack.
  2011  	if mp.gsignal != nil {
  2012  		stackfree(mp.gsignal.stack)
  2013  		if valgrindenabled {
  2014  			valgrindDeregisterStack(mp.gsignal.valgrindStackID)
  2015  			mp.gsignal.valgrindStackID = 0
  2016  		}
  2017  		// On some platforms, when calling into VDSO (e.g. nanotime)
  2018  		// we store our g on the gsignal stack, if there is one.
  2019  		// Now the stack is freed, unlink it from the m, so we
  2020  		// won't write to it when calling VDSO code.
  2021  		mp.gsignal = nil
  2022  	}
  2023  
  2024  	// Free vgetrandom state.
  2025  	vgetrandomDestroy(mp)
  2026  
  2027  	// Remove m from allm.
  2028  	lock(&sched.lock)
  2029  	for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
  2030  		if *pprev == mp {
  2031  			*pprev = mp.alllink
  2032  			goto found
  2033  		}
  2034  	}
  2035  	throw("m not found in allm")
  2036  found:
  2037  	// Events must not be traced after this point.
  2038  
  2039  	// Delay reaping m until it's done with the stack.
  2040  	//
  2041  	// Put mp on the free list, though it will not be reaped while freeWait
  2042  	// is freeMWait. mp is no longer reachable via allm, so even if it is
  2043  	// on an OS stack, we must keep a reference to mp alive so that the GC
  2044  	// doesn't free mp while we are still using it.
  2045  	//
  2046  	// Note that the free list must not be linked through alllink because
  2047  	// some functions walk allm without locking, so may be using alllink.
  2048  	//
  2049  	// N.B. It's important that the M appears on the free list simultaneously
  2050  	// with it being removed so that the tracer can find it.
  2051  	mp.freeWait.Store(freeMWait)
  2052  	mp.freelink = sched.freem
  2053  	sched.freem = mp
  2054  	unlock(&sched.lock)
  2055  
  2056  	atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
  2057  	sched.totalRuntimeLockWaitTime.Add(mp.mLockProfile.waitTime.Load())
  2058  
  2059  	// Release the P.
  2060  	handoffp(releasep())
  2061  	// After this point we must not have write barriers.
  2062  
  2063  	// Invoke the deadlock detector. This must happen after
  2064  	// handoffp because it may have started a new M to take our
  2065  	// P's work.
  2066  	lock(&sched.lock)
  2067  	sched.nmfreed++
  2068  	checkdead()
  2069  	unlock(&sched.lock)
  2070  
  2071  	if GOOS == "darwin" || GOOS == "ios" {
  2072  		// Make sure pendingPreemptSignals is correct when an M exits.
  2073  		// For #41702.
  2074  		if mp.signalPending.Load() != 0 {
  2075  			pendingPreemptSignals.Add(-1)
  2076  		}
  2077  	}
  2078  
  2079  	// Destroy all allocated resources. After this is called, we may no
  2080  	// longer take any locks.
  2081  	mdestroy(mp)
  2082  
  2083  	if osStack {
  2084  		// No more uses of mp, so it is safe to drop the reference.
  2085  		mp.freeWait.Store(freeMRef)
  2086  
  2087  		// Return from mstart and let the system thread
  2088  		// library free the g0 stack and terminate the thread.
  2089  		return
  2090  	}
  2091  
  2092  	// mstart is the thread's entry point, so there's nothing to
  2093  	// return to. Exit the thread directly. exitThread will clear
  2094  	// m.freeWait when it's done with the stack and the m can be
  2095  	// reaped.
  2096  	exitThread(&mp.freeWait)
  2097  }
  2098  
  2099  // forEachP calls fn(p) for every P p when p reaches a GC safe point.
  2100  // If a P is currently executing code, this will bring the P to a GC
  2101  // safe point and execute fn on that P. If the P is not executing code
  2102  // (it is idle or in a syscall), this will call fn(p) directly while
  2103  // preventing the P from exiting its state. This does not ensure that
  2104  // fn will run on every CPU executing Go code, but it acts as a global
  2105  // memory barrier. GC uses this as a "ragged barrier."
  2106  //
  2107  // The caller must hold worldsema. fn must not refer to any
  2108  // part of the current goroutine's stack, since the GC may move it.
  2109  func forEachP(reason waitReason, fn func(*p)) {
  2110  	systemstack(func() {
  2111  		gp := getg().m.curg
  2112  		// Mark the user stack as preemptible so that it may be scanned
  2113  		// by the GC or observed by the execution tracer. Otherwise, our
  2114  		// attempt to force all P's to a safepoint could result in a
  2115  		// deadlock as we attempt to preempt a goroutine that's trying
  2116  		// to preempt us (e.g. for a stack scan).
  2117  		//
  2118  		// casGToWaitingForSuspendG marks the goroutine as ineligible for a
  2119  		// stack shrink, effectively pinning the stack in memory for the duration.
  2120  		//
  2121  		// N.B. The execution tracer is not aware of this status transition and
  2122  		// handles it specially based on the wait reason.
  2123  		casGToWaitingForSuspendG(gp, _Grunning, reason)
  2124  		forEachPInternal(fn)
  2125  		casgstatus(gp, _Gwaiting, _Grunning)
  2126  	})
  2127  }
  2128  
  2129  // forEachPInternal calls fn(p) for every P p when p reaches a GC safe point.
  2130  // It is the internal implementation of forEachP.
  2131  //
  2132  // The caller must hold worldsema and either must ensure that a GC is not
  2133  // running (otherwise this may deadlock with the GC trying to preempt this P)
  2134  // or it must leave its goroutine in a preemptible state before it switches
  2135  // to the systemstack. Due to these restrictions, prefer forEachP when possible.
  2136  //
  2137  //go:systemstack
  2138  func forEachPInternal(fn func(*p)) {
  2139  	mp := acquirem()
  2140  	pp := getg().m.p.ptr()
  2141  
  2142  	lock(&sched.lock)
  2143  	if sched.safePointWait != 0 {
  2144  		throw("forEachP: sched.safePointWait != 0")
  2145  	}
  2146  	sched.safePointWait = gomaxprocs - 1
  2147  	sched.safePointFn = fn
  2148  
  2149  	// Ask all Ps to run the safe point function.
  2150  	for _, p2 := range allp {
  2151  		if p2 != pp {
  2152  			atomic.Store(&p2.runSafePointFn, 1)
  2153  		}
  2154  	}
  2155  	preemptall()
  2156  
  2157  	// Any P entering _Pidle or _Psyscall from now on will observe
  2158  	// p.runSafePointFn == 1 and will call runSafePointFn when
  2159  	// changing its status to _Pidle/_Psyscall.
  2160  
  2161  	// Run safe point function for all idle Ps. sched.pidle will
  2162  	// not change because we hold sched.lock.
  2163  	for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
  2164  		if atomic.Cas(&p.runSafePointFn, 1, 0) {
  2165  			fn(p)
  2166  			sched.safePointWait--
  2167  		}
  2168  	}
  2169  
  2170  	wait := sched.safePointWait > 0
  2171  	unlock(&sched.lock)
  2172  
  2173  	// Run fn for the current P.
  2174  	fn(pp)
  2175  
  2176  	// Force Ps currently in _Psyscall into _Pidle and hand them
  2177  	// off to induce safe point function execution.
  2178  	for _, p2 := range allp {
  2179  		s := p2.status
  2180  
  2181  		// We need to be fine-grained about tracing here, since handoffp
  2182  		// might call into the tracer, and the tracer is non-reentrant.
  2183  		trace := traceAcquire()
  2184  		if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
  2185  			if trace.ok() {
  2186  				// It's important that we traceRelease before we call handoffp, which may also traceAcquire.
  2187  				trace.ProcSteal(p2, false)
  2188  				traceRelease(trace)
  2189  			}
  2190  			sched.nGsyscallNoP.Add(1)
  2191  			p2.syscalltick++
  2192  			handoffp(p2)
  2193  		} else if trace.ok() {
  2194  			traceRelease(trace)
  2195  		}
  2196  	}
  2197  
  2198  	// Wait for remaining Ps to run fn.
  2199  	if wait {
  2200  		for {
  2201  			// Wait for 100us, then try to re-preempt in
  2202  			// case of any races.
  2203  			//
  2204  			// Requires system stack.
  2205  			if notetsleep(&sched.safePointNote, 100*1000) {
  2206  				noteclear(&sched.safePointNote)
  2207  				break
  2208  			}
  2209  			preemptall()
  2210  		}
  2211  	}
  2212  	if sched.safePointWait != 0 {
  2213  		throw("forEachP: not done")
  2214  	}
  2215  	for _, p2 := range allp {
  2216  		if p2.runSafePointFn != 0 {
  2217  			throw("forEachP: P did not run fn")
  2218  		}
  2219  	}
  2220  
  2221  	lock(&sched.lock)
  2222  	sched.safePointFn = nil
  2223  	unlock(&sched.lock)
  2224  	releasem(mp)
  2225  }
  2226  
  2227  // runSafePointFn runs the safe point function, if any, for this P.
  2228  // This should be called like
  2229  //
  2230  //	if getg().m.p.runSafePointFn != 0 {
  2231  //	    runSafePointFn()
  2232  //	}
  2233  //
  2234  // runSafePointFn must be checked on any transition in to _Pidle or
  2235  // _Psyscall to avoid a race where forEachP sees that the P is running
  2236  // just before the P goes into _Pidle/_Psyscall and neither forEachP
  2237  // nor the P run the safe-point function.
  2238  func runSafePointFn() {
  2239  	p := getg().m.p.ptr()
  2240  	// Resolve the race between forEachP running the safe-point
  2241  	// function on this P's behalf and this P running the
  2242  	// safe-point function directly.
  2243  	if !atomic.Cas(&p.runSafePointFn, 1, 0) {
  2244  		return
  2245  	}
  2246  	sched.safePointFn(p)
  2247  	lock(&sched.lock)
  2248  	sched.safePointWait--
  2249  	if sched.safePointWait == 0 {
  2250  		notewakeup(&sched.safePointNote)
  2251  	}
  2252  	unlock(&sched.lock)
  2253  }
  2254  
  2255  // When running with cgo, we call _cgo_thread_start
  2256  // to start threads for us so that we can play nicely with
  2257  // foreign code.
  2258  var cgoThreadStart unsafe.Pointer
  2259  
  2260  type cgothreadstart struct {
  2261  	g   guintptr
  2262  	tls *uint64
  2263  	fn  unsafe.Pointer
  2264  }
  2265  
  2266  // Allocate a new m unassociated with any thread.
  2267  // Can use p for allocation context if needed.
  2268  // fn is recorded as the new m's m.mstartfn.
  2269  // id is optional pre-allocated m ID. Omit by passing -1.
  2270  //
  2271  // This function is allowed to have write barriers even if the caller
  2272  // isn't because it borrows pp.
  2273  //
  2274  //go:yeswritebarrierrec
  2275  func allocm(pp *p, fn func(), id int64) *m {
  2276  	allocmLock.rlock()
  2277  
  2278  	// The caller owns pp, but we may borrow (i.e., acquirep) it. We must
  2279  	// disable preemption to ensure it is not stolen, which would make the
  2280  	// caller lose ownership.
  2281  	acquirem()
  2282  
  2283  	gp := getg()
  2284  	if gp.m.p == 0 {
  2285  		acquirep(pp) // temporarily borrow p for mallocs in this function
  2286  	}
  2287  
  2288  	// Release the free M list. We need to do this somewhere and
  2289  	// this may free up a stack we can use.
  2290  	if sched.freem != nil {
  2291  		lock(&sched.lock)
  2292  		var newList *m
  2293  		for freem := sched.freem; freem != nil; {
  2294  			// Wait for freeWait to indicate that freem's stack is unused.
  2295  			wait := freem.freeWait.Load()
  2296  			if wait == freeMWait {
  2297  				next := freem.freelink
  2298  				freem.freelink = newList
  2299  				newList = freem
  2300  				freem = next
  2301  				continue
  2302  			}
  2303  			// Drop any remaining trace resources.
  2304  			// Ms can continue to emit events all the way until wait != freeMWait,
  2305  			// so it's only safe to call traceThreadDestroy at this point.
  2306  			if traceEnabled() || traceShuttingDown() {
  2307  				traceThreadDestroy(freem)
  2308  			}
  2309  			// Free the stack if needed. For freeMRef, there is
  2310  			// nothing to do except drop freem from the sched.freem
  2311  			// list.
  2312  			if wait == freeMStack {
  2313  				// stackfree must be on the system stack, but allocm is
  2314  				// reachable off the system stack transitively from
  2315  				// startm.
  2316  				systemstack(func() {
  2317  					stackfree(freem.g0.stack)
  2318  					if valgrindenabled {
  2319  						valgrindDeregisterStack(freem.g0.valgrindStackID)
  2320  						freem.g0.valgrindStackID = 0
  2321  					}
  2322  				})
  2323  			}
  2324  			freem = freem.freelink
  2325  		}
  2326  		sched.freem = newList
  2327  		unlock(&sched.lock)
  2328  	}
  2329  
  2330  	mp := &new(mPadded).m
  2331  	mp.mstartfn = fn
  2332  	mcommoninit(mp, id)
  2333  
  2334  	// In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
  2335  	// Windows and Plan 9 will layout sched stack on OS stack.
  2336  	if iscgo || mStackIsSystemAllocated() {
  2337  		mp.g0 = malg(-1)
  2338  	} else {
  2339  		mp.g0 = malg(16384 * sys.StackGuardMultiplier)
  2340  	}
  2341  	mp.g0.m = mp
  2342  
  2343  	if pp == gp.m.p.ptr() {
  2344  		releasep()
  2345  	}
  2346  
  2347  	releasem(gp.m)
  2348  	allocmLock.runlock()
  2349  	return mp
  2350  }
  2351  
  2352  // needm is called when a cgo callback happens on a
  2353  // thread without an m (a thread not created by Go).
  2354  // In this case, needm is expected to find an m to use
  2355  // and return with m, g initialized correctly.
  2356  // Since m and g are not set now (likely nil, but see below)
  2357  // needm is limited in what routines it can call. In particular
  2358  // it can only call nosplit functions (textflag 7) and cannot
  2359  // do any scheduling that requires an m.
  2360  //
  2361  // In order to avoid needing heavy lifting here, we adopt
  2362  // the following strategy: there is a stack of available m's
  2363  // that can be stolen. Using compare-and-swap
  2364  // to pop from the stack has ABA races, so we simulate
  2365  // a lock by doing an exchange (via Casuintptr) to steal the stack
  2366  // head and replace the top pointer with MLOCKED (1).
  2367  // This serves as a simple spin lock that we can use even
  2368  // without an m. The thread that locks the stack in this way
  2369  // unlocks the stack by storing a valid stack head pointer.
  2370  //
  2371  // In order to make sure that there is always an m structure
  2372  // available to be stolen, we maintain the invariant that there
  2373  // is always one more than needed. At the beginning of the
  2374  // program (if cgo is in use) the list is seeded with a single m.
  2375  // If needm finds that it has taken the last m off the list, its job
  2376  // is - once it has installed its own m so that it can do things like
  2377  // allocate memory - to create a spare m and put it on the list.
  2378  //
  2379  // Each of these extra m's also has a g0 and a curg that are
  2380  // pressed into service as the scheduling stack and current
  2381  // goroutine for the duration of the cgo callback.
  2382  //
  2383  // It calls dropm to put the m back on the list,
  2384  // 1. when the callback is done with the m in non-pthread platforms,
  2385  // 2. or when the C thread exiting on pthread platforms.
  2386  //
  2387  // The signal argument indicates whether we're called from a signal
  2388  // handler.
  2389  //
  2390  //go:nosplit
  2391  func needm(signal bool) {
  2392  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  2393  		// Can happen if C/C++ code calls Go from a global ctor.
  2394  		// Can also happen on Windows if a global ctor uses a
  2395  		// callback created by syscall.NewCallback. See issue #6751
  2396  		// for details.
  2397  		//
  2398  		// Can not throw, because scheduler is not initialized yet.
  2399  		writeErrStr("fatal error: cgo callback before cgo call\n")
  2400  		exit(1)
  2401  	}
  2402  
  2403  	// Save and block signals before getting an M.
  2404  	// The signal handler may call needm itself,
  2405  	// and we must avoid a deadlock. Also, once g is installed,
  2406  	// any incoming signals will try to execute,
  2407  	// but we won't have the sigaltstack settings and other data
  2408  	// set up appropriately until the end of minit, which will
  2409  	// unblock the signals. This is the same dance as when
  2410  	// starting a new m to run Go code via newosproc.
  2411  	var sigmask sigset
  2412  	sigsave(&sigmask)
  2413  	sigblock(false)
  2414  
  2415  	// getExtraM is safe here because of the invariant above,
  2416  	// that the extra list always contains or will soon contain
  2417  	// at least one m.
  2418  	mp, last := getExtraM()
  2419  
  2420  	// Set needextram when we've just emptied the list,
  2421  	// so that the eventual call into cgocallbackg will
  2422  	// allocate a new m for the extra list. We delay the
  2423  	// allocation until then so that it can be done
  2424  	// after exitsyscall makes sure it is okay to be
  2425  	// running at all (that is, there's no garbage collection
  2426  	// running right now).
  2427  	mp.needextram = last
  2428  
  2429  	// Store the original signal mask for use by minit.
  2430  	mp.sigmask = sigmask
  2431  
  2432  	// Install TLS on some platforms (previously setg
  2433  	// would do this if necessary).
  2434  	osSetupTLS(mp)
  2435  
  2436  	// Install g (= m->g0) and set the stack bounds
  2437  	// to match the current stack.
  2438  	setg(mp.g0)
  2439  	sp := sys.GetCallerSP()
  2440  	callbackUpdateSystemStack(mp, sp, signal)
  2441  
  2442  	// Should mark we are already in Go now.
  2443  	// Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
  2444  	// which means the extram list may be empty, that will cause a deadlock.
  2445  	mp.isExtraInC = false
  2446  
  2447  	// Initialize this thread to use the m.
  2448  	asminit()
  2449  	minit()
  2450  
  2451  	// Emit a trace event for this dead -> syscall transition,
  2452  	// but only if we're not in a signal handler.
  2453  	//
  2454  	// N.B. the tracer can run on a bare M just fine, we just have
  2455  	// to make sure to do this before setg(nil) and unminit.
  2456  	var trace traceLocker
  2457  	if !signal {
  2458  		trace = traceAcquire()
  2459  	}
  2460  
  2461  	// mp.curg is now a real goroutine.
  2462  	casgstatus(mp.curg, _Gdead, _Gsyscall)
  2463  	sched.ngsys.Add(-1)
  2464  	sched.nGsyscallNoP.Add(1)
  2465  
  2466  	if !signal {
  2467  		if trace.ok() {
  2468  			trace.GoCreateSyscall(mp.curg)
  2469  			traceRelease(trace)
  2470  		}
  2471  	}
  2472  	mp.isExtraInSig = signal
  2473  }
  2474  
  2475  // Acquire an extra m and bind it to the C thread when a pthread key has been created.
  2476  //
  2477  //go:nosplit
  2478  func needAndBindM() {
  2479  	needm(false)
  2480  
  2481  	if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
  2482  		cgoBindM()
  2483  	}
  2484  }
  2485  
  2486  // newextram allocates m's and puts them on the extra list.
  2487  // It is called with a working local m, so that it can do things
  2488  // like call schedlock and allocate.
  2489  func newextram() {
  2490  	c := extraMWaiters.Swap(0)
  2491  	if c > 0 {
  2492  		for i := uint32(0); i < c; i++ {
  2493  			oneNewExtraM()
  2494  		}
  2495  	} else if extraMLength.Load() == 0 {
  2496  		// Make sure there is at least one extra M.
  2497  		oneNewExtraM()
  2498  	}
  2499  }
  2500  
  2501  // oneNewExtraM allocates an m and puts it on the extra list.
  2502  func oneNewExtraM() {
  2503  	// Create extra goroutine locked to extra m.
  2504  	// The goroutine is the context in which the cgo callback will run.
  2505  	// The sched.pc will never be returned to, but setting it to
  2506  	// goexit makes clear to the traceback routines where
  2507  	// the goroutine stack ends.
  2508  	mp := allocm(nil, nil, -1)
  2509  	gp := malg(4096)
  2510  	gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
  2511  	gp.sched.sp = gp.stack.hi
  2512  	gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
  2513  	gp.sched.lr = 0
  2514  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  2515  	gp.syscallpc = gp.sched.pc
  2516  	gp.syscallsp = gp.sched.sp
  2517  	gp.stktopsp = gp.sched.sp
  2518  	// malg returns status as _Gidle. Change to _Gdead before
  2519  	// adding to allg where GC can see it. We use _Gdead to hide
  2520  	// this from tracebacks and stack scans since it isn't a
  2521  	// "real" goroutine until needm grabs it.
  2522  	casgstatus(gp, _Gidle, _Gdead)
  2523  	gp.m = mp
  2524  	mp.curg = gp
  2525  	mp.isextra = true
  2526  	// mark we are in C by default.
  2527  	mp.isExtraInC = true
  2528  	mp.lockedInt++
  2529  	mp.lockedg.set(gp)
  2530  	gp.lockedm.set(mp)
  2531  	gp.goid = sched.goidgen.Add(1)
  2532  	if raceenabled {
  2533  		gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
  2534  	}
  2535  	// put on allg for garbage collector
  2536  	allgadd(gp)
  2537  
  2538  	// gp is now on the allg list, but we don't want it to be
  2539  	// counted by gcount. It would be more "proper" to increment
  2540  	// sched.ngfree, but that requires locking. Incrementing ngsys
  2541  	// has the same effect.
  2542  	sched.ngsys.Add(1)
  2543  
  2544  	// Add m to the extra list.
  2545  	addExtraM(mp)
  2546  }
  2547  
  2548  // dropm puts the current m back onto the extra list.
  2549  //
  2550  // 1. On systems without pthreads, like Windows
  2551  // dropm is called when a cgo callback has called needm but is now
  2552  // done with the callback and returning back into the non-Go thread.
  2553  //
  2554  // The main expense here is the call to signalstack to release the
  2555  // m's signal stack, and then the call to needm on the next callback
  2556  // from this thread. It is tempting to try to save the m for next time,
  2557  // which would eliminate both these costs, but there might not be
  2558  // a next time: the current thread (which Go does not control) might exit.
  2559  // If we saved the m for that thread, there would be an m leak each time
  2560  // such a thread exited. Instead, we acquire and release an m on each
  2561  // call. These should typically not be scheduling operations, just a few
  2562  // atomics, so the cost should be small.
  2563  //
  2564  // 2. On systems with pthreads
  2565  // dropm is called while a non-Go thread is exiting.
  2566  // We allocate a pthread per-thread variable using pthread_key_create,
  2567  // to register a thread-exit-time destructor.
  2568  // And store the g into a thread-specific value associated with the pthread key,
  2569  // when first return back to C.
  2570  // So that the destructor would invoke dropm while the non-Go thread is exiting.
  2571  // This is much faster since it avoids expensive signal-related syscalls.
  2572  //
  2573  // This always runs without a P, so //go:nowritebarrierrec is required.
  2574  //
  2575  // This may run with a different stack than was recorded in g0 (there is no
  2576  // call to callbackUpdateSystemStack prior to dropm), so this must be
  2577  // //go:nosplit to avoid the stack bounds check.
  2578  //
  2579  //go:nowritebarrierrec
  2580  //go:nosplit
  2581  func dropm() {
  2582  	// Clear m and g, and return m to the extra list.
  2583  	// After the call to setg we can only call nosplit functions
  2584  	// with no pointer manipulation.
  2585  	mp := getg().m
  2586  
  2587  	// Emit a trace event for this syscall -> dead transition.
  2588  	//
  2589  	// N.B. the tracer can run on a bare M just fine, we just have
  2590  	// to make sure to do this before setg(nil) and unminit.
  2591  	var trace traceLocker
  2592  	if !mp.isExtraInSig {
  2593  		trace = traceAcquire()
  2594  	}
  2595  
  2596  	// Return mp.curg to dead state.
  2597  	casgstatus(mp.curg, _Gsyscall, _Gdead)
  2598  	mp.curg.preemptStop = false
  2599  	sched.ngsys.Add(1)
  2600  	sched.nGsyscallNoP.Add(-1)
  2601  
  2602  	if !mp.isExtraInSig {
  2603  		if trace.ok() {
  2604  			trace.GoDestroySyscall()
  2605  			traceRelease(trace)
  2606  		}
  2607  	}
  2608  
  2609  	// Trash syscalltick so that it doesn't line up with mp.old.syscalltick anymore.
  2610  	//
  2611  	// In the new tracer, we model needm and dropm and a goroutine being created and
  2612  	// destroyed respectively. The m then might get reused with a different procid but
  2613  	// still with a reference to oldp, and still with the same syscalltick. The next
  2614  	// time a G is "created" in needm, it'll return and quietly reacquire its P from a
  2615  	// different m with a different procid, which will confuse the trace parser. By
  2616  	// trashing syscalltick, we ensure that it'll appear as if we lost the P to the
  2617  	// tracer parser and that we just reacquired it.
  2618  	//
  2619  	// Trash the value by decrementing because that gets us as far away from the value
  2620  	// the syscall exit code expects as possible. Setting to zero is risky because
  2621  	// syscalltick could already be zero (and in fact, is initialized to zero).
  2622  	mp.syscalltick--
  2623  
  2624  	// Reset trace state unconditionally. This goroutine is being 'destroyed'
  2625  	// from the perspective of the tracer.
  2626  	mp.curg.trace.reset()
  2627  
  2628  	// Flush all the M's buffers. This is necessary because the M might
  2629  	// be used on a different thread with a different procid, so we have
  2630  	// to make sure we don't write into the same buffer.
  2631  	if traceEnabled() || traceShuttingDown() {
  2632  		// Acquire sched.lock across thread destruction. One of the invariants of the tracer
  2633  		// is that a thread cannot disappear from the tracer's view (allm or freem) without
  2634  		// it noticing, so it requires that sched.lock be held over traceThreadDestroy.
  2635  		//
  2636  		// This isn't strictly necessary in this case, because this thread never leaves allm,
  2637  		// but the critical section is short and dropm is rare on pthread platforms, so just
  2638  		// take the lock and play it safe. traceThreadDestroy also asserts that the lock is held.
  2639  		lock(&sched.lock)
  2640  		traceThreadDestroy(mp)
  2641  		unlock(&sched.lock)
  2642  	}
  2643  	mp.isExtraInSig = false
  2644  
  2645  	// Block signals before unminit.
  2646  	// Unminit unregisters the signal handling stack (but needs g on some systems).
  2647  	// Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
  2648  	// It's important not to try to handle a signal between those two steps.
  2649  	sigmask := mp.sigmask
  2650  	sigblock(false)
  2651  	unminit()
  2652  
  2653  	setg(nil)
  2654  
  2655  	// Clear g0 stack bounds to ensure that needm always refreshes the
  2656  	// bounds when reusing this M.
