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  	for m := &firstmoduledata; m != nil; m = m.next {
   256  		doInit(m.inittasks)
   257  	}
   258  
   259  	// Disable init tracing after main init done to avoid overhead
   260  	// of collecting statistics in malloc and newproc
   261  	inittrace.active = false
   262  
   263  	close(main_init_done)
   264  
   265  	needUnlock = false
   266  	unlockOSThread()
   267  
   268  	if isarchive || islibrary {
   269  		// A program compiled with -buildmode=c-archive or c-shared
   270  		// has a main, but it is not executed.
   271  		if GOARCH == "wasm" {
   272  			// On Wasm, pause makes it return to the host.
   273  			// Unlike cgo callbacks where Ms are created on demand,
   274  			// on Wasm we have only one M. So we keep this M (and this
   275  			// G) for callbacks.
   276  			// Using the caller's SP unwinds this frame and backs to
   277  			// goexit. The -16 is: 8 for goexit's (fake) return PC,
   278  			// and pause's epilogue pops 8.
   279  			pause(sys.GetCallerSP() - 16) // should not return
   280  			panic("unreachable")
   281  		}
   282  		return
   283  	}
   284  	fn := main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
   285  	fn()
   286  
   287  	exitHooksRun := false
   288  	if raceenabled {
   289  		runExitHooks(0) // run hooks now, since racefini does not return
   290  		exitHooksRun = true
   291  		racefini()
   292  	}
   293  
   294  	// Check for C memory leaks if using ASAN and we've made cgo calls,
   295  	// or if we are running as a library in a C program.
   296  	// We always make one cgo call, above, to notify_runtime_init_done,
   297  	// so we ignore that one.
   298  	// No point in leak checking if no cgo calls, since leak checking
   299  	// just looks for objects allocated using malloc and friends.
   300  	// Just checking iscgo doesn't help because asan implies iscgo.
   301  	if asanenabled && (isarchive || islibrary || NumCgoCall() > 1) {
   302  		runExitHooks(0) // lsandoleakcheck may not return
   303  		exitHooksRun = true
   304  		lsandoleakcheck()
   305  	}
   306  
   307  	// Make racy client program work: if panicking on
   308  	// another goroutine at the same time as main returns,
   309  	// let the other goroutine finish printing the panic trace.
   310  	// Once it does, it will exit. See issues 3934 and 20018.
   311  	if runningPanicDefers.Load() != 0 {
   312  		// Running deferred functions should not take long.
   313  		for c := 0; c < 1000; c++ {
   314  			if runningPanicDefers.Load() == 0 {
   315  				break
   316  			}
   317  			Gosched()
   318  		}
   319  	}
   320  	if panicking.Load() != 0 {
   321  		gopark(nil, nil, waitReasonPanicWait, traceBlockForever, 1)
   322  	}
   323  	if !exitHooksRun {
   324  		runExitHooks(0)
   325  	}
   326  
   327  	exit(0)
   328  	for {
   329  		var x *int32
   330  		*x = 0
   331  	}
   332  }
   333  
   334  // os_beforeExit is called from os.Exit(0).
   335  //
   336  //go:linkname os_beforeExit os.runtime_beforeExit
   337  func os_beforeExit(exitCode int) {
   338  	runExitHooks(exitCode)
   339  	if exitCode == 0 && raceenabled {
   340  		racefini()
   341  	}
   342  
   343  	// See comment in main, above.
   344  	if exitCode == 0 && asanenabled && (isarchive || islibrary || NumCgoCall() > 1) {
   345  		lsandoleakcheck()
   346  	}
   347  }
   348  
   349  func init() {
   350  	exithook.Gosched = Gosched
   351  	exithook.Goid = func() uint64 { return getg().goid }
   352  	exithook.Throw = throw
   353  }
   354  
   355  func runExitHooks(code int) {
   356  	exithook.Run(code)
   357  }
   358  
   359  // start forcegc helper goroutine
   360  func init() {
   361  	go forcegchelper()
   362  }
   363  
   364  func forcegchelper() {
   365  	forcegc.g = getg()
   366  	lockInit(&forcegc.lock, lockRankForcegc)
   367  	for {
   368  		lock(&forcegc.lock)
   369  		if forcegc.idle.Load() {
   370  			throw("forcegc: phase error")
   371  		}
   372  		forcegc.idle.Store(true)
   373  		goparkunlock(&forcegc.lock, waitReasonForceGCIdle, traceBlockSystemGoroutine, 1)
   374  		// this goroutine is explicitly resumed by sysmon
   375  		if debug.gctrace > 0 {
   376  			println("GC forced")
   377  		}
   378  		// Time-triggered, fully concurrent.
   379  		gcStart(gcTrigger{kind: gcTriggerTime, now: nanotime()})
   380  	}
   381  }
   382  
   383  // Gosched yields the processor, allowing other goroutines to run. It does not
   384  // suspend the current goroutine, so execution resumes automatically.
   385  //
   386  //go:nosplit
   387  func Gosched() {
   388  	checkTimeouts()
   389  	mcall(gosched_m)
   390  }
   391  
   392  // goschedguarded yields the processor like gosched, but also checks
   393  // for forbidden states and opts out of the yield in those cases.
   394  //
   395  //go:nosplit
   396  func goschedguarded() {
   397  	mcall(goschedguarded_m)
   398  }
   399  
   400  // goschedIfBusy yields the processor like gosched, but only does so if
   401  // there are no idle Ps or if we're on the only P and there's nothing in
   402  // the run queue. In both cases, there is freely available idle time.
   403  //
   404  //go:nosplit
   405  func goschedIfBusy() {
   406  	gp := getg()
   407  	// Call gosched if gp.preempt is set; we may be in a tight loop that
   408  	// doesn't otherwise yield.
   409  	if !gp.preempt && sched.npidle.Load() > 0 {
   410  		return
   411  	}
   412  	mcall(gosched_m)
   413  }
   414  
   415  // Puts the current goroutine into a waiting state and calls unlockf on the
   416  // system stack.
   417  //
   418  // If unlockf returns false, the goroutine is resumed.
   419  //
   420  // unlockf must not access this G's stack, as it may be moved between
   421  // the call to gopark and the call to unlockf.
   422  //
   423  // Note that because unlockf is called after putting the G into a waiting
   424  // state, the G may have already been readied by the time unlockf is called
   425  // unless there is external synchronization preventing the G from being
   426  // readied. If unlockf returns false, it must guarantee that the G cannot be
   427  // externally readied.
   428  //
   429  // Reason explains why the goroutine has been parked. It is displayed in stack
   430  // traces and heap dumps. Reasons should be unique and descriptive. Do not
   431  // re-use reasons, add new ones.
   432  //
   433  // gopark should be an internal detail,
   434  // but widely used packages access it using linkname.
   435  // Notable members of the hall of shame include:
   436  //   - gvisor.dev/gvisor
   437  //   - github.com/sagernet/gvisor
   438  //
   439  // Do not remove or change the type signature.
   440  // See go.dev/issue/67401.
   441  //
   442  //go:linkname gopark
   443  func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason waitReason, traceReason traceBlockReason, traceskip int) {
   444  	if reason != waitReasonSleep {
   445  		checkTimeouts() // timeouts may expire while two goroutines keep the scheduler busy
   446  	}
   447  	mp := acquirem()
   448  	gp := mp.curg
   449  	status := readgstatus(gp)
   450  	if status != _Grunning && status != _Gscanrunning {
   451  		throw("gopark: bad g status")
   452  	}
   453  	mp.waitlock = lock
   454  	mp.waitunlockf = unlockf
   455  	gp.waitreason = reason
   456  	mp.waitTraceBlockReason = traceReason
   457  	mp.waitTraceSkip = traceskip
   458  	releasem(mp)
   459  	// can't do anything that might move the G between Ms here.
   460  	mcall(park_m)
   461  }
   462  
   463  // Puts the current goroutine into a waiting state and unlocks the lock.
   464  // The goroutine can be made runnable again by calling goready(gp).
   465  func goparkunlock(lock *mutex, reason waitReason, traceReason traceBlockReason, traceskip int) {
   466  	gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceReason, traceskip)
   467  }
   468  
   469  // goready should be an internal detail,
   470  // but widely used packages access it using linkname.
   471  // Notable members of the hall of shame include:
   472  //   - gvisor.dev/gvisor
   473  //   - github.com/sagernet/gvisor
   474  //
   475  // Do not remove or change the type signature.
   476  // See go.dev/issue/67401.
   477  //
   478  //go:linkname goready
   479  func goready(gp *g, traceskip int) {
   480  	systemstack(func() {
   481  		ready(gp, traceskip, true)
   482  	})
   483  }
   484  
   485  //go:nosplit
   486  func acquireSudog() *sudog {
   487  	// Delicate dance: the semaphore implementation calls
   488  	// acquireSudog, acquireSudog calls new(sudog),
   489  	// new calls malloc, malloc can call the garbage collector,
   490  	// and the garbage collector calls the semaphore implementation
   491  	// in stopTheWorld.
   492  	// Break the cycle by doing acquirem/releasem around new(sudog).
   493  	// The acquirem/releasem increments m.locks during new(sudog),
   494  	// which keeps the garbage collector from being invoked.
   495  	mp := acquirem()
   496  	pp := mp.p.ptr()
   497  	if len(pp.sudogcache) == 0 {
   498  		lock(&sched.sudoglock)
   499  		// First, try to grab a batch from central cache.
   500  		for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
   501  			s := sched.sudogcache
   502  			sched.sudogcache = s.next
   503  			s.next = nil
   504  			pp.sudogcache = append(pp.sudogcache, s)
   505  		}
   506  		unlock(&sched.sudoglock)
   507  		// If the central cache is empty, allocate a new one.
   508  		if len(pp.sudogcache) == 0 {
   509  			pp.sudogcache = append(pp.sudogcache, new(sudog))
   510  		}
   511  	}
   512  	n := len(pp.sudogcache)
   513  	s := pp.sudogcache[n-1]
   514  	pp.sudogcache[n-1] = nil
   515  	pp.sudogcache = pp.sudogcache[:n-1]
   516  	if s.elem != nil {
   517  		throw("acquireSudog: found s.elem != nil in cache")
   518  	}
   519  	releasem(mp)
   520  	return s
   521  }
   522  
   523  //go:nosplit
   524  func releaseSudog(s *sudog) {
   525  	if s.elem != nil {
   526  		throw("runtime: sudog with non-nil elem")
   527  	}
   528  	if s.isSelect {
   529  		throw("runtime: sudog with non-false isSelect")
   530  	}
   531  	if s.next != nil {
   532  		throw("runtime: sudog with non-nil next")
   533  	}
   534  	if s.prev != nil {
   535  		throw("runtime: sudog with non-nil prev")
   536  	}
   537  	if s.waitlink != nil {
   538  		throw("runtime: sudog with non-nil waitlink")
   539  	}
   540  	if s.c != nil {
   541  		throw("runtime: sudog with non-nil c")
   542  	}
   543  	gp := getg()
   544  	if gp.param != nil {
   545  		throw("runtime: releaseSudog with non-nil gp.param")
   546  	}
   547  	mp := acquirem() // avoid rescheduling to another P
   548  	pp := mp.p.ptr()
   549  	if len(pp.sudogcache) == cap(pp.sudogcache) {
   550  		// Transfer half of local cache to the central cache.
   551  		var first, last *sudog
   552  		for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
   553  			n := len(pp.sudogcache)
   554  			p := pp.sudogcache[n-1]
   555  			pp.sudogcache[n-1] = nil
   556  			pp.sudogcache = pp.sudogcache[:n-1]
   557  			if first == nil {
   558  				first = p
   559  			} else {
   560  				last.next = p
   561  			}
   562  			last = p
   563  		}
   564  		lock(&sched.sudoglock)
   565  		last.next = sched.sudogcache
   566  		sched.sudogcache = first
   567  		unlock(&sched.sudoglock)
   568  	}
   569  	pp.sudogcache = append(pp.sudogcache, s)
   570  	releasem(mp)
   571  }
   572  
   573  // called from assembly.
   574  func badmcall(fn func(*g)) {
   575  	throw("runtime: mcall called on m->g0 stack")
   576  }
   577  
   578  func badmcall2(fn func(*g)) {
   579  	throw("runtime: mcall function returned")
   580  }
   581  
   582  func badreflectcall() {
   583  	panic(plainError("arg size to reflect.call more than 1GB"))
   584  }
   585  
   586  //go:nosplit
   587  //go:nowritebarrierrec
   588  func badmorestackg0() {
   589  	if !crashStackImplemented {
   590  		writeErrStr("fatal: morestack on g0\n")
   591  		return
   592  	}
   593  
   594  	g := getg()
   595  	switchToCrashStack(func() {
   596  		print("runtime: morestack on g0, stack [", hex(g.stack.lo), " ", hex(g.stack.hi), "], sp=", hex(g.sched.sp), ", called from\n")
   597  		g.m.traceback = 2 // include pc and sp in stack trace
   598  		traceback1(g.sched.pc, g.sched.sp, g.sched.lr, g, 0)
   599  		print("\n")
   600  
   601  		throw("morestack on g0")
   602  	})
   603  }
   604  
   605  //go:nosplit
   606  //go:nowritebarrierrec
   607  func badmorestackgsignal() {
   608  	writeErrStr("fatal: morestack on gsignal\n")
   609  }
   610  
   611  //go:nosplit
   612  func badctxt() {
   613  	throw("ctxt != 0")
   614  }
   615  
   616  // gcrash is a fake g that can be used when crashing due to bad
   617  // stack conditions.
   618  var gcrash g
   619  
   620  var crashingG atomic.Pointer[g]
   621  
   622  // Switch to crashstack and call fn, with special handling of
   623  // concurrent and recursive cases.
   624  //
   625  // Nosplit as it is called in a bad stack condition (we know
   626  // morestack would fail).
   627  //
   628  //go:nosplit
   629  //go:nowritebarrierrec
   630  func switchToCrashStack(fn func()) {
   631  	me := getg()
   632  	if crashingG.CompareAndSwapNoWB(nil, me) {
   633  		switchToCrashStack0(fn) // should never return
   634  		abort()
   635  	}
   636  	if crashingG.Load() == me {
   637  		// recursive crashing. too bad.
   638  		writeErrStr("fatal: recursive switchToCrashStack\n")
   639  		abort()
   640  	}
   641  	// Another g is crashing. Give it some time, hopefully it will finish traceback.
   642  	usleep_no_g(100)
   643  	writeErrStr("fatal: concurrent switchToCrashStack\n")
   644  	abort()
   645  }
   646  
   647  // Disable crash stack on Windows for now. Apparently, throwing an exception
   648  // on a non-system-allocated crash stack causes EXCEPTION_STACK_OVERFLOW and
   649  // hangs the process (see issue 63938).
   650  const crashStackImplemented = GOOS != "windows"
   651  
   652  //go:noescape
   653  func switchToCrashStack0(fn func()) // in assembly
   654  
   655  func lockedOSThread() bool {
   656  	gp := getg()
   657  	return gp.lockedm != 0 && gp.m.lockedg != 0
   658  }
   659  
   660  var (
   661  	// allgs contains all Gs ever created (including dead Gs), and thus
   662  	// never shrinks.
   663  	//
   664  	// Access via the slice is protected by allglock or stop-the-world.
   665  	// Readers that cannot take the lock may (carefully!) use the atomic
   666  	// variables below.
   667  	allglock mutex
   668  	allgs    []*g
   669  
   670  	// allglen and allgptr are atomic variables that contain len(allgs) and
   671  	// &allgs[0] respectively. Proper ordering depends on totally-ordered
   672  	// loads and stores. Writes are protected by allglock.
   673  	//
   674  	// allgptr is updated before allglen. Readers should read allglen
   675  	// before allgptr to ensure that allglen is always <= len(allgptr). New
   676  	// Gs appended during the race can be missed. For a consistent view of
   677  	// all Gs, allglock must be held.
   678  	//
   679  	// allgptr copies should always be stored as a concrete type or
   680  	// unsafe.Pointer, not uintptr, to ensure that GC can still reach it
   681  	// even if it points to a stale array.
   682  	allglen uintptr
   683  	allgptr **g
   684  )
   685  
   686  func allgadd(gp *g) {
   687  	if readgstatus(gp) == _Gidle {
   688  		throw("allgadd: bad status Gidle")
   689  	}
   690  
   691  	lock(&allglock)
   692  	allgs = append(allgs, gp)
   693  	if &allgs[0] != allgptr {
   694  		atomicstorep(unsafe.Pointer(&allgptr), unsafe.Pointer(&allgs[0]))
   695  	}
   696  	atomic.Storeuintptr(&allglen, uintptr(len(allgs)))
   697  	unlock(&allglock)
   698  }
   699  
   700  // allGsSnapshot returns a snapshot of the slice of all Gs.
   701  //
   702  // The world must be stopped or allglock must be held.
   703  func allGsSnapshot() []*g {
   704  	assertWorldStoppedOrLockHeld(&allglock)
   705  
   706  	// Because the world is stopped or allglock is held, allgadd
   707  	// cannot happen concurrently with this. allgs grows
   708  	// monotonically and existing entries never change, so we can
   709  	// simply return a copy of the slice header. For added safety,
   710  	// we trim everything past len because that can still change.
   711  	return allgs[:len(allgs):len(allgs)]
   712  }
   713  
   714  // atomicAllG returns &allgs[0] and len(allgs) for use with atomicAllGIndex.
   715  func atomicAllG() (**g, uintptr) {
   716  	length := atomic.Loaduintptr(&allglen)
   717  	ptr := (**g)(atomic.Loadp(unsafe.Pointer(&allgptr)))
   718  	return ptr, length
   719  }
   720  
   721  // atomicAllGIndex returns ptr[i] with the allgptr returned from atomicAllG.
   722  func atomicAllGIndex(ptr **g, i uintptr) *g {
   723  	return *(**g)(add(unsafe.Pointer(ptr), i*goarch.PtrSize))
   724  }
   725  
   726  // forEachG calls fn on every G from allgs.
   727  //
   728  // forEachG takes a lock to exclude concurrent addition of new Gs.
   729  func forEachG(fn func(gp *g)) {
   730  	lock(&allglock)
   731  	for _, gp := range allgs {
   732  		fn(gp)
   733  	}
   734  	unlock(&allglock)
   735  }
   736  
   737  // forEachGRace calls fn on every G from allgs.
   738  //
   739  // forEachGRace avoids locking, but does not exclude addition of new Gs during
   740  // execution, which may be missed.
   741  func forEachGRace(fn func(gp *g)) {
   742  	ptr, length := atomicAllG()
   743  	for i := uintptr(0); i < length; i++ {
   744  		gp := atomicAllGIndex(ptr, i)
   745  		fn(gp)
   746  	}
   747  	return
   748  }
   749  
   750  const (
   751  	// Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
   752  	// 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
   753  	_GoidCacheBatch = 16
   754  )
   755  
   756  // cpuinit sets up CPU feature flags and calls internal/cpu.Initialize. env should be the complete
   757  // value of the GODEBUG environment variable.