  2657  	g0 := mp.g0
  2658  	g0.stack.hi = 0
  2659  	g0.stack.lo = 0
  2660  	g0.stackguard0 = 0
  2661  	g0.stackguard1 = 0
  2662  	mp.g0StackAccurate = false
  2663  
  2664  	putExtraM(mp)
  2665  
  2666  	msigrestore(sigmask)
  2667  }
  2668  
  2669  // bindm store the g0 of the current m into a thread-specific value.
  2670  //
  2671  // We allocate a pthread per-thread variable using pthread_key_create,
  2672  // to register a thread-exit-time destructor.
  2673  // We are here setting the thread-specific value of the pthread key, to enable the destructor.
  2674  // So that the pthread_key_destructor would dropm while the C thread is exiting.
  2675  //
  2676  // And the saved g will be used in pthread_key_destructor,
  2677  // since the g stored in the TLS by Go might be cleared in some platforms,
  2678  // before the destructor invoked, so, we restore g by the stored g, before dropm.
  2679  //
  2680  // We store g0 instead of m, to make the assembly code simpler,
  2681  // since we need to restore g0 in runtime.cgocallback.
  2682  //
  2683  // On systems without pthreads, like Windows, bindm shouldn't be used.
  2684  //
  2685  // NOTE: this always runs without a P, so, nowritebarrierrec required.
  2686  //
  2687  //go:nosplit
  2688  //go:nowritebarrierrec
  2689  func cgoBindM() {
  2690  	if GOOS == "windows" || GOOS == "plan9" {
  2691  		fatal("bindm in unexpected GOOS")
  2692  	}
  2693  	g := getg()
  2694  	if g.m.g0 != g {
  2695  		fatal("the current g is not g0")
  2696  	}
  2697  	if _cgo_bindm != nil {
  2698  		asmcgocall(_cgo_bindm, unsafe.Pointer(g))
  2699  	}
  2700  }
  2701  
  2702  // A helper function for EnsureDropM.
  2703  //
  2704  // getm should be an internal detail,
  2705  // but widely used packages access it using linkname.
  2706  // Notable members of the hall of shame include:
  2707  //   - fortio.org/log
  2708  //
  2709  // Do not remove or change the type signature.
  2710  // See go.dev/issue/67401.
  2711  //
  2712  //go:linkname getm
  2713  func getm() uintptr {
  2714  	return uintptr(unsafe.Pointer(getg().m))
  2715  }
  2716  
  2717  var (
  2718  	// Locking linked list of extra M's, via mp.schedlink. Must be accessed
  2719  	// only via lockextra/unlockextra.
  2720  	//
  2721  	// Can't be atomic.Pointer[m] because we use an invalid pointer as a
  2722  	// "locked" sentinel value. M's on this list remain visible to the GC
  2723  	// because their mp.curg is on allgs.
  2724  	extraM atomic.Uintptr
  2725  	// Number of M's in the extraM list.
  2726  	extraMLength atomic.Uint32
  2727  	// Number of waiters in lockextra.
  2728  	extraMWaiters atomic.Uint32
  2729  
  2730  	// Number of extra M's in use by threads.
  2731  	extraMInUse atomic.Uint32
  2732  )
  2733  
  2734  // lockextra locks the extra list and returns the list head.
  2735  // The caller must unlock the list by storing a new list head
  2736  // to extram. If nilokay is true, then lockextra will
  2737  // return a nil list head if that's what it finds. If nilokay is false,
  2738  // lockextra will keep waiting until the list head is no longer nil.
  2739  //
  2740  //go:nosplit
  2741  func lockextra(nilokay bool) *m {
  2742  	const locked = 1
  2743  
  2744  	incr := false
  2745  	for {
  2746  		old := extraM.Load()
  2747  		if old == locked {
  2748  			osyield_no_g()
  2749  			continue
  2750  		}
  2751  		if old == 0 && !nilokay {
  2752  			if !incr {
  2753  				// Add 1 to the number of threads
  2754  				// waiting for an M.
  2755  				// This is cleared by newextram.
  2756  				extraMWaiters.Add(1)
  2757  				incr = true
  2758  			}
  2759  			usleep_no_g(1)
  2760  			continue
  2761  		}
  2762  		if extraM.CompareAndSwap(old, locked) {
  2763  			return (*m)(unsafe.Pointer(old))
  2764  		}
  2765  		osyield_no_g()
  2766  		continue
  2767  	}
  2768  }
  2769  
  2770  //go:nosplit
  2771  func unlockextra(mp *m, delta int32) {
  2772  	extraMLength.Add(delta)
  2773  	extraM.Store(uintptr(unsafe.Pointer(mp)))
  2774  }
  2775  
  2776  // Return an M from the extra M list. Returns last == true if the list becomes
  2777  // empty because of this call.
  2778  //
  2779  // Spins waiting for an extra M, so caller must ensure that the list always
  2780  // contains or will soon contain at least one M.
  2781  //
  2782  //go:nosplit
  2783  func getExtraM() (mp *m, last bool) {
  2784  	mp = lockextra(false)
  2785  	extraMInUse.Add(1)
  2786  	unlockextra(mp.schedlink.ptr(), -1)
  2787  	return mp, mp.schedlink.ptr() == nil
  2788  }
  2789  
  2790  // Returns an extra M back to the list. mp must be from getExtraM. Newly
  2791  // allocated M's should use addExtraM.
  2792  //
  2793  //go:nosplit
  2794  func putExtraM(mp *m) {
  2795  	extraMInUse.Add(-1)
  2796  	addExtraM(mp)
  2797  }
  2798  
  2799  // Adds a newly allocated M to the extra M list.
  2800  //
  2801  //go:nosplit
  2802  func addExtraM(mp *m) {
  2803  	mnext := lockextra(true)
  2804  	mp.schedlink.set(mnext)
  2805  	unlockextra(mp, 1)
  2806  }
  2807  
  2808  var (
  2809  	// allocmLock is locked for read when creating new Ms in allocm and their
  2810  	// addition to allm. Thus acquiring this lock for write blocks the
  2811  	// creation of new Ms.
  2812  	allocmLock rwmutex
  2813  
  2814  	// execLock serializes exec and clone to avoid bugs or unspecified
  2815  	// behaviour around exec'ing while creating/destroying threads. See
  2816  	// issue #19546.
  2817  	execLock rwmutex
  2818  )
  2819  
  2820  // These errors are reported (via writeErrStr) by some OS-specific
  2821  // versions of newosproc and newosproc0.
  2822  const (
  2823  	failthreadcreate  = "runtime: failed to create new OS thread\n"
  2824  	failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
  2825  )
  2826  
  2827  // newmHandoff contains a list of m structures that need new OS threads.
  2828  // This is used by newm in situations where newm itself can't safely
  2829  // start an OS thread.
  2830  var newmHandoff struct {
  2831  	lock mutex
  2832  
  2833  	// newm points to a list of M structures that need new OS
  2834  	// threads. The list is linked through m.schedlink.
  2835  	newm muintptr
  2836  
  2837  	// waiting indicates that wake needs to be notified when an m
  2838  	// is put on the list.
  2839  	waiting bool
  2840  	wake    note
  2841  
  2842  	// haveTemplateThread indicates that the templateThread has
  2843  	// been started. This is not protected by lock. Use cas to set
  2844  	// to 1.
  2845  	haveTemplateThread uint32
  2846  }
  2847  
  2848  // Create a new m. It will start off with a call to fn, or else the scheduler.
  2849  // fn needs to be static and not a heap allocated closure.
  2850  // May run with m.p==nil, so write barriers are not allowed.
  2851  //
  2852  // id is optional pre-allocated m ID. Omit by passing -1.
  2853  //
  2854  //go:nowritebarrierrec
  2855  func newm(fn func(), pp *p, id int64) {
  2856  	// allocm adds a new M to allm, but they do not start until created by
  2857  	// the OS in newm1 or the template thread.
  2858  	//
  2859  	// doAllThreadsSyscall requires that every M in allm will eventually
  2860  	// start and be signal-able, even with a STW.
  2861  	//
  2862  	// Disable preemption here until we start the thread to ensure that
  2863  	// newm is not preempted between allocm and starting the new thread,
  2864  	// ensuring that anything added to allm is guaranteed to eventually
  2865  	// start.
  2866  	acquirem()
  2867  
  2868  	mp := allocm(pp, fn, id)
  2869  	mp.nextp.set(pp)
  2870  	mp.sigmask = initSigmask
  2871  	if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
  2872  		// We're on a locked M or a thread that may have been
  2873  		// started by C. The kernel state of this thread may
  2874  		// be strange (the user may have locked it for that
  2875  		// purpose). We don't want to clone that into another
  2876  		// thread. Instead, ask a known-good thread to create
  2877  		// the thread for us.
  2878  		//
  2879  		// This is disabled on Plan 9. See golang.org/issue/22227.
  2880  		//
  2881  		// TODO: This may be unnecessary on Windows, which
  2882  		// doesn't model thread creation off fork.
  2883  		lock(&newmHandoff.lock)
  2884  		if newmHandoff.haveTemplateThread == 0 {
  2885  			throw("on a locked thread with no template thread")
  2886  		}
  2887  		mp.schedlink = newmHandoff.newm
  2888  		newmHandoff.newm.set(mp)
  2889  		if newmHandoff.waiting {
  2890  			newmHandoff.waiting = false
  2891  			notewakeup(&newmHandoff.wake)
  2892  		}
  2893  		unlock(&newmHandoff.lock)
  2894  		// The M has not started yet, but the template thread does not
  2895  		// participate in STW, so it will always process queued Ms and
  2896  		// it is safe to releasem.
  2897  		releasem(getg().m)
  2898  		return
  2899  	}
  2900  	newm1(mp)
  2901  	releasem(getg().m)
  2902  }
  2903  
  2904  func newm1(mp *m) {
  2905  	if iscgo {
  2906  		var ts cgothreadstart
  2907  		if _cgo_thread_start == nil {
  2908  			throw("_cgo_thread_start missing")
  2909  		}
  2910  		ts.g.set(mp.g0)
  2911  		ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
  2912  		ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
  2913  		if msanenabled {
  2914  			msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2915  		}
  2916  		if asanenabled {
  2917  			asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2918  		}
  2919  		execLock.rlock() // Prevent process clone.
  2920  		asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
  2921  		execLock.runlock()
  2922  		return
  2923  	}
  2924  	execLock.rlock() // Prevent process clone.
  2925  	newosproc(mp)
  2926  	execLock.runlock()
  2927  }
  2928  
  2929  // startTemplateThread starts the template thread if it is not already
  2930  // running.
  2931  //
  2932  // The calling thread must itself be in a known-good state.
  2933  func startTemplateThread() {
  2934  	if GOARCH == "wasm" { // no threads on wasm yet
  2935  		return
  2936  	}
  2937  
  2938  	// Disable preemption to guarantee that the template thread will be
  2939  	// created before a park once haveTemplateThread is set.
  2940  	mp := acquirem()
  2941  	if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
  2942  		releasem(mp)
  2943  		return
  2944  	}
  2945  	newm(templateThread, nil, -1)
  2946  	releasem(mp)
  2947  }
  2948  
  2949  // templateThread is a thread in a known-good state that exists solely
  2950  // to start new threads in known-good states when the calling thread
  2951  // may not be in a good state.
  2952  //
  2953  // Many programs never need this, so templateThread is started lazily
  2954  // when we first enter a state that might lead to running on a thread
  2955  // in an unknown state.
  2956  //
  2957  // templateThread runs on an M without a P, so it must not have write
  2958  // barriers.
  2959  //
  2960  //go:nowritebarrierrec
  2961  func templateThread() {
  2962  	lock(&sched.lock)
  2963  	sched.nmsys++
  2964  	checkdead()
  2965  	unlock(&sched.lock)
  2966  
  2967  	for {
  2968  		lock(&newmHandoff.lock)
  2969  		for newmHandoff.newm != 0 {
  2970  			newm := newmHandoff.newm.ptr()
  2971  			newmHandoff.newm = 0
  2972  			unlock(&newmHandoff.lock)
  2973  			for newm != nil {
  2974  				next := newm.schedlink.ptr()
  2975  				newm.schedlink = 0
  2976  				newm1(newm)
  2977  				newm = next
  2978  			}
  2979  			lock(&newmHandoff.lock)
  2980  		}
  2981  		newmHandoff.waiting = true
  2982  		noteclear(&newmHandoff.wake)
  2983  		unlock(&newmHandoff.lock)
  2984  		notesleep(&newmHandoff.wake)
  2985  	}
  2986  }
  2987  
  2988  // Stops execution of the current m until new work is available.
  2989  // Returns with acquired P.
  2990  func stopm() {
  2991  	gp := getg()
  2992  
  2993  	if gp.m.locks != 0 {
  2994  		throw("stopm holding locks")
  2995  	}
  2996  	if gp.m.p != 0 {
  2997  		throw("stopm holding p")
  2998  	}
  2999  	if gp.m.spinning {
  3000  		throw("stopm spinning")
  3001  	}
  3002  
  3003  	lock(&sched.lock)
  3004  	mput(gp.m)
  3005  	unlock(&sched.lock)
  3006  	mPark()
  3007  	acquirep(gp.m.nextp.ptr())
  3008  	gp.m.nextp = 0
  3009  }
  3010  
  3011  func mspinning() {
  3012  	// startm's caller incremented nmspinning. Set the new M's spinning.
  3013  	getg().m.spinning = true
  3014  }
  3015  
  3016  // Schedules some M to run the p (creates an M if necessary).
  3017  // If p==nil, tries to get an idle P, if no idle P's does nothing.
  3018  // May run with m.p==nil, so write barriers are not allowed.
  3019  // If spinning is set, the caller has incremented nmspinning and must provide a
  3020  // P. startm will set m.spinning in the newly started M.
  3021  //
  3022  // Callers passing a non-nil P must call from a non-preemptible context. See
  3023  // comment on acquirem below.
  3024  //
  3025  // Argument lockheld indicates whether the caller already acquired the
  3026  // scheduler lock. Callers holding the lock when making the call must pass
  3027  // true. The lock might be temporarily dropped, but will be reacquired before
  3028  // returning.
  3029  //
  3030  // Must not have write barriers because this may be called without a P.
  3031  //
  3032  //go:nowritebarrierrec
  3033  func startm(pp *p, spinning, lockheld bool) {
  3034  	// Disable preemption.
  3035  	//
  3036  	// Every owned P must have an owner that will eventually stop it in the
  3037  	// event of a GC stop request. startm takes transient ownership of a P
  3038  	// (either from argument or pidleget below) and transfers ownership to
  3039  	// a started M, which will be responsible for performing the stop.
  3040  	//
  3041  	// Preemption must be disabled during this transient ownership,
  3042  	// otherwise the P this is running on may enter GC stop while still
  3043  	// holding the transient P, leaving that P in limbo and deadlocking the
  3044  	// STW.
  3045  	//
  3046  	// Callers passing a non-nil P must already be in non-preemptible
  3047  	// context, otherwise such preemption could occur on function entry to
  3048  	// startm. Callers passing a nil P may be preemptible, so we must
  3049  	// disable preemption before acquiring a P from pidleget below.
  3050  	mp := acquirem()
  3051  	if !lockheld {
  3052  		lock(&sched.lock)
  3053  	}
  3054  	if pp == nil {
  3055  		if spinning {
  3056  			// TODO(prattmic): All remaining calls to this function
  3057  			// with _p_ == nil could be cleaned up to find a P
  3058  			// before calling startm.
  3059  			throw("startm: P required for spinning=true")
  3060  		}
  3061  		pp, _ = pidleget(0)
  3062  		if pp == nil {
  3063  			if !lockheld {
  3064  				unlock(&sched.lock)
  3065  			}
  3066  			releasem(mp)
  3067  			return
  3068  		}
  3069  	}
  3070  	nmp := mget()
  3071  	if nmp == nil {
  3072  		// No M is available, we must drop sched.lock and call newm.
  3073  		// However, we already own a P to assign to the M.
  3074  		//
  3075  		// Once sched.lock is released, another G (e.g., in a syscall),
  3076  		// could find no idle P while checkdead finds a runnable G but
  3077  		// no running M's because this new M hasn't started yet, thus
  3078  		// throwing in an apparent deadlock.
  3079  		// This apparent deadlock is possible when startm is called
  3080  		// from sysmon, which doesn't count as a running M.
  3081  		//
  3082  		// Avoid this situation by pre-allocating the ID for the new M,
  3083  		// thus marking it as 'running' before we drop sched.lock. This
  3084  		// new M will eventually run the scheduler to execute any
  3085  		// queued G's.
  3086  		id := mReserveID()
  3087  		unlock(&sched.lock)
  3088  
  3089  		var fn func()
  3090  		if spinning {
  3091  			// The caller incremented nmspinning, so set m.spinning in the new M.
  3092  			fn = mspinning
  3093  		}
  3094  		newm(fn, pp, id)
  3095  
  3096  		if lockheld {
  3097  			lock(&sched.lock)
  3098  		}
  3099  		// Ownership transfer of pp committed by start in newm.
  3100  		// Preemption is now safe.
  3101  		releasem(mp)
  3102  		return
  3103  	}
  3104  	if !lockheld {
  3105  		unlock(&sched.lock)
  3106  	}
  3107  	if nmp.spinning {
  3108  		throw("startm: m is spinning")
  3109  	}
  3110  	if nmp.nextp != 0 {
  3111  		throw("startm: m has p")
  3112  	}
  3113  	if spinning && !runqempty(pp) {
  3114  		throw("startm: p has runnable gs")
  3115  	}
  3116  	// The caller incremented nmspinning, so set m.spinning in the new M.
  3117  	nmp.spinning = spinning
  3118  	nmp.nextp.set(pp)
  3119  	notewakeup(&nmp.park)
  3120  	// Ownership transfer of pp committed by wakeup. Preemption is now
  3121  	// safe.
  3122  	releasem(mp)
  3123  }
  3124  
  3125  // Hands off P from syscall or locked M.
  3126  // Always runs without a P, so write barriers are not allowed.
  3127  //
  3128  //go:nowritebarrierrec
  3129  func handoffp(pp *p) {
  3130  	// handoffp must start an M in any situation where
  3131  	// findrunnable would return a G to run on pp.
  3132  
  3133  	// if it has local work, start it straight away
  3134  	if !runqempty(pp) || !sched.runq.empty() {
  3135  		startm(pp, false, false)
  3136  		return
  3137  	}
  3138  	// if there's trace work to do, start it straight away
  3139  	if (traceEnabled() || traceShuttingDown()) && traceReaderAvailable() != nil {
  3140  		startm(pp, false, false)
  3141  		return
  3142  	}
  3143  	// if it has GC work, start it straight away
  3144  	if gcBlackenEnabled != 0 && gcShouldScheduleWorker(pp) {
  3145  		startm(pp, false, false)
  3146  		return
  3147  	}
  3148  	// no local work, check that there are no spinning/idle M's,
  3149  	// otherwise our help is not required
  3150  	if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
  3151  		sched.needspinning.Store(0)
  3152  		startm(pp, true, false)
  3153  		return
  3154  	}
  3155  	lock(&sched.lock)
  3156  	if sched.gcwaiting.Load() {
  3157  		pp.status = _Pgcstop
  3158  		pp.gcStopTime = nanotime()
  3159  		sched.stopwait--
  3160  		if sched.stopwait == 0 {
  3161  			notewakeup(&sched.stopnote)
  3162  		}
  3163  		unlock(&sched.lock)
  3164  		return
  3165  	}
  3166  	if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
  3167  		sched.safePointFn(pp)
  3168  		sched.safePointWait--
  3169  		if sched.safePointWait == 0 {
  3170  			notewakeup(&sched.safePointNote)
  3171  		}
  3172  	}
  3173  	if !sched.runq.empty() {
  3174  		unlock(&sched.lock)
  3175  		startm(pp, false, false)
  3176  		return
  3177  	}
  3178  	// If this is the last running P and nobody is polling network,
  3179  	// need to wakeup another M to poll network.
  3180  	if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
  3181  		unlock(&sched.lock)
  3182  		startm(pp, false, false)
  3183  		return
  3184  	}
  3185  
  3186  	// The scheduler lock cannot be held when calling wakeNetPoller below
  3187  	// because wakeNetPoller may call wakep which may call startm.
  3188  	when := pp.timers.wakeTime()
  3189  	pidleput(pp, 0)
  3190  	unlock(&sched.lock)
  3191  
  3192  	if when != 0 {
  3193  		wakeNetPoller(when)
  3194  	}
  3195  }
  3196  
  3197  // Tries to add one more P to execute G's.
  3198  // Called when a G is made runnable (newproc, ready).
  3199  // Must be called with a P.
  3200  //
  3201  // wakep should be an internal detail,
  3202  // but widely used packages access it using linkname.
  3203  // Notable members of the hall of shame include:
  3204  //   - gvisor.dev/gvisor
  3205  //
  3206  // Do not remove or change the type signature.
  3207  // See go.dev/issue/67401.
  3208  //
  3209  //go:linkname wakep
  3210  func wakep() {
  3211  	// Be conservative about spinning threads, only start one if none exist
  3212  	// already.
  3213  	if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
  3214  		return
  3215  	}
  3216  
  3217  	// Disable preemption until ownership of pp transfers to the next M in
  3218  	// startm. Otherwise preemption here would leave pp stuck waiting to
  3219  	// enter _Pgcstop.
  3220  	//
  3221  	// See preemption comment on acquirem in startm for more details.
  3222  	mp := acquirem()
  3223  
  3224  	var pp *p
  3225  	lock(&sched.lock)
  3226  	pp, _ = pidlegetSpinning(0)
  3227  	if pp == nil {
  3228  		if sched.nmspinning.Add(-1) < 0 {
  3229  			throw("wakep: negative nmspinning")
  3230  		}
  3231  		unlock(&sched.lock)
  3232  		releasem(mp)
  3233  		return
  3234  	}
  3235  	// Since we always have a P, the race in the "No M is available"
  3236  	// comment in startm doesn't apply during the small window between the
  3237  	// unlock here and lock in startm. A checkdead in between will always
  3238  	// see at least one running M (ours).
  3239  	unlock(&sched.lock)
  3240  
  3241  	startm(pp, true, false)
  3242  
  3243  	releasem(mp)
  3244  }
  3245  
  3246  // Stops execution of the current m that is locked to a g until the g is runnable again.
  3247  // Returns with acquired P.
  3248  func stoplockedm() {
  3249  	gp := getg()
  3250  
  3251  	if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
  3252  		throw("stoplockedm: inconsistent locking")
  3253  	}
  3254  	if gp.m.p != 0 {
  3255  		// Schedule another M to run this p.
  3256  		pp := releasep()
  3257  		handoffp(pp)
  3258  	}
  3259  	incidlelocked(1)
  3260  	// Wait until another thread schedules lockedg again.
  3261  	mPark()
  3262  	status := readgstatus(gp.m.lockedg.ptr())
  3263  	if status&^_Gscan != _Grunnable {
  3264  		print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
  3265  		dumpgstatus(gp.m.lockedg.ptr())
  3266  		throw("stoplockedm: not runnable")
  3267  	}
  3268  	acquirep(gp.m.nextp.ptr())
  3269  	gp.m.nextp = 0
  3270  }
  3271  
  3272  // Schedules the locked m to run the locked gp.
  3273  // May run during STW, so write barriers are not allowed.
  3274  //
  3275  //go:nowritebarrierrec
  3276  func startlockedm(gp *g) {
  3277  	mp := gp.lockedm.ptr()
  3278  	if mp == getg().m {
  3279  		throw("startlockedm: locked to me")
  3280  	}
  3281  	if mp.nextp != 0 {
  3282  		throw("startlockedm: m has p")
  3283  	}
  3284  	// directly handoff current P to the locked m
  3285  	incidlelocked(-1)
  3286  	pp := releasep()
  3287  	mp.nextp.set(pp)
  3288  	notewakeup(&mp.park)
  3289  	stopm()
  3290  }
  3291  
  3292  // Stops the current m for stopTheWorld.
  3293  // Returns when the world is restarted.
  3294  func gcstopm() {
  3295  	gp := getg()
  3296  
  3297  	if !sched.gcwaiting.Load() {
  3298  		throw("gcstopm: not waiting for gc")
  3299  	}
  3300  	if gp.m.spinning {
  3301  		gp.m.spinning = false
  3302  		// OK to just drop nmspinning here,
  3303  		// startTheWorld will unpark threads as necessary.
  3304  		if sched.nmspinning.Add(-1) < 0 {
  3305  			throw("gcstopm: negative nmspinning")
  3306  		}
  3307  	}
  3308  	pp := releasep()
  3309  	lock(&sched.lock)
  3310  	pp.status = _Pgcstop
  3311  	pp.gcStopTime = nanotime()
  3312  	sched.stopwait--
  3313  	if sched.stopwait == 0 {
  3314  		notewakeup(&sched.stopnote)
  3315  	}
  3316  	unlock(&sched.lock)
  3317  	stopm()
  3318  }
  3319  
  3320  // Schedules gp to run on the current M.
  3321  // If inheritTime is true, gp inherits the remaining time in the
  3322  // current time slice. Otherwise, it starts a new time slice.
  3323  // Never returns.
  3324  //
  3325  // Write barriers are allowed because this is called immediately after
  3326  // acquiring a P in several places.
  3327  //
  3328  //go:yeswritebarrierrec
  3329  func execute(gp *g, inheritTime bool) {
  3330  	mp := getg().m
  3331  
  3332  	if goroutineProfile.active {
  3333  		// Make sure that gp has had its stack written out to the goroutine
  3334  		// profile, exactly as it was when the goroutine profiler first stopped
  3335  		// the world.
  3336  		tryRecordGoroutineProfile(gp, nil, osyield)
  3337  	}
  3338  
  3339  	// Assign gp.m before entering _Grunning so running Gs have an M.
  3340  	mp.curg = gp
  3341  	gp.m = mp
  3342  	gp.syncSafePoint = false // Clear the flag, which may have been set by morestack.
  3343  	casgstatus(gp, _Grunnable, _Grunning)
  3344  	gp.waitsince = 0
  3345  	gp.preempt = false
  3346  	gp.stackguard0 = gp.stack.lo + stackGuard
  3347  	if !inheritTime {
  3348  		mp.p.ptr().schedtick++
  3349  	}
  3350  
  3351  	// Check whether the profiler needs to be turned on or off.
  3352  	hz := sched.profilehz
  3353  	if mp.profilehz != hz {
  3354  		setThreadCPUProfiler(hz)
  3355  	}
  3356  
  3357  	trace := traceAcquire()
  3358  	if trace.ok() {
  3359  		trace.GoStart()
  3360  		traceRelease(trace)
  3361  	}
  3362  
  3363  	gogo(&gp.sched)
  3364  }
  3365  
  3366  // Finds a runnable goroutine to execute.
  3367  // Tries to steal from other P's, get g from local or global queue, poll network.
  3368  // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
  3369  // reader) so the caller should try to wake a P.
  3370  func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
  3371  	mp := getg().m
  3372  
  3373  	// The conditions here and in handoffp must agree: if
  3374  	// findrunnable would return a G to run, handoffp must start
  3375  	// an M.