   758  func cpuinit(env string) {
   759  	switch GOOS {
   760  	case "aix", "darwin", "ios", "dragonfly", "freebsd", "netbsd", "openbsd", "illumos", "solaris", "linux":
   761  		cpu.DebugOptions = true
   762  	}
   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  	mallocinit()
   876  	cpuinit(godebug) // must run before alginit
   877  	randinit()       // must run before alginit, mcommoninit
   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  		_Gscanpreempted:
  1226  		if newval == oldval&^_Gscan {
  1227  			success = gp.atomicstatus.CompareAndSwap(oldval, newval)
  1228  		}
  1229  	}
  1230  	if !success {
  1231  		print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
  1232  		dumpgstatus(gp)
  1233  		throw("casfrom_Gscanstatus: gp->status is not in scan state")
  1234  	}
  1235  	releaseLockRankAndM(lockRankGscan)
  1236  }
  1237  
  1238  // This will return false if the gp is not in the expected status and the cas fails.
  1239  // This acts like a lock acquire while the casfromgstatus acts like a lock release.
  1240  func castogscanstatus(gp *g, oldval, newval uint32) bool {
  1241  	switch oldval {
  1242  	case _Grunnable,
  1243  		_Grunning,
  1244  		_Gwaiting,
  1245  		_Gsyscall:
  1246  		if newval == oldval|_Gscan {
  1247  			r := gp.atomicstatus.CompareAndSwap(oldval, newval)
  1248  			if r {
  1249  				acquireLockRankAndM(lockRankGscan)
  1250  			}
  1251  			return r
  1252  
  1253  		}
  1254  	}
  1255  	print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1256  	throw("castogscanstatus")
  1257  	panic("not reached")
  1258  }
  1259  
  1260  // casgstatusAlwaysTrack is a debug flag that causes casgstatus to always track
  1261  // various latencies on every transition instead of sampling them.
  1262  var casgstatusAlwaysTrack = false
  1263  
  1264  // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
  1265  // and casfrom_Gscanstatus instead.
  1266  // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
  1267  // put it in the Gscan state is finished.
  1268  //
  1269  //go:nosplit
  1270  func casgstatus(gp *g, oldval, newval uint32) {
  1271  	if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
  1272  		systemstack(func() {
  1273  			// Call on the systemstack to prevent print and throw from counting
  1274  			// against the nosplit stack reservation.
  1275  			print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
  1276  			throw("casgstatus: bad incoming values")
  1277  		})
  1278  	}
  1279  
  1280  	lockWithRankMayAcquire(nil, lockRankGscan)
  1281  
  1282  	// See https://golang.org/cl/21503 for justification of the yield delay.
  1283  	const yieldDelay = 5 * 1000
  1284  	var nextYield int64
  1285  
  1286  	// loop if gp->atomicstatus is in a scan state giving
  1287  	// GC time to finish and change the state to oldval.
  1288  	for i := 0; !gp.atomicstatus.CompareAndSwap(oldval, newval); i++ {
  1289  		if oldval == _Gwaiting && gp.atomicstatus.Load() == _Grunnable {
  1290  			systemstack(func() {
  1291  				// Call on the systemstack to prevent throw from counting
  1292  				// against the nosplit stack reservation.
  1293  				throw("casgstatus: waiting for Gwaiting but is Grunnable")
  1294  			})
  1295  		}
  1296  		if i == 0 {
  1297  			nextYield = nanotime() + yieldDelay
  1298  		}
  1299  		if nanotime() < nextYield {
  1300  			for x := 0; x < 10 && gp.atomicstatus.Load() != oldval; x++ {
  1301  				procyield(1)
  1302  			}
  1303  		} else {
  1304  			osyield()
  1305  			nextYield = nanotime() + yieldDelay/2
  1306  		}
  1307  	}
  1308  
  1309  	if gp.bubble != nil {
  1310  		systemstack(func() {
  1311  			gp.bubble.changegstatus(gp, oldval, newval)
  1312  		})
  1313  	}
  1314  
  1315  	if oldval == _Grunning {
  1316  		// Track every gTrackingPeriod time a goroutine transitions out of running.
  1317  		if casgstatusAlwaysTrack || gp.trackingSeq%gTrackingPeriod == 0 {
  1318  			gp.tracking = true
  1319  		}
  1320  		gp.trackingSeq++
  1321  	}
  1322  	if !gp.tracking {
  1323  		return
  1324  	}
  1325  
  1326  	// Handle various kinds of tracking.
  1327  	//
  1328  	// Currently:
  1329  	// - Time spent in runnable.
  1330  	// - Time spent blocked on a sync.Mutex or sync.RWMutex.
  1331  	switch oldval {
  1332  	case _Grunnable:
  1333  		// We transitioned out of runnable, so measure how much
  1334  		// time we spent in this state and add it to
  1335  		// runnableTime.
  1336  		now := nanotime()
  1337  		gp.runnableTime += now - gp.trackingStamp
  1338  		gp.trackingStamp = 0
  1339  	case _Gwaiting:
  1340  		if !gp.waitreason.isMutexWait() {
  1341  			// Not blocking on a lock.
  1342  			break
  1343  		}
  1344  		// Blocking on a lock, measure it. Note that because we're
  1345  		// sampling, we have to multiply by our sampling period to get
  1346  		// a more representative estimate of the absolute value.
  1347  		// gTrackingPeriod also represents an accurate sampling period
  1348  		// because we can only enter this state from _Grunning.
  1349  		now := nanotime()
  1350  		sched.totalMutexWaitTime.Add((now - gp.trackingStamp) * gTrackingPeriod)
  1351  		gp.trackingStamp = 0
  1352  	}
  1353  	switch newval {
  1354  	case _Gwaiting:
  1355  		if !gp.waitreason.isMutexWait() {
  1356  			// Not blocking on a lock.
  1357  			break
  1358  		}
  1359  		// Blocking on a lock. Write down the timestamp.
  1360  		now := nanotime()
  1361  		gp.trackingStamp = now
  1362  	case _Grunnable:
  1363  		// We just transitioned into runnable, so record what
  1364  		// time that happened.
  1365  		now := nanotime()
  1366  		gp.trackingStamp = now
  1367  	case _Grunning:
  1368  		// We're transitioning into running, so turn off
  1369  		// tracking and record how much time we spent in
  1370  		// runnable.
  1371  		gp.tracking = false
  1372  		sched.timeToRun.record(gp.runnableTime)
  1373  		gp.runnableTime = 0
  1374  	}
  1375  }
  1376  
  1377  // casGToWaiting transitions gp from old to _Gwaiting, and sets the wait reason.
  1378  //
  1379  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1380  func casGToWaiting(gp *g, old uint32, reason waitReason) {
  1381  	// Set the wait reason before calling casgstatus, because casgstatus will use it.
  1382  	gp.waitreason = reason
  1383  	casgstatus(gp, old, _Gwaiting)
  1384  }
  1385  
  1386  // casGToWaitingForSuspendG transitions gp from old to _Gwaiting, and sets the wait reason.
  1387  // The wait reason must be a valid isWaitingForSuspendG wait reason.
  1388  //
  1389  // Use this over casgstatus when possible to ensure that a waitreason is set.
  1390  func casGToWaitingForSuspendG(gp *g, old uint32, reason waitReason) {
  1391  	if !reason.isWaitingForSuspendG() {
  1392  		throw("casGToWaitingForSuspendG with non-isWaitingForSuspendG wait reason")
  1393  	}
  1394  	casGToWaiting(gp, old, reason)
  1395  }
  1396  
  1397  // casGToPreemptScan transitions gp from _Grunning to _Gscan|_Gpreempted.
  1398  //
  1399  // TODO(austin): This is the only status operation that both changes
  1400  // the status and locks the _Gscan bit. Rethink this.
  1401  func casGToPreemptScan(gp *g, old, new uint32) {
  1402  	if old != _Grunning || new != _Gscan|_Gpreempted {
  1403  		throw("bad g transition")
  1404  	}
  1405  	acquireLockRankAndM(lockRankGscan)
  1406  	for !gp.atomicstatus.CompareAndSwap(_Grunning, _Gscan|_Gpreempted) {
  1407  	}
  1408  	// We never notify gp.bubble that the goroutine state has moved
  1409  	// from _Grunning to _Gpreempted. We call bubble.changegstatus
  1410  	// after status changes happen, but doing so here would violate the
  1411  	// ordering between the gscan and synctest locks. The bubble doesn't
  1412  	// distinguish between _Grunning and _Gpreempted anyway, so not
  1413  	// notifying it is fine.
  1414  }
  1415  
  1416  // casGFromPreempted attempts to transition gp from _Gpreempted to
  1417  // _Gwaiting. If successful, the caller is responsible for
  1418  // re-scheduling gp.
  1419  func casGFromPreempted(gp *g, old, new uint32) bool {
  1420  	if old != _Gpreempted || new != _Gwaiting {
  1421  		throw("bad g transition")
  1422  	}
  1423  	gp.waitreason = waitReasonPreempted
  1424  	if !gp.atomicstatus.CompareAndSwap(_Gpreempted, _Gwaiting) {
  1425  		return false
  1426  	}
  1427  	if bubble := gp.bubble; bubble != nil {
  1428  		bubble.changegstatus(gp, _Gpreempted, _Gwaiting)
  1429  	}
  1430  	return true
  1431  }
  1432  
  1433  // stwReason is an enumeration of reasons the world is stopping.
  1434  type stwReason uint8
  1435  
  1436  // Reasons to stop-the-world.
  1437  //
  1438  // Avoid reusing reasons and add new ones instead.
  1439  const (
  1440  	stwUnknown                     stwReason = iota // "unknown"
  1441  	stwGCMarkTerm                                   // "GC mark termination"
  1442  	stwGCSweepTerm                                  // "GC sweep termination"
  1443  	stwWriteHeapDump                                // "write heap dump"
  1444  	stwGoroutineProfile                             // "goroutine profile"
  1445  	stwGoroutineProfileCleanup                      // "goroutine profile cleanup"
  1446  	stwAllGoroutinesStack                           // "all goroutines stack trace"
  1447  	stwReadMemStats                                 // "read mem stats"
  1448  	stwAllThreadsSyscall                            // "AllThreadsSyscall"
  1449  	stwGOMAXPROCS                                   // "GOMAXPROCS"
  1450  	stwStartTrace                                   // "start trace"
  1451  	stwStopTrace                                    // "stop trace"
  1452  	stwForTestCountPagesInUse                       // "CountPagesInUse (test)"
  1453  	stwForTestReadMetricsSlow                       // "ReadMetricsSlow (test)"
  1454  	stwForTestReadMemStatsSlow                      // "ReadMemStatsSlow (test)"
  1455  	stwForTestPageCachePagesLeaked                  // "PageCachePagesLeaked (test)"
  1456  	stwForTestResetDebugLog                         // "ResetDebugLog (test)"
  1457  )
  1458  
  1459  func (r stwReason) String() string {
  1460  	return stwReasonStrings[r]
  1461  }
  1462  
  1463  func (r stwReason) isGC() bool {
  1464  	return r == stwGCMarkTerm || r == stwGCSweepTerm
  1465  }
  1466  
  1467  // If you add to this list, also add it to src/internal/trace/parser.go.
  1468  // If you change the values of any of the stw* constants, bump the trace
  1469  // version number and make a copy of this.
  1470  var stwReasonStrings = [...]string{
  1471  	stwUnknown:                     "unknown",
  1472  	stwGCMarkTerm:                  "GC mark termination",
  1473  	stwGCSweepTerm:                 "GC sweep termination",
  1474  	stwWriteHeapDump:               "write heap dump",
  1475  	stwGoroutineProfile:            "goroutine profile",
  1476  	stwGoroutineProfileCleanup:     "goroutine profile cleanup",
  1477  	stwAllGoroutinesStack:          "all goroutines stack trace",
  1478  	stwReadMemStats:                "read mem stats",
  1479  	stwAllThreadsSyscall:           "AllThreadsSyscall",
  1480  	stwGOMAXPROCS:                  "GOMAXPROCS",
  1481  	stwStartTrace:                  "start trace",
  1482  	stwStopTrace:                   "stop trace",
  1483  	stwForTestCountPagesInUse:      "CountPagesInUse (test)",
  1484  	stwForTestReadMetricsSlow:      "ReadMetricsSlow (test)",
  1485  	stwForTestReadMemStatsSlow:     "ReadMemStatsSlow (test)",
  1486  	stwForTestPageCachePagesLeaked: "PageCachePagesLeaked (test)",
  1487  	stwForTestResetDebugLog:        "ResetDebugLog (test)",
  1488  }
  1489  
  1490  // worldStop provides context from the stop-the-world required by the
  1491  // start-the-world.
  1492  type worldStop struct {
  1493  	reason           stwReason
  1494  	startedStopping  int64
  1495  	finishedStopping int64
  1496  	stoppingCPUTime  int64
  1497  }
  1498  
  1499  // Temporary variable for stopTheWorld, when it can't write to the stack.
  1500  //
  1501  // Protected by worldsema.
  1502  var stopTheWorldContext worldStop
  1503  
  1504  // stopTheWorld stops all P's from executing goroutines, interrupting
  1505  // all goroutines at GC safe points and records reason as the reason
  1506  // for the stop. On return, only the current goroutine's P is running.
  1507  // stopTheWorld must not be called from a system stack and the caller
  1508  // must not hold worldsema. The caller must call startTheWorld when
  1509  // other P's should resume execution.
  1510  //
  1511  // stopTheWorld is safe for multiple goroutines to call at the
  1512  // same time. Each will execute its own stop, and the stops will
  1513  // be serialized.
  1514  //
  1515  // This is also used by routines that do stack dumps. If the system is
  1516  // in panic or being exited, this may not reliably stop all
  1517  // goroutines.
  1518  //
  1519  // Returns the STW context. When starting the world, this context must be
  1520  // passed to startTheWorld.
  1521  func stopTheWorld(reason stwReason) worldStop {
  1522  	semacquire(&worldsema)
  1523  	gp := getg()
  1524  	gp.m.preemptoff = reason.String()
  1525  	systemstack(func() {
  1526  		stopTheWorldContext = stopTheWorldWithSema(reason) // avoid write to stack
  1527  	})
  1528  	return stopTheWorldContext
  1529  }
  1530  
  1531  // startTheWorld undoes the effects of stopTheWorld.
  1532  //
  1533  // w must be the worldStop returned by stopTheWorld.
  1534  func startTheWorld(w worldStop) {
  1535  	systemstack(func() { startTheWorldWithSema(0, w) })
  1536  
  1537  	// worldsema must be held over startTheWorldWithSema to ensure
  1538  	// gomaxprocs cannot change while worldsema is held.
  1539  	//
  1540  	// Release worldsema with direct handoff to the next waiter, but
  1541  	// acquirem so that semrelease1 doesn't try to yield our time.
  1542  	//
  1543  	// Otherwise if e.g. ReadMemStats is being called in a loop,
  1544  	// it might stomp on other attempts to stop the world, such as
  1545  	// for starting or ending GC. The operation this blocks is
  1546  	// so heavy-weight that we should just try to be as fair as
  1547  	// possible here.
  1548  	//
  1549  	// We don't want to just allow us to get preempted between now
  1550  	// and releasing the semaphore because then we keep everyone
  1551  	// (including, for example, GCs) waiting longer.
  1552  	mp := acquirem()
  1553  	mp.preemptoff = ""
  1554  	semrelease1(&worldsema, true, 0)
  1555  	releasem(mp)
  1556  }
  1557  
  1558  // stopTheWorldGC has the same effect as stopTheWorld, but blocks
  1559  // until the GC is not running. It also blocks a GC from starting
  1560  // until startTheWorldGC is called.
  1561  func stopTheWorldGC(reason stwReason) worldStop {
  1562  	semacquire(&gcsema)
  1563  	return stopTheWorld(reason)
  1564  }
  1565  
  1566  // startTheWorldGC undoes the effects of stopTheWorldGC.
  1567  //
  1568  // w must be the worldStop returned by stopTheWorld.
  1569  func startTheWorldGC(w worldStop) {
  1570  	startTheWorld(w)
  1571  	semrelease(&gcsema)
  1572  }
  1573  
  1574  // Holding worldsema grants an M the right to try to stop the world.
  1575  var worldsema uint32 = 1
  1576  
  1577  // Holding gcsema grants the M the right to block a GC, and blocks
  1578  // until the current GC is done. In particular, it prevents gomaxprocs
  1579  // from changing concurrently.
  1580  //
  1581  // TODO(mknyszek): Once gomaxprocs and the execution tracer can handle
  1582  // being changed/enabled during a GC, remove this.
  1583  var gcsema uint32 = 1
  1584  
  1585  // stopTheWorldWithSema is the core implementation of stopTheWorld.
  1586  // The caller is responsible for acquiring worldsema and disabling
  1587  // preemption first and then should stopTheWorldWithSema on the system
  1588  // stack:
  1589  //
  1590  //	semacquire(&worldsema, 0)
  1591  //	m.preemptoff = "reason"
  1592  //	var stw worldStop
  1593  //	systemstack(func() {
  1594  //		stw = stopTheWorldWithSema(reason)
  1595  //	})
  1596  //
  1597  // When finished, the caller must either call startTheWorld or undo
  1598  // these three operations separately:
  1599  //
  1600  //	m.preemptoff = ""
  1601  //	systemstack(func() {
  1602  //		now = startTheWorldWithSema(stw)
  1603  //	})
  1604  //	semrelease(&worldsema)
  1605  //
  1606  // It is allowed to acquire worldsema once and then execute multiple
  1607  // startTheWorldWithSema/stopTheWorldWithSema pairs.
  1608  // Other P's are able to execute between successive calls to
  1609  // startTheWorldWithSema and stopTheWorldWithSema.
  1610  // Holding worldsema causes any other goroutines invoking
  1611  // stopTheWorld to block.
  1612  //
  1613  // Returns the STW context. When starting the world, this context must be
  1614  // passed to startTheWorldWithSema.
  1615  //
  1616  //go:systemstack
  1617  func stopTheWorldWithSema(reason stwReason) worldStop {
  1618  	// Mark the goroutine which called stopTheWorld preemptible so its
  1619  	// stack may be scanned by the GC or observed by the execution tracer.
  1620  	//
  1621  	// This lets a mark worker scan us or the execution tracer take our
  1622  	// stack while we try to stop the world since otherwise we could get
  1623  	// in a mutual preemption deadlock.
  1624  	//
  1625  	// We must not modify anything on the G stack because a stack shrink
  1626  	// may occur, now that we switched to _Gwaiting, specifically if we're
  1627  	// doing this during the mark phase (mark termination excepted, since
  1628  	// we know that stack scanning is done by that point). A stack shrink
  1629  	// is otherwise OK though because in order to return from this function
  1630  	// (and to leave the system stack) we must have preempted all
  1631  	// goroutines, including any attempting to scan our stack, in which
  1632  	// case, any stack shrinking will have already completed by the time we
  1633  	// exit.
  1634  	//
  1635  	// N.B. The execution tracer is not aware of this status transition and
  1636  	// andles it specially based on the wait reason.
  1637  	casGToWaitingForSuspendG(getg().m.curg, _Grunning, waitReasonStoppingTheWorld)
  1638  
  1639  	trace := traceAcquire()
  1640  	if trace.ok() {
  1641  		trace.STWStart(reason)
  1642  		traceRelease(trace)
  1643  	}
  1644  	gp := getg()
  1645  
  1646  	// If we hold a lock, then we won't be able to stop another M
  1647  	// that is blocked trying to acquire the lock.