  3376  
  3377  top:
  3378  	// We may have collected an allp snapshot below. The snapshot is only
  3379  	// required in each loop iteration. Clear it to all GC to collect the
  3380  	// slice.
  3381  	mp.clearAllpSnapshot()
  3382  
  3383  	pp := mp.p.ptr()
  3384  	if sched.gcwaiting.Load() {
  3385  		gcstopm()
  3386  		goto top
  3387  	}
  3388  	if pp.runSafePointFn != 0 {
  3389  		runSafePointFn()
  3390  	}
  3391  
  3392  	// now and pollUntil are saved for work stealing later,
  3393  	// which may steal timers. It's important that between now
  3394  	// and then, nothing blocks, so these numbers remain mostly
  3395  	// relevant.
  3396  	now, pollUntil, _ := pp.timers.check(0, nil)
  3397  
  3398  	// Try to schedule the trace reader.
  3399  	if traceEnabled() || traceShuttingDown() {
  3400  		gp := traceReader()
  3401  		if gp != nil {
  3402  			trace := traceAcquire()
  3403  			casgstatus(gp, _Gwaiting, _Grunnable)
  3404  			if trace.ok() {
  3405  				trace.GoUnpark(gp, 0)
  3406  				traceRelease(trace)
  3407  			}
  3408  			return gp, false, true
  3409  		}
  3410  	}
  3411  
  3412  	// Try to schedule a GC worker.
  3413  	if gcBlackenEnabled != 0 {
  3414  		gp, tnow := gcController.findRunnableGCWorker(pp, now)
  3415  		if gp != nil {
  3416  			return gp, false, true
  3417  		}
  3418  		now = tnow
  3419  	}
  3420  
  3421  	// Check the global runnable queue once in a while to ensure fairness.
  3422  	// Otherwise two goroutines can completely occupy the local runqueue
  3423  	// by constantly respawning each other.
  3424  	if pp.schedtick%61 == 0 && !sched.runq.empty() {
  3425  		lock(&sched.lock)
  3426  		gp := globrunqget()
  3427  		unlock(&sched.lock)
  3428  		if gp != nil {
  3429  			return gp, false, false
  3430  		}
  3431  	}
  3432  
  3433  	// Wake up the finalizer G.
  3434  	if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
  3435  		if gp := wakefing(); gp != nil {
  3436  			ready(gp, 0, true)
  3437  		}
  3438  	}
  3439  
  3440  	// Wake up one or more cleanup Gs.
  3441  	if gcCleanups.needsWake() {
  3442  		gcCleanups.wake()
  3443  	}
  3444  
  3445  	if *cgo_yield != nil {
  3446  		asmcgocall(*cgo_yield, nil)
  3447  	}
  3448  
  3449  	// local runq
  3450  	if gp, inheritTime := runqget(pp); gp != nil {
  3451  		return gp, inheritTime, false
  3452  	}
  3453  
  3454  	// global runq
  3455  	if !sched.runq.empty() {
  3456  		lock(&sched.lock)
  3457  		gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3458  		unlock(&sched.lock)
  3459  		if gp != nil {
  3460  			if runqputbatch(pp, &q); !q.empty() {
  3461  				throw("Couldn't put Gs into empty local runq")
  3462  			}
  3463  			return gp, false, false
  3464  		}
  3465  	}
  3466  
  3467  	// Poll network.
  3468  	// This netpoll is only an optimization before we resort to stealing.
  3469  	// We can safely skip it if there are no waiters or a thread is blocked
  3470  	// in netpoll already. If there is any kind of logical race with that
  3471  	// blocked thread (e.g. it has already returned from netpoll, but does
  3472  	// not set lastpoll yet), this thread will do blocking netpoll below
  3473  	// anyway.
  3474  	// We only poll from one thread at a time to avoid kernel contention
  3475  	// on machines with many cores.
  3476  	if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 && sched.pollingNet.Swap(1) == 0 {
  3477  		list, delta := netpoll(0)
  3478  		sched.pollingNet.Store(0)
  3479  		if !list.empty() { // non-blocking
  3480  			gp := list.pop()
  3481  			injectglist(&list)
  3482  			netpollAdjustWaiters(delta)
  3483  			trace := traceAcquire()
  3484  			casgstatus(gp, _Gwaiting, _Grunnable)
  3485  			if trace.ok() {
  3486  				trace.GoUnpark(gp, 0)
  3487  				traceRelease(trace)
  3488  			}
  3489  			return gp, false, false
  3490  		}
  3491  	}
  3492  
  3493  	// Spinning Ms: steal work from other Ps.
  3494  	//
  3495  	// Limit the number of spinning Ms to half the number of busy Ps.
  3496  	// This is necessary to prevent excessive CPU consumption when
  3497  	// GOMAXPROCS>>1 but the program parallelism is low.
  3498  	if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
  3499  		if !mp.spinning {
  3500  			mp.becomeSpinning()
  3501  		}
  3502  
  3503  		gp, inheritTime, tnow, w, newWork := stealWork(now)
  3504  		if gp != nil {
  3505  			// Successfully stole.
  3506  			return gp, inheritTime, false
  3507  		}
  3508  		if newWork {
  3509  			// There may be new timer or GC work; restart to
  3510  			// discover.
  3511  			goto top
  3512  		}
  3513  
  3514  		now = tnow
  3515  		if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3516  			// Earlier timer to wait for.
  3517  			pollUntil = w
  3518  		}
  3519  	}
  3520  
  3521  	// We have nothing to do.
  3522  	//
  3523  	// If we're in the GC mark phase, can safely scan and blacken objects,
  3524  	// and have work to do, run idle-time marking rather than give up the P.
  3525  	if gcBlackenEnabled != 0 && gcShouldScheduleWorker(pp) && gcController.addIdleMarkWorker() {
  3526  		node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3527  		if node != nil {
  3528  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3529  			gp := node.gp.ptr()
  3530  
  3531  			trace := traceAcquire()
  3532  			casgstatus(gp, _Gwaiting, _Grunnable)
  3533  			if trace.ok() {
  3534  				trace.GoUnpark(gp, 0)
  3535  				traceRelease(trace)
  3536  			}
  3537  			return gp, false, false
  3538  		}
  3539  		gcController.removeIdleMarkWorker()
  3540  	}
  3541  
  3542  	// wasm only:
  3543  	// If a callback returned and no other goroutine is awake,
  3544  	// then wake event handler goroutine which pauses execution
  3545  	// until a callback was triggered.
  3546  	gp, otherReady := beforeIdle(now, pollUntil)
  3547  	if gp != nil {
  3548  		trace := traceAcquire()
  3549  		casgstatus(gp, _Gwaiting, _Grunnable)
  3550  		if trace.ok() {
  3551  			trace.GoUnpark(gp, 0)
  3552  			traceRelease(trace)
  3553  		}
  3554  		return gp, false, false
  3555  	}
  3556  	if otherReady {
  3557  		goto top
  3558  	}
  3559  
  3560  	// Before we drop our P, make a snapshot of the allp slice,
  3561  	// which can change underfoot once we no longer block
  3562  	// safe-points. We don't need to snapshot the contents because
  3563  	// everything up to cap(allp) is immutable.
  3564  	//
  3565  	// We clear the snapshot from the M after return via
  3566  	// mp.clearAllpSnapshop (in schedule) and on each iteration of the top
  3567  	// loop.
  3568  	allpSnapshot := mp.snapshotAllp()
  3569  	// Also snapshot masks. Value changes are OK, but we can't allow
  3570  	// len to change out from under us.
  3571  	idlepMaskSnapshot := idlepMask
  3572  	timerpMaskSnapshot := timerpMask
  3573  
  3574  	// return P and block
  3575  	lock(&sched.lock)
  3576  	if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
  3577  		unlock(&sched.lock)
  3578  		goto top
  3579  	}
  3580  	if !sched.runq.empty() {
  3581  		gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3582  		unlock(&sched.lock)
  3583  		if gp == nil {
  3584  			throw("global runq empty with non-zero runqsize")
  3585  		}
  3586  		if runqputbatch(pp, &q); !q.empty() {
  3587  			throw("Couldn't put Gs into empty local runq")
  3588  		}
  3589  		return gp, false, false
  3590  	}
  3591  	if !mp.spinning && sched.needspinning.Load() == 1 {
  3592  		// See "Delicate dance" comment below.
  3593  		mp.becomeSpinning()
  3594  		unlock(&sched.lock)
  3595  		goto top
  3596  	}
  3597  	if releasep() != pp {
  3598  		throw("findrunnable: wrong p")
  3599  	}
  3600  	now = pidleput(pp, now)
  3601  	unlock(&sched.lock)
  3602  
  3603  	// Delicate dance: thread transitions from spinning to non-spinning
  3604  	// state, potentially concurrently with submission of new work. We must
  3605  	// drop nmspinning first and then check all sources again (with
  3606  	// #StoreLoad memory barrier in between). If we do it the other way
  3607  	// around, another thread can submit work after we've checked all
  3608  	// sources but before we drop nmspinning; as a result nobody will
  3609  	// unpark a thread to run the work.
  3610  	//
  3611  	// This applies to the following sources of work:
  3612  	//
  3613  	// * Goroutines added to the global or a per-P run queue.
  3614  	// * New/modified-earlier timers on a per-P timer heap.
  3615  	// * Idle-priority GC work (barring golang.org/issue/19112).
  3616  	//
  3617  	// If we discover new work below, we need to restore m.spinning as a
  3618  	// signal for resetspinning to unpark a new worker thread (because
  3619  	// there can be more than one starving goroutine).
  3620  	//
  3621  	// However, if after discovering new work we also observe no idle Ps
  3622  	// (either here or in resetspinning), we have a problem. We may be
  3623  	// racing with a non-spinning M in the block above, having found no
  3624  	// work and preparing to release its P and park. Allowing that P to go
  3625  	// idle will result in loss of work conservation (idle P while there is
  3626  	// runnable work). This could result in complete deadlock in the
  3627  	// unlikely event that we discover new work (from netpoll) right as we
  3628  	// are racing with _all_ other Ps going idle.
  3629  	//
  3630  	// We use sched.needspinning to synchronize with non-spinning Ms going
  3631  	// idle. If needspinning is set when they are about to drop their P,
  3632  	// they abort the drop and instead become a new spinning M on our
  3633  	// behalf. If we are not racing and the system is truly fully loaded
  3634  	// then no spinning threads are required, and the next thread to
  3635  	// naturally become spinning will clear the flag.
  3636  	//
  3637  	// Also see "Worker thread parking/unparking" comment at the top of the
  3638  	// file.
  3639  	wasSpinning := mp.spinning
  3640  	if mp.spinning {
  3641  		mp.spinning = false
  3642  		if sched.nmspinning.Add(-1) < 0 {
  3643  			throw("findrunnable: negative nmspinning")
  3644  		}
  3645  
  3646  		// Note the for correctness, only the last M transitioning from
  3647  		// spinning to non-spinning must perform these rechecks to
  3648  		// ensure no missed work. However, the runtime has some cases
  3649  		// of transient increments of nmspinning that are decremented
  3650  		// without going through this path, so we must be conservative
  3651  		// and perform the check on all spinning Ms.
  3652  		//
  3653  		// See https://go.dev/issue/43997.
  3654  
  3655  		// Check global and P runqueues again.
  3656  
  3657  		lock(&sched.lock)
  3658  		if !sched.runq.empty() {
  3659  			pp, _ := pidlegetSpinning(0)
  3660  			if pp != nil {
  3661  				gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3662  				unlock(&sched.lock)
  3663  				if gp == nil {
  3664  					throw("global runq empty with non-zero runqsize")
  3665  				}
  3666  				if runqputbatch(pp, &q); !q.empty() {
  3667  					throw("Couldn't put Gs into empty local runq")
  3668  				}
  3669  				acquirep(pp)
  3670  				mp.becomeSpinning()
  3671  				return gp, false, false
  3672  			}
  3673  		}
  3674  		unlock(&sched.lock)
  3675  
  3676  		pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
  3677  		if pp != nil {
  3678  			acquirep(pp)
  3679  			mp.becomeSpinning()
  3680  			goto top
  3681  		}
  3682  
  3683  		// Check for idle-priority GC work again.
  3684  		pp, gp := checkIdleGCNoP()
  3685  		if pp != nil {
  3686  			acquirep(pp)
  3687  			mp.becomeSpinning()
  3688  
  3689  			// Run the idle worker.
  3690  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3691  			trace := traceAcquire()
  3692  			casgstatus(gp, _Gwaiting, _Grunnable)
  3693  			if trace.ok() {
  3694  				trace.GoUnpark(gp, 0)
  3695  				traceRelease(trace)
  3696  			}
  3697  			return gp, false, false
  3698  		}
  3699  
  3700  		// Finally, check for timer creation or expiry concurrently with
  3701  		// transitioning from spinning to non-spinning.
  3702  		//
  3703  		// Note that we cannot use checkTimers here because it calls
  3704  		// adjusttimers which may need to allocate memory, and that isn't
  3705  		// allowed when we don't have an active P.
  3706  		pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
  3707  	}
  3708  
  3709  	// We don't need allp anymore at this pointer, but can't clear the
  3710  	// snapshot without a P for the write barrier..
  3711  
  3712  	// Poll network until next timer.
  3713  	if netpollinited() && (netpollAnyWaiters() || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
  3714  		sched.pollUntil.Store(pollUntil)
  3715  		if mp.p != 0 {
  3716  			throw("findrunnable: netpoll with p")
  3717  		}
  3718  		if mp.spinning {
  3719  			throw("findrunnable: netpoll with spinning")
  3720  		}
  3721  		delay := int64(-1)
  3722  		if pollUntil != 0 {
  3723  			if now == 0 {
  3724  				now = nanotime()
  3725  			}
  3726  			delay = pollUntil - now
  3727  			if delay < 0 {
  3728  				delay = 0
  3729  			}
  3730  		}
  3731  		if faketime != 0 {
  3732  			// When using fake time, just poll.
  3733  			delay = 0
  3734  		}
  3735  		list, delta := netpoll(delay) // block until new work is available
  3736  		// Refresh now again, after potentially blocking.
  3737  		now = nanotime()
  3738  		sched.pollUntil.Store(0)
  3739  		sched.lastpoll.Store(now)
  3740  		if faketime != 0 && list.empty() {
  3741  			// Using fake time and nothing is ready; stop M.
  3742  			// When all M's stop, checkdead will call timejump.
  3743  			stopm()
  3744  			goto top
  3745  		}
  3746  		lock(&sched.lock)
  3747  		pp, _ := pidleget(now)
  3748  		unlock(&sched.lock)
  3749  		if pp == nil {
  3750  			injectglist(&list)
  3751  			netpollAdjustWaiters(delta)
  3752  		} else {
  3753  			acquirep(pp)
  3754  			if !list.empty() {
  3755  				gp := list.pop()
  3756  				injectglist(&list)
  3757  				netpollAdjustWaiters(delta)
  3758  				trace := traceAcquire()
  3759  				casgstatus(gp, _Gwaiting, _Grunnable)
  3760  				if trace.ok() {
  3761  					trace.GoUnpark(gp, 0)
  3762  					traceRelease(trace)
  3763  				}
  3764  				return gp, false, false
  3765  			}
  3766  			if wasSpinning {
  3767  				mp.becomeSpinning()
  3768  			}
  3769  			goto top
  3770  		}
  3771  	} else if pollUntil != 0 && netpollinited() {
  3772  		pollerPollUntil := sched.pollUntil.Load()
  3773  		if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
  3774  			netpollBreak()
  3775  		}
  3776  	}
  3777  	stopm()
  3778  	goto top
  3779  }
  3780  
  3781  // pollWork reports whether there is non-background work this P could
  3782  // be doing. This is a fairly lightweight check to be used for
  3783  // background work loops, like idle GC. It checks a subset of the
  3784  // conditions checked by the actual scheduler.
  3785  func pollWork() bool {
  3786  	if !sched.runq.empty() {
  3787  		return true
  3788  	}
  3789  	p := getg().m.p.ptr()
  3790  	if !runqempty(p) {
  3791  		return true
  3792  	}
  3793  	if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 {
  3794  		if list, delta := netpoll(0); !list.empty() {
  3795  			injectglist(&list)
  3796  			netpollAdjustWaiters(delta)
  3797  			return true
  3798  		}
  3799  	}
  3800  	return false
  3801  }
  3802  
  3803  // stealWork attempts to steal a runnable goroutine or timer from any P.
  3804  //
  3805  // If newWork is true, new work may have been readied.
  3806  //
  3807  // If now is not 0 it is the current time. stealWork returns the passed time or
  3808  // the current time if now was passed as 0.
  3809  func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
  3810  	pp := getg().m.p.ptr()
  3811  
  3812  	ranTimer := false
  3813  
  3814  	const stealTries = 4
  3815  	for i := 0; i < stealTries; i++ {
  3816  		stealTimersOrRunNextG := i == stealTries-1
  3817  
  3818  		for enum := stealOrder.start(cheaprand()); !enum.done(); enum.next() {
  3819  			if sched.gcwaiting.Load() {
  3820  				// GC work may be available.
  3821  				return nil, false, now, pollUntil, true
  3822  			}
  3823  			p2 := allp[enum.position()]
  3824  			if pp == p2 {
  3825  				continue
  3826  			}
  3827  
  3828  			// Steal timers from p2. This call to checkTimers is the only place
  3829  			// where we might hold a lock on a different P's timers. We do this
  3830  			// once on the last pass before checking runnext because stealing
  3831  			// from the other P's runnext should be the last resort, so if there
  3832  			// are timers to steal do that first.
  3833  			//
  3834  			// We only check timers on one of the stealing iterations because
  3835  			// the time stored in now doesn't change in this loop and checking
  3836  			// the timers for each P more than once with the same value of now
  3837  			// is probably a waste of time.
  3838  			//
  3839  			// timerpMask tells us whether the P may have timers at all. If it
  3840  			// can't, no need to check at all.
  3841  			if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
  3842  				tnow, w, ran := p2.timers.check(now, nil)
  3843  				now = tnow
  3844  				if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3845  					pollUntil = w
  3846  				}
  3847  				if ran {
  3848  					// Running the timers may have
  3849  					// made an arbitrary number of G's
  3850  					// ready and added them to this P's
  3851  					// local run queue. That invalidates
  3852  					// the assumption of runqsteal
  3853  					// that it always has room to add
  3854  					// stolen G's. So check now if there
  3855  					// is a local G to run.
  3856  					if gp, inheritTime := runqget(pp); gp != nil {
  3857  						return gp, inheritTime, now, pollUntil, ranTimer
  3858  					}
  3859  					ranTimer = true
  3860  				}
  3861  			}
  3862  
  3863  			// Don't bother to attempt to steal if p2 is idle.
  3864  			if !idlepMask.read(enum.position()) {
  3865  				if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
  3866  					return gp, false, now, pollUntil, ranTimer
  3867  				}
  3868  			}
  3869  		}
  3870  	}
  3871  
  3872  	// No goroutines found to steal. Regardless, running a timer may have
  3873  	// made some goroutine ready that we missed. Indicate the next timer to
  3874  	// wait for.
  3875  	return nil, false, now, pollUntil, ranTimer
  3876  }
  3877  
  3878  // Check all Ps for a runnable G to steal.
  3879  //
  3880  // On entry we have no P. If a G is available to steal and a P is available,
  3881  // the P is returned which the caller should acquire and attempt to steal the
  3882  // work to.
  3883  func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
  3884  	for id, p2 := range allpSnapshot {
  3885  		if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
  3886  			lock(&sched.lock)
  3887  			pp, _ := pidlegetSpinning(0)
  3888  			if pp == nil {
  3889  				// Can't get a P, don't bother checking remaining Ps.
  3890  				unlock(&sched.lock)
  3891  				return nil
  3892  			}
  3893  			unlock(&sched.lock)
  3894  			return pp
  3895  		}
  3896  	}
  3897  
  3898  	// No work available.
  3899  	return nil
  3900  }
  3901  
  3902  // Check all Ps for a timer expiring sooner than pollUntil.
  3903  //
  3904  // Returns updated pollUntil value.
  3905  func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
  3906  	for id, p2 := range allpSnapshot {
  3907  		if timerpMaskSnapshot.read(uint32(id)) {
  3908  			w := p2.timers.wakeTime()
  3909  			if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3910  				pollUntil = w
  3911  			}
  3912  		}
  3913  	}
  3914  
  3915  	return pollUntil
  3916  }
  3917  
  3918  // Check for idle-priority GC, without a P on entry.
  3919  //
  3920  // If some GC work, a P, and a worker G are all available, the P and G will be
  3921  // returned. The returned P has not been wired yet.
  3922  func checkIdleGCNoP() (*p, *g) {
  3923  	// N.B. Since we have no P, gcBlackenEnabled may change at any time; we
  3924  	// must check again after acquiring a P. As an optimization, we also check
  3925  	// if an idle mark worker is needed at all. This is OK here, because if we
  3926  	// observe that one isn't needed, at least one is currently running. Even if
  3927  	// it stops running, its own journey into the scheduler should schedule it
  3928  	// again, if need be (at which point, this check will pass, if relevant).
  3929  	if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
  3930  		return nil, nil
  3931  	}
  3932  	if !gcShouldScheduleWorker(nil) {
  3933  		return nil, nil
  3934  	}
  3935  
  3936  	// Work is available; we can start an idle GC worker only if there is
  3937  	// an available P and available worker G.
  3938  	//
  3939  	// We can attempt to acquire these in either order, though both have
  3940  	// synchronization concerns (see below). Workers are almost always
  3941  	// available (see comment in findRunnableGCWorker for the one case
  3942  	// there may be none). Since we're slightly less likely to find a P,
  3943  	// check for that first.
  3944  	//
  3945  	// Synchronization: note that we must hold sched.lock until we are
  3946  	// committed to keeping it. Otherwise we cannot put the unnecessary P
  3947  	// back in sched.pidle without performing the full set of idle
  3948  	// transition checks.
  3949  	//
  3950  	// If we were to check gcBgMarkWorkerPool first, we must somehow handle
  3951  	// the assumption in gcControllerState.findRunnableGCWorker that an
  3952  	// empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
  3953  	lock(&sched.lock)
  3954  	pp, now := pidlegetSpinning(0)
  3955  	if pp == nil {
  3956  		unlock(&sched.lock)
  3957  		return nil, nil
  3958  	}
  3959  
  3960  	// Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
  3961  	if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
  3962  		pidleput(pp, now)
  3963  		unlock(&sched.lock)
  3964  		return nil, nil
  3965  	}
  3966  
  3967  	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3968  	if node == nil {
  3969  		pidleput(pp, now)
  3970  		unlock(&sched.lock)
  3971  		gcController.removeIdleMarkWorker()
  3972  		return nil, nil
  3973  	}
  3974  
  3975  	unlock(&sched.lock)
  3976  
  3977  	return pp, node.gp.ptr()
  3978  }
  3979  
  3980  // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
  3981  // going to wake up before the when argument; or it wakes an idle P to service
  3982  // timers and the network poller if there isn't one already.
  3983  func wakeNetPoller(when int64) {
  3984  	if sched.lastpoll.Load() == 0 {
  3985  		// In findrunnable we ensure that when polling the pollUntil
  3986  		// field is either zero or the time to which the current
  3987  		// poll is expected to run. This can have a spurious wakeup
  3988  		// but should never miss a wakeup.
  3989  		pollerPollUntil := sched.pollUntil.Load()
  3990  		if pollerPollUntil == 0 || pollerPollUntil > when {
  3991  			netpollBreak()
  3992  		}
  3993  	} else {
  3994  		// There are no threads in the network poller, try to get
  3995  		// one there so it can handle new timers.
  3996  		if GOOS != "plan9" { // Temporary workaround - see issue #42303.
  3997  			wakep()
  3998  		}
  3999  	}
  4000  }
  4001  
  4002  func resetspinning() {
  4003  	gp := getg()
  4004  	if !gp.m.spinning {
  4005  		throw("resetspinning: not a spinning m")
  4006  	}
  4007  	gp.m.spinning = false
  4008  	nmspinning := sched.nmspinning.Add(-1)
  4009  	if nmspinning < 0 {
  4010  		throw("findrunnable: negative nmspinning")
  4011  	}
  4012  	// M wakeup policy is deliberately somewhat conservative, so check if we
  4013  	// need to wakeup another P here. See "Worker thread parking/unparking"
  4014  	// comment at the top of the file for details.
  4015  	wakep()
  4016  }
  4017  
  4018  // injectglist adds each runnable G on the list to some run queue,
  4019  // and clears glist. If there is no current P, they are added to the
  4020  // global queue, and up to npidle M's are started to run them.
  4021  // Otherwise, for each idle P, this adds a G to the global queue
  4022  // and starts an M. Any remaining G's are added to the current P's
  4023  // local run queue.
  4024  // This may temporarily acquire sched.lock.
  4025  // Can run concurrently with GC.
  4026  func injectglist(glist *gList) {
  4027  	if glist.empty() {
  4028  		return
  4029  	}
  4030  
  4031  	// Mark all the goroutines as runnable before we put them
  4032  	// on the run queues.
  4033  	var tail *g
  4034  	trace := traceAcquire()
  4035  	for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
  4036  		tail = gp
  4037  		casgstatus(gp, _Gwaiting, _Grunnable)
  4038  		if trace.ok() {
  4039  			trace.GoUnpark(gp, 0)
  4040  		}
  4041  	}
  4042  	if trace.ok() {
  4043  		traceRelease(trace)
  4044  	}
  4045  
  4046  	// Turn the gList into a gQueue.
  4047  	q := gQueue{glist.head, tail.guintptr(), glist.size}
  4048  	*glist = gList{}
  4049  
  4050  	startIdle := func(n int32) {
  4051  		for ; n > 0; n-- {
  4052  			mp := acquirem() // See comment in startm.
  4053  			lock(&sched.lock)
  4054  
  4055  			pp, _ := pidlegetSpinning(0)
  4056  			if pp == nil {
  4057  				unlock(&sched.lock)
  4058  				releasem(mp)
  4059  				break
  4060  			}
  4061  
  4062  			startm(pp, false, true)
  4063  			unlock(&sched.lock)
  4064  			releasem(mp)
  4065  		}
  4066  	}
  4067  
  4068  	pp := getg().m.p.ptr()
  4069  	if pp == nil {
  4070  		n := q.size
  4071  		lock(&sched.lock)
  4072  		globrunqputbatch(&q)
  4073  		unlock(&sched.lock)
  4074  		startIdle(n)
  4075  		return
  4076  	}
  4077  
  4078  	var globq gQueue
  4079  	npidle := sched.npidle.Load()
  4080  	for ; npidle > 0 && !q.empty(); npidle-- {
  4081  		g := q.pop()
  4082  		globq.pushBack(g)
  4083  	}
  4084  	if !globq.empty() {
  4085  		n := globq.size
  4086  		lock(&sched.lock)
  4087  		globrunqputbatch(&globq)
  4088  		unlock(&sched.lock)
  4089  		startIdle(n)
  4090  	}
  4091  
  4092  	if runqputbatch(pp, &q); !q.empty() {
  4093  		lock(&sched.lock)
  4094  		globrunqputbatch(&q)
  4095  		unlock(&sched.lock)
  4096  	}
  4097  
  4098  	// Some P's might have become idle after we loaded `sched.npidle`
  4099  	// but before any goroutines were added to the queue, which could
  4100  	// lead to idle P's when there is work available in the global queue.