  1648  	if gp.m.locks > 0 {
  1649  		throw("stopTheWorld: holding locks")
  1650  	}
  1651  
  1652  	lock(&sched.lock)
  1653  	start := nanotime() // exclude time waiting for sched.lock from start and total time metrics.
  1654  	sched.stopwait = gomaxprocs
  1655  	sched.gcwaiting.Store(true)
  1656  	preemptall()
  1657  	// stop current P
  1658  	gp.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
  1659  	gp.m.p.ptr().gcStopTime = start
  1660  	sched.stopwait--
  1661  	// try to retake all P's in Psyscall status
  1662  	trace = traceAcquire()
  1663  	for _, pp := range allp {
  1664  		s := pp.status
  1665  		if s == _Psyscall && atomic.Cas(&pp.status, s, _Pgcstop) {
  1666  			if trace.ok() {
  1667  				trace.ProcSteal(pp, false)
  1668  			}
  1669  			pp.syscalltick++
  1670  			pp.gcStopTime = nanotime()
  1671  			sched.stopwait--
  1672  		}
  1673  	}
  1674  	if trace.ok() {
  1675  		traceRelease(trace)
  1676  	}
  1677  
  1678  	// stop idle P's
  1679  	now := nanotime()
  1680  	for {
  1681  		pp, _ := pidleget(now)
  1682  		if pp == nil {
  1683  			break
  1684  		}
  1685  		pp.status = _Pgcstop
  1686  		pp.gcStopTime = nanotime()
  1687  		sched.stopwait--
  1688  	}
  1689  	wait := sched.stopwait > 0
  1690  	unlock(&sched.lock)
  1691  
  1692  	// wait for remaining P's to stop voluntarily
  1693  	if wait {
  1694  		for {
  1695  			// wait for 100us, then try to re-preempt in case of any races
  1696  			if notetsleep(&sched.stopnote, 100*1000) {
  1697  				noteclear(&sched.stopnote)
  1698  				break
  1699  			}
  1700  			preemptall()
  1701  		}
  1702  	}
  1703  
  1704  	finish := nanotime()
  1705  	startTime := finish - start
  1706  	if reason.isGC() {
  1707  		sched.stwStoppingTimeGC.record(startTime)
  1708  	} else {
  1709  		sched.stwStoppingTimeOther.record(startTime)
  1710  	}
  1711  
  1712  	// Double-check we actually stopped everything, and all the invariants hold.
  1713  	// Also accumulate all the time spent by each P in _Pgcstop up to the point
  1714  	// where everything was stopped. This will be accumulated into the total pause
  1715  	// CPU time by the caller.
  1716  	stoppingCPUTime := int64(0)
  1717  	bad := ""
  1718  	if sched.stopwait != 0 {
  1719  		bad = "stopTheWorld: not stopped (stopwait != 0)"
  1720  	} else {
  1721  		for _, pp := range allp {
  1722  			if pp.status != _Pgcstop {
  1723  				bad = "stopTheWorld: not stopped (status != _Pgcstop)"
  1724  			}
  1725  			if pp.gcStopTime == 0 && bad == "" {
  1726  				bad = "stopTheWorld: broken CPU time accounting"
  1727  			}
  1728  			stoppingCPUTime += finish - pp.gcStopTime
  1729  			pp.gcStopTime = 0
  1730  		}
  1731  	}
  1732  	if freezing.Load() {
  1733  		// Some other thread is panicking. This can cause the
  1734  		// sanity checks above to fail if the panic happens in
  1735  		// the signal handler on a stopped thread. Either way,
  1736  		// we should halt this thread.
  1737  		lock(&deadlock)
  1738  		lock(&deadlock)
  1739  	}
  1740  	if bad != "" {
  1741  		throw(bad)
  1742  	}
  1743  
  1744  	worldStopped()
  1745  
  1746  	// Switch back to _Grunning, now that the world is stopped.
  1747  	casgstatus(getg().m.curg, _Gwaiting, _Grunning)
  1748  
  1749  	return worldStop{
  1750  		reason:           reason,
  1751  		startedStopping:  start,
  1752  		finishedStopping: finish,
  1753  		stoppingCPUTime:  stoppingCPUTime,
  1754  	}
  1755  }
  1756  
  1757  // reason is the same STW reason passed to stopTheWorld. start is the start
  1758  // time returned by stopTheWorld.
  1759  //
  1760  // now is the current time; prefer to pass 0 to capture a fresh timestamp.
  1761  //
  1762  // stattTheWorldWithSema returns now.
  1763  func startTheWorldWithSema(now int64, w worldStop) int64 {
  1764  	assertWorldStopped()
  1765  
  1766  	mp := acquirem() // disable preemption because it can be holding p in a local var
  1767  	if netpollinited() {
  1768  		list, delta := netpoll(0) // non-blocking
  1769  		injectglist(&list)
  1770  		netpollAdjustWaiters(delta)
  1771  	}
  1772  	lock(&sched.lock)
  1773  
  1774  	procs := gomaxprocs
  1775  	if newprocs != 0 {
  1776  		procs = newprocs
  1777  		newprocs = 0
  1778  	}
  1779  	p1 := procresize(procs)
  1780  	sched.gcwaiting.Store(false)
  1781  	if sched.sysmonwait.Load() {
  1782  		sched.sysmonwait.Store(false)
  1783  		notewakeup(&sched.sysmonnote)
  1784  	}
  1785  	unlock(&sched.lock)
  1786  
  1787  	worldStarted()
  1788  
  1789  	for p1 != nil {
  1790  		p := p1
  1791  		p1 = p1.link.ptr()
  1792  		if p.m != 0 {
  1793  			mp := p.m.ptr()
  1794  			p.m = 0
  1795  			if mp.nextp != 0 {
  1796  				throw("startTheWorld: inconsistent mp->nextp")
  1797  			}
  1798  			mp.nextp.set(p)
  1799  			notewakeup(&mp.park)
  1800  		} else {
  1801  			// Start M to run P.  Do not start another M below.
  1802  			newm(nil, p, -1)
  1803  		}
  1804  	}
  1805  
  1806  	// Capture start-the-world time before doing clean-up tasks.
  1807  	if now == 0 {
  1808  		now = nanotime()
  1809  	}
  1810  	totalTime := now - w.startedStopping
  1811  	if w.reason.isGC() {
  1812  		sched.stwTotalTimeGC.record(totalTime)
  1813  	} else {
  1814  		sched.stwTotalTimeOther.record(totalTime)
  1815  	}
  1816  	trace := traceAcquire()
  1817  	if trace.ok() {
  1818  		trace.STWDone()
  1819  		traceRelease(trace)
  1820  	}
  1821  
  1822  	// Wakeup an additional proc in case we have excessive runnable goroutines
  1823  	// in local queues or in the global queue. If we don't, the proc will park itself.
  1824  	// If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
  1825  	wakep()
  1826  
  1827  	releasem(mp)
  1828  
  1829  	return now
  1830  }
  1831  
  1832  // usesLibcall indicates whether this runtime performs system calls
  1833  // via libcall.
  1834  func usesLibcall() bool {
  1835  	switch GOOS {
  1836  	case "aix", "darwin", "illumos", "ios", "solaris", "windows":
  1837  		return true
  1838  	case "openbsd":
  1839  		return GOARCH != "mips64"
  1840  	}
  1841  	return false
  1842  }
  1843  
  1844  // mStackIsSystemAllocated indicates whether this runtime starts on a
  1845  // system-allocated stack.
  1846  func mStackIsSystemAllocated() bool {
  1847  	switch GOOS {
  1848  	case "aix", "darwin", "plan9", "illumos", "ios", "solaris", "windows":
  1849  		return true
  1850  	case "openbsd":
  1851  		return GOARCH != "mips64"
  1852  	}
  1853  	return false
  1854  }
  1855  
  1856  // mstart is the entry-point for new Ms.
  1857  // It is written in assembly, uses ABI0, is marked TOPFRAME, and calls mstart0.
  1858  func mstart()
  1859  
  1860  // mstart0 is the Go entry-point for new Ms.
  1861  // This must not split the stack because we may not even have stack
  1862  // bounds set up yet.
  1863  //
  1864  // May run during STW (because it doesn't have a P yet), so write
  1865  // barriers are not allowed.
  1866  //
  1867  //go:nosplit
  1868  //go:nowritebarrierrec
  1869  func mstart0() {
  1870  	gp := getg()
  1871  
  1872  	osStack := gp.stack.lo == 0
  1873  	if osStack {
  1874  		// Initialize stack bounds from system stack.
  1875  		// Cgo may have left stack size in stack.hi.
  1876  		// minit may update the stack bounds.
  1877  		//
  1878  		// Note: these bounds may not be very accurate.
  1879  		// We set hi to &size, but there are things above
  1880  		// it. The 1024 is supposed to compensate this,
  1881  		// but is somewhat arbitrary.
  1882  		size := gp.stack.hi
  1883  		if size == 0 {
  1884  			size = 16384 * sys.StackGuardMultiplier
  1885  		}
  1886  		gp.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
  1887  		gp.stack.lo = gp.stack.hi - size + 1024
  1888  	}
  1889  	// Initialize stack guard so that we can start calling regular
  1890  	// Go code.
  1891  	gp.stackguard0 = gp.stack.lo + stackGuard
  1892  	// This is the g0, so we can also call go:systemstack
  1893  	// functions, which check stackguard1.
  1894  	gp.stackguard1 = gp.stackguard0
  1895  	mstart1()
  1896  
  1897  	// Exit this thread.
  1898  	if mStackIsSystemAllocated() {
  1899  		// Windows, Solaris, illumos, Darwin, AIX and Plan 9 always system-allocate
  1900  		// the stack, but put it in gp.stack before mstart,
  1901  		// so the logic above hasn't set osStack yet.
  1902  		osStack = true
  1903  	}
  1904  	mexit(osStack)
  1905  }
  1906  
  1907  // The go:noinline is to guarantee the sys.GetCallerPC/sys.GetCallerSP below are safe,
  1908  // so that we can set up g0.sched to return to the call of mstart1 above.
  1909  //
  1910  //go:noinline
  1911  func mstart1() {
  1912  	gp := getg()
  1913  
  1914  	if gp != gp.m.g0 {
  1915  		throw("bad runtime·mstart")
  1916  	}
  1917  
  1918  	// Set up m.g0.sched as a label returning to just
  1919  	// after the mstart1 call in mstart0 above, for use by goexit0 and mcall.
  1920  	// We're never coming back to mstart1 after we call schedule,
  1921  	// so other calls can reuse the current frame.
  1922  	// And goexit0 does a gogo that needs to return from mstart1
  1923  	// and let mstart0 exit the thread.
  1924  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1925  	gp.sched.pc = sys.GetCallerPC()
  1926  	gp.sched.sp = sys.GetCallerSP()
  1927  
  1928  	asminit()
  1929  	minit()
  1930  
  1931  	// Install signal handlers; after minit so that minit can
  1932  	// prepare the thread to be able to handle the signals.
  1933  	if gp.m == &m0 {
  1934  		mstartm0()
  1935  	}
  1936  
  1937  	if debug.dataindependenttiming == 1 {
  1938  		sys.EnableDIT()
  1939  	}
  1940  
  1941  	if fn := gp.m.mstartfn; fn != nil {
  1942  		fn()
  1943  	}
  1944  
  1945  	if gp.m != &m0 {
  1946  		acquirep(gp.m.nextp.ptr())
  1947  		gp.m.nextp = 0
  1948  	}
  1949  	schedule()
  1950  }
  1951  
  1952  // mstartm0 implements part of mstart1 that only runs on the m0.
  1953  //
  1954  // Write barriers are allowed here because we know the GC can't be
  1955  // running yet, so they'll be no-ops.
  1956  //
  1957  //go:yeswritebarrierrec
  1958  func mstartm0() {
  1959  	// Create an extra M for callbacks on threads not created by Go.
  1960  	// An extra M is also needed on Windows for callbacks created by
  1961  	// syscall.NewCallback. See issue #6751 for details.
  1962  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  1963  		cgoHasExtraM = true
  1964  		newextram()
  1965  	}
  1966  	initsig(false)
  1967  }
  1968  
  1969  // mPark causes a thread to park itself, returning once woken.
  1970  //
  1971  //go:nosplit
  1972  func mPark() {
  1973  	gp := getg()
  1974  	notesleep(&gp.m.park)
  1975  	noteclear(&gp.m.park)
  1976  }
  1977  
  1978  // mexit tears down and exits the current thread.
  1979  //
  1980  // Don't call this directly to exit the thread, since it must run at
  1981  // the top of the thread stack. Instead, use gogo(&gp.m.g0.sched) to
  1982  // unwind the stack to the point that exits the thread.
  1983  //
  1984  // It is entered with m.p != nil, so write barriers are allowed. It
  1985  // will release the P before exiting.
  1986  //
  1987  //go:yeswritebarrierrec
  1988  func mexit(osStack bool) {
  1989  	mp := getg().m
  1990  
  1991  	if mp == &m0 {
  1992  		// This is the main thread. Just wedge it.
  1993  		//
  1994  		// On Linux, exiting the main thread puts the process
  1995  		// into a non-waitable zombie state. On Plan 9,
  1996  		// exiting the main thread unblocks wait even though
  1997  		// other threads are still running. On Solaris we can
  1998  		// neither exitThread nor return from mstart. Other
  1999  		// bad things probably happen on other platforms.
  2000  		//
  2001  		// We could try to clean up this M more before wedging
  2002  		// it, but that complicates signal handling.
  2003  		handoffp(releasep())
  2004  		lock(&sched.lock)
  2005  		sched.nmfreed++
  2006  		checkdead()
  2007  		unlock(&sched.lock)
  2008  		mPark()
  2009  		throw("locked m0 woke up")
  2010  	}
  2011  
  2012  	sigblock(true)
  2013  	unminit()
  2014  
  2015  	// Free the gsignal stack.
  2016  	if mp.gsignal != nil {
  2017  		stackfree(mp.gsignal.stack)
  2018  		if valgrindenabled {
  2019  			valgrindDeregisterStack(mp.gsignal.valgrindStackID)
  2020  			mp.gsignal.valgrindStackID = 0
  2021  		}
  2022  		// On some platforms, when calling into VDSO (e.g. nanotime)
  2023  		// we store our g on the gsignal stack, if there is one.
  2024  		// Now the stack is freed, unlink it from the m, so we
  2025  		// won't write to it when calling VDSO code.
  2026  		mp.gsignal = nil
  2027  	}
  2028  
  2029  	// Free vgetrandom state.
  2030  	vgetrandomDestroy(mp)
  2031  
  2032  	// Remove m from allm.
  2033  	lock(&sched.lock)
  2034  	for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
  2035  		if *pprev == mp {
  2036  			*pprev = mp.alllink
  2037  			goto found
  2038  		}
  2039  	}
  2040  	throw("m not found in allm")
  2041  found:
  2042  	// Events must not be traced after this point.
  2043  
  2044  	// Delay reaping m until it's done with the stack.
  2045  	//
  2046  	// Put mp on the free list, though it will not be reaped while freeWait
  2047  	// is freeMWait. mp is no longer reachable via allm, so even if it is
  2048  	// on an OS stack, we must keep a reference to mp alive so that the GC
  2049  	// doesn't free mp while we are still using it.
  2050  	//
  2051  	// Note that the free list must not be linked through alllink because
  2052  	// some functions walk allm without locking, so may be using alllink.
  2053  	//
  2054  	// N.B. It's important that the M appears on the free list simultaneously
  2055  	// with it being removed so that the tracer can find it.
  2056  	mp.freeWait.Store(freeMWait)
  2057  	mp.freelink = sched.freem
  2058  	sched.freem = mp
  2059  	unlock(&sched.lock)
  2060  
  2061  	atomic.Xadd64(&ncgocall, int64(mp.ncgocall))
  2062  	sched.totalRuntimeLockWaitTime.Add(mp.mLockProfile.waitTime.Load())
  2063  
  2064  	// Release the P.
  2065  	handoffp(releasep())
  2066  	// After this point we must not have write barriers.
  2067  
  2068  	// Invoke the deadlock detector. This must happen after
  2069  	// handoffp because it may have started a new M to take our
  2070  	// P's work.
  2071  	lock(&sched.lock)
  2072  	sched.nmfreed++
  2073  	checkdead()
  2074  	unlock(&sched.lock)
  2075  
  2076  	if GOOS == "darwin" || GOOS == "ios" {
  2077  		// Make sure pendingPreemptSignals is correct when an M exits.
  2078  		// For #41702.
  2079  		if mp.signalPending.Load() != 0 {
  2080  			pendingPreemptSignals.Add(-1)
  2081  		}
  2082  	}
  2083  
  2084  	// Destroy all allocated resources. After this is called, we may no
  2085  	// longer take any locks.
  2086  	mdestroy(mp)
  2087  
  2088  	if osStack {
  2089  		// No more uses of mp, so it is safe to drop the reference.
  2090  		mp.freeWait.Store(freeMRef)
  2091  
  2092  		// Return from mstart and let the system thread
  2093  		// library free the g0 stack and terminate the thread.
  2094  		return
  2095  	}
  2096  
  2097  	// mstart is the thread's entry point, so there's nothing to
  2098  	// return to. Exit the thread directly. exitThread will clear
  2099  	// m.freeWait when it's done with the stack and the m can be
  2100  	// reaped.
  2101  	exitThread(&mp.freeWait)
  2102  }
  2103  
  2104  // forEachP calls fn(p) for every P p when p reaches a GC safe point.
  2105  // If a P is currently executing code, this will bring the P to a GC
  2106  // safe point and execute fn on that P. If the P is not executing code
  2107  // (it is idle or in a syscall), this will call fn(p) directly while
  2108  // preventing the P from exiting its state. This does not ensure that
  2109  // fn will run on every CPU executing Go code, but it acts as a global
  2110  // memory barrier. GC uses this as a "ragged barrier."
  2111  //
  2112  // The caller must hold worldsema. fn must not refer to any
  2113  // part of the current goroutine's stack, since the GC may move it.
  2114  func forEachP(reason waitReason, fn func(*p)) {
  2115  	systemstack(func() {
  2116  		gp := getg().m.curg
  2117  		// Mark the user stack as preemptible so that it may be scanned
  2118  		// by the GC or observed by the execution tracer. Otherwise, our
  2119  		// attempt to force all P's to a safepoint could result in a
  2120  		// deadlock as we attempt to preempt a goroutine that's trying
  2121  		// to preempt us (e.g. for a stack scan).
  2122  		//
  2123  		// We must not modify anything on the G stack because a stack shrink
  2124  		// may occur. A stack shrink is otherwise OK though because in order
  2125  		// to return from this function (and to leave the system stack) we
  2126  		// must have preempted all goroutines, including any attempting
  2127  		// to scan our stack, in which case, any stack shrinking will
  2128  		// have already completed by the time we exit.
  2129  		//
  2130  		// N.B. The execution tracer is not aware of this status
  2131  		// transition and handles it specially based on the
  2132  		// wait reason.
  2133  		casGToWaitingForSuspendG(gp, _Grunning, reason)
  2134  		forEachPInternal(fn)
  2135  		casgstatus(gp, _Gwaiting, _Grunning)
  2136  	})
  2137  }
  2138  
  2139  // forEachPInternal calls fn(p) for every P p when p reaches a GC safe point.