  4101  	// That could potentially last until other goroutines become ready
  4102  	// to run. That said, we need to find a way to hedge
  4103  	//
  4104  	// Calling wakep() here is the best bet, it will do nothing in the
  4105  	// common case (no racing on `sched.npidle`), while it could wake one
  4106  	// more P to execute G's, which might end up with >1 P's: the first one
  4107  	// wakes another P and so forth until there is no more work, but this
  4108  	// ought to be an extremely rare case.
  4109  	//
  4110  	// Also see "Worker thread parking/unparking" comment at the top of the file for details.
  4111  	wakep()
  4112  }
  4113  
  4114  // One round of scheduler: find a runnable goroutine and execute it.
  4115  // Never returns.
  4116  func schedule() {
  4117  	mp := getg().m
  4118  
  4119  	if mp.locks != 0 {
  4120  		throw("schedule: holding locks")
  4121  	}
  4122  
  4123  	if mp.lockedg != 0 {
  4124  		stoplockedm()
  4125  		execute(mp.lockedg.ptr(), false) // Never returns.
  4126  	}
  4127  
  4128  	// We should not schedule away from a g that is executing a cgo call,
  4129  	// since the cgo call is using the m's g0 stack.
  4130  	if mp.incgo {
  4131  		throw("schedule: in cgo")
  4132  	}
  4133  
  4134  top:
  4135  	pp := mp.p.ptr()
  4136  	pp.preempt = false
  4137  
  4138  	// Safety check: if we are spinning, the run queue should be empty.
  4139  	// Check this before calling checkTimers, as that might call
  4140  	// goready to put a ready goroutine on the local run queue.
  4141  	if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
  4142  		throw("schedule: spinning with local work")
  4143  	}
  4144  
  4145  	gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
  4146  
  4147  	// findRunnable may have collected an allp snapshot. The snapshot is
  4148  	// only required within findRunnable. Clear it to all GC to collect the
  4149  	// slice.
  4150  	mp.clearAllpSnapshot()
  4151  
  4152  	if debug.dontfreezetheworld > 0 && freezing.Load() {
  4153  		// See comment in freezetheworld. We don't want to perturb
  4154  		// scheduler state, so we didn't gcstopm in findRunnable, but
  4155  		// also don't want to allow new goroutines to run.
  4156  		//
  4157  		// Deadlock here rather than in the findRunnable loop so if
  4158  		// findRunnable is stuck in a loop we don't perturb that
  4159  		// either.
  4160  		lock(&deadlock)
  4161  		lock(&deadlock)
  4162  	}
  4163  
  4164  	// This thread is going to run a goroutine and is not spinning anymore,
  4165  	// so if it was marked as spinning we need to reset it now and potentially
  4166  	// start a new spinning M.
  4167  	if mp.spinning {
  4168  		resetspinning()
  4169  	}
  4170  
  4171  	if sched.disable.user && !schedEnabled(gp) {
  4172  		// Scheduling of this goroutine is disabled. Put it on
  4173  		// the list of pending runnable goroutines for when we
  4174  		// re-enable user scheduling and look again.
  4175  		lock(&sched.lock)
  4176  		if schedEnabled(gp) {
  4177  			// Something re-enabled scheduling while we
  4178  			// were acquiring the lock.
  4179  			unlock(&sched.lock)
  4180  		} else {
  4181  			sched.disable.runnable.pushBack(gp)
  4182  			unlock(&sched.lock)
  4183  			goto top
  4184  		}
  4185  	}
  4186  
  4187  	// If about to schedule a not-normal goroutine (a GCworker or tracereader),
  4188  	// wake a P if there is one.
  4189  	if tryWakeP {
  4190  		wakep()
  4191  	}
  4192  	if gp.lockedm != 0 {
  4193  		// Hands off own p to the locked m,
  4194  		// then blocks waiting for a new p.
  4195  		startlockedm(gp)
  4196  		goto top
  4197  	}
  4198  
  4199  	execute(gp, inheritTime)
  4200  }
  4201  
  4202  // dropg removes the association between m and the current goroutine m->curg (gp for short).
  4203  // Typically a caller sets gp's status away from Grunning and then
  4204  // immediately calls dropg to finish the job. The caller is also responsible
  4205  // for arranging that gp will be restarted using ready at an
  4206  // appropriate time. After calling dropg and arranging for gp to be
  4207  // readied later, the caller can do other work but eventually should
  4208  // call schedule to restart the scheduling of goroutines on this m.
  4209  func dropg() {
  4210  	gp := getg()
  4211  
  4212  	setMNoWB(&gp.m.curg.m, nil)
  4213  	setGNoWB(&gp.m.curg, nil)
  4214  }
  4215  
  4216  func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
  4217  	unlock((*mutex)(lock))
  4218  	return true
  4219  }
  4220  
  4221  // park continuation on g0.
  4222  func park_m(gp *g) {
  4223  	mp := getg().m
  4224  
  4225  	trace := traceAcquire()
  4226  
  4227  	// If g is in a synctest group, we don't want to let the group
  4228  	// become idle until after the waitunlockf (if any) has confirmed
  4229  	// that the park is happening.
  4230  	// We need to record gp.bubble here, since waitunlockf can change it.
  4231  	bubble := gp.bubble
  4232  	if bubble != nil {
  4233  		bubble.incActive()
  4234  	}
  4235  
  4236  	if trace.ok() {
  4237  		// Trace the event before the transition. It may take a
  4238  		// stack trace, but we won't own the stack after the
  4239  		// transition anymore.
  4240  		trace.GoPark(mp.waitTraceBlockReason, mp.waitTraceSkip)
  4241  	}
  4242  	// N.B. Not using casGToWaiting here because the waitreason is
  4243  	// set by park_m's caller.
  4244  	casgstatus(gp, _Grunning, _Gwaiting)
  4245  	if trace.ok() {
  4246  		traceRelease(trace)
  4247  	}
  4248  
  4249  	dropg()
  4250  
  4251  	if fn := mp.waitunlockf; fn != nil {
  4252  		ok := fn(gp, mp.waitlock)
  4253  		mp.waitunlockf = nil
  4254  		mp.waitlock = nil
  4255  		if !ok {
  4256  			trace := traceAcquire()
  4257  			casgstatus(gp, _Gwaiting, _Grunnable)
  4258  			if bubble != nil {
  4259  				bubble.decActive()
  4260  			}
  4261  			if trace.ok() {
  4262  				trace.GoUnpark(gp, 2)
  4263  				traceRelease(trace)
  4264  			}
  4265  			execute(gp, true) // Schedule it back, never returns.
  4266  		}
  4267  	}
  4268  
  4269  	if bubble != nil {
  4270  		bubble.decActive()
  4271  	}
  4272  
  4273  	schedule()
  4274  }
  4275  
  4276  func goschedImpl(gp *g, preempted bool) {
  4277  	trace := traceAcquire()
  4278  	status := readgstatus(gp)
  4279  	if status&^_Gscan != _Grunning {
  4280  		dumpgstatus(gp)
  4281  		throw("bad g status")
  4282  	}
  4283  	if trace.ok() {
  4284  		// Trace the event before the transition. It may take a
  4285  		// stack trace, but we won't own the stack after the
  4286  		// transition anymore.
  4287  		if preempted {
  4288  			trace.GoPreempt()
  4289  		} else {
  4290  			trace.GoSched()
  4291  		}
  4292  	}
  4293  	casgstatus(gp, _Grunning, _Grunnable)
  4294  	if trace.ok() {
  4295  		traceRelease(trace)
  4296  	}
  4297  
  4298  	dropg()
  4299  	lock(&sched.lock)
  4300  	globrunqput(gp)
  4301  	unlock(&sched.lock)
  4302  
  4303  	if mainStarted {
  4304  		wakep()
  4305  	}
  4306  
  4307  	schedule()
  4308  }
  4309  
  4310  // Gosched continuation on g0.
  4311  func gosched_m(gp *g) {
  4312  	goschedImpl(gp, false)
  4313  }
  4314  
  4315  // goschedguarded is a forbidden-states-avoided version of gosched_m.
  4316  func goschedguarded_m(gp *g) {
  4317  	if !canPreemptM(gp.m) {
  4318  		gogo(&gp.sched) // never return
  4319  	}
  4320  	goschedImpl(gp, false)
  4321  }
  4322  
  4323  func gopreempt_m(gp *g) {
  4324  	goschedImpl(gp, true)
  4325  }
  4326  
  4327  // preemptPark parks gp and puts it in _Gpreempted.
  4328  //
  4329  //go:systemstack
  4330  func preemptPark(gp *g) {
  4331  	status := readgstatus(gp)
  4332  	if status&^_Gscan != _Grunning {
  4333  		dumpgstatus(gp)
  4334  		throw("bad g status")
  4335  	}
  4336  
  4337  	if gp.asyncSafePoint {
  4338  		// Double-check that async preemption does not
  4339  		// happen in SPWRITE assembly functions.
  4340  		// isAsyncSafePoint must exclude this case.
  4341  		f := findfunc(gp.sched.pc)
  4342  		if !f.valid() {
  4343  			throw("preempt at unknown pc")
  4344  		}
  4345  		if f.flag&abi.FuncFlagSPWrite != 0 {
  4346  			println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
  4347  			throw("preempt SPWRITE")
  4348  		}
  4349  	}
  4350  
  4351  	// Transition from _Grunning to _Gscan|_Gpreempted. We can't
  4352  	// be in _Grunning when we dropg because then we'd be running
  4353  	// without an M, but the moment we're in _Gpreempted,
  4354  	// something could claim this G before we've fully cleaned it
  4355  	// up. Hence, we set the scan bit to lock down further
  4356  	// transitions until we can dropg.
  4357  	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
  4358  	dropg()
  4359  
  4360  	// Be careful about how we trace this next event. The ordering
  4361  	// is subtle.
  4362  	//
  4363  	// The moment we CAS into _Gpreempted, suspendG could CAS to
  4364  	// _Gwaiting, do its work, and ready the goroutine. All of
  4365  	// this could happen before we even get the chance to emit
  4366  	// an event. The end result is that the events could appear
  4367  	// out of order, and the tracer generally assumes the scheduler
  4368  	// takes care of the ordering between GoPark and GoUnpark.
  4369  	//
  4370  	// The answer here is simple: emit the event while we still hold
  4371  	// the _Gscan bit on the goroutine. We still need to traceAcquire
  4372  	// and traceRelease across the CAS because the tracer could be
  4373  	// what's calling suspendG in the first place, and we want the
  4374  	// CAS and event emission to appear atomic to the tracer.
  4375  	trace := traceAcquire()
  4376  	if trace.ok() {
  4377  		trace.GoPark(traceBlockPreempted, 0)
  4378  	}
  4379  	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
  4380  	if trace.ok() {
  4381  		traceRelease(trace)
  4382  	}
  4383  	schedule()
  4384  }
  4385  
  4386  // goyield is like Gosched, but it:
  4387  // - emits a GoPreempt trace event instead of a GoSched trace event
  4388  // - puts the current G on the runq of the current P instead of the globrunq
  4389  //
  4390  // goyield should be an internal detail,
  4391  // but widely used packages access it using linkname.
  4392  // Notable members of the hall of shame include:
  4393  //   - gvisor.dev/gvisor
  4394  //   - github.com/sagernet/gvisor
  4395  //
  4396  // Do not remove or change the type signature.
  4397  // See go.dev/issue/67401.
  4398  //
  4399  //go:linkname goyield
  4400  func goyield() {
  4401  	checkTimeouts()
  4402  	mcall(goyield_m)
  4403  }
  4404  
  4405  func goyield_m(gp *g) {
  4406  	trace := traceAcquire()
  4407  	pp := gp.m.p.ptr()
  4408  	if trace.ok() {
  4409  		// Trace the event before the transition. It may take a
  4410  		// stack trace, but we won't own the stack after the
  4411  		// transition anymore.
  4412  		trace.GoPreempt()
  4413  	}
  4414  	casgstatus(gp, _Grunning, _Grunnable)
  4415  	if trace.ok() {
  4416  		traceRelease(trace)
  4417  	}
  4418  	dropg()
  4419  	runqput(pp, gp, false)
  4420  	schedule()
  4421  }
  4422  
  4423  // Finishes execution of the current goroutine.
  4424  func goexit1() {
  4425  	if raceenabled {
  4426  		if gp := getg(); gp.bubble != nil {
  4427  			racereleasemergeg(gp, gp.bubble.raceaddr())
  4428  		}
  4429  		racegoend()
  4430  	}
  4431  	trace := traceAcquire()
  4432  	if trace.ok() {
  4433  		trace.GoEnd()
  4434  		traceRelease(trace)
  4435  	}
  4436  	mcall(goexit0)
  4437  }
  4438  
  4439  // goexit continuation on g0.
  4440  func goexit0(gp *g) {
  4441  	gdestroy(gp)
  4442  	schedule()
  4443  }
  4444  
  4445  func gdestroy(gp *g) {
  4446  	mp := getg().m
  4447  	pp := mp.p.ptr()
  4448  
  4449  	casgstatus(gp, _Grunning, _Gdead)
  4450  	gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
  4451  	if isSystemGoroutine(gp, false) {
  4452  		sched.ngsys.Add(-1)
  4453  	}
  4454  	gp.m = nil
  4455  	locked := gp.lockedm != 0
  4456  	gp.lockedm = 0
  4457  	mp.lockedg = 0
  4458  	gp.preemptStop = false
  4459  	gp.paniconfault = false
  4460  	gp._defer = nil // should be true already but just in case.
  4461  	gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
  4462  	gp.writebuf = nil
  4463  	gp.waitreason = waitReasonZero
  4464  	gp.param = nil
  4465  	gp.labels = nil
  4466  	gp.timer = nil
  4467  	gp.bubble = nil
  4468  
  4469  	if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
  4470  		// Flush assist credit to the global pool. This gives
  4471  		// better information to pacing if the application is
  4472  		// rapidly creating an exiting goroutines.
  4473  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
  4474  		scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
  4475  		gcController.bgScanCredit.Add(scanCredit)
  4476  		gp.gcAssistBytes = 0
  4477  	}
  4478  
  4479  	dropg()
  4480  
  4481  	if GOARCH == "wasm" { // no threads yet on wasm
  4482  		gfput(pp, gp)
  4483  		return
  4484  	}
  4485  
  4486  	if locked && mp.lockedInt != 0 {
  4487  		print("runtime: mp.lockedInt = ", mp.lockedInt, "\n")
  4488  		if mp.isextra {
  4489  			throw("runtime.Goexit called in a thread that was not created by the Go runtime")
  4490  		}
  4491  		throw("exited a goroutine internally locked to the OS thread")
  4492  	}
  4493  	gfput(pp, gp)
  4494  	if locked {
  4495  		// The goroutine may have locked this thread because
  4496  		// it put it in an unusual kernel state. Kill it
  4497  		// rather than returning it to the thread pool.
  4498  
  4499  		// Return to mstart, which will release the P and exit
  4500  		// the thread.
  4501  		if GOOS != "plan9" { // See golang.org/issue/22227.
  4502  			gogo(&mp.g0.sched)
  4503  		} else {
  4504  			// Clear lockedExt on plan9 since we may end up re-using
  4505  			// this thread.
  4506  			mp.lockedExt = 0
  4507  		}
  4508  	}
  4509  }
  4510  
  4511  // save updates getg().sched to refer to pc and sp so that a following
  4512  // gogo will restore pc and sp.
  4513  //
  4514  // save must not have write barriers because invoking a write barrier
  4515  // can clobber getg().sched.
  4516  //
  4517  //go:nosplit
  4518  //go:nowritebarrierrec
  4519  func save(pc, sp, bp uintptr) {
  4520  	gp := getg()
  4521  
  4522  	if gp == gp.m.g0 || gp == gp.m.gsignal {
  4523  		// m.g0.sched is special and must describe the context
  4524  		// for exiting the thread. mstart1 writes to it directly.
  4525  		// m.gsignal.sched should not be used at all.
  4526  		// This check makes sure save calls do not accidentally
  4527  		// run in contexts where they'd write to system g's.
  4528  		throw("save on system g not allowed")
  4529  	}
  4530  
  4531  	gp.sched.pc = pc
  4532  	gp.sched.sp = sp
  4533  	gp.sched.lr = 0
  4534  	gp.sched.bp = bp
  4535  	// We need to ensure ctxt is zero, but can't have a write
  4536  	// barrier here. However, it should always already be zero.
  4537  	// Assert that.
  4538  	if gp.sched.ctxt != nil {
  4539  		badctxt()
  4540  	}
  4541  }
  4542  
  4543  // The goroutine g is about to enter a system call.
  4544  // Record that it's not using the cpu anymore.
  4545  // This is called only from the go syscall library and cgocall,
  4546  // not from the low-level system calls used by the runtime.
  4547  //
  4548  // Entersyscall cannot split the stack: the save must
  4549  // make g->sched refer to the caller's stack segment, because
  4550  // entersyscall is going to return immediately after.
  4551  //
  4552  // Nothing entersyscall calls can split the stack either.
  4553  // We cannot safely move the stack during an active call to syscall,
  4554  // because we do not know which of the uintptr arguments are
  4555  // really pointers (back into the stack).
  4556  // In practice, this means that we make the fast path run through
  4557  // entersyscall doing no-split things, and the slow path has to use systemstack
  4558  // to run bigger things on the system stack.
  4559  //
  4560  // reentersyscall is the entry point used by cgo callbacks, where explicitly
  4561  // saved SP and PC are restored. This is needed when exitsyscall will be called
  4562  // from a function further up in the call stack than the parent, as g->syscallsp
  4563  // must always point to a valid stack frame. entersyscall below is the normal
  4564  // entry point for syscalls, which obtains the SP and PC from the caller.
  4565  //
  4566  //go:nosplit
  4567  func reentersyscall(pc, sp, bp uintptr) {
  4568  	trace := traceAcquire()
  4569  	gp := getg()
  4570  
  4571  	// Disable preemption because during this function g is in Gsyscall status,
  4572  	// but can have inconsistent g->sched, do not let GC observe it.
  4573  	gp.m.locks++
  4574  
  4575  	// Entersyscall must not call any function that might split/grow the stack.
  4576  	// (See details in comment above.)
  4577  	// Catch calls that might, by replacing the stack guard with something that
  4578  	// will trip any stack check and leaving a flag to tell newstack to die.
  4579  	gp.stackguard0 = stackPreempt
  4580  	gp.throwsplit = true
  4581  
  4582  	// Leave SP around for GC and traceback.
  4583  	save(pc, sp, bp)
  4584  	gp.syscallsp = sp
  4585  	gp.syscallpc = pc
  4586  	gp.syscallbp = bp
  4587  	casgstatus(gp, _Grunning, _Gsyscall)
  4588  	if staticLockRanking {
  4589  		// When doing static lock ranking casgstatus can call
  4590  		// systemstack which clobbers g.sched.
  4591  		save(pc, sp, bp)
  4592  	}
  4593  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4594  		systemstack(func() {
  4595  			print("entersyscall inconsistent sp ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4596  			throw("entersyscall")
  4597  		})
  4598  	}
  4599  	if gp.syscallbp != 0 && gp.syscallbp < gp.stack.lo || gp.stack.hi < gp.syscallbp {
  4600  		systemstack(func() {
  4601  			print("entersyscall inconsistent bp ", hex(gp.syscallbp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4602  			throw("entersyscall")
  4603  		})
  4604  	}
  4605  
  4606  	if trace.ok() {
  4607  		systemstack(func() {
  4608  			trace.GoSysCall()
  4609  			traceRelease(trace)
  4610  		})
  4611  		// systemstack itself clobbers g.sched.{pc,sp} and we might
  4612  		// need them later when the G is genuinely blocked in a
  4613  		// syscall
  4614  		save(pc, sp, bp)
  4615  	}
  4616  
  4617  	if sched.sysmonwait.Load() {
  4618  		systemstack(entersyscall_sysmon)
  4619  		save(pc, sp, bp)
  4620  	}
  4621  
  4622  	if gp.m.p.ptr().runSafePointFn != 0 {
  4623  		// runSafePointFn may stack split if run on this stack
  4624  		systemstack(runSafePointFn)
  4625  		save(pc, sp, bp)
  4626  	}
  4627  
  4628  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4629  	pp := gp.m.p.ptr()
  4630  	pp.m = 0
  4631  	gp.m.oldp.set(pp)
  4632  	gp.m.p = 0
  4633  	atomic.Store(&pp.status, _Psyscall)
  4634  	if sched.gcwaiting.Load() {
  4635  		systemstack(entersyscall_gcwait)
  4636  		save(pc, sp, bp)
  4637  	}
  4638  
  4639  	gp.m.locks--
  4640  }
  4641  
  4642  // Standard syscall entry used by the go syscall library and normal cgo calls.
  4643  //
  4644  // This is exported via linkname to assembly in the syscall package and x/sys.
  4645  //
  4646  // Other packages should not be accessing entersyscall directly,
  4647  // but widely used packages access it using linkname.
  4648  // Notable members of the hall of shame include:
  4649  //   - gvisor.dev/gvisor
  4650  //
  4651  // Do not remove or change the type signature.
  4652  // See go.dev/issue/67401.
  4653  //
  4654  //go:nosplit
  4655  //go:linkname entersyscall
  4656  func entersyscall() {
  4657  	// N.B. getcallerfp cannot be written directly as argument in the call
  4658  	// to reentersyscall because it forces spilling the other arguments to
  4659  	// the stack. This results in exceeding the nosplit stack requirements
  4660  	// on some platforms.
  4661  	fp := getcallerfp()
  4662  	reentersyscall(sys.GetCallerPC(), sys.GetCallerSP(), fp)
  4663  }
  4664  
  4665  func entersyscall_sysmon() {
  4666  	lock(&sched.lock)
  4667  	if sched.sysmonwait.Load() {
  4668  		sched.sysmonwait.Store(false)
  4669  		notewakeup(&sched.sysmonnote)
  4670  	}
  4671  	unlock(&sched.lock)
  4672  }
  4673  
  4674  func entersyscall_gcwait() {
  4675  	gp := getg()
  4676  	pp := gp.m.oldp.ptr()
  4677  
  4678  	lock(&sched.lock)
  4679  	trace := traceAcquire()
  4680  	if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
  4681  		if trace.ok() {
  4682  			// This is a steal in the new tracer. While it's very likely
  4683  			// that we were the ones to put this P into _Psyscall, between
  4684  			// then and now it's totally possible it had been stolen and
  4685  			// then put back into _Psyscall for us to acquire here. In such
  4686  			// case ProcStop would be incorrect.
  4687  			//
  4688  			// TODO(mknyszek): Consider emitting a ProcStop instead when
  4689  			// gp.m.syscalltick == pp.syscalltick, since then we know we never
  4690  			// lost the P.
  4691  			trace.ProcSteal(pp, true)
  4692  			traceRelease(trace)
  4693  		}
  4694  		sched.nGsyscallNoP.Add(1)
  4695  		pp.gcStopTime = nanotime()
  4696  		pp.syscalltick++
  4697  		if sched.stopwait--; sched.stopwait == 0 {
  4698  			notewakeup(&sched.stopnote)
  4699  		}
  4700  	} else if trace.ok() {
  4701  		traceRelease(trace)
  4702  	}
  4703  	unlock(&sched.lock)
  4704  }
  4705  
  4706  // The same as entersyscall(), but with a hint that the syscall is blocking.
  4707  
  4708  // entersyscallblock should be an internal detail,
  4709  // but widely used packages access it using linkname.
  4710  // Notable members of the hall of shame include:
  4711  //   - gvisor.dev/gvisor
  4712  //
  4713  // Do not remove or change the type signature.
  4714  // See go.dev/issue/67401.
  4715  //
  4716  //go:linkname entersyscallblock
  4717  //go:nosplit
  4718  func entersyscallblock() {
  4719  	gp := getg()
  4720  
  4721  	gp.m.locks++ // see comment in entersyscall
  4722  	gp.throwsplit = true
  4723  	gp.stackguard0 = stackPreempt // see comment in entersyscall
  4724  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4725  	gp.m.p.ptr().syscalltick++
  4726  
  4727  	sched.nGsyscallNoP.Add(1)
  4728  
  4729  	// Leave SP around for GC and traceback.
  4730  	pc := sys.GetCallerPC()
  4731  	sp := sys.GetCallerSP()
  4732  	bp := getcallerfp()
  4733  	save(pc, sp, bp)
  4734  	gp.syscallsp = gp.sched.sp
  4735  	gp.syscallpc = gp.sched.pc
  4736  	gp.syscallbp = gp.sched.bp
  4737  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4738  		sp1 := sp
  4739  		sp2 := gp.sched.sp
  4740  		sp3 := gp.syscallsp
  4741  		systemstack(func() {
  4742  			print("entersyscallblock inconsistent sp ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4743  			throw("entersyscallblock")
  4744  		})
  4745  	}
  4746  	casgstatus(gp, _Grunning, _Gsyscall)
  4747  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4748  		systemstack(func() {
  4749  			print("entersyscallblock inconsistent sp ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4750  			throw("entersyscallblock")
  4751  		})
  4752  	}
  4753  	if gp.syscallbp != 0 && gp.syscallbp < gp.stack.lo || gp.stack.hi < gp.syscallbp {
  4754  		systemstack(func() {
  4755  			print("entersyscallblock inconsistent bp ", hex(bp), " ", hex(gp.sched.bp), " ", hex(gp.syscallbp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4756  			throw("entersyscallblock")
  4757  		})
  4758  	}
  4759  
  4760  	systemstack(entersyscallblock_handoff)
  4761  
  4762  	// Resave for traceback during blocked call.
  4763  	save(sys.GetCallerPC(), sys.GetCallerSP(), getcallerfp())
  4764  
  4765  	gp.m.locks--
  4766  }
  4767  
  4768  func entersyscallblock_handoff() {
  4769  	trace := traceAcquire()
  4770  	if trace.ok() {
  4771  		trace.GoSysCall()
  4772  		traceRelease(trace)
  4773  	}
  4774  	handoffp(releasep())
  4775  }
  4776  
  4777  // The goroutine g exited its system call.
  4778  // Arrange for it to run on a cpu again.
  4779  // This is called only from the go syscall library, not
  4780  // from the low-level system calls used by the runtime.
  4781  //
  4782  // Write barriers are not allowed because our P may have been stolen.
  4783  //
  4784  // This is exported via linkname to assembly in the syscall package.
  4785  //
  4786  // exitsyscall should be an internal detail,
  4787  // but widely used packages access it using linkname.
  4788  // Notable members of the hall of shame include:
  4789  //   - gvisor.dev/gvisor
  4790  //
  4791  // Do not remove or change the type signature.
  4792  // See go.dev/issue/67401.
  4793  //
  4794  //go:nosplit
  4795  //go:nowritebarrierrec
  4796  //go:linkname exitsyscall
  4797  func exitsyscall() {
  4798  	gp := getg()
  4799  
  4800  	gp.m.locks++ // see comment in entersyscall
  4801  	if sys.GetCallerSP() > gp.syscallsp {
  4802  		throw("exitsyscall: syscall frame is no longer valid")
  4803  	}
  4804  
  4805  	gp.waitsince = 0
  4806  	oldp := gp.m.oldp.ptr()
  4807  	gp.m.oldp = 0
  4808  	if exitsyscallfast(oldp) {
  4809  		// When exitsyscallfast returns success, we have a P so can now use
  4810  		// write barriers
  4811  		if goroutineProfile.active {
  4812  			// Make sure that gp has had its stack written out to the goroutine
  4813  			// profile, exactly as it was when the goroutine profiler first
  4814  			// stopped the world.
  4815  			systemstack(func() {
  4816  				tryRecordGoroutineProfileWB(gp)
  4817  			})
  4818  		}
  4819  		trace := traceAcquire()
  4820  		if trace.ok() {
  4821  			lostP := oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick
  4822  			systemstack(func() {
  4823  				// Write out syscall exit eagerly.
  4824  				//
  4825  				// It's important that we write this *after* we know whether we
  4826  				// lost our P or not (determined by exitsyscallfast).
  4827  				trace.GoSysExit(lostP)
  4828  				if lostP {
  4829  					// We lost the P at some point, even though we got it back here.
  4830  					// Trace that we're starting again, because there was a tracev2.GoSysBlock
  4831  					// call somewhere in exitsyscallfast (indicating that this goroutine
  4832  					// had blocked) and we're about to start running again.
  4833  					trace.GoStart()
  4834  				}
  4835  			})
  4836  		}
  4837  		// There's a cpu for us, so we can run.
  4838  		gp.m.p.ptr().syscalltick++
  4839  		// We need to cas the status and scan before resuming...
  4840  		casgstatus(gp, _Gsyscall, _Grunning)
  4841  		if trace.ok() {
  4842  			traceRelease(trace)
  4843  		}
  4844  
  4845  		// Garbage collector isn't running (since we are),
  4846  		// so okay to clear syscallsp.
  4847  		gp.syscallsp = 0
  4848  		gp.m.locks--
  4849  		if gp.preempt {
  4850  			// restore the preemption request in case we've cleared it in newstack
  4851  			gp.stackguard0 = stackPreempt
  4852  		} else {
  4853  			// otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
  4854  			gp.stackguard0 = gp.stack.lo + stackGuard
  4855  		}
  4856  		gp.throwsplit = false
  4857  
  4858  		if sched.disable.user && !schedEnabled(gp) {
  4859  			// Scheduling of this goroutine is disabled.
  4860  			Gosched()
  4861  		}
  4862  
  4863  		return
  4864  	}
  4865  
  4866  	gp.m.locks--
  4867  
  4868  	// Call the scheduler.
  4869  	mcall(exitsyscall0)
  4870  
  4871  	// Scheduler returned, so we're allowed to run now.
  4872  	// Delete the syscallsp information that we left for
  4873  	// the garbage collector during the system call.
  4874  	// Must wait until now because until gosched returns
  4875  	// we don't know for sure that the garbage collector
  4876  	// is not running.
  4877  	gp.syscallsp = 0
  4878  	gp.m.p.ptr().syscalltick++
  4879  	gp.throwsplit = false
  4880  }
  4881  
  4882  //go:nosplit
  4883  func exitsyscallfast(oldp *p) bool {
  4884  	// Freezetheworld sets stopwait but does not retake P's.
  4885  	if sched.stopwait == freezeStopWait {
  4886  		return false
  4887  	}
  4888  
  4889  	// Try to re-acquire the last P.
  4890  	trace := traceAcquire()
  4891  	if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
  4892  		// There's a cpu for us, so we can run.
  4893  		wirep(oldp)
  4894  		exitsyscallfast_reacquired(trace)
  4895  		if trace.ok() {
  4896  			traceRelease(trace)
  4897  		}
  4898  		return true
  4899  	}
  4900  	if trace.ok() {
  4901  		traceRelease(trace)
  4902  	}
  4903  
  4904  	// Try to get any other idle P.
  4905  	if sched.pidle != 0 {
  4906  		var ok bool
  4907  		systemstack(func() {
  4908  			ok = exitsyscallfast_pidle()
  4909  		})
  4910  		if ok {
  4911  			return true
  4912  		}
  4913  	}
  4914  	return false
  4915  }
  4916  
  4917  // exitsyscallfast_reacquired is the exitsyscall path on which this G
  4918  // has successfully reacquired the P it was running on before the
  4919  // syscall.
  4920  //
  4921  //go:nosplit
  4922  func exitsyscallfast_reacquired(trace traceLocker) {
  4923  	gp := getg()
  4924  	if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  4925  		if trace.ok() {
  4926  			// The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
  4927  			// tracev2.GoSysBlock for this syscall was already emitted,
  4928  			// but here we effectively retake the p from the new syscall running on the same p.
  4929  			systemstack(func() {
  4930  				// We're stealing the P. It's treated
  4931  				// as if it temporarily stopped running. Then, start running.
  4932  				trace.ProcSteal(gp.m.p.ptr(), true)
  4933  				trace.ProcStart()
  4934  			})
  4935  		}
  4936  		gp.m.p.ptr().syscalltick++
  4937  	}
  4938  }
  4939  
  4940  func exitsyscallfast_pidle() bool {
  4941  	lock(&sched.lock)
  4942  	pp, _ := pidleget(0)
  4943  	if pp != nil && sched.sysmonwait.Load() {
  4944  		sched.sysmonwait.Store(false)
  4945  		notewakeup(&sched.sysmonnote)
  4946  	}
  4947  	unlock(&sched.lock)
  4948  	if pp != nil {
  4949  		sched.nGsyscallNoP.Add(-1)
  4950  		acquirep(pp)
  4951  		return true
  4952  	}
  4953  	return false
  4954  }
  4955  
  4956  // exitsyscall slow path on g0.
  4957  // Failed to acquire P, enqueue gp as runnable.
  4958  //
  4959  // Called via mcall, so gp is the calling g from this M.
  4960  //
  4961  //go:nowritebarrierrec
  4962  func exitsyscall0(gp *g) {
  4963  	var trace traceLocker
  4964  	traceExitingSyscall()
  4965  	trace = traceAcquire()
  4966  	casgstatus(gp, _Gsyscall, _Grunnable)
  4967  	traceExitedSyscall()
  4968  	if trace.ok() {
  4969  		// Write out syscall exit eagerly.
  4970  		//
  4971  		// It's important that we write this *after* we know whether we
  4972  		// lost our P or not (determined by exitsyscallfast).
  4973  		trace.GoSysExit(true)
  4974  		traceRelease(trace)
  4975  	}
  4976  	sched.nGsyscallNoP.Add(-1)
  4977  	dropg()
  4978  	lock(&sched.lock)
  4979  	var pp *p
  4980  	if schedEnabled(gp) {
  4981  		pp, _ = pidleget(0)
  4982  	}
  4983  	var locked bool
  4984  	if pp == nil {
  4985  		globrunqput(gp)
  4986  
  4987  		// Below, we stoplockedm if gp is locked. globrunqput releases
  4988  		// ownership of gp, so we must check if gp is locked prior to
  4989  		// committing the release by unlocking sched.lock, otherwise we
  4990  		// could race with another M transitioning gp from unlocked to
  4991  		// locked.
  4992  		locked = gp.lockedm != 0
  4993  	} else if sched.sysmonwait.Load() {
  4994  		sched.sysmonwait.Store(false)
  4995  		notewakeup(&sched.sysmonnote)
  4996  	}
  4997  	unlock(&sched.lock)
  4998  	if pp != nil {
  4999  		acquirep(pp)
  5000  		execute(gp, false) // Never returns.
  5001  	}
  5002  	if locked {
  5003  		// Wait until another thread schedules gp and so m again.
  5004  		//
  5005  		// N.B. lockedm must be this M, as this g was running on this M
  5006  		// before entersyscall.
  5007  		stoplockedm()
  5008  		execute(gp, false) // Never returns.
  5009  	}
  5010  	stopm()
  5011  	schedule() // Never returns.
  5012  }
  5013  
  5014  // Called from syscall package before fork.
  5015  //
  5016  // syscall_runtime_BeforeFork is for package syscall,
  5017  // but widely used packages access it using linkname.
  5018  // Notable members of the hall of shame include:
  5019  //   - gvisor.dev/gvisor
  5020  //
  5021  // Do not remove or change the type signature.
  5022  // See go.dev/issue/67401.
  5023  //
  5024  //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
  5025  //go:nosplit
  5026  func syscall_runtime_BeforeFork() {
  5027  	gp := getg().m.curg
  5028  
  5029  	// Block signals during a fork, so that the child does not run
  5030  	// a signal handler before exec if a signal is sent to the process
  5031  	// group. See issue #18600.
  5032  	gp.m.locks++
  5033  	sigsave(&gp.m.sigmask)
  5034  	sigblock(false)
  5035  
  5036  	// This function is called before fork in syscall package.
  5037  	// Code between fork and exec must not allocate memory nor even try to grow stack.
  5038  	// Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
  5039  	// runtime_AfterFork will undo this in parent process, but not in child.
  5040  	gp.stackguard0 = stackFork
  5041  }
  5042  
  5043  // Called from syscall package after fork in parent.
  5044  //
  5045  // syscall_runtime_AfterFork is for package syscall,
  5046  // but widely used packages access it using linkname.
  5047  // Notable members of the hall of shame include:
  5048  //   - gvisor.dev/gvisor
  5049  //
  5050  // Do not remove or change the type signature.
  5051  // See go.dev/issue/67401.
  5052  //
  5053  //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
  5054  //go:nosplit
  5055  func syscall_runtime_AfterFork() {
  5056  	gp := getg().m.curg
  5057  
  5058  	// See the comments in beforefork.
  5059  	gp.stackguard0 = gp.stack.lo + stackGuard
  5060  
  5061  	msigrestore(gp.m.sigmask)
  5062  
  5063  	gp.m.locks--
  5064  }
  5065  
  5066  // inForkedChild is true while manipulating signals in the child process.
  5067  // This is used to avoid calling libc functions in case we are using vfork.
  5068  var inForkedChild bool
  5069  
  5070  // Called from syscall package after fork in child.
  5071  // It resets non-sigignored signals to the default handler, and
  5072  // restores the signal mask in preparation for the exec.
  5073  //
  5074  // Because this might be called during a vfork, and therefore may be
  5075  // temporarily sharing address space with the parent process, this must
  5076  // not change any global variables or calling into C code that may do so.
  5077  //
  5078  // syscall_runtime_AfterForkInChild is for package syscall,
  5079  // but widely used packages access it using linkname.
  5080  // Notable members of the hall of shame include:
  5081  //   - gvisor.dev/gvisor
  5082  //
  5083  // Do not remove or change the type signature.
  5084  // See go.dev/issue/67401.
  5085  //
  5086  //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
  5087  //go:nosplit
  5088  //go:nowritebarrierrec
  5089  func syscall_runtime_AfterForkInChild() {
  5090  	// It's OK to change the global variable inForkedChild here
  5091  	// because we are going to change it back. There is no race here,
  5092  	// because if we are sharing address space with the parent process,
  5093  	// then the parent process can not be running concurrently.
  5094  	inForkedChild = true
  5095  
  5096  	clearSignalHandlers()
  5097  
  5098  	// When we are the child we are the only thread running,
  5099  	// so we know that nothing else has changed gp.m.sigmask.
  5100  	msigrestore(getg().m.sigmask)
  5101  
  5102  	inForkedChild = false
  5103  }
  5104  
  5105  // pendingPreemptSignals is the number of preemption signals
  5106  // that have been sent but not received. This is only used on Darwin.
  5107  // For #41702.
  5108  var pendingPreemptSignals atomic.Int32
  5109  
  5110  // Called from syscall package before Exec.
  5111  //
  5112  //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
  5113  func syscall_runtime_BeforeExec() {
  5114  	// Prevent thread creation during exec.
  5115  	execLock.lock()
  5116  
  5117  	// On Darwin, wait for all pending preemption signals to
  5118  	// be received. See issue #41702.
  5119  	if GOOS == "darwin" || GOOS == "ios" {
  5120  		for pendingPreemptSignals.Load() > 0 {
  5121  			osyield()
  5122  		}
  5123  	}
  5124  }
  5125  
  5126  // Called from syscall package after Exec.
  5127  //
  5128  //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
  5129  func syscall_runtime_AfterExec() {
  5130  	execLock.unlock()
  5131  }
  5132  
  5133  // Allocate a new g, with a stack big enough for stacksize bytes.
  5134  func malg(stacksize int32) *g {
  5135  	newg := new(g)
  5136  	if stacksize >= 0 {
  5137  		stacksize = round2(stackSystem + stacksize)
  5138  		systemstack(func() {
  5139  			newg.stack = stackalloc(uint32(stacksize))
  5140  			if valgrindenabled {
  5141  				newg.valgrindStackID = valgrindRegisterStack(unsafe.Pointer(newg.stack.lo), unsafe.Pointer(newg.stack.hi))
  5142  			}
  5143  		})
  5144  		newg.stackguard0 = newg.stack.lo + stackGuard
  5145  		newg.stackguard1 = ^uintptr(0)
  5146  		// Clear the bottom word of the stack. We record g
  5147  		// there on gsignal stack during VDSO on ARM and ARM64.
  5148  		*(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
  5149  	}
  5150  	return newg
  5151  }
  5152  
  5153  // Create a new g running fn.
  5154  // Put it on the queue of g's waiting to run.
  5155  // The compiler turns a go statement into a call to this.
  5156  func newproc(fn *funcval) {
  5157  	gp := getg()
  5158  	pc := sys.GetCallerPC()
  5159  	systemstack(func() {
  5160  		newg := newproc1(fn, gp, pc, false, waitReasonZero)
  5161  
  5162  		pp := getg().m.p.ptr()
  5163  		runqput(pp, newg, true)
  5164  
  5165  		if mainStarted {
  5166  			wakep()
  5167  		}
  5168  	})
  5169  }
  5170  
  5171  // Create a new g in state _Grunnable (or _Gwaiting if parked is true), starting at fn.
  5172  // callerpc is the address of the go statement that created this. The caller is responsible
  5173  // for adding the new g to the scheduler. If parked is true, waitreason must be non-zero.
  5174  func newproc1(fn *funcval, callergp *g, callerpc uintptr, parked bool, waitreason waitReason) *g {
  5175  	if fn == nil {
  5176  		fatal("go of nil func value")
  5177  	}
  5178  
  5179  	mp := acquirem() // disable preemption because we hold M and P in local vars.
  5180  	pp := mp.p.ptr()
  5181  	newg := gfget(pp)
  5182  	if newg == nil {
  5183  		newg = malg(stackMin)
  5184  		casgstatus(newg, _Gidle, _Gdead)
  5185  		allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
  5186  	}
  5187  	if newg.stack.hi == 0 {
  5188  		throw("newproc1: newg missing stack")
  5189  	}
  5190  
  5191  	if readgstatus(newg) != _Gdead {
  5192  		throw("newproc1: new g is not Gdead")
  5193  	}
  5194  
  5195  	totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
  5196  	totalSize = alignUp(totalSize, sys.StackAlign)
  5197  	sp := newg.stack.hi - totalSize
  5198  	if usesLR {
  5199  		// caller's LR
  5200  		*(*uintptr)(unsafe.Pointer(sp)) = 0
  5201  		prepGoExitFrame(sp)
  5202  	}
  5203  	if GOARCH == "arm64" {
  5204  		// caller's FP
  5205  		*(*uintptr)(unsafe.Pointer(sp - goarch.PtrSize)) = 0
  5206  	}
  5207  
  5208  	memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
  5209  	newg.sched.sp = sp
  5210  	newg.stktopsp = sp
  5211  	newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
  5212  	newg.sched.g = guintptr(unsafe.Pointer(newg))
  5213  	gostartcallfn(&newg.sched, fn)
  5214  	newg.parentGoid = callergp.goid
  5215  	newg.gopc = callerpc
  5216  	newg.ancestors = saveAncestors(callergp)
  5217  	newg.startpc = fn.fn
  5218  	newg.runningCleanups.Store(false)
  5219  	if isSystemGoroutine(newg, false) {
  5220  		sched.ngsys.Add(1)
  5221  	} else {
  5222  		// Only user goroutines inherit synctest groups and pprof labels.
  5223  		newg.bubble = callergp.bubble
  5224  		if mp.curg != nil {
  5225  			newg.labels = mp.curg.labels
  5226  		}
  5227  		if goroutineProfile.active {
  5228  			// A concurrent goroutine profile is running. It should include
  5229  			// exactly the set of goroutines that were alive when the goroutine
  5230  			// profiler first stopped the world. That does not include newg, so
  5231  			// mark it as not needing a profile before transitioning it from
  5232  			// _Gdead.
  5233  			newg.goroutineProfiled.Store(goroutineProfileSatisfied)
  5234  		}
  5235  	}
  5236  	// Track initial transition?
  5237  	newg.trackingSeq = uint8(cheaprand())
  5238  	if newg.trackingSeq%gTrackingPeriod == 0 {
  5239  		newg.tracking = true
  5240  	}
  5241  	gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
  5242  
  5243  	// Get a goid and switch to runnable. Make all this atomic to the tracer.
  5244  	trace := traceAcquire()
  5245  	var status uint32 = _Grunnable
  5246  	if parked {
  5247  		status = _Gwaiting
  5248  		newg.waitreason = waitreason
  5249  	}
  5250  	if pp.goidcache == pp.goidcacheend {
  5251  		// Sched.goidgen is the last allocated id,
  5252  		// this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
  5253  		// At startup sched.goidgen=0, so main goroutine receives goid=1.
  5254  		pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
  5255  		pp.goidcache -= _GoidCacheBatch - 1
  5256  		pp.goidcacheend = pp.goidcache + _GoidCacheBatch
  5257  	}
  5258  	newg.goid = pp.goidcache
  5259  	casgstatus(newg, _Gdead, status)
  5260  	pp.goidcache++
  5261  	newg.trace.reset()
  5262  	if trace.ok() {
  5263  		trace.GoCreate(newg, newg.startpc, parked)
  5264  		traceRelease(trace)
  5265  	}
  5266  
  5267  	// Set up race context.
  5268  	if raceenabled {
  5269  		newg.racectx = racegostart(callerpc)
  5270  		newg.raceignore = 0
  5271  		if newg.labels != nil {
  5272  			// See note in proflabel.go on labelSync's role in synchronizing
  5273  			// with the reads in the signal handler.
  5274  			racereleasemergeg(newg, unsafe.Pointer(&labelSync))
  5275  		}
  5276  	}
  5277  	pp.goroutinesCreated++
  5278  	releasem(mp)
  5279  
  5280  	return newg
  5281  }
  5282  
  5283  // saveAncestors copies previous ancestors of the given caller g and
  5284  // includes info for the current caller into a new set of tracebacks for
  5285  // a g being created.
  5286  func saveAncestors(callergp *g) *[]ancestorInfo {
  5287  	// Copy all prior info, except for the root goroutine (goid 0).
  5288  	if debug.tracebackancestors <= 0 || callergp.goid == 0 {
  5289  		return nil
  5290  	}
  5291  	var callerAncestors []ancestorInfo
  5292  	if callergp.ancestors != nil {
  5293  		callerAncestors = *callergp.ancestors
  5294  	}
  5295  	n := int32(len(callerAncestors)) + 1
  5296  	if n > debug.tracebackancestors {
  5297  		n = debug.tracebackancestors
  5298  	}
  5299  	ancestors := make([]ancestorInfo, n)
  5300  	copy(ancestors[1:], callerAncestors)
  5301  
  5302  	var pcs [tracebackInnerFrames]uintptr
  5303  	npcs := gcallers(callergp, 0, pcs[:])
  5304  	ipcs := make([]uintptr, npcs)
  5305  	copy(ipcs, pcs[:])
  5306  	ancestors[0] = ancestorInfo{
  5307  		pcs:  ipcs,
  5308  		goid: callergp.goid,
  5309  		gopc: callergp.gopc,
  5310  	}
  5311  
  5312  	ancestorsp := new([]ancestorInfo)
  5313  	*ancestorsp = ancestors
  5314  	return ancestorsp
  5315  }
  5316  
  5317  // Put on gfree list.
  5318  // If local list is too long, transfer a batch to the global list.
  5319  func gfput(pp *p, gp *g) {
  5320  	if readgstatus(gp) != _Gdead {
  5321  		throw("gfput: bad status (not Gdead)")
  5322  	}
  5323  
  5324  	stksize := gp.stack.hi - gp.stack.lo
  5325  
  5326  	if stksize != uintptr(startingStackSize) {
  5327  		// non-standard stack size - free it.
  5328  		stackfree(gp.stack)
  5329  		gp.stack.lo = 0
  5330  		gp.stack.hi = 0
  5331  		gp.stackguard0 = 0
  5332  		if valgrindenabled {
  5333  			valgrindDeregisterStack(gp.valgrindStackID)
  5334  			gp.valgrindStackID = 0
  5335  		}
  5336  	}
  5337  
  5338  	pp.gFree.push(gp)
  5339  	if pp.gFree.size >= 64 {
  5340  		var (
  5341  			stackQ   gQueue
  5342  			noStackQ gQueue
  5343  		)
  5344  		for pp.gFree.size >= 32 {
  5345  			gp := pp.gFree.pop()
  5346  			if gp.stack.lo == 0 {
  5347  				noStackQ.push(gp)
  5348  			} else {
  5349  				stackQ.push(gp)
  5350  			}
  5351  		}
  5352  		lock(&sched.gFree.lock)
  5353  		sched.gFree.noStack.pushAll(noStackQ)
  5354  		sched.gFree.stack.pushAll(stackQ)
  5355  		unlock(&sched.gFree.lock)
  5356  	}
  5357  }
  5358  
  5359  // Get from gfree list.
  5360  // If local list is empty, grab a batch from global list.
  5361  func gfget(pp *p) *g {
  5362  retry:
  5363  	if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
  5364  		lock(&sched.gFree.lock)
  5365  		// Move a batch of free Gs to the P.
  5366  		for pp.gFree.size < 32 {
  5367  			// Prefer Gs with stacks.
  5368  			gp := sched.gFree.stack.pop()
  5369  			if gp == nil {
  5370  				gp = sched.gFree.noStack.pop()
  5371  				if gp == nil {
  5372  					break
  5373  				}
  5374  			}
  5375  			pp.gFree.push(gp)
  5376  		}
  5377  		unlock(&sched.gFree.lock)
  5378  		goto retry
  5379  	}
  5380  	gp := pp.gFree.pop()
  5381  	if gp == nil {
  5382  		return nil
  5383  	}
  5384  	if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
  5385  		// Deallocate old stack. We kept it in gfput because it was the
  5386  		// right size when the goroutine was put on the free list, but
  5387  		// the right size has changed since then.
  5388  		systemstack(func() {
  5389  			stackfree(gp.stack)
  5390  			gp.stack.lo = 0
  5391  			gp.stack.hi = 0
  5392  			gp.stackguard0 = 0
  5393  			if valgrindenabled {
  5394  				valgrindDeregisterStack(gp.valgrindStackID)
  5395  				gp.valgrindStackID = 0
  5396  			}
  5397  		})
  5398  	}
  5399  	if gp.stack.lo == 0 {
  5400  		// Stack was deallocated in gfput or just above. Allocate a new one.
  5401  		systemstack(func() {
  5402  			gp.stack = stackalloc(startingStackSize)
  5403  			if valgrindenabled {
  5404  				gp.valgrindStackID = valgrindRegisterStack(unsafe.Pointer(gp.stack.lo), unsafe.Pointer(gp.stack.hi))
  5405  			}
  5406  		})
  5407  		gp.stackguard0 = gp.stack.lo + stackGuard
  5408  	} else {
  5409  		if raceenabled {
  5410  			racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5411  		}
  5412  		if msanenabled {
  5413  			msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5414  		}
  5415  		if asanenabled {
  5416  			asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5417  		}
  5418  	}
  5419  	return gp
  5420  }
  5421  
  5422  // Purge all cached G's from gfree list to the global list.
  5423  func gfpurge(pp *p) {
  5424  	var (
  5425  		stackQ   gQueue
  5426  		noStackQ gQueue
  5427  	)
  5428  	for !pp.gFree.empty() {
  5429  		gp := pp.gFree.pop()
  5430  		if gp.stack.lo == 0 {
  5431  			noStackQ.push(gp)
  5432  		} else {
  5433  			stackQ.push(gp)
  5434  		}
  5435  	}
  5436  	lock(&sched.gFree.lock)
  5437  	sched.gFree.noStack.pushAll(noStackQ)
  5438  	sched.gFree.stack.pushAll(stackQ)
  5439  	unlock(&sched.gFree.lock)
  5440  }
  5441  
  5442  // Breakpoint executes a breakpoint trap.
  5443  func Breakpoint() {
  5444  	breakpoint()
  5445  }
  5446  
  5447  // dolockOSThread is called by LockOSThread and lockOSThread below
  5448  // after they modify m.locked. Do not allow preemption during this call,
  5449  // or else the m might be different in this function than in the caller.
  5450  //
  5451  //go:nosplit
  5452  func dolockOSThread() {
  5453  	if GOARCH == "wasm" {
  5454  		return // no threads on wasm yet
  5455  	}
  5456  	gp := getg()
  5457  	gp.m.lockedg.set(gp)
  5458  	gp.lockedm.set(gp.m)
  5459  }
  5460  
  5461  // LockOSThread wires the calling goroutine to its current operating system thread.
  5462  // The calling goroutine will always execute in that thread,
  5463  // and no other goroutine will execute in it,
  5464  // until the calling goroutine has made as many calls to
  5465  // [UnlockOSThread] as to LockOSThread.
  5466  // If the calling goroutine exits without unlocking the thread,
  5467  // the thread will be terminated.
  5468  //
  5469  // All init functions are run on the startup thread. Calling LockOSThread
  5470  // from an init function will cause the main function to be invoked on
  5471  // that thread.
  5472  //
  5473  // A goroutine should call LockOSThread before calling OS services or
  5474  // non-Go library functions that depend on per-thread state.
  5475  //
  5476  //go:nosplit
  5477  func LockOSThread() {
  5478  	if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
  5479  		// If we need to start a new thread from the locked
  5480  		// thread, we need the template thread. Start it now
  5481  		// while we're in a known-good state.
  5482  		startTemplateThread()
  5483  	}
  5484  	gp := getg()
  5485  	gp.m.lockedExt++
  5486  	if gp.m.lockedExt == 0 {
  5487  		gp.m.lockedExt--
  5488  		panic("LockOSThread nesting overflow")
  5489  	}
  5490  	dolockOSThread()
  5491  }
  5492  
  5493  //go:nosplit
  5494  func lockOSThread() {
  5495  	getg().m.lockedInt++
  5496  	dolockOSThread()
  5497  }
  5498  
  5499  // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
  5500  // after they update m->locked. Do not allow preemption during this call,
  5501  // or else the m might be in different in this function than in the caller.