  2140  // It is the internal implementation of forEachP.
  2141  //
  2142  // The caller must hold worldsema and either must ensure that a GC is not
  2143  // running (otherwise this may deadlock with the GC trying to preempt this P)
  2144  // or it must leave its goroutine in a preemptible state before it switches
  2145  // to the systemstack. Due to these restrictions, prefer forEachP when possible.
  2146  //
  2147  //go:systemstack
  2148  func forEachPInternal(fn func(*p)) {
  2149  	mp := acquirem()
  2150  	pp := getg().m.p.ptr()
  2151  
  2152  	lock(&sched.lock)
  2153  	if sched.safePointWait != 0 {
  2154  		throw("forEachP: sched.safePointWait != 0")
  2155  	}
  2156  	sched.safePointWait = gomaxprocs - 1
  2157  	sched.safePointFn = fn
  2158  
  2159  	// Ask all Ps to run the safe point function.
  2160  	for _, p2 := range allp {
  2161  		if p2 != pp {
  2162  			atomic.Store(&p2.runSafePointFn, 1)
  2163  		}
  2164  	}
  2165  	preemptall()
  2166  
  2167  	// Any P entering _Pidle or _Psyscall from now on will observe
  2168  	// p.runSafePointFn == 1 and will call runSafePointFn when
  2169  	// changing its status to _Pidle/_Psyscall.
  2170  
  2171  	// Run safe point function for all idle Ps. sched.pidle will
  2172  	// not change because we hold sched.lock.
  2173  	for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
  2174  		if atomic.Cas(&p.runSafePointFn, 1, 0) {
  2175  			fn(p)
  2176  			sched.safePointWait--
  2177  		}
  2178  	}
  2179  
  2180  	wait := sched.safePointWait > 0
  2181  	unlock(&sched.lock)
  2182  
  2183  	// Run fn for the current P.
  2184  	fn(pp)
  2185  
  2186  	// Force Ps currently in _Psyscall into _Pidle and hand them
  2187  	// off to induce safe point function execution.
  2188  	for _, p2 := range allp {
  2189  		s := p2.status
  2190  
  2191  		// We need to be fine-grained about tracing here, since handoffp
  2192  		// might call into the tracer, and the tracer is non-reentrant.
  2193  		trace := traceAcquire()
  2194  		if s == _Psyscall && p2.runSafePointFn == 1 && atomic.Cas(&p2.status, s, _Pidle) {
  2195  			if trace.ok() {
  2196  				// It's important that we traceRelease before we call handoffp, which may also traceAcquire.
  2197  				trace.ProcSteal(p2, false)
  2198  				traceRelease(trace)
  2199  			}
  2200  			p2.syscalltick++
  2201  			handoffp(p2)
  2202  		} else if trace.ok() {
  2203  			traceRelease(trace)
  2204  		}
  2205  	}
  2206  
  2207  	// Wait for remaining Ps to run fn.
  2208  	if wait {
  2209  		for {
  2210  			// Wait for 100us, then try to re-preempt in
  2211  			// case of any races.
  2212  			//
  2213  			// Requires system stack.
  2214  			if notetsleep(&sched.safePointNote, 100*1000) {
  2215  				noteclear(&sched.safePointNote)
  2216  				break
  2217  			}
  2218  			preemptall()
  2219  		}
  2220  	}
  2221  	if sched.safePointWait != 0 {
  2222  		throw("forEachP: not done")
  2223  	}
  2224  	for _, p2 := range allp {
  2225  		if p2.runSafePointFn != 0 {
  2226  			throw("forEachP: P did not run fn")
  2227  		}
  2228  	}
  2229  
  2230  	lock(&sched.lock)
  2231  	sched.safePointFn = nil
  2232  	unlock(&sched.lock)
  2233  	releasem(mp)
  2234  }
  2235  
  2236  // runSafePointFn runs the safe point function, if any, for this P.
  2237  // This should be called like
  2238  //
  2239  //	if getg().m.p.runSafePointFn != 0 {
  2240  //	    runSafePointFn()
  2241  //	}
  2242  //
  2243  // runSafePointFn must be checked on any transition in to _Pidle or
  2244  // _Psyscall to avoid a race where forEachP sees that the P is running
  2245  // just before the P goes into _Pidle/_Psyscall and neither forEachP
  2246  // nor the P run the safe-point function.
  2247  func runSafePointFn() {
  2248  	p := getg().m.p.ptr()
  2249  	// Resolve the race between forEachP running the safe-point
  2250  	// function on this P's behalf and this P running the
  2251  	// safe-point function directly.
  2252  	if !atomic.Cas(&p.runSafePointFn, 1, 0) {
  2253  		return
  2254  	}
  2255  	sched.safePointFn(p)
  2256  	lock(&sched.lock)
  2257  	sched.safePointWait--
  2258  	if sched.safePointWait == 0 {
  2259  		notewakeup(&sched.safePointNote)
  2260  	}
  2261  	unlock(&sched.lock)
  2262  }
  2263  
  2264  // When running with cgo, we call _cgo_thread_start
  2265  // to start threads for us so that we can play nicely with
  2266  // foreign code.
  2267  var cgoThreadStart unsafe.Pointer
  2268  
  2269  type cgothreadstart struct {
  2270  	g   guintptr
  2271  	tls *uint64
  2272  	fn  unsafe.Pointer
  2273  }
  2274  
  2275  // Allocate a new m unassociated with any thread.
  2276  // Can use p for allocation context if needed.
  2277  // fn is recorded as the new m's m.mstartfn.
  2278  // id is optional pre-allocated m ID. Omit by passing -1.
  2279  //
  2280  // This function is allowed to have write barriers even if the caller
  2281  // isn't because it borrows pp.
  2282  //
  2283  //go:yeswritebarrierrec
  2284  func allocm(pp *p, fn func(), id int64) *m {
  2285  	allocmLock.rlock()
  2286  
  2287  	// The caller owns pp, but we may borrow (i.e., acquirep) it. We must
  2288  	// disable preemption to ensure it is not stolen, which would make the
  2289  	// caller lose ownership.
  2290  	acquirem()
  2291  
  2292  	gp := getg()
  2293  	if gp.m.p == 0 {
  2294  		acquirep(pp) // temporarily borrow p for mallocs in this function
  2295  	}
  2296  
  2297  	// Release the free M list. We need to do this somewhere and
  2298  	// this may free up a stack we can use.
  2299  	if sched.freem != nil {
  2300  		lock(&sched.lock)
  2301  		var newList *m
  2302  		for freem := sched.freem; freem != nil; {
  2303  			// Wait for freeWait to indicate that freem's stack is unused.
  2304  			wait := freem.freeWait.Load()
  2305  			if wait == freeMWait {
  2306  				next := freem.freelink
  2307  				freem.freelink = newList
  2308  				newList = freem
  2309  				freem = next
  2310  				continue
  2311  			}
  2312  			// Drop any remaining trace resources.
  2313  			// Ms can continue to emit events all the way until wait != freeMWait,
  2314  			// so it's only safe to call traceThreadDestroy at this point.
  2315  			if traceEnabled() || traceShuttingDown() {
  2316  				traceThreadDestroy(freem)
  2317  			}
  2318  			// Free the stack if needed. For freeMRef, there is
  2319  			// nothing to do except drop freem from the sched.freem
  2320  			// list.
  2321  			if wait == freeMStack {
  2322  				// stackfree must be on the system stack, but allocm is
  2323  				// reachable off the system stack transitively from
  2324  				// startm.
  2325  				systemstack(func() {
  2326  					stackfree(freem.g0.stack)
  2327  					if valgrindenabled {
  2328  						valgrindDeregisterStack(freem.g0.valgrindStackID)
  2329  						freem.g0.valgrindStackID = 0
  2330  					}
  2331  				})
  2332  			}
  2333  			freem = freem.freelink
  2334  		}
  2335  		sched.freem = newList
  2336  		unlock(&sched.lock)
  2337  	}
  2338  
  2339  	mp := &new(mPadded).m
  2340  	mp.mstartfn = fn
  2341  	mcommoninit(mp, id)
  2342  
  2343  	// In case of cgo or Solaris or illumos or Darwin, pthread_create will make us a stack.
  2344  	// Windows and Plan 9 will layout sched stack on OS stack.
  2345  	if iscgo || mStackIsSystemAllocated() {
  2346  		mp.g0 = malg(-1)
  2347  	} else {
  2348  		mp.g0 = malg(16384 * sys.StackGuardMultiplier)
  2349  	}
  2350  	mp.g0.m = mp
  2351  
  2352  	if pp == gp.m.p.ptr() {
  2353  		releasep()
  2354  	}
  2355  
  2356  	releasem(gp.m)
  2357  	allocmLock.runlock()
  2358  	return mp
  2359  }
  2360  
  2361  // needm is called when a cgo callback happens on a
  2362  // thread without an m (a thread not created by Go).
  2363  // In this case, needm is expected to find an m to use
  2364  // and return with m, g initialized correctly.
  2365  // Since m and g are not set now (likely nil, but see below)
  2366  // needm is limited in what routines it can call. In particular
  2367  // it can only call nosplit functions (textflag 7) and cannot
  2368  // do any scheduling that requires an m.
  2369  //
  2370  // In order to avoid needing heavy lifting here, we adopt
  2371  // the following strategy: there is a stack of available m's
  2372  // that can be stolen. Using compare-and-swap
  2373  // to pop from the stack has ABA races, so we simulate
  2374  // a lock by doing an exchange (via Casuintptr) to steal the stack
  2375  // head and replace the top pointer with MLOCKED (1).
  2376  // This serves as a simple spin lock that we can use even
  2377  // without an m. The thread that locks the stack in this way
  2378  // unlocks the stack by storing a valid stack head pointer.
  2379  //
  2380  // In order to make sure that there is always an m structure
  2381  // available to be stolen, we maintain the invariant that there
  2382  // is always one more than needed. At the beginning of the
  2383  // program (if cgo is in use) the list is seeded with a single m.
  2384  // If needm finds that it has taken the last m off the list, its job
  2385  // is - once it has installed its own m so that it can do things like
  2386  // allocate memory - to create a spare m and put it on the list.
  2387  //
  2388  // Each of these extra m's also has a g0 and a curg that are
  2389  // pressed into service as the scheduling stack and current
  2390  // goroutine for the duration of the cgo callback.
  2391  //
  2392  // It calls dropm to put the m back on the list,
  2393  // 1. when the callback is done with the m in non-pthread platforms,
  2394  // 2. or when the C thread exiting on pthread platforms.
  2395  //
  2396  // The signal argument indicates whether we're called from a signal
  2397  // handler.
  2398  //
  2399  //go:nosplit
  2400  func needm(signal bool) {
  2401  	if (iscgo || GOOS == "windows") && !cgoHasExtraM {
  2402  		// Can happen if C/C++ code calls Go from a global ctor.
  2403  		// Can also happen on Windows if a global ctor uses a
  2404  		// callback created by syscall.NewCallback. See issue #6751
  2405  		// for details.
  2406  		//
  2407  		// Can not throw, because scheduler is not initialized yet.
  2408  		writeErrStr("fatal error: cgo callback before cgo call\n")
  2409  		exit(1)
  2410  	}
  2411  
  2412  	// Save and block signals before getting an M.
  2413  	// The signal handler may call needm itself,
  2414  	// and we must avoid a deadlock. Also, once g is installed,
  2415  	// any incoming signals will try to execute,
  2416  	// but we won't have the sigaltstack settings and other data
  2417  	// set up appropriately until the end of minit, which will
  2418  	// unblock the signals. This is the same dance as when
  2419  	// starting a new m to run Go code via newosproc.
  2420  	var sigmask sigset
  2421  	sigsave(&sigmask)
  2422  	sigblock(false)
  2423  
  2424  	// getExtraM is safe here because of the invariant above,
  2425  	// that the extra list always contains or will soon contain
  2426  	// at least one m.
  2427  	mp, last := getExtraM()
  2428  
  2429  	// Set needextram when we've just emptied the list,
  2430  	// so that the eventual call into cgocallbackg will
  2431  	// allocate a new m for the extra list. We delay the
  2432  	// allocation until then so that it can be done
  2433  	// after exitsyscall makes sure it is okay to be
  2434  	// running at all (that is, there's no garbage collection
  2435  	// running right now).
  2436  	mp.needextram = last
  2437  
  2438  	// Store the original signal mask for use by minit.
  2439  	mp.sigmask = sigmask
  2440  
  2441  	// Install TLS on some platforms (previously setg
  2442  	// would do this if necessary).
  2443  	osSetupTLS(mp)
  2444  
  2445  	// Install g (= m->g0) and set the stack bounds
  2446  	// to match the current stack.
  2447  	setg(mp.g0)
  2448  	sp := sys.GetCallerSP()
  2449  	callbackUpdateSystemStack(mp, sp, signal)
  2450  
  2451  	// Should mark we are already in Go now.
  2452  	// Otherwise, we may call needm again when we get a signal, before cgocallbackg1,
  2453  	// which means the extram list may be empty, that will cause a deadlock.
  2454  	mp.isExtraInC = false
  2455  
  2456  	// Initialize this thread to use the m.
  2457  	asminit()
  2458  	minit()
  2459  
  2460  	// Emit a trace event for this dead -> syscall transition,
  2461  	// but only if we're not in a signal handler.
  2462  	//
  2463  	// N.B. the tracer can run on a bare M just fine, we just have
  2464  	// to make sure to do this before setg(nil) and unminit.
  2465  	var trace traceLocker
  2466  	if !signal {
  2467  		trace = traceAcquire()
  2468  	}
  2469  
  2470  	// mp.curg is now a real goroutine.
  2471  	casgstatus(mp.curg, _Gdead, _Gsyscall)
  2472  	sched.ngsys.Add(-1)
  2473  
  2474  	if !signal {
  2475  		if trace.ok() {
  2476  			trace.GoCreateSyscall(mp.curg)
  2477  			traceRelease(trace)
  2478  		}
  2479  	}
  2480  	mp.isExtraInSig = signal
  2481  }
  2482  
  2483  // Acquire an extra m and bind it to the C thread when a pthread key has been created.
  2484  //
  2485  //go:nosplit
  2486  func needAndBindM() {
  2487  	needm(false)
  2488  
  2489  	if _cgo_pthread_key_created != nil && *(*uintptr)(_cgo_pthread_key_created) != 0 {
  2490  		cgoBindM()
  2491  	}
  2492  }
  2493  
  2494  // newextram allocates m's and puts them on the extra list.
  2495  // It is called with a working local m, so that it can do things
  2496  // like call schedlock and allocate.
  2497  func newextram() {
  2498  	c := extraMWaiters.Swap(0)
  2499  	if c > 0 {
  2500  		for i := uint32(0); i < c; i++ {
  2501  			oneNewExtraM()
  2502  		}
  2503  	} else if extraMLength.Load() == 0 {
  2504  		// Make sure there is at least one extra M.
  2505  		oneNewExtraM()
  2506  	}
  2507  }
  2508  
  2509  // oneNewExtraM allocates an m and puts it on the extra list.
  2510  func oneNewExtraM() {
  2511  	// Create extra goroutine locked to extra m.
  2512  	// The goroutine is the context in which the cgo callback will run.
  2513  	// The sched.pc will never be returned to, but setting it to
  2514  	// goexit makes clear to the traceback routines where
  2515  	// the goroutine stack ends.
  2516  	mp := allocm(nil, nil, -1)
  2517  	gp := malg(4096)
  2518  	gp.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum
  2519  	gp.sched.sp = gp.stack.hi
  2520  	gp.sched.sp -= 4 * goarch.PtrSize // extra space in case of reads slightly beyond frame
  2521  	gp.sched.lr = 0
  2522  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  2523  	gp.syscallpc = gp.sched.pc
  2524  	gp.syscallsp = gp.sched.sp
  2525  	gp.stktopsp = gp.sched.sp
  2526  	// malg returns status as _Gidle. Change to _Gdead before
  2527  	// adding to allg where GC can see it. We use _Gdead to hide
  2528  	// this from tracebacks and stack scans since it isn't a
  2529  	// "real" goroutine until needm grabs it.
  2530  	casgstatus(gp, _Gidle, _Gdead)
  2531  	gp.m = mp
  2532  	mp.curg = gp
  2533  	mp.isextra = true
  2534  	// mark we are in C by default.
  2535  	mp.isExtraInC = true
  2536  	mp.lockedInt++
  2537  	mp.lockedg.set(gp)
  2538  	gp.lockedm.set(mp)
  2539  	gp.goid = sched.goidgen.Add(1)
  2540  	if raceenabled {
  2541  		gp.racectx = racegostart(abi.FuncPCABIInternal(newextram) + sys.PCQuantum)
  2542  	}
  2543  	// put on allg for garbage collector
  2544  	allgadd(gp)
  2545  
  2546  	// gp is now on the allg list, but we don't want it to be
  2547  	// counted by gcount. It would be more "proper" to increment
  2548  	// sched.ngfree, but that requires locking. Incrementing ngsys
  2549  	// has the same effect.
  2550  	sched.ngsys.Add(1)
  2551  
  2552  	// Add m to the extra list.
  2553  	addExtraM(mp)
  2554  }
  2555  
  2556  // dropm puts the current m back onto the extra list.
  2557  //
  2558  // 1. On systems without pthreads, like Windows
  2559  // dropm is called when a cgo callback has called needm but is now
  2560  // done with the callback and returning back into the non-Go thread.
  2561  //
  2562  // The main expense here is the call to signalstack to release the
  2563  // m's signal stack, and then the call to needm on the next callback
  2564  // from this thread. It is tempting to try to save the m for next time,
  2565  // which would eliminate both these costs, but there might not be
  2566  // a next time: the current thread (which Go does not control) might exit.
  2567  // If we saved the m for that thread, there would be an m leak each time
  2568  // such a thread exited. Instead, we acquire and release an m on each
  2569  // call. These should typically not be scheduling operations, just a few
  2570  // atomics, so the cost should be small.
  2571  //
  2572  // 2. On systems with pthreads
  2573  // dropm is called while a non-Go thread is exiting.
  2574  // We allocate a pthread per-thread variable using pthread_key_create,
  2575  // to register a thread-exit-time destructor.
  2576  // And store the g into a thread-specific value associated with the pthread key,
  2577  // when first return back to C.
  2578  // So that the destructor would invoke dropm while the non-Go thread is exiting.
  2579  // This is much faster since it avoids expensive signal-related syscalls.
  2580  //
  2581  // This always runs without a P, so //go:nowritebarrierrec is required.
  2582  //
  2583  // This may run with a different stack than was recorded in g0 (there is no
  2584  // call to callbackUpdateSystemStack prior to dropm), so this must be
  2585  // //go:nosplit to avoid the stack bounds check.
  2586  //
  2587  //go:nowritebarrierrec
  2588  //go:nosplit
  2589  func dropm() {
  2590  	// Clear m and g, and return m to the extra list.
  2591  	// After the call to setg we can only call nosplit functions
  2592  	// with no pointer manipulation.
  2593  	mp := getg().m
  2594  
  2595  	// Emit a trace event for this syscall -> dead transition.
  2596  	//
  2597  	// N.B. the tracer can run on a bare M just fine, we just have
  2598  	// to make sure to do this before setg(nil) and unminit.