  5502  //
  5503  //go:nosplit
  5504  func dounlockOSThread() {
  5505  	if GOARCH == "wasm" {
  5506  		return // no threads on wasm yet
  5507  	}
  5508  	gp := getg()
  5509  	if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
  5510  		return
  5511  	}
  5512  	gp.m.lockedg = 0
  5513  	gp.lockedm = 0
  5514  }
  5515  
  5516  // UnlockOSThread undoes an earlier call to LockOSThread.
  5517  // If this drops the number of active LockOSThread calls on the
  5518  // calling goroutine to zero, it unwires the calling goroutine from
  5519  // its fixed operating system thread.
  5520  // If there are no active LockOSThread calls, this is a no-op.
  5521  //
  5522  // Before calling UnlockOSThread, the caller must ensure that the OS
  5523  // thread is suitable for running other goroutines. If the caller made
  5524  // any permanent changes to the state of the thread that would affect
  5525  // other goroutines, it should not call this function and thus leave
  5526  // the goroutine locked to the OS thread until the goroutine (and
  5527  // hence the thread) exits.
  5528  //
  5529  //go:nosplit
  5530  func UnlockOSThread() {
  5531  	gp := getg()
  5532  	if gp.m.lockedExt == 0 {
  5533  		return
  5534  	}
  5535  	gp.m.lockedExt--
  5536  	dounlockOSThread()
  5537  }
  5538  
  5539  //go:nosplit
  5540  func unlockOSThread() {
  5541  	gp := getg()
  5542  	if gp.m.lockedInt == 0 {
  5543  		systemstack(badunlockosthread)
  5544  	}
  5545  	gp.m.lockedInt--
  5546  	dounlockOSThread()
  5547  }
  5548  
  5549  func badunlockosthread() {
  5550  	throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
  5551  }
  5552  
  5553  func gcount(includeSys bool) int32 {
  5554  	n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.stack.size - sched.gFree.noStack.size
  5555  	if !includeSys {
  5556  		n -= sched.ngsys.Load()
  5557  	}
  5558  	for _, pp := range allp {
  5559  		n -= pp.gFree.size
  5560  	}
  5561  
  5562  	// All these variables can be changed concurrently, so the result can be inconsistent.
  5563  	// But at least the current goroutine is running.
  5564  	if n < 1 {
  5565  		n = 1
  5566  	}
  5567  	return n
  5568  }
  5569  
  5570  // goroutineleakcount returns the number of leaked goroutines last reported by
  5571  // the runtime.
  5572  //
  5573  //go:linkname goroutineleakcount runtime/pprof.runtime_goroutineleakcount
  5574  func goroutineleakcount() int {
  5575  	return work.goroutineLeak.count
  5576  }
  5577  
  5578  func mcount() int32 {
  5579  	return int32(sched.mnext - sched.nmfreed)
  5580  }
  5581  
  5582  var prof struct {
  5583  	signalLock atomic.Uint32
  5584  
  5585  	// Must hold signalLock to write. Reads may be lock-free, but
  5586  	// signalLock should be taken to synchronize with changes.
  5587  	hz atomic.Int32
  5588  }
  5589  
  5590  func _System()                    { _System() }
  5591  func _ExternalCode()              { _ExternalCode() }
  5592  func _LostExternalCode()          { _LostExternalCode() }
  5593  func _GC()                        { _GC() }
  5594  func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
  5595  func _LostContendedRuntimeLock()  { _LostContendedRuntimeLock() }
  5596  func _VDSO()                      { _VDSO() }
  5597  
  5598  // Called if we receive a SIGPROF signal.
  5599  // Called by the signal handler, may run during STW.
  5600  //
  5601  //go:nowritebarrierrec
  5602  func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
  5603  	if prof.hz.Load() == 0 {
  5604  		return
  5605  	}
  5606  
  5607  	// If mp.profilehz is 0, then profiling is not enabled for this thread.
  5608  	// We must check this to avoid a deadlock between setcpuprofilerate
  5609  	// and the call to cpuprof.add, below.
  5610  	if mp != nil && mp.profilehz == 0 {
  5611  		return
  5612  	}
  5613  
  5614  	// On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
  5615  	// internal/runtime/atomic. If SIGPROF arrives while the program is inside
  5616  	// the critical section, it creates a deadlock (when writing the sample).
  5617  	// As a workaround, create a counter of SIGPROFs while in critical section
  5618  	// to store the count, and pass it to sigprof.add() later when SIGPROF is
  5619  	// received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
  5620  	if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
  5621  		if f := findfunc(pc); f.valid() {
  5622  			if stringslite.HasPrefix(funcname(f), "internal/runtime/atomic") {
  5623  				cpuprof.lostAtomic++
  5624  				return
  5625  			}
  5626  		}
  5627  		if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
  5628  			// internal/runtime/atomic functions call into kernel
  5629  			// helpers on arm < 7. See
  5630  			// internal/runtime/atomic/sys_linux_arm.s.
  5631  			cpuprof.lostAtomic++
  5632  			return
  5633  		}
  5634  	}
  5635  
  5636  	// Profiling runs concurrently with GC, so it must not allocate.
  5637  	// Set a trap in case the code does allocate.
  5638  	// Note that on windows, one thread takes profiles of all the
  5639  	// other threads, so mp is usually not getg().m.
  5640  	// In fact mp may not even be stopped.
  5641  	// See golang.org/issue/17165.
  5642  	getg().m.mallocing++
  5643  
  5644  	var u unwinder
  5645  	var stk [maxCPUProfStack]uintptr
  5646  	n := 0
  5647  	if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
  5648  		cgoOff := 0
  5649  		// Check cgoCallersUse to make sure that we are not
  5650  		// interrupting other code that is fiddling with
  5651  		// cgoCallers.  We are running in a signal handler
  5652  		// with all signals blocked, so we don't have to worry
  5653  		// about any other code interrupting us.
  5654  		if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
  5655  			for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
  5656  				cgoOff++
  5657  			}
  5658  			n += copy(stk[:], mp.cgoCallers[:cgoOff])
  5659  			mp.cgoCallers[0] = 0
  5660  		}
  5661  
  5662  		// Collect Go stack that leads to the cgo call.
  5663  		u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
  5664  	} else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
  5665  		// Libcall, i.e. runtime syscall on windows.
  5666  		// Collect Go stack that leads to the call.
  5667  		u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
  5668  	} else if mp != nil && mp.vdsoSP != 0 {
  5669  		// VDSO call, e.g. nanotime1 on Linux.
  5670  		// Collect Go stack that leads to the call.
  5671  		u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
  5672  	} else {
  5673  		u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
  5674  	}
  5675  	n += tracebackPCs(&u, 0, stk[n:])
  5676  
  5677  	if n <= 0 {
  5678  		// Normal traceback is impossible or has failed.
  5679  		// Account it against abstract "System" or "GC".
  5680  		n = 2
  5681  		if inVDSOPage(pc) {
  5682  			pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
  5683  		} else if pc > firstmoduledata.etext {
  5684  			// "ExternalCode" is better than "etext".
  5685  			pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
  5686  		}
  5687  		stk[0] = pc
  5688  		if mp.preemptoff != "" {
  5689  			stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
  5690  		} else {
  5691  			stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
  5692  		}
  5693  	}
  5694  
  5695  	if prof.hz.Load() != 0 {
  5696  		// Note: it can happen on Windows that we interrupted a system thread
  5697  		// with no g, so gp could nil. The other nil checks are done out of
  5698  		// caution, but not expected to be nil in practice.
  5699  		var tagPtr *unsafe.Pointer
  5700  		if gp != nil && gp.m != nil && gp.m.curg != nil {
  5701  			tagPtr = &gp.m.curg.labels
  5702  		}
  5703  		cpuprof.add(tagPtr, stk[:n])
  5704  
  5705  		gprof := gp
  5706  		var mp *m
  5707  		var pp *p
  5708  		if gp != nil && gp.m != nil {
  5709  			if gp.m.curg != nil {
  5710  				gprof = gp.m.curg
  5711  			}
  5712  			mp = gp.m
  5713  			pp = gp.m.p.ptr()
  5714  		}
  5715  		traceCPUSample(gprof, mp, pp, stk[:n])
  5716  	}
  5717  	getg().m.mallocing--
  5718  }
  5719  
  5720  // setcpuprofilerate sets the CPU profiling rate to hz times per second.
  5721  // If hz <= 0, setcpuprofilerate turns off CPU profiling.
  5722  func setcpuprofilerate(hz int32) {
  5723  	// Force sane arguments.
  5724  	if hz < 0 {
  5725  		hz = 0
  5726  	}
  5727  
  5728  	// Disable preemption, otherwise we can be rescheduled to another thread
  5729  	// that has profiling enabled.
  5730  	gp := getg()
  5731  	gp.m.locks++
  5732  
  5733  	// Stop profiler on this thread so that it is safe to lock prof.
  5734  	// if a profiling signal came in while we had prof locked,
  5735  	// it would deadlock.
  5736  	setThreadCPUProfiler(0)
  5737  
  5738  	for !prof.signalLock.CompareAndSwap(0, 1) {
  5739  		osyield()
  5740  	}
  5741  	if prof.hz.Load() != hz {
  5742  		setProcessCPUProfiler(hz)
  5743  		prof.hz.Store(hz)
  5744  	}
  5745  	prof.signalLock.Store(0)
  5746  
  5747  	lock(&sched.lock)
  5748  	sched.profilehz = hz
  5749  	unlock(&sched.lock)
  5750  
  5751  	if hz != 0 {
  5752  		setThreadCPUProfiler(hz)
  5753  	}
  5754  
  5755  	gp.m.locks--
  5756  }
  5757  
  5758  // init initializes pp, which may be a freshly allocated p or a
  5759  // previously destroyed p, and transitions it to status _Pgcstop.
  5760  func (pp *p) init(id int32) {
  5761  	pp.id = id
  5762  	pp.gcw.id = id
  5763  	pp.status = _Pgcstop
  5764  	pp.sudogcache = pp.sudogbuf[:0]
  5765  	pp.deferpool = pp.deferpoolbuf[:0]
  5766  	pp.wbBuf.reset()
  5767  	if pp.mcache == nil {
  5768  		if id == 0 {
  5769  			if mcache0 == nil {
  5770  				throw("missing mcache?")
  5771  			}
  5772  			// Use the bootstrap mcache0. Only one P will get
  5773  			// mcache0: the one with ID 0.
  5774  			pp.mcache = mcache0
  5775  		} else {
  5776  			pp.mcache = allocmcache()
  5777  		}
  5778  	}
  5779  	if raceenabled && pp.raceprocctx == 0 {
  5780  		if id == 0 {
  5781  			pp.raceprocctx = raceprocctx0
  5782  			raceprocctx0 = 0 // bootstrap
  5783  		} else {
  5784  			pp.raceprocctx = raceproccreate()
  5785  		}
  5786  	}
  5787  	lockInit(&pp.timers.mu, lockRankTimers)
  5788  
  5789  	// This P may get timers when it starts running. Set the mask here
  5790  	// since the P may not go through pidleget (notably P 0 on startup).
  5791  	timerpMask.set(id)
  5792  	// Similarly, we may not go through pidleget before this P starts
  5793  	// running if it is P 0 on startup.
  5794  	idlepMask.clear(id)
  5795  }
  5796  
  5797  // destroy releases all of the resources associated with pp and
  5798  // transitions it to status _Pdead.
  5799  //
  5800  // sched.lock must be held and the world must be stopped.
  5801  func (pp *p) destroy() {
  5802  	assertLockHeld(&sched.lock)
  5803  	assertWorldStopped()
  5804  
  5805  	// Move all runnable goroutines to the global queue
  5806  	for pp.runqhead != pp.runqtail {
  5807  		// Pop from tail of local queue
  5808  		pp.runqtail--
  5809  		gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
  5810  		// Push onto head of global queue
  5811  		globrunqputhead(gp)
  5812  	}
  5813  	if pp.runnext != 0 {
  5814  		globrunqputhead(pp.runnext.ptr())
  5815  		pp.runnext = 0
  5816  	}
  5817  
  5818  	// Move all timers to the local P.
  5819  	getg().m.p.ptr().timers.take(&pp.timers)
  5820  
  5821  	// No need to flush p's write barrier buffer or span queue, as Ps
  5822  	// cannot be destroyed during the mark phase.
  5823  	if phase := gcphase; phase != _GCoff {
  5824  		println("runtime: p id", pp.id, "destroyed during GC phase", phase)
  5825  		throw("P destroyed while GC is running")
  5826  	}
  5827  	// We should free the queues though.
  5828  	pp.gcw.spanq.destroy()
  5829  
  5830  	clear(pp.sudogbuf[:])
  5831  	pp.sudogcache = pp.sudogbuf[:0]
  5832  	pp.pinnerCache = nil
  5833  	clear(pp.deferpoolbuf[:])
  5834  	pp.deferpool = pp.deferpoolbuf[:0]
  5835  	systemstack(func() {
  5836  		for i := 0; i < pp.mspancache.len; i++ {
  5837  			// Safe to call since the world is stopped.
  5838  			mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
  5839  		}
  5840  		pp.mspancache.len = 0
  5841  		lock(&mheap_.lock)
  5842  		pp.pcache.flush(&mheap_.pages)
  5843  		unlock(&mheap_.lock)
  5844  	})
  5845  	freemcache(pp.mcache)
  5846  	pp.mcache = nil
  5847  	gfpurge(pp)
  5848  	if raceenabled {
  5849  		if pp.timers.raceCtx != 0 {
  5850  			// The race detector code uses a callback to fetch
  5851  			// the proc context, so arrange for that callback
  5852  			// to see the right thing.
  5853  			// This hack only works because we are the only
  5854  			// thread running.
  5855  			mp := getg().m
  5856  			phold := mp.p.ptr()
  5857  			mp.p.set(pp)
  5858  
  5859  			racectxend(pp.timers.raceCtx)
  5860  			pp.timers.raceCtx = 0
  5861  
  5862  			mp.p.set(phold)
  5863  		}
  5864  		raceprocdestroy(pp.raceprocctx)
  5865  		pp.raceprocctx = 0
  5866  	}
  5867  	pp.gcAssistTime = 0
  5868  	gcCleanups.queued += pp.cleanupsQueued
  5869  	pp.cleanupsQueued = 0
  5870  	sched.goroutinesCreated.Add(int64(pp.goroutinesCreated))
  5871  	pp.goroutinesCreated = 0
  5872  	pp.xRegs.free()
  5873  	pp.status = _Pdead
  5874  }
  5875  
  5876  // Change number of processors.
  5877  //
  5878  // sched.lock must be held, and the world must be stopped.
  5879  //
  5880  // gcworkbufs must not be being modified by either the GC or the write barrier
  5881  // code, so the GC must not be running if the number of Ps actually changes.
  5882  //
  5883  // Returns list of Ps with local work, they need to be scheduled by the caller.
  5884  func procresize(nprocs int32) *p {
  5885  	assertLockHeld(&sched.lock)
  5886  	assertWorldStopped()
  5887  
  5888  	old := gomaxprocs
  5889  	if old < 0 || nprocs <= 0 {
  5890  		throw("procresize: invalid arg")
  5891  	}
  5892  	trace := traceAcquire()
  5893  	if trace.ok() {
  5894  		trace.Gomaxprocs(nprocs)
  5895  		traceRelease(trace)
  5896  	}
  5897  
  5898  	// update statistics
  5899  	now := nanotime()
  5900  	if sched.procresizetime != 0 {
  5901  		sched.totaltime += int64(old) * (now - sched.procresizetime)
  5902  	}
  5903  	sched.procresizetime = now
  5904  
  5905  	// Grow allp if necessary.
  5906  	if nprocs > int32(len(allp)) {
  5907  		// Synchronize with retake, which could be running
  5908  		// concurrently since it doesn't run on a P.
  5909  		lock(&allpLock)
  5910  		if nprocs <= int32(cap(allp)) {
  5911  			allp = allp[:nprocs]
  5912  		} else {
  5913  			nallp := make([]*p, nprocs)
  5914  			// Copy everything up to allp's cap so we
  5915  			// never lose old allocated Ps.
  5916  			copy(nallp, allp[:cap(allp)])
  5917  			allp = nallp
  5918  		}
  5919  
  5920  		idlepMask = idlepMask.resize(nprocs)
  5921  		timerpMask = timerpMask.resize(nprocs)
  5922  		work.spanqMask = work.spanqMask.resize(nprocs)
  5923  		unlock(&allpLock)
  5924  	}
  5925  
  5926  	// initialize new P's
  5927  	for i := old; i < nprocs; i++ {
  5928  		pp := allp[i]
  5929  		if pp == nil {
  5930  			pp = new(p)
  5931  		}
  5932  		pp.init(i)
  5933  		atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
  5934  	}
  5935  
  5936  	gp := getg()
  5937  	if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
  5938  		// continue to use the current P
  5939  		gp.m.p.ptr().status = _Prunning
  5940  		gp.m.p.ptr().mcache.prepareForSweep()
  5941  	} else {
  5942  		// release the current P and acquire allp[0].
  5943  		//
  5944  		// We must do this before destroying our current P
  5945  		// because p.destroy itself has write barriers, so we
  5946  		// need to do that from a valid P.
  5947  		if gp.m.p != 0 {
  5948  			trace := traceAcquire()
  5949  			if trace.ok() {
  5950  				// Pretend that we were descheduled
  5951  				// and then scheduled again to keep
  5952  				// the trace consistent.
  5953  				trace.GoSched()
  5954  				trace.ProcStop(gp.m.p.ptr())
  5955  				traceRelease(trace)
  5956  			}
  5957  			gp.m.p.ptr().m = 0
  5958  		}
  5959  		gp.m.p = 0
  5960  		pp := allp[0]
  5961  		pp.m = 0
  5962  		pp.status = _Pidle
  5963  		acquirep(pp)
  5964  		trace := traceAcquire()
  5965  		if trace.ok() {
  5966  			trace.GoStart()
  5967  			traceRelease(trace)
  5968  		}
  5969  	}
  5970  
  5971  	// g.m.p is now set, so we no longer need mcache0 for bootstrapping.
  5972  	mcache0 = nil
  5973  
  5974  	// release resources from unused P's
  5975  	for i := nprocs; i < old; i++ {
  5976  		pp := allp[i]
  5977  		pp.destroy()
  5978  		// can't free P itself because it can be referenced by an M in syscall
  5979  	}
  5980  
  5981  	// Trim allp.
  5982  	if int32(len(allp)) != nprocs {
  5983  		lock(&allpLock)
  5984  		allp = allp[:nprocs]
  5985  		idlepMask = idlepMask.resize(nprocs)
  5986  		timerpMask = timerpMask.resize(nprocs)
  5987  		work.spanqMask = work.spanqMask.resize(nprocs)
  5988  		unlock(&allpLock)
  5989  	}
  5990  
  5991  	var runnablePs *p
  5992  	for i := nprocs - 1; i >= 0; i-- {
  5993  		pp := allp[i]
  5994  		if gp.m.p.ptr() == pp {
  5995  			continue
  5996  		}
  5997  		pp.status = _Pidle
  5998  		if runqempty(pp) {
  5999  			pidleput(pp, now)
  6000  		} else {
  6001  			pp.m.set(mget())
  6002  			pp.link.set(runnablePs)
  6003  			runnablePs = pp
  6004  		}
  6005  	}
  6006  	stealOrder.reset(uint32(nprocs))
  6007  	var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
  6008  	atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
  6009  	if old != nprocs {
  6010  		// Notify the limiter that the amount of procs has changed.
  6011  		gcCPULimiter.resetCapacity(now, nprocs)
  6012  	}
  6013  	return runnablePs
  6014  }
  6015  
  6016  // Associate p and the current m.
  6017  //
  6018  // This function is allowed to have write barriers even if the caller
  6019  // isn't because it immediately acquires pp.
  6020  //
  6021  //go:yeswritebarrierrec
  6022  func acquirep(pp *p) {
  6023  	// Do the part that isn't allowed to have write barriers.
  6024  	wirep(pp)
  6025  
  6026  	// Have p; write barriers now allowed.
  6027  
  6028  	// Perform deferred mcache flush before this P can allocate
  6029  	// from a potentially stale mcache.
  6030  	pp.mcache.prepareForSweep()
  6031  
  6032  	trace := traceAcquire()
  6033  	if trace.ok() {
  6034  		trace.ProcStart()
  6035  		traceRelease(trace)
  6036  	}
  6037  }
  6038  
  6039  // wirep is the first step of acquirep, which actually associates the
  6040  // current M to pp. This is broken out so we can disallow write
  6041  // barriers for this part, since we don't yet have a P.
  6042  //
  6043  //go:nowritebarrierrec
  6044  //go:nosplit
  6045  func wirep(pp *p) {
  6046  	gp := getg()
  6047  
  6048  	if gp.m.p != 0 {
  6049  		// Call on the systemstack to avoid a nosplit overflow build failure
  6050  		// on some platforms when built with -N -l. See #64113.
  6051  		systemstack(func() {
  6052  			throw("wirep: already in go")
  6053  		})
  6054  	}
  6055  	if pp.m != 0 || pp.status != _Pidle {
  6056  		// Call on the systemstack to avoid a nosplit overflow build failure
  6057  		// on some platforms when built with -N -l. See #64113.
  6058  		systemstack(func() {
  6059  			id := int64(0)
  6060  			if pp.m != 0 {
  6061  				id = pp.m.ptr().id
  6062  			}
  6063  			print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
  6064  			throw("wirep: invalid p state")
  6065  		})
  6066  	}
  6067  	gp.m.p.set(pp)
  6068  	pp.m.set(gp.m)
  6069  	pp.status = _Prunning
  6070  }
  6071  
  6072  // Disassociate p and the current m.
  6073  func releasep() *p {
  6074  	trace := traceAcquire()
  6075  	if trace.ok() {
  6076  		trace.ProcStop(getg().m.p.ptr())
  6077  		traceRelease(trace)
  6078  	}
  6079  	return releasepNoTrace()
  6080  }
  6081  
  6082  // Disassociate p and the current m without tracing an event.
  6083  func releasepNoTrace() *p {
  6084  	gp := getg()
  6085  
  6086  	if gp.m.p == 0 {
  6087  		throw("releasep: invalid arg")
  6088  	}
  6089  	pp := gp.m.p.ptr()
  6090  	if pp.m.ptr() != gp.m || pp.status != _Prunning {
  6091  		print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
  6092  		throw("releasep: invalid p state")
  6093  	}
  6094  	gp.m.p = 0
  6095  	pp.m = 0
  6096  	pp.status = _Pidle
  6097  	return pp
  6098  }
  6099  
  6100  func incidlelocked(v int32) {
  6101  	lock(&sched.lock)
  6102  	sched.nmidlelocked += v
  6103  	if v > 0 {
  6104  		checkdead()
  6105  	}
  6106  	unlock(&sched.lock)
  6107  }
  6108  
  6109  // Check for deadlock situation.
  6110  // The check is based on number of running M's, if 0 -> deadlock.
  6111  // sched.lock must be held.
  6112  func checkdead() {
  6113  	assertLockHeld(&sched.lock)
  6114  
  6115  	// For -buildmode=c-shared or -buildmode=c-archive it's OK if
  6116  	// there are no running goroutines. The calling program is
  6117  	// assumed to be running.
  6118  	// One exception is Wasm, which is single-threaded. If we are
  6119  	// in Go and all goroutines are blocked, it deadlocks.
  6120  	if (islibrary || isarchive) && GOARCH != "wasm" {
  6121  		return
  6122  	}
  6123  
  6124  	// If we are dying because of a signal caught on an already idle thread,
  6125  	// freezetheworld will cause all running threads to block.
  6126  	// And runtime will essentially enter into deadlock state,
  6127  	// except that there is a thread that will call exit soon.
  6128  	if panicking.Load() > 0 {
  6129  		return
  6130  	}
  6131  
  6132  	// If we are not running under cgo, but we have an extra M then account
  6133  	// for it. (It is possible to have an extra M on Windows without cgo to
  6134  	// accommodate callbacks created by syscall.NewCallback. See issue #6751
  6135  	// for details.)
  6136  	var run0 int32
  6137  	if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
  6138  		run0 = 1
  6139  	}
  6140  
  6141  	run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
  6142  	if run > run0 {
  6143  		return
  6144  	}
  6145  	if run < 0 {
  6146  		print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
  6147  		unlock(&sched.lock)
  6148  		throw("checkdead: inconsistent counts")
  6149  	}
  6150  
  6151  	grunning := 0
  6152  	forEachG(func(gp *g) {
  6153  		if isSystemGoroutine(gp, false) {
  6154  			return
  6155  		}
  6156  		s := readgstatus(gp)
  6157  		switch s &^ _Gscan {
  6158  		case _Gwaiting,
  6159  			_Gpreempted:
  6160  			grunning++
  6161  		case _Grunnable,
  6162  			_Grunning,
  6163  			_Gsyscall:
  6164  			print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
  6165  			unlock(&sched.lock)
  6166  			throw("checkdead: runnable g")
  6167  		}
  6168  	})
  6169  	if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
  6170  		unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  6171  		fatal("no goroutines (main called runtime.Goexit) - deadlock!")
  6172  	}
  6173  
  6174  	// Maybe jump time forward for playground.
  6175  	if faketime != 0 {
  6176  		if when := timeSleepUntil(); when < maxWhen {
  6177  			faketime = when
  6178  
  6179  			// Start an M to steal the timer.
  6180  			pp, _ := pidleget(faketime)
  6181  			if pp == nil {
  6182  				// There should always be a free P since
  6183  				// nothing is running.
  6184  				unlock(&sched.lock)
  6185  				throw("checkdead: no p for timer")
  6186  			}
  6187  			mp := mget()
  6188  			if mp == nil {
  6189  				// There should always be a free M since
  6190  				// nothing is running.
  6191  				unlock(&sched.lock)
  6192  				throw("checkdead: no m for timer")
  6193  			}
  6194  			// M must be spinning to steal. We set this to be
  6195  			// explicit, but since this is the only M it would
  6196  			// become spinning on its own anyways.
  6197  			sched.nmspinning.Add(1)
  6198  			mp.spinning = true
  6199  			mp.nextp.set(pp)
  6200  			notewakeup(&mp.park)
  6201  			return
  6202  		}
  6203  	}
  6204  
  6205  	// There are no goroutines running, so we can look at the P's.
  6206  	for _, pp := range allp {
  6207  		if len(pp.timers.heap) > 0 {
  6208  			return
  6209  		}
  6210  	}
  6211  
  6212  	unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  6213  	fatal("all goroutines are asleep - deadlock!")
  6214  }
  6215  
  6216  // forcegcperiod is the maximum time in nanoseconds between garbage
  6217  // collections. If we go this long without a garbage collection, one
  6218  // is forced to run.
  6219  //
  6220  // This is a variable for testing purposes. It normally doesn't change.