  2599  	var trace traceLocker
  2600  	if !mp.isExtraInSig {
  2601  		trace = traceAcquire()
  2602  	}
  2603  
  2604  	// Return mp.curg to dead state.
  2605  	casgstatus(mp.curg, _Gsyscall, _Gdead)
  2606  	mp.curg.preemptStop = false
  2607  	sched.ngsys.Add(1)
  2608  
  2609  	if !mp.isExtraInSig {
  2610  		if trace.ok() {
  2611  			trace.GoDestroySyscall()
  2612  			traceRelease(trace)
  2613  		}
  2614  	}
  2615  
  2616  	// Trash syscalltick so that it doesn't line up with mp.old.syscalltick anymore.
  2617  	//
  2618  	// In the new tracer, we model needm and dropm and a goroutine being created and
  2619  	// destroyed respectively. The m then might get reused with a different procid but
  2620  	// still with a reference to oldp, and still with the same syscalltick. The next
  2621  	// time a G is "created" in needm, it'll return and quietly reacquire its P from a
  2622  	// different m with a different procid, which will confuse the trace parser. By
  2623  	// trashing syscalltick, we ensure that it'll appear as if we lost the P to the
  2624  	// tracer parser and that we just reacquired it.
  2625  	//
  2626  	// Trash the value by decrementing because that gets us as far away from the value
  2627  	// the syscall exit code expects as possible. Setting to zero is risky because
  2628  	// syscalltick could already be zero (and in fact, is initialized to zero).
  2629  	mp.syscalltick--
  2630  
  2631  	// Reset trace state unconditionally. This goroutine is being 'destroyed'
  2632  	// from the perspective of the tracer.
  2633  	mp.curg.trace.reset()
  2634  
  2635  	// Flush all the M's buffers. This is necessary because the M might
  2636  	// be used on a different thread with a different procid, so we have
  2637  	// to make sure we don't write into the same buffer.
  2638  	if traceEnabled() || traceShuttingDown() {
  2639  		// Acquire sched.lock across thread destruction. One of the invariants of the tracer
  2640  		// is that a thread cannot disappear from the tracer's view (allm or freem) without
  2641  		// it noticing, so it requires that sched.lock be held over traceThreadDestroy.
  2642  		//
  2643  		// This isn't strictly necessary in this case, because this thread never leaves allm,
  2644  		// but the critical section is short and dropm is rare on pthread platforms, so just
  2645  		// take the lock and play it safe. traceThreadDestroy also asserts that the lock is held.
  2646  		lock(&sched.lock)
  2647  		traceThreadDestroy(mp)
  2648  		unlock(&sched.lock)
  2649  	}
  2650  	mp.isExtraInSig = false
  2651  
  2652  	// Block signals before unminit.
  2653  	// Unminit unregisters the signal handling stack (but needs g on some systems).
  2654  	// Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
  2655  	// It's important not to try to handle a signal between those two steps.
  2656  	sigmask := mp.sigmask
  2657  	sigblock(false)
  2658  	unminit()
  2659  
  2660  	setg(nil)
  2661  
  2662  	// Clear g0 stack bounds to ensure that needm always refreshes the
  2663  	// bounds when reusing this M.
  2664  	g0 := mp.g0
  2665  	g0.stack.hi = 0
  2666  	g0.stack.lo = 0
  2667  	g0.stackguard0 = 0
  2668  	g0.stackguard1 = 0
  2669  	mp.g0StackAccurate = false
  2670  
  2671  	putExtraM(mp)
  2672  
  2673  	msigrestore(sigmask)
  2674  }
  2675  
  2676  // bindm store the g0 of the current m into a thread-specific value.
  2677  //
  2678  // We allocate a pthread per-thread variable using pthread_key_create,
  2679  // to register a thread-exit-time destructor.
  2680  // We are here setting the thread-specific value of the pthread key, to enable the destructor.
  2681  // So that the pthread_key_destructor would dropm while the C thread is exiting.
  2682  //
  2683  // And the saved g will be used in pthread_key_destructor,
  2684  // since the g stored in the TLS by Go might be cleared in some platforms,
  2685  // before the destructor invoked, so, we restore g by the stored g, before dropm.
  2686  //
  2687  // We store g0 instead of m, to make the assembly code simpler,
  2688  // since we need to restore g0 in runtime.cgocallback.
  2689  //
  2690  // On systems without pthreads, like Windows, bindm shouldn't be used.
  2691  //
  2692  // NOTE: this always runs without a P, so, nowritebarrierrec required.
  2693  //
  2694  //go:nosplit
  2695  //go:nowritebarrierrec
  2696  func cgoBindM() {
  2697  	if GOOS == "windows" || GOOS == "plan9" {
  2698  		fatal("bindm in unexpected GOOS")
  2699  	}
  2700  	g := getg()
  2701  	if g.m.g0 != g {
  2702  		fatal("the current g is not g0")
  2703  	}
  2704  	if _cgo_bindm != nil {
  2705  		asmcgocall(_cgo_bindm, unsafe.Pointer(g))
  2706  	}
  2707  }
  2708  
  2709  // A helper function for EnsureDropM.
  2710  //
  2711  // getm should be an internal detail,
  2712  // but widely used packages access it using linkname.
  2713  // Notable members of the hall of shame include:
  2714  //   - fortio.org/log
  2715  //
  2716  // Do not remove or change the type signature.
  2717  // See go.dev/issue/67401.
  2718  //
  2719  //go:linkname getm
  2720  func getm() uintptr {
  2721  	return uintptr(unsafe.Pointer(getg().m))
  2722  }
  2723  
  2724  var (
  2725  	// Locking linked list of extra M's, via mp.schedlink. Must be accessed
  2726  	// only via lockextra/unlockextra.
  2727  	//
  2728  	// Can't be atomic.Pointer[m] because we use an invalid pointer as a
  2729  	// "locked" sentinel value. M's on this list remain visible to the GC
  2730  	// because their mp.curg is on allgs.
  2731  	extraM atomic.Uintptr
  2732  	// Number of M's in the extraM list.
  2733  	extraMLength atomic.Uint32
  2734  	// Number of waiters in lockextra.
  2735  	extraMWaiters atomic.Uint32
  2736  
  2737  	// Number of extra M's in use by threads.
  2738  	extraMInUse atomic.Uint32
  2739  )
  2740  
  2741  // lockextra locks the extra list and returns the list head.
  2742  // The caller must unlock the list by storing a new list head
  2743  // to extram. If nilokay is true, then lockextra will
  2744  // return a nil list head if that's what it finds. If nilokay is false,
  2745  // lockextra will keep waiting until the list head is no longer nil.
  2746  //
  2747  //go:nosplit
  2748  func lockextra(nilokay bool) *m {
  2749  	const locked = 1
  2750  
  2751  	incr := false
  2752  	for {
  2753  		old := extraM.Load()
  2754  		if old == locked {
  2755  			osyield_no_g()
  2756  			continue
  2757  		}
  2758  		if old == 0 && !nilokay {
  2759  			if !incr {
  2760  				// Add 1 to the number of threads
  2761  				// waiting for an M.
  2762  				// This is cleared by newextram.
  2763  				extraMWaiters.Add(1)
  2764  				incr = true
  2765  			}
  2766  			usleep_no_g(1)
  2767  			continue
  2768  		}
  2769  		if extraM.CompareAndSwap(old, locked) {
  2770  			return (*m)(unsafe.Pointer(old))
  2771  		}
  2772  		osyield_no_g()
  2773  		continue
  2774  	}
  2775  }
  2776  
  2777  //go:nosplit
  2778  func unlockextra(mp *m, delta int32) {
  2779  	extraMLength.Add(delta)
  2780  	extraM.Store(uintptr(unsafe.Pointer(mp)))
  2781  }
  2782  
  2783  // Return an M from the extra M list. Returns last == true if the list becomes
  2784  // empty because of this call.
  2785  //
  2786  // Spins waiting for an extra M, so caller must ensure that the list always
  2787  // contains or will soon contain at least one M.
  2788  //
  2789  //go:nosplit
  2790  func getExtraM() (mp *m, last bool) {
  2791  	mp = lockextra(false)
  2792  	extraMInUse.Add(1)
  2793  	unlockextra(mp.schedlink.ptr(), -1)
  2794  	return mp, mp.schedlink.ptr() == nil
  2795  }
  2796  
  2797  // Returns an extra M back to the list. mp must be from getExtraM. Newly
  2798  // allocated M's should use addExtraM.
  2799  //
  2800  //go:nosplit
  2801  func putExtraM(mp *m) {
  2802  	extraMInUse.Add(-1)
  2803  	addExtraM(mp)
  2804  }
  2805  
  2806  // Adds a newly allocated M to the extra M list.
  2807  //
  2808  //go:nosplit
  2809  func addExtraM(mp *m) {
  2810  	mnext := lockextra(true)
  2811  	mp.schedlink.set(mnext)
  2812  	unlockextra(mp, 1)
  2813  }
  2814  
  2815  var (
  2816  	// allocmLock is locked for read when creating new Ms in allocm and their
  2817  	// addition to allm. Thus acquiring this lock for write blocks the
  2818  	// creation of new Ms.
  2819  	allocmLock rwmutex
  2820  
  2821  	// execLock serializes exec and clone to avoid bugs or unspecified
  2822  	// behaviour around exec'ing while creating/destroying threads. See
  2823  	// issue #19546.
  2824  	execLock rwmutex
  2825  )
  2826  
  2827  // These errors are reported (via writeErrStr) by some OS-specific
  2828  // versions of newosproc and newosproc0.
  2829  const (
  2830  	failthreadcreate  = "runtime: failed to create new OS thread\n"
  2831  	failallocatestack = "runtime: failed to allocate stack for the new OS thread\n"
  2832  )
  2833  
  2834  // newmHandoff contains a list of m structures that need new OS threads.
  2835  // This is used by newm in situations where newm itself can't safely
  2836  // start an OS thread.
  2837  var newmHandoff struct {
  2838  	lock mutex
  2839  
  2840  	// newm points to a list of M structures that need new OS
  2841  	// threads. The list is linked through m.schedlink.
  2842  	newm muintptr
  2843  
  2844  	// waiting indicates that wake needs to be notified when an m
  2845  	// is put on the list.
  2846  	waiting bool
  2847  	wake    note
  2848  
  2849  	// haveTemplateThread indicates that the templateThread has
  2850  	// been started. This is not protected by lock. Use cas to set
  2851  	// to 1.
  2852  	haveTemplateThread uint32
  2853  }
  2854  
  2855  // Create a new m. It will start off with a call to fn, or else the scheduler.
  2856  // fn needs to be static and not a heap allocated closure.
  2857  // May run with m.p==nil, so write barriers are not allowed.
  2858  //
  2859  // id is optional pre-allocated m ID. Omit by passing -1.
  2860  //
  2861  //go:nowritebarrierrec
  2862  func newm(fn func(), pp *p, id int64) {
  2863  	// allocm adds a new M to allm, but they do not start until created by
  2864  	// the OS in newm1 or the template thread.
  2865  	//
  2866  	// doAllThreadsSyscall requires that every M in allm will eventually
  2867  	// start and be signal-able, even with a STW.
  2868  	//
  2869  	// Disable preemption here until we start the thread to ensure that
  2870  	// newm is not preempted between allocm and starting the new thread,
  2871  	// ensuring that anything added to allm is guaranteed to eventually
  2872  	// start.
  2873  	acquirem()
  2874  
  2875  	mp := allocm(pp, fn, id)
  2876  	mp.nextp.set(pp)
  2877  	mp.sigmask = initSigmask
  2878  	if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
  2879  		// We're on a locked M or a thread that may have been
  2880  		// started by C. The kernel state of this thread may
  2881  		// be strange (the user may have locked it for that
  2882  		// purpose). We don't want to clone that into another
  2883  		// thread. Instead, ask a known-good thread to create
  2884  		// the thread for us.
  2885  		//
  2886  		// This is disabled on Plan 9. See golang.org/issue/22227.
  2887  		//
  2888  		// TODO: This may be unnecessary on Windows, which
  2889  		// doesn't model thread creation off fork.
  2890  		lock(&newmHandoff.lock)
  2891  		if newmHandoff.haveTemplateThread == 0 {
  2892  			throw("on a locked thread with no template thread")
  2893  		}
  2894  		mp.schedlink = newmHandoff.newm
  2895  		newmHandoff.newm.set(mp)
  2896  		if newmHandoff.waiting {
  2897  			newmHandoff.waiting = false
  2898  			notewakeup(&newmHandoff.wake)
  2899  		}
  2900  		unlock(&newmHandoff.lock)
  2901  		// The M has not started yet, but the template thread does not
  2902  		// participate in STW, so it will always process queued Ms and
  2903  		// it is safe to releasem.
  2904  		releasem(getg().m)
  2905  		return
  2906  	}
  2907  	newm1(mp)
  2908  	releasem(getg().m)
  2909  }
  2910  
  2911  func newm1(mp *m) {
  2912  	if iscgo {
  2913  		var ts cgothreadstart
  2914  		if _cgo_thread_start == nil {
  2915  			throw("_cgo_thread_start missing")
  2916  		}
  2917  		ts.g.set(mp.g0)
  2918  		ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
  2919  		ts.fn = unsafe.Pointer(abi.FuncPCABI0(mstart))
  2920  		if msanenabled {
  2921  			msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2922  		}
  2923  		if asanenabled {
  2924  			asanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  2925  		}
  2926  		execLock.rlock() // Prevent process clone.
  2927  		asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
  2928  		execLock.runlock()
  2929  		return
  2930  	}
  2931  	execLock.rlock() // Prevent process clone.
  2932  	newosproc(mp)
  2933  	execLock.runlock()
  2934  }
  2935  
  2936  // startTemplateThread starts the template thread if it is not already
  2937  // running.
  2938  //
  2939  // The calling thread must itself be in a known-good state.
  2940  func startTemplateThread() {
  2941  	if GOARCH == "wasm" { // no threads on wasm yet
  2942  		return
  2943  	}
  2944  
  2945  	// Disable preemption to guarantee that the template thread will be
  2946  	// created before a park once haveTemplateThread is set.
  2947  	mp := acquirem()
  2948  	if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
  2949  		releasem(mp)
  2950  		return
  2951  	}
  2952  	newm(templateThread, nil, -1)
  2953  	releasem(mp)
  2954  }
  2955  
  2956  // templateThread is a thread in a known-good state that exists solely
  2957  // to start new threads in known-good states when the calling thread
  2958  // may not be in a good state.
  2959  //
  2960  // Many programs never need this, so templateThread is started lazily
  2961  // when we first enter a state that might lead to running on a thread
  2962  // in an unknown state.
  2963  //
  2964  // templateThread runs on an M without a P, so it must not have write
  2965  // barriers.
  2966  //
  2967  //go:nowritebarrierrec
  2968  func templateThread() {
  2969  	lock(&sched.lock)
  2970  	sched.nmsys++
  2971  	checkdead()
  2972  	unlock(&sched.lock)
  2973  
  2974  	for {
  2975  		lock(&newmHandoff.lock)
  2976  		for newmHandoff.newm != 0 {
  2977  			newm := newmHandoff.newm.ptr()
  2978  			newmHandoff.newm = 0
  2979  			unlock(&newmHandoff.lock)
  2980  			for newm != nil {
  2981  				next := newm.schedlink.ptr()
  2982  				newm.schedlink = 0
  2983  				newm1(newm)
  2984  				newm = next
  2985  			}
  2986  			lock(&newmHandoff.lock)
  2987  		}
  2988  		newmHandoff.waiting = true
  2989  		noteclear(&newmHandoff.wake)
  2990  		unlock(&newmHandoff.lock)
  2991  		notesleep(&newmHandoff.wake)
  2992  	}
  2993  }
  2994  
  2995  // Stops execution of the current m until new work is available.
  2996  // Returns with acquired P.
  2997  func stopm() {
  2998  	gp := getg()
  2999  
  3000  	if gp.m.locks != 0 {
  3001  		throw("stopm holding locks")
  3002  	}
  3003  	if gp.m.p != 0 {
  3004  		throw("stopm holding p")
  3005  	}
  3006  	if gp.m.spinning {
  3007  		throw("stopm spinning")
  3008  	}
  3009  
  3010  	lock(&sched.lock)
  3011  	mput(gp.m)
  3012  	unlock(&sched.lock)
  3013  	mPark()
  3014  	acquirep(gp.m.nextp.ptr())
  3015  	gp.m.nextp = 0
  3016  }
  3017  
  3018  func mspinning() {
  3019  	// startm's caller incremented nmspinning. Set the new M's spinning.
  3020  	getg().m.spinning = true
  3021  }
  3022  
  3023  // Schedules some M to run the p (creates an M if necessary).
  3024  // If p==nil, tries to get an idle P, if no idle P's does nothing.
  3025  // May run with m.p==nil, so write barriers are not allowed.
  3026  // If spinning is set, the caller has incremented nmspinning and must provide a
  3027  // P. startm will set m.spinning in the newly started M.
  3028  //
  3029  // Callers passing a non-nil P must call from a non-preemptible context. See
  3030  // comment on acquirem below.
  3031  //
  3032  // Argument lockheld indicates whether the caller already acquired the
  3033  // scheduler lock. Callers holding the lock when making the call must pass
  3034  // true. The lock might be temporarily dropped, but will be reacquired before
  3035  // returning.
  3036  //
  3037  // Must not have write barriers because this may be called without a P.
  3038  //
  3039  //go:nowritebarrierrec
  3040  func startm(pp *p, spinning, lockheld bool) {
  3041  	// Disable preemption.
  3042  	//
  3043  	// Every owned P must have an owner that will eventually stop it in the
  3044  	// event of a GC stop request. startm takes transient ownership of a P
  3045  	// (either from argument or pidleget below) and transfers ownership to
  3046  	// a started M, which will be responsible for performing the stop.
  3047  	//
  3048  	// Preemption must be disabled during this transient ownership,
  3049  	// otherwise the P this is running on may enter GC stop while still
  3050  	// holding the transient P, leaving that P in limbo and deadlocking the
  3051  	// STW.
  3052  	//
  3053  	// Callers passing a non-nil P must already be in non-preemptible
  3054  	// context, otherwise such preemption could occur on function entry to
  3055  	// startm. Callers passing a nil P may be preemptible, so we must
  3056  	// disable preemption before acquiring a P from pidleget below.
  3057  	mp := acquirem()
  3058  	if !lockheld {
  3059  		lock(&sched.lock)
  3060  	}
  3061  	if pp == nil {
  3062  		if spinning {
  3063  			// TODO(prattmic): All remaining calls to this function
  3064  			// with _p_ == nil could be cleaned up to find a P
  3065  			// before calling startm.
  3066  			throw("startm: P required for spinning=true")
  3067  		}
  3068  		pp, _ = pidleget(0)
  3069  		if pp == nil {
  3070  			if !lockheld {
  3071  				unlock(&sched.lock)
  3072  			}
  3073  			releasem(mp)
  3074  			return
  3075  		}
  3076  	}
  3077  	nmp := mget()
  3078  	if nmp == nil {
  3079  		// No M is available, we must drop sched.lock and call newm.
  3080  		// However, we already own a P to assign to the M.
  3081  		//
  3082  		// Once sched.lock is released, another G (e.g., in a syscall),
  3083  		// could find no idle P while checkdead finds a runnable G but
  3084  		// no running M's because this new M hasn't started yet, thus
  3085  		// throwing in an apparent deadlock.