  6221  var forcegcperiod int64 = 2 * 60 * 1e9
  6222  
  6223  // haveSysmon indicates whether there is sysmon thread support.
  6224  //
  6225  // No threads on wasm yet, so no sysmon.
  6226  const haveSysmon = GOARCH != "wasm"
  6227  
  6228  // Always runs without a P, so write barriers are not allowed.
  6229  //
  6230  //go:nowritebarrierrec
  6231  func sysmon() {
  6232  	lock(&sched.lock)
  6233  	sched.nmsys++
  6234  	checkdead()
  6235  	unlock(&sched.lock)
  6236  
  6237  	lastgomaxprocs := int64(0)
  6238  	lasttrace := int64(0)
  6239  	idle := 0 // how many cycles in succession we had not wokeup somebody
  6240  	delay := uint32(0)
  6241  
  6242  	for {
  6243  		if idle == 0 { // start with 20us sleep...
  6244  			delay = 20
  6245  		} else if idle > 50 { // start doubling the sleep after 1ms...
  6246  			delay *= 2
  6247  		}
  6248  		if delay > 10*1000 { // up to 10ms
  6249  			delay = 10 * 1000
  6250  		}
  6251  		usleep(delay)
  6252  
  6253  		// sysmon should not enter deep sleep if schedtrace is enabled so that
  6254  		// it can print that information at the right time.
  6255  		//
  6256  		// It should also not enter deep sleep if there are any active P's so
  6257  		// that it can retake P's from syscalls, preempt long running G's, and
  6258  		// poll the network if all P's are busy for long stretches.
  6259  		//
  6260  		// It should wakeup from deep sleep if any P's become active either due
  6261  		// to exiting a syscall or waking up due to a timer expiring so that it
  6262  		// can resume performing those duties. If it wakes from a syscall it
  6263  		// resets idle and delay as a bet that since it had retaken a P from a
  6264  		// syscall before, it may need to do it again shortly after the
  6265  		// application starts work again. It does not reset idle when waking
  6266  		// from a timer to avoid adding system load to applications that spend
  6267  		// most of their time sleeping.
  6268  		now := nanotime()
  6269  		if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
  6270  			lock(&sched.lock)
  6271  			if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
  6272  				syscallWake := false
  6273  				next := timeSleepUntil()
  6274  				if next > now {
  6275  					sched.sysmonwait.Store(true)
  6276  					unlock(&sched.lock)
  6277  					// Make wake-up period small enough
  6278  					// for the sampling to be correct.
  6279  					sleep := forcegcperiod / 2
  6280  					if next-now < sleep {
  6281  						sleep = next - now
  6282  					}
  6283  					shouldRelax := sleep >= osRelaxMinNS
  6284  					if shouldRelax {
  6285  						osRelax(true)
  6286  					}
  6287  					syscallWake = notetsleep(&sched.sysmonnote, sleep)
  6288  					if shouldRelax {
  6289  						osRelax(false)
  6290  					}
  6291  					lock(&sched.lock)
  6292  					sched.sysmonwait.Store(false)
  6293  					noteclear(&sched.sysmonnote)
  6294  				}
  6295  				if syscallWake {
  6296  					idle = 0
  6297  					delay = 20
  6298  				}
  6299  			}
  6300  			unlock(&sched.lock)
  6301  		}
  6302  
  6303  		lock(&sched.sysmonlock)
  6304  		// Update now in case we blocked on sysmonnote or spent a long time
  6305  		// blocked on schedlock or sysmonlock above.
  6306  		now = nanotime()
  6307  
  6308  		// trigger libc interceptors if needed
  6309  		if *cgo_yield != nil {
  6310  			asmcgocall(*cgo_yield, nil)
  6311  		}
  6312  		// poll network if not polled for more than 10ms
  6313  		lastpoll := sched.lastpoll.Load()
  6314  		if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
  6315  			sched.lastpoll.CompareAndSwap(lastpoll, now)
  6316  			list, delta := netpoll(0) // non-blocking - returns list of goroutines
  6317  			if !list.empty() {
  6318  				// Need to decrement number of idle locked M's
  6319  				// (pretending that one more is running) before injectglist.
  6320  				// Otherwise it can lead to the following situation:
  6321  				// injectglist grabs all P's but before it starts M's to run the P's,
  6322  				// another M returns from syscall, finishes running its G,
  6323  				// observes that there is no work to do and no other running M's
  6324  				// and reports deadlock.
  6325  				incidlelocked(-1)
  6326  				injectglist(&list)
  6327  				incidlelocked(1)
  6328  				netpollAdjustWaiters(delta)
  6329  			}
  6330  		}
  6331  		// Check if we need to update GOMAXPROCS at most once per second.
  6332  		if debug.updatemaxprocs != 0 && lastgomaxprocs+1e9 <= now {
  6333  			sysmonUpdateGOMAXPROCS()
  6334  			lastgomaxprocs = now
  6335  		}
  6336  		if scavenger.sysmonWake.Load() != 0 {
  6337  			// Kick the scavenger awake if someone requested it.
  6338  			scavenger.wake()
  6339  		}
  6340  		// retake P's blocked in syscalls
  6341  		// and preempt long running G's
  6342  		if retake(now) != 0 {
  6343  			idle = 0
  6344  		} else {
  6345  			idle++
  6346  		}
  6347  		// check if we need to force a GC
  6348  		if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
  6349  			lock(&forcegc.lock)
  6350  			forcegc.idle.Store(false)
  6351  			var list gList
  6352  			list.push(forcegc.g)
  6353  			injectglist(&list)
  6354  			unlock(&forcegc.lock)
  6355  		}
  6356  		if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
  6357  			lasttrace = now
  6358  			schedtrace(debug.scheddetail > 0)
  6359  		}
  6360  		unlock(&sched.sysmonlock)
  6361  	}
  6362  }
  6363  
  6364  type sysmontick struct {
  6365  	schedtick   uint32
  6366  	syscalltick uint32
  6367  	schedwhen   int64
  6368  	syscallwhen int64
  6369  }
  6370  
  6371  // forcePreemptNS is the time slice given to a G before it is
  6372  // preempted.
  6373  const forcePreemptNS = 10 * 1000 * 1000 // 10ms
  6374  
  6375  func retake(now int64) uint32 {
  6376  	n := 0
  6377  	// Prevent allp slice changes. This lock will be completely
  6378  	// uncontended unless we're already stopping the world.
  6379  	lock(&allpLock)
  6380  	// We can't use a range loop over allp because we may
  6381  	// temporarily drop the allpLock. Hence, we need to re-fetch
  6382  	// allp each time around the loop.
  6383  	for i := 0; i < len(allp); i++ {
  6384  		pp := allp[i]
  6385  		if pp == nil {
  6386  			// This can happen if procresize has grown
  6387  			// allp but not yet created new Ps.
  6388  			continue
  6389  		}
  6390  		pd := &pp.sysmontick
  6391  		s := pp.status
  6392  		sysretake := false
  6393  		if s == _Prunning || s == _Psyscall {
  6394  			// Preempt G if it's running on the same schedtick for
  6395  			// too long. This could be from a single long-running
  6396  			// goroutine or a sequence of goroutines run via
  6397  			// runnext, which share a single schedtick time slice.
  6398  			t := int64(pp.schedtick)
  6399  			if int64(pd.schedtick) != t {
  6400  				pd.schedtick = uint32(t)
  6401  				pd.schedwhen = now
  6402  			} else if pd.schedwhen+forcePreemptNS <= now {
  6403  				preemptone(pp)
  6404  				// In case of syscall, preemptone() doesn't
  6405  				// work, because there is no M wired to P.
  6406  				sysretake = true
  6407  			}
  6408  		}
  6409  		if s == _Psyscall {
  6410  			// Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
  6411  			t := int64(pp.syscalltick)
  6412  			if !sysretake && int64(pd.syscalltick) != t {
  6413  				pd.syscalltick = uint32(t)
  6414  				pd.syscallwhen = now
  6415  				continue
  6416  			}
  6417  			// On the one hand we don't want to retake Ps if there is no other work to do,
  6418  			// but on the other hand we want to retake them eventually
  6419  			// because they can prevent the sysmon thread from deep sleep.
  6420  			if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
  6421  				continue
  6422  			}
  6423  			// Drop allpLock so we can take sched.lock.
  6424  			unlock(&allpLock)
  6425  			// Need to decrement number of idle locked M's
  6426  			// (pretending that one more is running) before the CAS.
  6427  			// Otherwise the M from which we retake can exit the syscall,
  6428  			// increment nmidle and report deadlock.
  6429  			incidlelocked(-1)
  6430  			trace := traceAcquire()
  6431  			if atomic.Cas(&pp.status, s, _Pidle) {
  6432  				if trace.ok() {
  6433  					trace.ProcSteal(pp, false)
  6434  					traceRelease(trace)
  6435  				}
  6436  				sched.nGsyscallNoP.Add(1)
  6437  				n++
  6438  				pp.syscalltick++
  6439  				handoffp(pp)
  6440  			} else if trace.ok() {
  6441  				traceRelease(trace)
  6442  			}
  6443  			incidlelocked(1)
  6444  			lock(&allpLock)
  6445  		}
  6446  	}
  6447  	unlock(&allpLock)
  6448  	return uint32(n)
  6449  }
  6450  
  6451  // Tell all goroutines that they have been preempted and they should stop.
  6452  // This function is purely best-effort. It can fail to inform a goroutine if a
  6453  // processor just started running it.
  6454  // No locks need to be held.
  6455  // Returns true if preemption request was issued to at least one goroutine.
  6456  func preemptall() bool {
  6457  	res := false
  6458  	for _, pp := range allp {
  6459  		if pp.status != _Prunning {
  6460  			continue
  6461  		}
  6462  		if preemptone(pp) {
  6463  			res = true
  6464  		}
  6465  	}
  6466  	return res
  6467  }
  6468  
  6469  // Tell the goroutine running on processor P to stop.
  6470  // This function is purely best-effort. It can incorrectly fail to inform the
  6471  // goroutine. It can inform the wrong goroutine. Even if it informs the
  6472  // correct goroutine, that goroutine might ignore the request if it is
  6473  // simultaneously executing newstack.
  6474  // No lock needs to be held.
  6475  // Returns true if preemption request was issued.
  6476  // The actual preemption will happen at some point in the future
  6477  // and will be indicated by the gp->status no longer being
  6478  // Grunning
  6479  func preemptone(pp *p) bool {
  6480  	mp := pp.m.ptr()
  6481  	if mp == nil || mp == getg().m {
  6482  		return false
  6483  	}
  6484  	gp := mp.curg
  6485  	if gp == nil || gp == mp.g0 {
  6486  		return false
  6487  	}
  6488  
  6489  	gp.preempt = true
  6490  
  6491  	// Every call in a goroutine checks for stack overflow by
  6492  	// comparing the current stack pointer to gp->stackguard0.
  6493  	// Setting gp->stackguard0 to StackPreempt folds
  6494  	// preemption into the normal stack overflow check.
  6495  	gp.stackguard0 = stackPreempt
  6496  
  6497  	// Request an async preemption of this P.
  6498  	if preemptMSupported && debug.asyncpreemptoff == 0 {
  6499  		pp.preempt = true
  6500  		preemptM(mp)
  6501  	}
  6502  
  6503  	return true
  6504  }
  6505  
  6506  var starttime int64
  6507  
  6508  func schedtrace(detailed bool) {
  6509  	now := nanotime()
  6510  	if starttime == 0 {
  6511  		starttime = now
  6512  	}
  6513  
  6514  	lock(&sched.lock)
  6515  	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.runq.size)
  6516  	if detailed {
  6517  		print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
  6518  	}
  6519  	// We must be careful while reading data from P's, M's and G's.
  6520  	// Even if we hold schedlock, most data can be changed concurrently.
  6521  	// E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
  6522  	for i, pp := range allp {
  6523  		h := atomic.Load(&pp.runqhead)
  6524  		t := atomic.Load(&pp.runqtail)
  6525  		if detailed {
  6526  			print("  P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
  6527  			mp := pp.m.ptr()
  6528  			if mp != nil {
  6529  				print(mp.id)
  6530  			} else {
  6531  				print("nil")
  6532  			}
  6533  			print(" runqsize=", t-h, " gfreecnt=", pp.gFree.size, " timerslen=", len(pp.timers.heap), "\n")
  6534  		} else {
  6535  			// In non-detailed mode format lengths of per-P run queues as:
  6536  			// [ len1 len2 len3 len4 ]
  6537  			print(" ")
  6538  			if i == 0 {
  6539  				print("[ ")
  6540  			}
  6541  			print(t - h)
  6542  			if i == len(allp)-1 {
  6543  				print(" ]")
  6544  			}
  6545  		}
  6546  	}
  6547  
  6548  	if !detailed {
  6549  		// Format per-P schedticks as: schedticks=[ tick1 tick2 tick3 tick4 ].
  6550  		print(" schedticks=[ ")
  6551  		for _, pp := range allp {
  6552  			print(pp.schedtick)
  6553  			print(" ")
  6554  		}
  6555  		print("]\n")
  6556  	}
  6557  
  6558  	if !detailed {
  6559  		unlock(&sched.lock)
  6560  		return
  6561  	}
  6562  
  6563  	for mp := allm; mp != nil; mp = mp.alllink {
  6564  		pp := mp.p.ptr()
  6565  		print("  M", mp.id, ": p=")
  6566  		if pp != nil {
  6567  			print(pp.id)
  6568  		} else {
  6569  			print("nil")
  6570  		}
  6571  		print(" curg=")
  6572  		if mp.curg != nil {
  6573  			print(mp.curg.goid)
  6574  		} else {
  6575  			print("nil")
  6576  		}
  6577  		print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
  6578  		if lockedg := mp.lockedg.ptr(); lockedg != nil {
  6579  			print(lockedg.goid)
  6580  		} else {
  6581  			print("nil")
  6582  		}
  6583  		print("\n")
  6584  	}
  6585  
  6586  	forEachG(func(gp *g) {
  6587  		print("  G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
  6588  		if gp.m != nil {
  6589  			print(gp.m.id)
  6590  		} else {
  6591  			print("nil")
  6592  		}
  6593  		print(" lockedm=")
  6594  		if lockedm := gp.lockedm.ptr(); lockedm != nil {
  6595  			print(lockedm.id)
  6596  		} else {
  6597  			print("nil")
  6598  		}
  6599  		print("\n")
  6600  	})
  6601  	unlock(&sched.lock)
  6602  }
  6603  
  6604  type updateMaxProcsGState struct {
  6605  	lock mutex
  6606  	g    *g
  6607  	idle atomic.Bool
  6608  
  6609  	// Readable when idle == false, writable when idle == true.
  6610  	procs int32 // new GOMAXPROCS value
  6611  }
  6612  
  6613  var (
  6614  	// GOMAXPROCS update godebug metric. Incremented if automatic
  6615  	// GOMAXPROCS updates actually change the value of GOMAXPROCS.
  6616  	updatemaxprocs = &godebugInc{name: "updatemaxprocs"}
  6617  
  6618  	// Synchronization and state between updateMaxProcsGoroutine and
  6619  	// sysmon.
  6620  	updateMaxProcsG updateMaxProcsGState
  6621  
  6622  	// Synchronization between GOMAXPROCS and sysmon.
  6623  	//
  6624  	// Setting GOMAXPROCS via a call to GOMAXPROCS disables automatic
  6625  	// GOMAXPROCS updates.
  6626  	//
  6627  	// We want to make two guarantees to callers of GOMAXPROCS. After
  6628  	// GOMAXPROCS returns:
  6629  	//
  6630  	// 1. The runtime will not make any automatic changes to GOMAXPROCS.
  6631  	//
  6632  	// 2. The runtime will not perform any of the system calls used to
  6633  	//    determine the appropriate value of GOMAXPROCS (i.e., it won't
  6634  	//    call defaultGOMAXPROCS).
  6635  	//
  6636  	// (1) is the baseline guarantee that everyone needs. The GOMAXPROCS
  6637  	// API isn't useful to anyone if automatic updates may occur after it
  6638  	// returns. This is easily achieved by double-checking the state under
  6639  	// STW before committing an automatic GOMAXPROCS update.
  6640  	//
  6641  	// (2) doesn't matter to most users, as it is isn't observable as long
  6642  	// as (1) holds. However, it can be important to users sandboxing Go.
  6643  	// They want disable these system calls and need some way to know when
  6644  	// they are guaranteed the calls will stop.
  6645  	//
  6646  	// This would be simple to achieve if we simply called
  6647  	// defaultGOMAXPROCS under STW in updateMaxProcsGoroutine below.
  6648  	// However, we would like to avoid scheduling this goroutine every
  6649  	// second when it will almost never do anything. Instead, sysmon calls
  6650  	// defaultGOMAXPROCS to decide whether to schedule
  6651  	// updateMaxProcsGoroutine. Thus we need to synchronize between sysmon
  6652  	// and GOMAXPROCS calls.
  6653  	//
  6654  	// GOMAXPROCS can't hold a runtime mutex across STW. It could hold a
  6655  	// semaphore, but sysmon cannot take semaphores. Instead, we have a
  6656  	// more complex scheme:
  6657  	//
  6658  	// * sysmon holds computeMaxProcsLock while calling defaultGOMAXPROCS.
  6659  	// * sysmon skips the current update if sched.customGOMAXPROCS is
  6660  	//   set.
  6661  	// * GOMAXPROCS sets sched.customGOMAXPROCS once it is committed to
  6662  	//   changing GOMAXPROCS.
  6663  	// * GOMAXPROCS takes computeMaxProcsLock to wait for outstanding
  6664  	//   defaultGOMAXPROCS calls to complete.
  6665  	//
  6666  	// N.B. computeMaxProcsLock could simply be sched.lock, but we want to
  6667  	// avoid holding that lock during the potentially slow
  6668  	// defaultGOMAXPROCS.
  6669  	computeMaxProcsLock mutex
  6670  )
  6671  
  6672  // Start GOMAXPROCS update helper goroutine.
  6673  //
  6674  // This is based on forcegchelper.
  6675  func defaultGOMAXPROCSUpdateEnable() {
  6676  	if debug.updatemaxprocs == 0 {
  6677  		// Unconditionally increment the metric when updates are disabled.
  6678  		//
  6679  		// It would be more descriptive if we did a dry run of the
  6680  		// complete update, determining the appropriate value of
  6681  		// GOMAXPROCS and the bailing out and just incrementing the
  6682  		// metric if a change would occur.
  6683  		//
  6684  		// Not only is that a lot of ongoing work for a disabled
  6685  		// feature, but some users need to be able to completely
  6686  		// disable the update system calls (such as sandboxes).
  6687  		// Currently, updatemaxprocs=0 serves that purpose.
  6688  		updatemaxprocs.IncNonDefault()
  6689  		return
  6690  	}
  6691  
  6692  	go updateMaxProcsGoroutine()
  6693  }
  6694  
  6695  func updateMaxProcsGoroutine() {
  6696  	updateMaxProcsG.g = getg()
  6697  	lockInit(&updateMaxProcsG.lock, lockRankUpdateMaxProcsG)
  6698  	for {
  6699  		lock(&updateMaxProcsG.lock)
  6700  		if updateMaxProcsG.idle.Load() {
  6701  			throw("updateMaxProcsGoroutine: phase error")
  6702  		}
  6703  		updateMaxProcsG.idle.Store(true)
  6704  		goparkunlock(&updateMaxProcsG.lock, waitReasonUpdateGOMAXPROCSIdle, traceBlockSystemGoroutine, 1)
  6705  		// This goroutine is explicitly resumed by sysmon.
  6706  
  6707  		stw := stopTheWorldGC(stwGOMAXPROCS)
  6708  
  6709  		// Still OK to update?
  6710  		lock(&sched.lock)
  6711  		custom := sched.customGOMAXPROCS
  6712  		unlock(&sched.lock)
  6713  		if custom {
  6714  			startTheWorldGC(stw)
  6715  			return
  6716  		}
  6717  
  6718  		// newprocs will be processed by startTheWorld
  6719  		//
  6720  		// TODO(prattmic): this could use a nicer API. Perhaps add it to the
  6721  		// stw parameter?
  6722  		newprocs = updateMaxProcsG.procs
  6723  		lock(&sched.lock)
  6724  		sched.customGOMAXPROCS = false
  6725  		unlock(&sched.lock)
  6726  
  6727  		startTheWorldGC(stw)
  6728  	}
  6729  }
  6730  
  6731  func sysmonUpdateGOMAXPROCS() {
  6732  	// Synchronize with GOMAXPROCS. See comment on computeMaxProcsLock.
  6733  	lock(&computeMaxProcsLock)
  6734  
  6735  	// No update if GOMAXPROCS was set manually.
  6736  	lock(&sched.lock)
  6737  	custom := sched.customGOMAXPROCS
  6738  	curr := gomaxprocs
  6739  	unlock(&sched.lock)
  6740  	if custom {
  6741  		unlock(&computeMaxProcsLock)
  6742  		return
  6743  	}
  6744  
  6745  	// Don't hold sched.lock while we read the filesystem.
  6746  	procs := defaultGOMAXPROCS(0)
  6747  	unlock(&computeMaxProcsLock)
  6748  	if procs == curr {
  6749  		// Nothing to do.
  6750  		return
  6751  	}
  6752  
  6753  	// Sysmon can't directly stop the world. Run the helper to do so on our
  6754  	// behalf. If updateGOMAXPROCS.idle is false, then a previous update is
  6755  	// still pending.
  6756  	if updateMaxProcsG.idle.Load() {
  6757  		lock(&updateMaxProcsG.lock)
  6758  		updateMaxProcsG.procs = procs
  6759  		updateMaxProcsG.idle.Store(false)
  6760  		var list gList
  6761  		list.push(updateMaxProcsG.g)
  6762  		injectglist(&list)
  6763  		unlock(&updateMaxProcsG.lock)
  6764  	}
  6765  }
  6766  
  6767  // schedEnableUser enables or disables the scheduling of user
  6768  // goroutines.
  6769  //
  6770  // This does not stop already running user goroutines, so the caller
  6771  // should first stop the world when disabling user goroutines.
  6772  func schedEnableUser(enable bool) {
  6773  	lock(&sched.lock)
  6774  	if sched.disable.user == !enable {
  6775  		unlock(&sched.lock)
  6776  		return
  6777  	}
  6778  	sched.disable.user = !enable
  6779  	if enable {
  6780  		n := sched.disable.runnable.size
  6781  		globrunqputbatch(&sched.disable.runnable)
  6782  		unlock(&sched.lock)
  6783  		for ; n != 0 && sched.npidle.Load() != 0; n-- {
  6784  			startm(nil, false, false)
  6785  		}
  6786  	} else {
  6787  		unlock(&sched.lock)
  6788  	}
  6789  }
  6790  
  6791  // schedEnabled reports whether gp should be scheduled. It returns
  6792  // false is scheduling of gp is disabled.
  6793  //
  6794  // sched.lock must be held.
  6795  func schedEnabled(gp *g) bool {
  6796  	assertLockHeld(&sched.lock)
  6797  
  6798  	if sched.disable.user {
  6799  		return isSystemGoroutine(gp, true)
  6800  	}
  6801  	return true
  6802  }
  6803  
  6804  // Put mp on midle list.
  6805  // sched.lock must be held.
  6806  // May run during STW, so write barriers are not allowed.
  6807  //
  6808  //go:nowritebarrierrec
  6809  func mput(mp *m) {
  6810  	assertLockHeld(&sched.lock)
  6811  
  6812  	mp.schedlink = sched.midle
  6813  	sched.midle.set(mp)
  6814  	sched.nmidle++
  6815  	checkdead()
  6816  }
  6817  
  6818  // Try to get an m from midle list.
  6819  // sched.lock must be held.
  6820  // May run during STW, so write barriers are not allowed.
  6821  //
  6822  //go:nowritebarrierrec
  6823  func mget() *m {
  6824  	assertLockHeld(&sched.lock)
  6825  
  6826  	mp := sched.midle.ptr()
  6827  	if mp != nil {
  6828  		sched.midle = mp.schedlink
  6829  		sched.nmidle--
  6830  	}
  6831  	return mp
  6832  }
  6833  
  6834  // Put gp on the global runnable queue.
  6835  // sched.lock must be held.
  6836  // May run during STW, so write barriers are not allowed.
  6837  //
  6838  //go:nowritebarrierrec
  6839  func globrunqput(gp *g) {
  6840  	assertLockHeld(&sched.lock)
  6841  
  6842  	sched.runq.pushBack(gp)
  6843  }
  6844  
  6845  // Put gp at the head of the global runnable queue.
  6846  // sched.lock must be held.
  6847  // May run during STW, so write barriers are not allowed.
  6848  //
  6849  //go:nowritebarrierrec
  6850  func globrunqputhead(gp *g) {
  6851  	assertLockHeld(&sched.lock)
  6852  
  6853  	sched.runq.push(gp)
  6854  }
  6855  
  6856  // Put a batch of runnable goroutines on the global runnable queue.
  6857  // This clears *batch.
  6858  // sched.lock must be held.
  6859  // May run during STW, so write barriers are not allowed.
  6860  //
  6861  //go:nowritebarrierrec
  6862  func globrunqputbatch(batch *gQueue) {
  6863  	assertLockHeld(&sched.lock)
  6864  
  6865  	sched.runq.pushBackAll(*batch)
  6866  	*batch = gQueue{}
  6867  }
  6868  
  6869  // Try get a single G from the global runnable queue.
  6870  // sched.lock must be held.
  6871  func globrunqget() *g {
  6872  	assertLockHeld(&sched.lock)
  6873  
  6874  	if sched.runq.size == 0 {
  6875  		return nil
  6876  	}
  6877  
  6878  	return sched.runq.pop()
  6879  }
  6880  
  6881  // Try get a batch of G's from the global runnable queue.
  6882  // sched.lock must be held.
  6883  func globrunqgetbatch(n int32) (gp *g, q gQueue) {
  6884  	assertLockHeld(&sched.lock)
  6885  
  6886  	if sched.runq.size == 0 {
  6887  		return
  6888  	}
  6889  
  6890  	n = min(n, sched.runq.size, sched.runq.size/gomaxprocs+1)
  6891  
  6892  	gp = sched.runq.pop()
  6893  	n--
  6894  
  6895  	for ; n > 0; n-- {
  6896  		gp1 := sched.runq.pop()
  6897  		q.pushBack(gp1)
  6898  	}
  6899  	return
  6900  }
  6901  
  6902  // pMask is an atomic bitstring with one bit per P.
  6903  type pMask []uint32
  6904  
  6905  // read returns true if P id's bit is set.
  6906  func (p pMask) read(id uint32) bool {
  6907  	word := id / 32
  6908  	mask := uint32(1) << (id % 32)
  6909  	return (atomic.Load(&p[word]) & mask) != 0
  6910  }
  6911  
  6912  // set sets P id's bit.
  6913  func (p pMask) set(id int32) {
  6914  	word := id / 32
  6915  	mask := uint32(1) << (id % 32)
  6916  	atomic.Or(&p[word], mask)
  6917  }
  6918  
  6919  // clear clears P id's bit.
  6920  func (p pMask) clear(id int32) {
  6921  	word := id / 32
  6922  	mask := uint32(1) << (id % 32)
  6923  	atomic.And(&p[word], ^mask)
  6924  }
  6925  
  6926  // any returns true if any bit in p is set.
  6927  func (p pMask) any() bool {
  6928  	for i := range p {
  6929  		if atomic.Load(&p[i]) != 0 {
  6930  			return true
  6931  		}
  6932  	}
  6933  	return false
  6934  }
  6935  
  6936  // resize resizes the pMask and returns a new one.
  6937  //
  6938  // The result may alias p, so callers are encouraged to
  6939  // discard p. Not safe for concurrent use.
  6940  func (p pMask) resize(nprocs int32) pMask {
  6941  	maskWords := (nprocs + 31) / 32
  6942  
  6943  	if maskWords <= int32(cap(p)) {
  6944  		return p[:maskWords]
  6945  	}
  6946  	newMask := make([]uint32, maskWords)
  6947  	// No need to copy beyond len, old Ps are irrelevant.
  6948  	copy(newMask, p)
  6949  	return newMask
  6950  }
  6951  
  6952  // pidleput puts p on the _Pidle list. now must be a relatively recent call
  6953  // to nanotime or zero. Returns now or the current time if now was zero.
  6954  //
  6955  // This releases ownership of p. Once sched.lock is released it is no longer
  6956  // safe to use p.
  6957  //
  6958  // sched.lock must be held.
  6959  //
  6960  // May run during STW, so write barriers are not allowed.
  6961  //
  6962  //go:nowritebarrierrec
  6963  func pidleput(pp *p, now int64) int64 {
  6964  	assertLockHeld(&sched.lock)
  6965  
  6966  	if !runqempty(pp) {
  6967  		throw("pidleput: P has non-empty run queue")
  6968  	}
  6969  	if now == 0 {
  6970  		now = nanotime()
  6971  	}
  6972  	if pp.timers.len.Load() == 0 {
  6973  		timerpMask.clear(pp.id)
  6974  	}
  6975  	idlepMask.set(pp.id)
  6976  	pp.link = sched.pidle
  6977  	sched.pidle.set(pp)
  6978  	sched.npidle.Add(1)
  6979  	if !pp.limiterEvent.start(limiterEventIdle, now) {
  6980  		throw("must be able to track idle limiter event")
  6981  	}
  6982  	return now
  6983  }
  6984  
  6985  // pidleget tries to get a p from the _Pidle list, acquiring ownership.
  6986  //
  6987  // sched.lock must be held.
  6988  //
  6989  // May run during STW, so write barriers are not allowed.
  6990  //
  6991  //go:nowritebarrierrec
  6992  func pidleget(now int64) (*p, int64) {
  6993  	assertLockHeld(&sched.lock)
  6994  
  6995  	pp := sched.pidle.ptr()
  6996  	if pp != nil {
  6997  		// Timer may get added at any time now.
  6998  		if now == 0 {
  6999  			now = nanotime()
  7000  		}
  7001  		timerpMask.set(pp.id)
  7002  		idlepMask.clear(pp.id)
  7003  		sched.pidle = pp.link
  7004  		sched.npidle.Add(-1)
  7005  		pp.limiterEvent.stop(limiterEventIdle, now)
  7006  	}
  7007  	return pp, now
  7008  }
  7009  
  7010  // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
  7011  // This is called by spinning Ms (or callers than need a spinning M) that have
  7012  // found work. If no P is available, this must synchronized with non-spinning
  7013  // Ms that may be preparing to drop their P without discovering this work.
  7014  //
  7015  // sched.lock must be held.
  7016  //
  7017  // May run during STW, so write barriers are not allowed.
  7018  //
  7019  //go:nowritebarrierrec
  7020  func pidlegetSpinning(now int64) (*p, int64) {
  7021  	assertLockHeld(&sched.lock)
  7022  
  7023  	pp, now := pidleget(now)
  7024  	if pp == nil {
  7025  		// See "Delicate dance" comment in findrunnable. We found work
  7026  		// that we cannot take, we must synchronize with non-spinning
  7027  		// Ms that may be preparing to drop their P.
  7028  		sched.needspinning.Store(1)
  7029  		return nil, now
  7030  	}
  7031  
  7032  	return pp, now
  7033  }
  7034  
  7035  // runqempty reports whether pp has no Gs on its local run queue.
  7036  // It never returns true spuriously.
  7037  func runqempty(pp *p) bool {
  7038  	// Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
  7039  	// 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
  7040  	// Simply observing that runqhead == runqtail and then observing that runqnext == nil
  7041  	// does not mean the queue is empty.
  7042  	for {
  7043  		head := atomic.Load(&pp.runqhead)
  7044  		tail := atomic.Load(&pp.runqtail)
  7045  		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
  7046  		if tail == atomic.Load(&pp.runqtail) {
  7047  			return head == tail && runnext == 0
  7048  		}
  7049  	}
  7050  }
  7051  
  7052  // To shake out latent assumptions about scheduling order,
  7053  // we introduce some randomness into scheduling decisions
  7054  // when running with the race detector.
  7055  // The need for this was made obvious by changing the
  7056  // (deterministic) scheduling order in Go 1.5 and breaking
  7057  // many poorly-written tests.
  7058  // With the randomness here, as long as the tests pass
  7059  // consistently with -race, they shouldn't have latent scheduling
  7060  // assumptions.
  7061  const randomizeScheduler = raceenabled
  7062  
  7063  // runqput tries to put g on the local runnable queue.
  7064  // If next is false, runqput adds g to the tail of the runnable queue.
  7065  // If next is true, runqput puts g in the pp.runnext slot.
  7066  // If the run queue is full, runnext puts g on the global queue.
  7067  // Executed only by the owner P.
  7068  func runqput(pp *p, gp *g, next bool) {
  7069  	if !haveSysmon && next {
  7070  		// A runnext goroutine shares the same time slice as the
  7071  		// current goroutine (inheritTime from runqget). To prevent a
  7072  		// ping-pong pair of goroutines from starving all others, we
  7073  		// depend on sysmon to preempt "long-running goroutines". That
  7074  		// is, any set of goroutines sharing the same time slice.
  7075  		//
  7076  		// If there is no sysmon, we must avoid runnext entirely or
  7077  		// risk starvation.
  7078  		next = false
  7079  	}
  7080  	if randomizeScheduler && next && randn(2) == 0 {
  7081  		next = false
  7082  	}
  7083  
  7084  	if next {
  7085  	retryNext:
  7086  		oldnext := pp.runnext
  7087  		if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
  7088  			goto retryNext
  7089  		}
  7090  		if oldnext == 0 {
  7091  			return
  7092  		}
  7093  		// Kick the old runnext out to the regular run queue.
  7094  		gp = oldnext.ptr()
  7095  	}
  7096  
  7097  retry:
  7098  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  7099  	t := pp.runqtail
  7100  	if t-h < uint32(len(pp.runq)) {
  7101  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  7102  		atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
  7103  		return
  7104  	}
  7105  	if runqputslow(pp, gp, h, t) {
  7106  		return
  7107  	}
  7108  	// the queue is not full, now the put above must succeed
  7109  	goto retry
  7110  }
  7111  
  7112  // Put g and a batch of work from local runnable queue on global queue.
  7113  // Executed only by the owner P.
  7114  func runqputslow(pp *p, gp *g, h, t uint32) bool {
  7115  	var batch [len(pp.runq)/2 + 1]*g
  7116  
  7117  	// First, grab a batch from local queue.
  7118  	n := t - h
  7119  	n = n / 2
  7120  	if n != uint32(len(pp.runq)/2) {
  7121  		throw("runqputslow: queue is not full")
  7122  	}
  7123  	for i := uint32(0); i < n; i++ {
  7124  		batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  7125  	}
  7126  	if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  7127  		return false
  7128  	}
  7129  	batch[n] = gp
  7130  
  7131  	if randomizeScheduler {
  7132  		for i := uint32(1); i <= n; i++ {
  7133  			j := cheaprandn(i + 1)
  7134  			batch[i], batch[j] = batch[j], batch[i]
  7135  		}
  7136  	}
  7137  
  7138  	// Link the goroutines.
  7139  	for i := uint32(0); i < n; i++ {
  7140  		batch[i].schedlink.set(batch[i+1])
  7141  	}
  7142  
  7143  	q := gQueue{batch[0].guintptr(), batch[n].guintptr(), int32(n + 1)}
  7144  
  7145  	// Now put the batch on global queue.
  7146  	lock(&sched.lock)
  7147  	globrunqputbatch(&q)
  7148  	unlock(&sched.lock)
  7149  	return true
  7150  }
  7151  
  7152  // runqputbatch tries to put all the G's on q on the local runnable queue.
  7153  // If the local runq is full the input queue still contains unqueued Gs.
  7154  // Executed only by the owner P.
  7155  func runqputbatch(pp *p, q *gQueue) {
  7156  	if q.empty() {
  7157  		return
  7158  	}
  7159  	h := atomic.LoadAcq(&pp.runqhead)
  7160  	t := pp.runqtail
  7161  	n := uint32(0)
  7162  	for !q.empty() && t-h < uint32(len(pp.runq)) {
  7163  		gp := q.pop()
  7164  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  7165  		t++
  7166  		n++
  7167  	}
  7168  
  7169  	if randomizeScheduler {
  7170  		off := func(o uint32) uint32 {
  7171  			return (pp.runqtail + o) % uint32(len(pp.runq))
  7172  		}
  7173  		for i := uint32(1); i < n; i++ {
  7174  			j := cheaprandn(i + 1)
  7175  			pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
  7176  		}
  7177  	}
  7178  
  7179  	atomic.StoreRel(&pp.runqtail, t)
  7180  
  7181  	return
  7182  }
  7183  
  7184  // Get g from local runnable queue.
  7185  // If inheritTime is true, gp should inherit the remaining time in the
  7186  // current time slice. Otherwise, it should start a new time slice.
  7187  // Executed only by the owner P.
  7188  func runqget(pp *p) (gp *g, inheritTime bool) {
  7189  	// If there's a runnext, it's the next G to run.
  7190  	next := pp.runnext
  7191  	// If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
  7192  	// because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
  7193  	// Hence, there's no need to retry this CAS if it fails.
  7194  	if next != 0 && pp.runnext.cas(next, 0) {
  7195  		return next.ptr(), true
  7196  	}
  7197  
  7198  	for {
  7199  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7200  		t := pp.runqtail
  7201  		if t == h {
  7202  			return nil, false
  7203  		}
  7204  		gp := pp.runq[h%uint32(len(pp.runq))].ptr()
  7205  		if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
  7206  			return gp, false
  7207  		}
  7208  	}
  7209  }
  7210  
  7211  // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
  7212  // Executed only by the owner P.
  7213  func runqdrain(pp *p) (drainQ gQueue) {
  7214  	oldNext := pp.runnext
  7215  	if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
  7216  		drainQ.pushBack(oldNext.ptr())
  7217  	}
  7218  
  7219  retry:
  7220  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7221  	t := pp.runqtail
  7222  	qn := t - h
  7223  	if qn == 0 {
  7224  		return
  7225  	}
  7226  	if qn > uint32(len(pp.runq)) { // read inconsistent h and t
  7227  		goto retry
  7228  	}
  7229  
  7230  	if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
  7231  		goto retry
  7232  	}
  7233  
  7234  	// We've inverted the order in which it gets G's from the local P's runnable queue
  7235  	// and then advances the head pointer because we don't want to mess up the statuses of G's
  7236  	// while runqdrain() and runqsteal() are running in parallel.
  7237  	// Thus we should advance the head pointer before draining the local P into a gQueue,
  7238  	// so that we can update any gp.schedlink only after we take the full ownership of G,
  7239  	// meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
  7240  	// See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
  7241  	for i := uint32(0); i < qn; i++ {
  7242  		gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  7243  		drainQ.pushBack(gp)
  7244  	}
  7245  	return
  7246  }
  7247  
  7248  // Grabs a batch of goroutines from pp's runnable queue into batch.
  7249  // Batch is a ring buffer starting at batchHead.
  7250  // Returns number of grabbed goroutines.
  7251  // Can be executed by any P.
  7252  func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
  7253  	for {
  7254  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7255  		t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
  7256  		n := t - h
  7257  		n = n - n/2
  7258  		if n == 0 {
  7259  			if stealRunNextG {
  7260  				// Try to steal from pp.runnext.
  7261  				if next := pp.runnext; next != 0 {
  7262  					if pp.status == _Prunning {
  7263  						// Sleep to ensure that pp isn't about to run the g
  7264  						// we are about to steal.
  7265  						// The important use case here is when the g running
  7266  						// on pp ready()s another g and then almost
  7267  						// immediately blocks. Instead of stealing runnext
  7268  						// in this window, back off to give pp a chance to
  7269  						// schedule runnext. This will avoid thrashing gs
  7270  						// between different Ps.
  7271  						// A sync chan send/recv takes ~50ns as of time of
  7272  						// writing, so 3us gives ~50x overshoot.
  7273  						if !osHasLowResTimer {
  7274  							usleep(3)
  7275  						} else {
  7276  							// On some platforms system timer granularity is
  7277  							// 1-15ms, which is way too much for this
  7278  							// optimization. So just yield.
  7279  							osyield()
  7280  						}
  7281  					}
  7282  					if !pp.runnext.cas(next, 0) {
  7283  						continue
  7284  					}
  7285  					batch[batchHead%uint32(len(batch))] = next
  7286  					return 1
  7287  				}
  7288  			}
  7289  			return 0
  7290  		}
  7291  		if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
  7292  			continue
  7293  		}
  7294  		for i := uint32(0); i < n; i++ {
  7295  			g := pp.runq[(h+i)%uint32(len(pp.runq))]
  7296  			batch[(batchHead+i)%uint32(len(batch))] = g
  7297  		}
  7298  		if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  7299  			return n
  7300  		}
  7301  	}
  7302  }
  7303  
  7304  // Steal half of elements from local runnable queue of p2
  7305  // and put onto local runnable queue of p.
  7306  // Returns one of the stolen elements (or nil if failed).
  7307  func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
  7308  	t := pp.runqtail
  7309  	n := runqgrab(p2, &pp.runq, t, stealRunNextG)
  7310  	if n == 0 {
  7311  		return nil
  7312  	}
  7313  	n--
  7314  	gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
  7315  	if n == 0 {
  7316  		return gp
  7317  	}
  7318  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  7319  	if t-h+n >= uint32(len(pp.runq)) {
  7320  		throw("runqsteal: runq overflow")
  7321  	}
  7322  	atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
  7323  	return gp
  7324  }
  7325  
  7326  // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
  7327  // be on one gQueue or gList at a time.
  7328  type gQueue struct {
  7329  	head guintptr
  7330  	tail guintptr
  7331  	size int32
  7332  }
  7333  
  7334  // empty reports whether q is empty.
  7335  func (q *gQueue) empty() bool {
  7336  	return q.head == 0
  7337  }
  7338  
  7339  // push adds gp to the head of q.
  7340  func (q *gQueue) push(gp *g) {
  7341  	gp.schedlink = q.head
  7342  	q.head.set(gp)
  7343  	if q.tail == 0 {
  7344  		q.tail.set(gp)
  7345  	}
  7346  	q.size++
  7347  }
  7348  
  7349  // pushBack adds gp to the tail of q.
  7350  func (q *gQueue) pushBack(gp *g) {
  7351  	gp.schedlink = 0
  7352  	if q.tail != 0 {
  7353  		q.tail.ptr().schedlink.set(gp)
  7354  	} else {
  7355  		q.head.set(gp)
  7356  	}
  7357  	q.tail.set(gp)
  7358  	q.size++
  7359  }
  7360  
  7361  // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
  7362  // not be used.
  7363  func (q *gQueue) pushBackAll(q2 gQueue) {
  7364  	if q2.tail == 0 {
  7365  		return
  7366  	}
  7367  	q2.tail.ptr().schedlink = 0
  7368  	if q.tail != 0 {
  7369  		q.tail.ptr().schedlink = q2.head
  7370  	} else {
  7371  		q.head = q2.head
  7372  	}
  7373  	q.tail = q2.tail
  7374  	q.size += q2.size
  7375  }
  7376  
  7377  // pop removes and returns the head of queue q. It returns nil if
  7378  // q is empty.
  7379  func (q *gQueue) pop() *g {
  7380  	gp := q.head.ptr()
  7381  	if gp != nil {
  7382  		q.head = gp.schedlink
  7383  		if q.head == 0 {
  7384  			q.tail = 0
  7385  		}
  7386  		q.size--
  7387  	}
  7388  	return gp
  7389  }
  7390  
  7391  // popList takes all Gs in q and returns them as a gList.
  7392  func (q *gQueue) popList() gList {
  7393  	stack := gList{q.head, q.size}
  7394  	*q = gQueue{}
  7395  	return stack
  7396  }
  7397  
  7398  // A gList is a list of Gs linked through g.schedlink. A G can only be
  7399  // on one gQueue or gList at a time.
  7400  type gList struct {
  7401  	head guintptr
  7402  	size int32
  7403  }
  7404  
  7405  // empty reports whether l is empty.
  7406  func (l *gList) empty() bool {
  7407  	return l.head == 0
  7408  }
  7409  
  7410  // push adds gp to the head of l.
  7411  func (l *gList) push(gp *g) {
  7412  	gp.schedlink = l.head
  7413  	l.head.set(gp)
  7414  	l.size++
  7415  }
  7416  
  7417  // pushAll prepends all Gs in q to l. After this q must not be used.
  7418  func (l *gList) pushAll(q gQueue) {
  7419  	if !q.empty() {
  7420  		q.tail.ptr().schedlink = l.head
  7421  		l.head = q.head
  7422  		l.size += q.size
  7423  	}
  7424  }
  7425  
  7426  // pop removes and returns the head of l. If l is empty, it returns nil.
  7427  func (l *gList) pop() *g {
  7428  	gp := l.head.ptr()
  7429  	if gp != nil {
  7430  		l.head = gp.schedlink
  7431  		l.size--
  7432  	}
  7433  	return gp
  7434  }
  7435  
  7436  //go:linkname setMaxThreads runtime/debug.setMaxThreads
  7437  func setMaxThreads(in int) (out int) {
  7438  	lock(&sched.lock)
  7439  	out = int(sched.maxmcount)
  7440  	if in > 0x7fffffff { // MaxInt32
  7441  		sched.maxmcount = 0x7fffffff
  7442  	} else {
  7443  		sched.maxmcount = int32(in)
  7444  	}
  7445  	checkmcount()
  7446  	unlock(&sched.lock)
  7447  	return
  7448  }
  7449  
  7450  // procPin should be an internal detail,
  7451  // but widely used packages access it using linkname.
  7452  // Notable members of the hall of shame include:
  7453  //   - github.com/bytedance/gopkg
  7454  //   - github.com/choleraehyq/pid
  7455  //   - github.com/songzhibin97/gkit
  7456  //
  7457  // Do not remove or change the type signature.
  7458  // See go.dev/issue/67401.
  7459  //
  7460  //go:linkname procPin
  7461  //go:nosplit
  7462  func procPin() int {
  7463  	gp := getg()
  7464  	mp := gp.m
  7465  
  7466  	mp.locks++
  7467  	return int(mp.p.ptr().id)
  7468  }
  7469  
  7470  // procUnpin should be an internal detail,
  7471  // but widely used packages access it using linkname.
  7472  // Notable members of the hall of shame include:
  7473  //   - github.com/bytedance/gopkg
  7474  //   - github.com/choleraehyq/pid
  7475  //   - github.com/songzhibin97/gkit
  7476  //
  7477  // Do not remove or change the type signature.
  7478  // See go.dev/issue/67401.
  7479  //
  7480  //go:linkname procUnpin
  7481  //go:nosplit
  7482  func procUnpin() {
  7483  	gp := getg()
  7484  	gp.m.locks--
  7485  }
  7486  
  7487  //go:linkname sync_runtime_procPin sync.runtime_procPin
  7488  //go:nosplit
  7489  func sync_runtime_procPin() int {
  7490  	return procPin()
  7491  }
  7492  
  7493  //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
  7494  //go:nosplit
  7495  func sync_runtime_procUnpin() {
  7496  	procUnpin()
  7497  }
  7498  
  7499  //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
  7500  //go:nosplit
  7501  func sync_atomic_runtime_procPin() int {
  7502  	return procPin()
  7503  }
  7504  
  7505  //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
  7506  //go:nosplit
  7507  func sync_atomic_runtime_procUnpin() {
  7508  	procUnpin()
  7509  }
  7510  
  7511  // Active spinning for sync.Mutex.
  7512  //
  7513  //go:linkname internal_sync_runtime_canSpin internal/sync.runtime_canSpin
  7514  //go:nosplit
  7515  func internal_sync_runtime_canSpin(i int) bool {
  7516  	// sync.Mutex is cooperative, so we are conservative with spinning.
  7517  	// Spin only few times and only if running on a multicore machine and
  7518  	// GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
  7519  	// As opposed to runtime mutex we don't do passive spinning here,
  7520  	// because there can be work on global runq or on other Ps.
  7521  	if i >= active_spin || numCPUStartup <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
  7522  		return false
  7523  	}
  7524  	if p := getg().m.p.ptr(); !runqempty(p) {
  7525  		return false
  7526  	}
  7527  	return true
  7528  }
  7529  
  7530  //go:linkname internal_sync_runtime_doSpin internal/sync.runtime_doSpin
  7531  //go:nosplit
  7532  func internal_sync_runtime_doSpin() {
  7533  	procyield(active_spin_cnt)
  7534  }
  7535  
  7536  // Active spinning for sync.Mutex.
  7537  //
  7538  // sync_runtime_canSpin should be an internal detail,
  7539  // but widely used packages access it using linkname.
  7540  // Notable members of the hall of shame include:
  7541  //   - github.com/livekit/protocol
  7542  //   - github.com/sagernet/gvisor
  7543  //   - gvisor.dev/gvisor
  7544  //
  7545  // Do not remove or change the type signature.
  7546  // See go.dev/issue/67401.
  7547  //
  7548  //go:linkname sync_runtime_canSpin sync.runtime_canSpin
  7549  //go:nosplit
  7550  func sync_runtime_canSpin(i int) bool {
  7551  	return internal_sync_runtime_canSpin(i)
  7552  }
  7553  
  7554  // sync_runtime_doSpin should be an internal detail,
  7555  // but widely used packages access it using linkname.
  7556  // Notable members of the hall of shame include:
  7557  //   - github.com/livekit/protocol
  7558  //   - github.com/sagernet/gvisor
  7559  //   - gvisor.dev/gvisor
  7560  //
  7561  // Do not remove or change the type signature.
  7562  // See go.dev/issue/67401.
  7563  //
  7564  //go:linkname sync_runtime_doSpin sync.runtime_doSpin
  7565  //go:nosplit
  7566  func sync_runtime_doSpin() {
  7567  	internal_sync_runtime_doSpin()
  7568  }
  7569  
  7570  var stealOrder randomOrder
  7571  
  7572  // randomOrder/randomEnum are helper types for randomized work stealing.
  7573  // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
  7574  // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
  7575  // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
  7576  type randomOrder struct {
  7577  	count    uint32
  7578  	coprimes []uint32
  7579  }
  7580  
  7581  type randomEnum struct {
  7582  	i     uint32
  7583  	count uint32
  7584  	pos   uint32
  7585  	inc   uint32
  7586  }
  7587  
  7588  func (ord *randomOrder) reset(count uint32) {
  7589  	ord.count = count
  7590  	ord.coprimes = ord.coprimes[:0]
  7591  	for i := uint32(1); i <= count; i++ {
  7592  		if gcd(i, count) == 1 {
  7593  			ord.coprimes = append(ord.coprimes, i)
  7594  		}
  7595  	}
  7596  }
  7597  
  7598  func (ord *randomOrder) start(i uint32) randomEnum {
  7599  	return randomEnum{
  7600  		count: ord.count,
  7601  		pos:   i % ord.count,
  7602  		inc:   ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
  7603  	}
  7604  }
  7605  
  7606  func (enum *randomEnum) done() bool {
  7607  	return enum.i == enum.count
  7608  }
  7609  
  7610  func (enum *randomEnum) next() {
  7611  	enum.i++
  7612  	enum.pos = (enum.pos + enum.inc) % enum.count
  7613  }
  7614  
  7615  func (enum *randomEnum) position() uint32 {
  7616  	return enum.pos
  7617  }
  7618  
  7619  func gcd(a, b uint32) uint32 {
  7620  	for b != 0 {
  7621  		a, b = b, a%b
  7622  	}
  7623  	return a
  7624  }
  7625  
  7626  // An initTask represents the set of initializations that need to be done for a package.
  7627  // Keep in sync with ../../test/noinit.go:initTask
  7628  type initTask struct {
  7629  	state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
  7630  	nfns  uint32
  7631  	// followed by nfns pcs, uintptr sized, one per init function to run
  7632  }
  7633  
  7634  // inittrace stores statistics for init functions which are
  7635  // updated by malloc and newproc when active is true.
  7636  var inittrace tracestat
  7637  
  7638  type tracestat struct {
  7639  	active bool   // init tracing activation status
  7640  	id     uint64 // init goroutine id
  7641  	allocs uint64 // heap allocations
  7642  	bytes  uint64 // heap allocated bytes
  7643  }
  7644  
  7645  func doInit(ts []*initTask) {
  7646  	for _, t := range ts {
  7647  		doInit1(t)
  7648  	}
  7649  }
  7650  
  7651  func doInit1(t *initTask) {
  7652  	switch t.state {
  7653  	case 2: // fully initialized
  7654  		return
  7655  	case 1: // initialization in progress
  7656  		throw("recursive call during initialization - linker skew")
  7657  	default: // not initialized yet
  7658  		t.state = 1 // initialization in progress
  7659  
  7660  		var (
  7661  			start  int64
  7662  			before tracestat
  7663  		)
  7664  
  7665  		if inittrace.active {
  7666  			start = nanotime()
  7667  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  7668  			before = inittrace
  7669  		}
  7670  
  7671  		if t.nfns == 0 {
  7672  			// We should have pruned all of these in the linker.
  7673  			throw("inittask with no functions")
  7674  		}
  7675  
  7676  		firstFunc := add(unsafe.Pointer(t), 8)
  7677  		for i := uint32(0); i < t.nfns; i++ {
  7678  			p := add(firstFunc, uintptr(i)*goarch.PtrSize)
  7679  			f := *(*func())(unsafe.Pointer(&p))
  7680  			f()
  7681  		}
  7682  
  7683  		if inittrace.active {
  7684  			end := nanotime()
  7685  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  7686  			after := inittrace
  7687  
  7688  			f := *(*func())(unsafe.Pointer(&firstFunc))
  7689  			pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
  7690  
  7691  			var sbuf [24]byte
  7692  			print("init ", pkg, " @")
  7693  			print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
  7694  			print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
  7695  			print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
  7696  			print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
  7697  			print("\n")
  7698  		}
  7699  
  7700  		t.state = 2 // initialization done
  7701  	}
  7702  }
  7703