  3086  		// This apparent deadlock is possible when startm is called
  3087  		// from sysmon, which doesn't count as a running M.
  3088  		//
  3089  		// Avoid this situation by pre-allocating the ID for the new M,
  3090  		// thus marking it as 'running' before we drop sched.lock. This
  3091  		// new M will eventually run the scheduler to execute any
  3092  		// queued G's.
  3093  		id := mReserveID()
  3094  		unlock(&sched.lock)
  3095  
  3096  		var fn func()
  3097  		if spinning {
  3098  			// The caller incremented nmspinning, so set m.spinning in the new M.
  3099  			fn = mspinning
  3100  		}
  3101  		newm(fn, pp, id)
  3102  
  3103  		if lockheld {
  3104  			lock(&sched.lock)
  3105  		}
  3106  		// Ownership transfer of pp committed by start in newm.
  3107  		// Preemption is now safe.
  3108  		releasem(mp)
  3109  		return
  3110  	}
  3111  	if !lockheld {
  3112  		unlock(&sched.lock)
  3113  	}
  3114  	if nmp.spinning {
  3115  		throw("startm: m is spinning")
  3116  	}
  3117  	if nmp.nextp != 0 {
  3118  		throw("startm: m has p")
  3119  	}
  3120  	if spinning && !runqempty(pp) {
  3121  		throw("startm: p has runnable gs")
  3122  	}
  3123  	// The caller incremented nmspinning, so set m.spinning in the new M.
  3124  	nmp.spinning = spinning
  3125  	nmp.nextp.set(pp)
  3126  	notewakeup(&nmp.park)
  3127  	// Ownership transfer of pp committed by wakeup. Preemption is now
  3128  	// safe.
  3129  	releasem(mp)
  3130  }
  3131  
  3132  // Hands off P from syscall or locked M.
  3133  // Always runs without a P, so write barriers are not allowed.
  3134  //
  3135  //go:nowritebarrierrec
  3136  func handoffp(pp *p) {
  3137  	// handoffp must start an M in any situation where
  3138  	// findrunnable would return a G to run on pp.
  3139  
  3140  	// if it has local work, start it straight away
  3141  	if !runqempty(pp) || !sched.runq.empty() {
  3142  		startm(pp, false, false)
  3143  		return
  3144  	}
  3145  	// if there's trace work to do, start it straight away
  3146  	if (traceEnabled() || traceShuttingDown()) && traceReaderAvailable() != nil {
  3147  		startm(pp, false, false)
  3148  		return
  3149  	}
  3150  	// if it has GC work, start it straight away
  3151  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) {
  3152  		startm(pp, false, false)
  3153  		return
  3154  	}
  3155  	// no local work, check that there are no spinning/idle M's,
  3156  	// otherwise our help is not required
  3157  	if sched.nmspinning.Load()+sched.npidle.Load() == 0 && sched.nmspinning.CompareAndSwap(0, 1) { // TODO: fast atomic
  3158  		sched.needspinning.Store(0)
  3159  		startm(pp, true, false)
  3160  		return
  3161  	}
  3162  	lock(&sched.lock)
  3163  	if sched.gcwaiting.Load() {
  3164  		pp.status = _Pgcstop
  3165  		pp.gcStopTime = nanotime()
  3166  		sched.stopwait--
  3167  		if sched.stopwait == 0 {
  3168  			notewakeup(&sched.stopnote)
  3169  		}
  3170  		unlock(&sched.lock)
  3171  		return
  3172  	}
  3173  	if pp.runSafePointFn != 0 && atomic.Cas(&pp.runSafePointFn, 1, 0) {
  3174  		sched.safePointFn(pp)
  3175  		sched.safePointWait--
  3176  		if sched.safePointWait == 0 {
  3177  			notewakeup(&sched.safePointNote)
  3178  		}
  3179  	}
  3180  	if !sched.runq.empty() {
  3181  		unlock(&sched.lock)
  3182  		startm(pp, false, false)
  3183  		return
  3184  	}
  3185  	// If this is the last running P and nobody is polling network,
  3186  	// need to wakeup another M to poll network.
  3187  	if sched.npidle.Load() == gomaxprocs-1 && sched.lastpoll.Load() != 0 {
  3188  		unlock(&sched.lock)
  3189  		startm(pp, false, false)
  3190  		return
  3191  	}
  3192  
  3193  	// The scheduler lock cannot be held when calling wakeNetPoller below
  3194  	// because wakeNetPoller may call wakep which may call startm.
  3195  	when := pp.timers.wakeTime()
  3196  	pidleput(pp, 0)
  3197  	unlock(&sched.lock)
  3198  
  3199  	if when != 0 {
  3200  		wakeNetPoller(when)
  3201  	}
  3202  }
  3203  
  3204  // Tries to add one more P to execute G's.
  3205  // Called when a G is made runnable (newproc, ready).
  3206  // Must be called with a P.
  3207  //
  3208  // wakep should be an internal detail,
  3209  // but widely used packages access it using linkname.
  3210  // Notable members of the hall of shame include:
  3211  //   - gvisor.dev/gvisor
  3212  //
  3213  // Do not remove or change the type signature.
  3214  // See go.dev/issue/67401.
  3215  //
  3216  //go:linkname wakep
  3217  func wakep() {
  3218  	// Be conservative about spinning threads, only start one if none exist
  3219  	// already.
  3220  	if sched.nmspinning.Load() != 0 || !sched.nmspinning.CompareAndSwap(0, 1) {
  3221  		return
  3222  	}
  3223  
  3224  	// Disable preemption until ownership of pp transfers to the next M in
  3225  	// startm. Otherwise preemption here would leave pp stuck waiting to
  3226  	// enter _Pgcstop.
  3227  	//
  3228  	// See preemption comment on acquirem in startm for more details.
  3229  	mp := acquirem()
  3230  
  3231  	var pp *p
  3232  	lock(&sched.lock)
  3233  	pp, _ = pidlegetSpinning(0)
  3234  	if pp == nil {
  3235  		if sched.nmspinning.Add(-1) < 0 {
  3236  			throw("wakep: negative nmspinning")
  3237  		}
  3238  		unlock(&sched.lock)
  3239  		releasem(mp)
  3240  		return
  3241  	}
  3242  	// Since we always have a P, the race in the "No M is available"
  3243  	// comment in startm doesn't apply during the small window between the
  3244  	// unlock here and lock in startm. A checkdead in between will always
  3245  	// see at least one running M (ours).
  3246  	unlock(&sched.lock)
  3247  
  3248  	startm(pp, true, false)
  3249  
  3250  	releasem(mp)
  3251  }
  3252  
  3253  // Stops execution of the current m that is locked to a g until the g is runnable again.
  3254  // Returns with acquired P.
  3255  func stoplockedm() {
  3256  	gp := getg()
  3257  
  3258  	if gp.m.lockedg == 0 || gp.m.lockedg.ptr().lockedm.ptr() != gp.m {
  3259  		throw("stoplockedm: inconsistent locking")
  3260  	}
  3261  	if gp.m.p != 0 {
  3262  		// Schedule another M to run this p.
  3263  		pp := releasep()
  3264  		handoffp(pp)
  3265  	}
  3266  	incidlelocked(1)
  3267  	// Wait until another thread schedules lockedg again.
  3268  	mPark()
  3269  	status := readgstatus(gp.m.lockedg.ptr())
  3270  	if status&^_Gscan != _Grunnable {
  3271  		print("runtime:stoplockedm: lockedg (atomicstatus=", status, ") is not Grunnable or Gscanrunnable\n")
  3272  		dumpgstatus(gp.m.lockedg.ptr())
  3273  		throw("stoplockedm: not runnable")
  3274  	}
  3275  	acquirep(gp.m.nextp.ptr())
  3276  	gp.m.nextp = 0
  3277  }
  3278  
  3279  // Schedules the locked m to run the locked gp.
  3280  // May run during STW, so write barriers are not allowed.
  3281  //
  3282  //go:nowritebarrierrec
  3283  func startlockedm(gp *g) {
  3284  	mp := gp.lockedm.ptr()
  3285  	if mp == getg().m {
  3286  		throw("startlockedm: locked to me")
  3287  	}
  3288  	if mp.nextp != 0 {
  3289  		throw("startlockedm: m has p")
  3290  	}
  3291  	// directly handoff current P to the locked m
  3292  	incidlelocked(-1)
  3293  	pp := releasep()
  3294  	mp.nextp.set(pp)
  3295  	notewakeup(&mp.park)
  3296  	stopm()
  3297  }
  3298  
  3299  // Stops the current m for stopTheWorld.
  3300  // Returns when the world is restarted.
  3301  func gcstopm() {
  3302  	gp := getg()
  3303  
  3304  	if !sched.gcwaiting.Load() {
  3305  		throw("gcstopm: not waiting for gc")
  3306  	}
  3307  	if gp.m.spinning {
  3308  		gp.m.spinning = false
  3309  		// OK to just drop nmspinning here,
  3310  		// startTheWorld will unpark threads as necessary.
  3311  		if sched.nmspinning.Add(-1) < 0 {
  3312  			throw("gcstopm: negative nmspinning")
  3313  		}
  3314  	}
  3315  	pp := releasep()
  3316  	lock(&sched.lock)
  3317  	pp.status = _Pgcstop
  3318  	pp.gcStopTime = nanotime()
  3319  	sched.stopwait--
  3320  	if sched.stopwait == 0 {
  3321  		notewakeup(&sched.stopnote)
  3322  	}
  3323  	unlock(&sched.lock)
  3324  	stopm()
  3325  }
  3326  
  3327  // Schedules gp to run on the current M.
  3328  // If inheritTime is true, gp inherits the remaining time in the
  3329  // current time slice. Otherwise, it starts a new time slice.
  3330  // Never returns.
  3331  //
  3332  // Write barriers are allowed because this is called immediately after
  3333  // acquiring a P in several places.
  3334  //
  3335  //go:yeswritebarrierrec
  3336  func execute(gp *g, inheritTime bool) {
  3337  	mp := getg().m
  3338  
  3339  	if goroutineProfile.active {
  3340  		// Make sure that gp has had its stack written out to the goroutine
  3341  		// profile, exactly as it was when the goroutine profiler first stopped
  3342  		// the world.
  3343  		tryRecordGoroutineProfile(gp, nil, osyield)
  3344  	}
  3345  
  3346  	// Assign gp.m before entering _Grunning so running Gs have an M.
  3347  	mp.curg = gp
  3348  	gp.m = mp
  3349  	gp.syncSafePoint = false // Clear the flag, which may have been set by morestack.
  3350  	casgstatus(gp, _Grunnable, _Grunning)
  3351  	gp.waitsince = 0
  3352  	gp.preempt = false
  3353  	gp.stackguard0 = gp.stack.lo + stackGuard
  3354  	if !inheritTime {
  3355  		mp.p.ptr().schedtick++
  3356  	}
  3357  
  3358  	// Check whether the profiler needs to be turned on or off.
  3359  	hz := sched.profilehz
  3360  	if mp.profilehz != hz {
  3361  		setThreadCPUProfiler(hz)
  3362  	}
  3363  
  3364  	trace := traceAcquire()
  3365  	if trace.ok() {
  3366  		trace.GoStart()
  3367  		traceRelease(trace)
  3368  	}
  3369  
  3370  	gogo(&gp.sched)
  3371  }
  3372  
  3373  // Finds a runnable goroutine to execute.
  3374  // Tries to steal from other P's, get g from local or global queue, poll network.
  3375  // tryWakeP indicates that the returned goroutine is not normal (GC worker, trace
  3376  // reader) so the caller should try to wake a P.
  3377  func findRunnable() (gp *g, inheritTime, tryWakeP bool) {
  3378  	mp := getg().m
  3379  
  3380  	// The conditions here and in handoffp must agree: if
  3381  	// findrunnable would return a G to run, handoffp must start
  3382  	// an M.
  3383  
  3384  top:
  3385  	// We may have collected an allp snapshot below. The snapshot is only
  3386  	// required in each loop iteration. Clear it to all GC to collect the
  3387  	// slice.
  3388  	mp.clearAllpSnapshot()
  3389  
  3390  	pp := mp.p.ptr()
  3391  	if sched.gcwaiting.Load() {
  3392  		gcstopm()
  3393  		goto top
  3394  	}
  3395  	if pp.runSafePointFn != 0 {
  3396  		runSafePointFn()
  3397  	}
  3398  
  3399  	// now and pollUntil are saved for work stealing later,
  3400  	// which may steal timers. It's important that between now
  3401  	// and then, nothing blocks, so these numbers remain mostly
  3402  	// relevant.
  3403  	now, pollUntil, _ := pp.timers.check(0, nil)
  3404  
  3405  	// Try to schedule the trace reader.
  3406  	if traceEnabled() || traceShuttingDown() {
  3407  		gp := traceReader()
  3408  		if gp != nil {
  3409  			trace := traceAcquire()
  3410  			casgstatus(gp, _Gwaiting, _Grunnable)
  3411  			if trace.ok() {
  3412  				trace.GoUnpark(gp, 0)
  3413  				traceRelease(trace)
  3414  			}
  3415  			return gp, false, true
  3416  		}
  3417  	}
  3418  
  3419  	// Try to schedule a GC worker.
  3420  	if gcBlackenEnabled != 0 {
  3421  		gp, tnow := gcController.findRunnableGCWorker(pp, now)
  3422  		if gp != nil {
  3423  			return gp, false, true
  3424  		}
  3425  		now = tnow
  3426  	}
  3427  
  3428  	// Check the global runnable queue once in a while to ensure fairness.
  3429  	// Otherwise two goroutines can completely occupy the local runqueue
  3430  	// by constantly respawning each other.
  3431  	if pp.schedtick%61 == 0 && !sched.runq.empty() {
  3432  		lock(&sched.lock)
  3433  		gp := globrunqget()
  3434  		unlock(&sched.lock)
  3435  		if gp != nil {
  3436  			return gp, false, false
  3437  		}
  3438  	}
  3439  
  3440  	// Wake up the finalizer G.
  3441  	if fingStatus.Load()&(fingWait|fingWake) == fingWait|fingWake {
  3442  		if gp := wakefing(); gp != nil {
  3443  			ready(gp, 0, true)
  3444  		}
  3445  	}
  3446  
  3447  	// Wake up one or more cleanup Gs.
  3448  	if gcCleanups.needsWake() {
  3449  		gcCleanups.wake()
  3450  	}
  3451  
  3452  	if *cgo_yield != nil {
  3453  		asmcgocall(*cgo_yield, nil)
  3454  	}
  3455  
  3456  	// local runq
  3457  	if gp, inheritTime := runqget(pp); gp != nil {
  3458  		return gp, inheritTime, false
  3459  	}
  3460  
  3461  	// global runq
  3462  	if !sched.runq.empty() {
  3463  		lock(&sched.lock)
  3464  		gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3465  		unlock(&sched.lock)
  3466  		if gp != nil {
  3467  			if runqputbatch(pp, &q); !q.empty() {
  3468  				throw("Couldn't put Gs into empty local runq")
  3469  			}
  3470  			return gp, false, false
  3471  		}
  3472  	}
  3473  
  3474  	// Poll network.
  3475  	// This netpoll is only an optimization before we resort to stealing.
  3476  	// We can safely skip it if there are no waiters or a thread is blocked
  3477  	// in netpoll already. If there is any kind of logical race with that
  3478  	// blocked thread (e.g. it has already returned from netpoll, but does
  3479  	// not set lastpoll yet), this thread will do blocking netpoll below
  3480  	// anyway.
  3481  	// We only poll from one thread at a time to avoid kernel contention
  3482  	// on machines with many cores.
  3483  	if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 && sched.pollingNet.Swap(1) == 0 {
  3484  		list, delta := netpoll(0)
  3485  		sched.pollingNet.Store(0)
  3486  		if !list.empty() { // non-blocking
  3487  			gp := list.pop()
  3488  			injectglist(&list)
  3489  			netpollAdjustWaiters(delta)
  3490  			trace := traceAcquire()
  3491  			casgstatus(gp, _Gwaiting, _Grunnable)
  3492  			if trace.ok() {
  3493  				trace.GoUnpark(gp, 0)
  3494  				traceRelease(trace)
  3495  			}
  3496  			return gp, false, false
  3497  		}
  3498  	}
  3499  
  3500  	// Spinning Ms: steal work from other Ps.
  3501  	//
  3502  	// Limit the number of spinning Ms to half the number of busy Ps.
  3503  	// This is necessary to prevent excessive CPU consumption when
  3504  	// GOMAXPROCS>>1 but the program parallelism is low.
  3505  	if mp.spinning || 2*sched.nmspinning.Load() < gomaxprocs-sched.npidle.Load() {
  3506  		if !mp.spinning {
  3507  			mp.becomeSpinning()
  3508  		}
  3509  
  3510  		gp, inheritTime, tnow, w, newWork := stealWork(now)
  3511  		if gp != nil {
  3512  			// Successfully stole.
  3513  			return gp, inheritTime, false
  3514  		}
  3515  		if newWork {
  3516  			// There may be new timer or GC work; restart to
  3517  			// discover.
  3518  			goto top
  3519  		}
  3520  
  3521  		now = tnow
  3522  		if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3523  			// Earlier timer to wait for.
  3524  			pollUntil = w
  3525  		}
  3526  	}
  3527  
  3528  	// We have nothing to do.
  3529  	//
  3530  	// If we're in the GC mark phase, can safely scan and blacken objects,
  3531  	// and have work to do, run idle-time marking rather than give up the P.
  3532  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(pp) && gcController.addIdleMarkWorker() {
  3533  		node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3534  		if node != nil {
  3535  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3536  			gp := node.gp.ptr()
  3537  
  3538  			trace := traceAcquire()
  3539  			casgstatus(gp, _Gwaiting, _Grunnable)
  3540  			if trace.ok() {
  3541  				trace.GoUnpark(gp, 0)
  3542  				traceRelease(trace)
  3543  			}
  3544  			return gp, false, false
  3545  		}
  3546  		gcController.removeIdleMarkWorker()
  3547  	}
  3548  
  3549  	// wasm only:
  3550  	// If a callback returned and no other goroutine is awake,
  3551  	// then wake event handler goroutine which pauses execution
  3552  	// until a callback was triggered.
  3553  	gp, otherReady := beforeIdle(now, pollUntil)
  3554  	if gp != nil {
  3555  		trace := traceAcquire()
  3556  		casgstatus(gp, _Gwaiting, _Grunnable)
  3557  		if trace.ok() {
  3558  			trace.GoUnpark(gp, 0)
  3559  			traceRelease(trace)
  3560  		}
  3561  		return gp, false, false
  3562  	}
  3563  	if otherReady {
  3564  		goto top
  3565  	}
  3566  
  3567  	// Before we drop our P, make a snapshot of the allp slice,
  3568  	// which can change underfoot once we no longer block
  3569  	// safe-points. We don't need to snapshot the contents because
  3570  	// everything up to cap(allp) is immutable.
  3571  	//
  3572  	// We clear the snapshot from the M after return via
  3573  	// mp.clearAllpSnapshop (in schedule) and on each iteration of the top
  3574  	// loop.
  3575  	allpSnapshot := mp.snapshotAllp()
  3576  	// Also snapshot masks. Value changes are OK, but we can't allow
  3577  	// len to change out from under us.
  3578  	idlepMaskSnapshot := idlepMask
  3579  	timerpMaskSnapshot := timerpMask
  3580  
  3581  	// return P and block
  3582  	lock(&sched.lock)
  3583  	if sched.gcwaiting.Load() || pp.runSafePointFn != 0 {
  3584  		unlock(&sched.lock)
  3585  		goto top
  3586  	}
  3587  	if !sched.runq.empty() {
  3588  		gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3589  		unlock(&sched.lock)
  3590  		if gp == nil {
  3591  			throw("global runq empty with non-zero runqsize")
  3592  		}
  3593  		if runqputbatch(pp, &q); !q.empty() {
  3594  			throw("Couldn't put Gs into empty local runq")
  3595  		}
  3596  		return gp, false, false
  3597  	}
  3598  	if !mp.spinning && sched.needspinning.Load() == 1 {
  3599  		// See "Delicate dance" comment below.
  3600  		mp.becomeSpinning()
  3601  		unlock(&sched.lock)
  3602  		goto top
  3603  	}
  3604  	if releasep() != pp {
  3605  		throw("findrunnable: wrong p")
  3606  	}
  3607  	now = pidleput(pp, now)
  3608  	unlock(&sched.lock)
  3609  
  3610  	// Delicate dance: thread transitions from spinning to non-spinning
  3611  	// state, potentially concurrently with submission of new work. We must
  3612  	// drop nmspinning first and then check all sources again (with
  3613  	// #StoreLoad memory barrier in between). If we do it the other way
  3614  	// around, another thread can submit work after we've checked all
  3615  	// sources but before we drop nmspinning; as a result nobody will
  3616  	// unpark a thread to run the work.
  3617  	//
  3618  	// This applies to the following sources of work:
  3619  	//
  3620  	// * Goroutines added to the global or a per-P run queue.
  3621  	// * New/modified-earlier timers on a per-P timer heap.
  3622  	// * Idle-priority GC work (barring golang.org/issue/19112).
  3623  	//
  3624  	// If we discover new work below, we need to restore m.spinning as a
  3625  	// signal for resetspinning to unpark a new worker thread (because
  3626  	// there can be more than one starving goroutine).
  3627  	//
  3628  	// However, if after discovering new work we also observe no idle Ps
  3629  	// (either here or in resetspinning), we have a problem. We may be
  3630  	// racing with a non-spinning M in the block above, having found no
  3631  	// work and preparing to release its P and park. Allowing that P to go
  3632  	// idle will result in loss of work conservation (idle P while there is
  3633  	// runnable work). This could result in complete deadlock in the
  3634  	// unlikely event that we discover new work (from netpoll) right as we
  3635  	// are racing with _all_ other Ps going idle.
  3636  	//
  3637  	// We use sched.needspinning to synchronize with non-spinning Ms going
  3638  	// idle. If needspinning is set when they are about to drop their P,
  3639  	// they abort the drop and instead become a new spinning M on our
  3640  	// behalf. If we are not racing and the system is truly fully loaded
  3641  	// then no spinning threads are required, and the next thread to
  3642  	// naturally become spinning will clear the flag.
  3643  	//
  3644  	// Also see "Worker thread parking/unparking" comment at the top of the
  3645  	// file.
  3646  	wasSpinning := mp.spinning
  3647  	if mp.spinning {
  3648  		mp.spinning = false
  3649  		if sched.nmspinning.Add(-1) < 0 {
  3650  			throw("findrunnable: negative nmspinning")
  3651  		}
  3652  
  3653  		// Note the for correctness, only the last M transitioning from
  3654  		// spinning to non-spinning must perform these rechecks to
  3655  		// ensure no missed work. However, the runtime has some cases
  3656  		// of transient increments of nmspinning that are decremented
  3657  		// without going through this path, so we must be conservative
  3658  		// and perform the check on all spinning Ms.
  3659  		//
  3660  		// See https://go.dev/issue/43997.
  3661  
  3662  		// Check global and P runqueues again.
  3663  
  3664  		lock(&sched.lock)
  3665  		if !sched.runq.empty() {
  3666  			pp, _ := pidlegetSpinning(0)
  3667  			if pp != nil {
  3668  				gp, q := globrunqgetbatch(int32(len(pp.runq)) / 2)
  3669  				unlock(&sched.lock)
  3670  				if gp == nil {
  3671  					throw("global runq empty with non-zero runqsize")
  3672  				}
  3673  				if runqputbatch(pp, &q); !q.empty() {
  3674  					throw("Couldn't put Gs into empty local runq")
  3675  				}
  3676  				acquirep(pp)
  3677  				mp.becomeSpinning()
  3678  				return gp, false, false
  3679  			}
  3680  		}
  3681  		unlock(&sched.lock)
  3682  
  3683  		pp := checkRunqsNoP(allpSnapshot, idlepMaskSnapshot)
  3684  		if pp != nil {
  3685  			acquirep(pp)
  3686  			mp.becomeSpinning()
  3687  			goto top
  3688  		}
  3689  
  3690  		// Check for idle-priority GC work again.
  3691  		pp, gp := checkIdleGCNoP()
  3692  		if pp != nil {
  3693  			acquirep(pp)
  3694  			mp.becomeSpinning()
  3695  
  3696  			// Run the idle worker.
  3697  			pp.gcMarkWorkerMode = gcMarkWorkerIdleMode
  3698  			trace := traceAcquire()
  3699  			casgstatus(gp, _Gwaiting, _Grunnable)
  3700  			if trace.ok() {
  3701  				trace.GoUnpark(gp, 0)
  3702  				traceRelease(trace)
  3703  			}
  3704  			return gp, false, false
  3705  		}
  3706  
  3707  		// Finally, check for timer creation or expiry concurrently with
  3708  		// transitioning from spinning to non-spinning.
  3709  		//
  3710  		// Note that we cannot use checkTimers here because it calls
  3711  		// adjusttimers which may need to allocate memory, and that isn't
  3712  		// allowed when we don't have an active P.
  3713  		pollUntil = checkTimersNoP(allpSnapshot, timerpMaskSnapshot, pollUntil)
  3714  	}
  3715  
  3716  	// We don't need allp anymore at this pointer, but can't clear the
  3717  	// snapshot without a P for the write barrier..
  3718  
  3719  	// Poll network until next timer.
  3720  	if netpollinited() && (netpollAnyWaiters() || pollUntil != 0) && sched.lastpoll.Swap(0) != 0 {
  3721  		sched.pollUntil.Store(pollUntil)
  3722  		if mp.p != 0 {
  3723  			throw("findrunnable: netpoll with p")
  3724  		}
  3725  		if mp.spinning {
  3726  			throw("findrunnable: netpoll with spinning")
  3727  		}
  3728  		delay := int64(-1)
  3729  		if pollUntil != 0 {
  3730  			if now == 0 {
  3731  				now = nanotime()
  3732  			}
  3733  			delay = pollUntil - now
  3734  			if delay < 0 {
  3735  				delay = 0
  3736  			}
  3737  		}
  3738  		if faketime != 0 {
  3739  			// When using fake time, just poll.
  3740  			delay = 0
  3741  		}
  3742  		list, delta := netpoll(delay) // block until new work is available
  3743  		// Refresh now again, after potentially blocking.
  3744  		now = nanotime()
  3745  		sched.pollUntil.Store(0)
  3746  		sched.lastpoll.Store(now)
  3747  		if faketime != 0 && list.empty() {
  3748  			// Using fake time and nothing is ready; stop M.
  3749  			// When all M's stop, checkdead will call timejump.
  3750  			stopm()
  3751  			goto top
  3752  		}
  3753  		lock(&sched.lock)
  3754  		pp, _ := pidleget(now)
  3755  		unlock(&sched.lock)
  3756  		if pp == nil {
  3757  			injectglist(&list)
  3758  			netpollAdjustWaiters(delta)
  3759  		} else {
  3760  			acquirep(pp)
  3761  			if !list.empty() {
  3762  				gp := list.pop()
  3763  				injectglist(&list)
  3764  				netpollAdjustWaiters(delta)
  3765  				trace := traceAcquire()
  3766  				casgstatus(gp, _Gwaiting, _Grunnable)
  3767  				if trace.ok() {
  3768  					trace.GoUnpark(gp, 0)
  3769  					traceRelease(trace)
  3770  				}
  3771  				return gp, false, false
  3772  			}
  3773  			if wasSpinning {
  3774  				mp.becomeSpinning()
  3775  			}
  3776  			goto top
  3777  		}
  3778  	} else if pollUntil != 0 && netpollinited() {
  3779  		pollerPollUntil := sched.pollUntil.Load()
  3780  		if pollerPollUntil == 0 || pollerPollUntil > pollUntil {
  3781  			netpollBreak()
  3782  		}
  3783  	}
  3784  	stopm()
  3785  	goto top
  3786  }
  3787  
  3788  // pollWork reports whether there is non-background work this P could
  3789  // be doing. This is a fairly lightweight check to be used for
  3790  // background work loops, like idle GC. It checks a subset of the
  3791  // conditions checked by the actual scheduler.
  3792  func pollWork() bool {
  3793  	if !sched.runq.empty() {
  3794  		return true
  3795  	}
  3796  	p := getg().m.p.ptr()
  3797  	if !runqempty(p) {
  3798  		return true
  3799  	}
  3800  	if netpollinited() && netpollAnyWaiters() && sched.lastpoll.Load() != 0 {
  3801  		if list, delta := netpoll(0); !list.empty() {
  3802  			injectglist(&list)
  3803  			netpollAdjustWaiters(delta)
  3804  			return true
  3805  		}
  3806  	}
  3807  	return false
  3808  }
  3809  
  3810  // stealWork attempts to steal a runnable goroutine or timer from any P.
  3811  //
  3812  // If newWork is true, new work may have been readied.
  3813  //
  3814  // If now is not 0 it is the current time. stealWork returns the passed time or
  3815  // the current time if now was passed as 0.
  3816  func stealWork(now int64) (gp *g, inheritTime bool, rnow, pollUntil int64, newWork bool) {
  3817  	pp := getg().m.p.ptr()
  3818  
  3819  	ranTimer := false
  3820  
  3821  	const stealTries = 4
  3822  	for i := 0; i < stealTries; i++ {
  3823  		stealTimersOrRunNextG := i == stealTries-1
  3824  
  3825  		for enum := stealOrder.start(cheaprand()); !enum.done(); enum.next() {
  3826  			if sched.gcwaiting.Load() {
  3827  				// GC work may be available.
  3828  				return nil, false, now, pollUntil, true
  3829  			}
  3830  			p2 := allp[enum.position()]
  3831  			if pp == p2 {
  3832  				continue
  3833  			}
  3834  
  3835  			// Steal timers from p2. This call to checkTimers is the only place
  3836  			// where we might hold a lock on a different P's timers. We do this
  3837  			// once on the last pass before checking runnext because stealing
  3838  			// from the other P's runnext should be the last resort, so if there
  3839  			// are timers to steal do that first.
  3840  			//
  3841  			// We only check timers on one of the stealing iterations because
  3842  			// the time stored in now doesn't change in this loop and checking
  3843  			// the timers for each P more than once with the same value of now
  3844  			// is probably a waste of time.
  3845  			//
  3846  			// timerpMask tells us whether the P may have timers at all. If it
  3847  			// can't, no need to check at all.
  3848  			if stealTimersOrRunNextG && timerpMask.read(enum.position()) {
  3849  				tnow, w, ran := p2.timers.check(now, nil)
  3850  				now = tnow
  3851  				if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3852  					pollUntil = w
  3853  				}
  3854  				if ran {
  3855  					// Running the timers may have
  3856  					// made an arbitrary number of G's
  3857  					// ready and added them to this P's
  3858  					// local run queue. That invalidates
  3859  					// the assumption of runqsteal
  3860  					// that it always has room to add
  3861  					// stolen G's. So check now if there
  3862  					// is a local G to run.
  3863  					if gp, inheritTime := runqget(pp); gp != nil {
  3864  						return gp, inheritTime, now, pollUntil, ranTimer
  3865  					}
  3866  					ranTimer = true
  3867  				}
  3868  			}
  3869  
  3870  			// Don't bother to attempt to steal if p2 is idle.
  3871  			if !idlepMask.read(enum.position()) {
  3872  				if gp := runqsteal(pp, p2, stealTimersOrRunNextG); gp != nil {
  3873  					return gp, false, now, pollUntil, ranTimer
  3874  				}
  3875  			}
  3876  		}
  3877  	}
  3878  
  3879  	// No goroutines found to steal. Regardless, running a timer may have
  3880  	// made some goroutine ready that we missed. Indicate the next timer to
  3881  	// wait for.
  3882  	return nil, false, now, pollUntil, ranTimer
  3883  }
  3884  
  3885  // Check all Ps for a runnable G to steal.
  3886  //
  3887  // On entry we have no P. If a G is available to steal and a P is available,
  3888  // the P is returned which the caller should acquire and attempt to steal the
  3889  // work to.
  3890  func checkRunqsNoP(allpSnapshot []*p, idlepMaskSnapshot pMask) *p {
  3891  	for id, p2 := range allpSnapshot {
  3892  		if !idlepMaskSnapshot.read(uint32(id)) && !runqempty(p2) {
  3893  			lock(&sched.lock)
  3894  			pp, _ := pidlegetSpinning(0)
  3895  			if pp == nil {
  3896  				// Can't get a P, don't bother checking remaining Ps.
  3897  				unlock(&sched.lock)
  3898  				return nil
  3899  			}
  3900  			unlock(&sched.lock)
  3901  			return pp
  3902  		}
  3903  	}
  3904  
  3905  	// No work available.
  3906  	return nil
  3907  }
  3908  
  3909  // Check all Ps for a timer expiring sooner than pollUntil.
  3910  //
  3911  // Returns updated pollUntil value.
  3912  func checkTimersNoP(allpSnapshot []*p, timerpMaskSnapshot pMask, pollUntil int64) int64 {
  3913  	for id, p2 := range allpSnapshot {
  3914  		if timerpMaskSnapshot.read(uint32(id)) {
  3915  			w := p2.timers.wakeTime()
  3916  			if w != 0 && (pollUntil == 0 || w < pollUntil) {
  3917  				pollUntil = w
  3918  			}
  3919  		}
  3920  	}
  3921  
  3922  	return pollUntil
  3923  }
  3924  
  3925  // Check for idle-priority GC, without a P on entry.
  3926  //
  3927  // If some GC work, a P, and a worker G are all available, the P and G will be
  3928  // returned. The returned P has not been wired yet.
  3929  func checkIdleGCNoP() (*p, *g) {
  3930  	// N.B. Since we have no P, gcBlackenEnabled may change at any time; we
  3931  	// must check again after acquiring a P. As an optimization, we also check
  3932  	// if an idle mark worker is needed at all. This is OK here, because if we
  3933  	// observe that one isn't needed, at least one is currently running. Even if
  3934  	// it stops running, its own journey into the scheduler should schedule it
  3935  	// again, if need be (at which point, this check will pass, if relevant).
  3936  	if atomic.Load(&gcBlackenEnabled) == 0 || !gcController.needIdleMarkWorker() {
  3937  		return nil, nil
  3938  	}
  3939  	if !gcMarkWorkAvailable(nil) {
  3940  		return nil, nil
  3941  	}
  3942  
  3943  	// Work is available; we can start an idle GC worker only if there is
  3944  	// an available P and available worker G.
  3945  	//
  3946  	// We can attempt to acquire these in either order, though both have
  3947  	// synchronization concerns (see below). Workers are almost always
  3948  	// available (see comment in findRunnableGCWorker for the one case
  3949  	// there may be none). Since we're slightly less likely to find a P,
  3950  	// check for that first.
  3951  	//
  3952  	// Synchronization: note that we must hold sched.lock until we are
  3953  	// committed to keeping it. Otherwise we cannot put the unnecessary P
  3954  	// back in sched.pidle without performing the full set of idle
  3955  	// transition checks.
  3956  	//
  3957  	// If we were to check gcBgMarkWorkerPool first, we must somehow handle
  3958  	// the assumption in gcControllerState.findRunnableGCWorker that an
  3959  	// empty gcBgMarkWorkerPool is only possible if gcMarkDone is running.
  3960  	lock(&sched.lock)
  3961  	pp, now := pidlegetSpinning(0)
  3962  	if pp == nil {
  3963  		unlock(&sched.lock)
  3964  		return nil, nil
  3965  	}
  3966  
  3967  	// Now that we own a P, gcBlackenEnabled can't change (as it requires STW).
  3968  	if gcBlackenEnabled == 0 || !gcController.addIdleMarkWorker() {
  3969  		pidleput(pp, now)
  3970  		unlock(&sched.lock)
  3971  		return nil, nil
  3972  	}
  3973  
  3974  	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
  3975  	if node == nil {
  3976  		pidleput(pp, now)
  3977  		unlock(&sched.lock)
  3978  		gcController.removeIdleMarkWorker()
  3979  		return nil, nil
  3980  	}
  3981  
  3982  	unlock(&sched.lock)
  3983  
  3984  	return pp, node.gp.ptr()
  3985  }
  3986  
  3987  // wakeNetPoller wakes up the thread sleeping in the network poller if it isn't
  3988  // going to wake up before the when argument; or it wakes an idle P to service
  3989  // timers and the network poller if there isn't one already.
  3990  func wakeNetPoller(when int64) {
  3991  	if sched.lastpoll.Load() == 0 {
  3992  		// In findrunnable we ensure that when polling the pollUntil
  3993  		// field is either zero or the time to which the current
  3994  		// poll is expected to run. This can have a spurious wakeup
  3995  		// but should never miss a wakeup.
  3996  		pollerPollUntil := sched.pollUntil.Load()
  3997  		if pollerPollUntil == 0 || pollerPollUntil > when {
  3998  			netpollBreak()
  3999  		}
  4000  	} else {
  4001  		// There are no threads in the network poller, try to get
  4002  		// one there so it can handle new timers.
  4003  		if GOOS != "plan9" { // Temporary workaround - see issue #42303.
  4004  			wakep()
  4005  		}
  4006  	}
  4007  }
  4008  
  4009  func resetspinning() {
  4010  	gp := getg()
  4011  	if !gp.m.spinning {
  4012  		throw("resetspinning: not a spinning m")
  4013  	}
  4014  	gp.m.spinning = false
  4015  	nmspinning := sched.nmspinning.Add(-1)
  4016  	if nmspinning < 0 {
  4017  		throw("findrunnable: negative nmspinning")
  4018  	}
  4019  	// M wakeup policy is deliberately somewhat conservative, so check if we
  4020  	// need to wakeup another P here. See "Worker thread parking/unparking"
  4021  	// comment at the top of the file for details.
  4022  	wakep()
  4023  }
  4024  
  4025  // injectglist adds each runnable G on the list to some run queue,
  4026  // and clears glist. If there is no current P, they are added to the
  4027  // global queue, and up to npidle M's are started to run them.
  4028  // Otherwise, for each idle P, this adds a G to the global queue
  4029  // and starts an M. Any remaining G's are added to the current P's
  4030  // local run queue.
  4031  // This may temporarily acquire sched.lock.
  4032  // Can run concurrently with GC.
  4033  func injectglist(glist *gList) {
  4034  	if glist.empty() {
  4035  		return
  4036  	}
  4037  
  4038  	// Mark all the goroutines as runnable before we put them
  4039  	// on the run queues.
  4040  	var tail *g
  4041  	trace := traceAcquire()
  4042  	for gp := glist.head.ptr(); gp != nil; gp = gp.schedlink.ptr() {
  4043  		tail = gp
  4044  		casgstatus(gp, _Gwaiting, _Grunnable)
  4045  		if trace.ok() {
  4046  			trace.GoUnpark(gp, 0)
  4047  		}
  4048  	}
  4049  	if trace.ok() {
  4050  		traceRelease(trace)
  4051  	}
  4052  
  4053  	// Turn the gList into a gQueue.
  4054  	q := gQueue{glist.head, tail.guintptr(), glist.size}
  4055  	*glist = gList{}
  4056  
  4057  	startIdle := func(n int32) {
  4058  		for ; n > 0; n-- {
  4059  			mp := acquirem() // See comment in startm.
  4060  			lock(&sched.lock)
  4061  
  4062  			pp, _ := pidlegetSpinning(0)
  4063  			if pp == nil {
  4064  				unlock(&sched.lock)
  4065  				releasem(mp)
  4066  				break
  4067  			}
  4068  
  4069  			startm(pp, false, true)
  4070  			unlock(&sched.lock)
  4071  			releasem(mp)
  4072  		}
  4073  	}
  4074  
  4075  	pp := getg().m.p.ptr()
  4076  	if pp == nil {
  4077  		n := q.size
  4078  		lock(&sched.lock)
  4079  		globrunqputbatch(&q)
  4080  		unlock(&sched.lock)
  4081  		startIdle(n)
  4082  		return
  4083  	}
  4084  
  4085  	var globq gQueue
  4086  	npidle := sched.npidle.Load()
  4087  	for ; npidle > 0 && !q.empty(); npidle-- {
  4088  		g := q.pop()
  4089  		globq.pushBack(g)
  4090  	}
  4091  	if !globq.empty() {
  4092  		n := globq.size
  4093  		lock(&sched.lock)
  4094  		globrunqputbatch(&globq)
  4095  		unlock(&sched.lock)
  4096  		startIdle(n)
  4097  	}
  4098  
  4099  	if runqputbatch(pp, &q); !q.empty() {
  4100  		lock(&sched.lock)
  4101  		globrunqputbatch(&q)
  4102  		unlock(&sched.lock)
  4103  	}
  4104  
  4105  	// Some P's might have become idle after we loaded `sched.npidle`
  4106  	// but before any goroutines were added to the queue, which could
  4107  	// lead to idle P's when there is work available in the global queue.
  4108  	// That could potentially last until other goroutines become ready
  4109  	// to run. That said, we need to find a way to hedge
  4110  	//
  4111  	// Calling wakep() here is the best bet, it will do nothing in the
  4112  	// common case (no racing on `sched.npidle`), while it could wake one
  4113  	// more P to execute G's, which might end up with >1 P's: the first one
  4114  	// wakes another P and so forth until there is no more work, but this
  4115  	// ought to be an extremely rare case.
  4116  	//
  4117  	// Also see "Worker thread parking/unparking" comment at the top of the file for details.
  4118  	wakep()
  4119  }
  4120  
  4121  // One round of scheduler: find a runnable goroutine and execute it.
  4122  // Never returns.
  4123  func schedule() {
  4124  	mp := getg().m
  4125  
  4126  	if mp.locks != 0 {
  4127  		throw("schedule: holding locks")
  4128  	}
  4129  
  4130  	if mp.lockedg != 0 {
  4131  		stoplockedm()
  4132  		execute(mp.lockedg.ptr(), false) // Never returns.
  4133  	}
  4134  
  4135  	// We should not schedule away from a g that is executing a cgo call,
  4136  	// since the cgo call is using the m's g0 stack.
  4137  	if mp.incgo {
  4138  		throw("schedule: in cgo")
  4139  	}
  4140  
  4141  top:
  4142  	pp := mp.p.ptr()
  4143  	pp.preempt = false
  4144  
  4145  	// Safety check: if we are spinning, the run queue should be empty.
  4146  	// Check this before calling checkTimers, as that might call
  4147  	// goready to put a ready goroutine on the local run queue.
  4148  	if mp.spinning && (pp.runnext != 0 || pp.runqhead != pp.runqtail) {
  4149  		throw("schedule: spinning with local work")
  4150  	}
  4151  
  4152  	gp, inheritTime, tryWakeP := findRunnable() // blocks until work is available
  4153  
  4154  	// findRunnable may have collected an allp snapshot. The snapshot is
  4155  	// only required within findRunnable. Clear it to all GC to collect the
  4156  	// slice.
  4157  	mp.clearAllpSnapshot()
  4158  
  4159  	if debug.dontfreezetheworld > 0 && freezing.Load() {
  4160  		// See comment in freezetheworld. We don't want to perturb
  4161  		// scheduler state, so we didn't gcstopm in findRunnable, but
  4162  		// also don't want to allow new goroutines to run.
  4163  		//
  4164  		// Deadlock here rather than in the findRunnable loop so if
  4165  		// findRunnable is stuck in a loop we don't perturb that
  4166  		// either.
  4167  		lock(&deadlock)
  4168  		lock(&deadlock)
  4169  	}
  4170  
  4171  	// This thread is going to run a goroutine and is not spinning anymore,
  4172  	// so if it was marked as spinning we need to reset it now and potentially
  4173  	// start a new spinning M.
  4174  	if mp.spinning {
  4175  		resetspinning()
  4176  	}
  4177  
  4178  	if sched.disable.user && !schedEnabled(gp) {
  4179  		// Scheduling of this goroutine is disabled. Put it on
  4180  		// the list of pending runnable goroutines for when we
  4181  		// re-enable user scheduling and look again.
  4182  		lock(&sched.lock)
  4183  		if schedEnabled(gp) {
  4184  			// Something re-enabled scheduling while we
  4185  			// were acquiring the lock.
  4186  			unlock(&sched.lock)
  4187  		} else {
  4188  			sched.disable.runnable.pushBack(gp)
  4189  			unlock(&sched.lock)
  4190  			goto top
  4191  		}
  4192  	}
  4193  
  4194  	// If about to schedule a not-normal goroutine (a GCworker or tracereader),
  4195  	// wake a P if there is one.
  4196  	if tryWakeP {
  4197  		wakep()
  4198  	}
  4199  	if gp.lockedm != 0 {
  4200  		// Hands off own p to the locked m,
  4201  		// then blocks waiting for a new p.
  4202  		startlockedm(gp)
  4203  		goto top
  4204  	}
  4205  
  4206  	execute(gp, inheritTime)
  4207  }
  4208  
  4209  // dropg removes the association between m and the current goroutine m->curg (gp for short).
  4210  // Typically a caller sets gp's status away from Grunning and then
  4211  // immediately calls dropg to finish the job. The caller is also responsible
  4212  // for arranging that gp will be restarted using ready at an
  4213  // appropriate time. After calling dropg and arranging for gp to be
  4214  // readied later, the caller can do other work but eventually should
  4215  // call schedule to restart the scheduling of goroutines on this m.
  4216  func dropg() {
  4217  	gp := getg()
  4218  
  4219  	setMNoWB(&gp.m.curg.m, nil)
  4220  	setGNoWB(&gp.m.curg, nil)
  4221  }
  4222  
  4223  func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
  4224  	unlock((*mutex)(lock))
  4225  	return true
  4226  }
  4227  
  4228  // park continuation on g0.
  4229  func park_m(gp *g) {
  4230  	mp := getg().m
  4231  
  4232  	trace := traceAcquire()
  4233  
  4234  	// If g is in a synctest group, we don't want to let the group
  4235  	// become idle until after the waitunlockf (if any) has confirmed
  4236  	// that the park is happening.
  4237  	// We need to record gp.bubble here, since waitunlockf can change it.
  4238  	bubble := gp.bubble
  4239  	if bubble != nil {
  4240  		bubble.incActive()
  4241  	}
  4242  
  4243  	if trace.ok() {
  4244  		// Trace the event before the transition. It may take a
  4245  		// stack trace, but we won't own the stack after the
  4246  		// transition anymore.
  4247  		trace.GoPark(mp.waitTraceBlockReason, mp.waitTraceSkip)
  4248  	}
  4249  	// N.B. Not using casGToWaiting here because the waitreason is
  4250  	// set by park_m's caller.
  4251  	casgstatus(gp, _Grunning, _Gwaiting)
  4252  	if trace.ok() {
  4253  		traceRelease(trace)
  4254  	}
  4255  
  4256  	dropg()
  4257  
  4258  	if fn := mp.waitunlockf; fn != nil {
  4259  		ok := fn(gp, mp.waitlock)
  4260  		mp.waitunlockf = nil
  4261  		mp.waitlock = nil
  4262  		if !ok {
  4263  			trace := traceAcquire()
  4264  			casgstatus(gp, _Gwaiting, _Grunnable)
  4265  			if bubble != nil {
  4266  				bubble.decActive()
  4267  			}
  4268  			if trace.ok() {
  4269  				trace.GoUnpark(gp, 2)
  4270  				traceRelease(trace)
  4271  			}
  4272  			execute(gp, true) // Schedule it back, never returns.
  4273  		}
  4274  	}
  4275  
  4276  	if bubble != nil {
  4277  		bubble.decActive()
  4278  	}
  4279  
  4280  	schedule()
  4281  }
  4282  
  4283  func goschedImpl(gp *g, preempted bool) {
  4284  	trace := traceAcquire()
  4285  	status := readgstatus(gp)
  4286  	if status&^_Gscan != _Grunning {
  4287  		dumpgstatus(gp)
  4288  		throw("bad g status")
  4289  	}
  4290  	if trace.ok() {
  4291  		// Trace the event before the transition. It may take a
  4292  		// stack trace, but we won't own the stack after the
  4293  		// transition anymore.
  4294  		if preempted {
  4295  			trace.GoPreempt()
  4296  		} else {
  4297  			trace.GoSched()
  4298  		}
  4299  	}
  4300  	casgstatus(gp, _Grunning, _Grunnable)
  4301  	if trace.ok() {
  4302  		traceRelease(trace)
  4303  	}
  4304  
  4305  	dropg()
  4306  	lock(&sched.lock)
  4307  	globrunqput(gp)
  4308  	unlock(&sched.lock)
  4309  
  4310  	if mainStarted {
  4311  		wakep()
  4312  	}
  4313  
  4314  	schedule()
  4315  }
  4316  
  4317  // Gosched continuation on g0.
  4318  func gosched_m(gp *g) {
  4319  	goschedImpl(gp, false)
  4320  }
  4321  
  4322  // goschedguarded is a forbidden-states-avoided version of gosched_m.
  4323  func goschedguarded_m(gp *g) {
  4324  	if !canPreemptM(gp.m) {
  4325  		gogo(&gp.sched) // never return
  4326  	}
  4327  	goschedImpl(gp, false)
  4328  }
  4329  
  4330  func gopreempt_m(gp *g) {
  4331  	goschedImpl(gp, true)
  4332  }
  4333  
  4334  // preemptPark parks gp and puts it in _Gpreempted.
  4335  //
  4336  //go:systemstack
  4337  func preemptPark(gp *g) {
  4338  	status := readgstatus(gp)
  4339  	if status&^_Gscan != _Grunning {
  4340  		dumpgstatus(gp)
  4341  		throw("bad g status")
  4342  	}
  4343  
  4344  	if gp.asyncSafePoint {
  4345  		// Double-check that async preemption does not
  4346  		// happen in SPWRITE assembly functions.
  4347  		// isAsyncSafePoint must exclude this case.
  4348  		f := findfunc(gp.sched.pc)
  4349  		if !f.valid() {
  4350  			throw("preempt at unknown pc")
  4351  		}
  4352  		if f.flag&abi.FuncFlagSPWrite != 0 {
  4353  			println("runtime: unexpected SPWRITE function", funcname(f), "in async preempt")
  4354  			throw("preempt SPWRITE")
  4355  		}
  4356  	}
  4357  
  4358  	// Transition from _Grunning to _Gscan|_Gpreempted. We can't
  4359  	// be in _Grunning when we dropg because then we'd be running
  4360  	// without an M, but the moment we're in _Gpreempted,
  4361  	// something could claim this G before we've fully cleaned it
  4362  	// up. Hence, we set the scan bit to lock down further
  4363  	// transitions until we can dropg.
  4364  	casGToPreemptScan(gp, _Grunning, _Gscan|_Gpreempted)
  4365  	dropg()
  4366  
  4367  	// Be careful about how we trace this next event. The ordering
  4368  	// is subtle.
  4369  	//
  4370  	// The moment we CAS into _Gpreempted, suspendG could CAS to
  4371  	// _Gwaiting, do its work, and ready the goroutine. All of
  4372  	// this could happen before we even get the chance to emit
  4373  	// an event. The end result is that the events could appear
  4374  	// out of order, and the tracer generally assumes the scheduler
  4375  	// takes care of the ordering between GoPark and GoUnpark.
  4376  	//
  4377  	// The answer here is simple: emit the event while we still hold
  4378  	// the _Gscan bit on the goroutine. We still need to traceAcquire
  4379  	// and traceRelease across the CAS because the tracer could be
  4380  	// what's calling suspendG in the first place, and we want the
  4381  	// CAS and event emission to appear atomic to the tracer.
  4382  	trace := traceAcquire()
  4383  	if trace.ok() {
  4384  		trace.GoPark(traceBlockPreempted, 0)
  4385  	}
  4386  	casfrom_Gscanstatus(gp, _Gscan|_Gpreempted, _Gpreempted)
  4387  	if trace.ok() {
  4388  		traceRelease(trace)
  4389  	}
  4390  	schedule()
  4391  }
  4392  
  4393  // goyield is like Gosched, but it:
  4394  // - emits a GoPreempt trace event instead of a GoSched trace event
  4395  // - puts the current G on the runq of the current P instead of the globrunq
  4396  //
  4397  // goyield should be an internal detail,
  4398  // but widely used packages access it using linkname.
  4399  // Notable members of the hall of shame include:
  4400  //   - gvisor.dev/gvisor
  4401  //   - github.com/sagernet/gvisor
  4402  //
  4403  // Do not remove or change the type signature.
  4404  // See go.dev/issue/67401.
  4405  //
  4406  //go:linkname goyield
  4407  func goyield() {
  4408  	checkTimeouts()
  4409  	mcall(goyield_m)
  4410  }
  4411  
  4412  func goyield_m(gp *g) {
  4413  	trace := traceAcquire()
  4414  	pp := gp.m.p.ptr()
  4415  	if trace.ok() {
  4416  		// Trace the event before the transition. It may take a
  4417  		// stack trace, but we won't own the stack after the
  4418  		// transition anymore.
  4419  		trace.GoPreempt()
  4420  	}
  4421  	casgstatus(gp, _Grunning, _Grunnable)
  4422  	if trace.ok() {
  4423  		traceRelease(trace)
  4424  	}
  4425  	dropg()
  4426  	runqput(pp, gp, false)
  4427  	schedule()
  4428  }
  4429  
  4430  // Finishes execution of the current goroutine.
  4431  func goexit1() {
  4432  	if raceenabled {
  4433  		if gp := getg(); gp.bubble != nil {
  4434  			racereleasemergeg(gp, gp.bubble.raceaddr())
  4435  		}
  4436  		racegoend()
  4437  	}
  4438  	trace := traceAcquire()
  4439  	if trace.ok() {
  4440  		trace.GoEnd()
  4441  		traceRelease(trace)
  4442  	}
  4443  	mcall(goexit0)
  4444  }
  4445  
  4446  // goexit continuation on g0.
  4447  func goexit0(gp *g) {
  4448  	gdestroy(gp)
  4449  	schedule()
  4450  }
  4451  
  4452  func gdestroy(gp *g) {
  4453  	mp := getg().m
  4454  	pp := mp.p.ptr()
  4455  
  4456  	casgstatus(gp, _Grunning, _Gdead)
  4457  	gcController.addScannableStack(pp, -int64(gp.stack.hi-gp.stack.lo))
  4458  	if isSystemGoroutine(gp, false) {
  4459  		sched.ngsys.Add(-1)
  4460  	}
  4461  	gp.m = nil
  4462  	locked := gp.lockedm != 0
  4463  	gp.lockedm = 0
  4464  	mp.lockedg = 0
  4465  	gp.preemptStop = false
  4466  	gp.paniconfault = false
  4467  	gp._defer = nil // should be true already but just in case.
  4468  	gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
  4469  	gp.writebuf = nil
  4470  	gp.waitreason = waitReasonZero
  4471  	gp.param = nil
  4472  	gp.labels = nil
  4473  	gp.timer = nil
  4474  	gp.bubble = nil
  4475  
  4476  	if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
  4477  		// Flush assist credit to the global pool. This gives
  4478  		// better information to pacing if the application is
  4479  		// rapidly creating an exiting goroutines.
  4480  		assistWorkPerByte := gcController.assistWorkPerByte.Load()
  4481  		scanCredit := int64(assistWorkPerByte * float64(gp.gcAssistBytes))
  4482  		gcController.bgScanCredit.Add(scanCredit)
  4483  		gp.gcAssistBytes = 0
  4484  	}
  4485  
  4486  	dropg()
  4487  
  4488  	if GOARCH == "wasm" { // no threads yet on wasm
  4489  		gfput(pp, gp)
  4490  		return
  4491  	}
  4492  
  4493  	if locked && mp.lockedInt != 0 {
  4494  		print("runtime: mp.lockedInt = ", mp.lockedInt, "\n")
  4495  		if mp.isextra {
  4496  			throw("runtime.Goexit called in a thread that was not created by the Go runtime")
  4497  		}
  4498  		throw("exited a goroutine internally locked to the OS thread")
  4499  	}
  4500  	gfput(pp, gp)
  4501  	if locked {
  4502  		// The goroutine may have locked this thread because
  4503  		// it put it in an unusual kernel state. Kill it
  4504  		// rather than returning it to the thread pool.
  4505  
  4506  		// Return to mstart, which will release the P and exit
  4507  		// the thread.
  4508  		if GOOS != "plan9" { // See golang.org/issue/22227.
  4509  			gogo(&mp.g0.sched)
  4510  		} else {
  4511  			// Clear lockedExt on plan9 since we may end up re-using
  4512  			// this thread.
  4513  			mp.lockedExt = 0
  4514  		}
  4515  	}
  4516  }
  4517  
  4518  // save updates getg().sched to refer to pc and sp so that a following
  4519  // gogo will restore pc and sp.
  4520  //
  4521  // save must not have write barriers because invoking a write barrier
  4522  // can clobber getg().sched.
  4523  //
  4524  //go:nosplit
  4525  //go:nowritebarrierrec
  4526  func save(pc, sp, bp uintptr) {
  4527  	gp := getg()
  4528  
  4529  	if gp == gp.m.g0 || gp == gp.m.gsignal {
  4530  		// m.g0.sched is special and must describe the context
  4531  		// for exiting the thread. mstart1 writes to it directly.
  4532  		// m.gsignal.sched should not be used at all.
  4533  		// This check makes sure save calls do not accidentally
  4534  		// run in contexts where they'd write to system g's.
  4535  		throw("save on system g not allowed")
  4536  	}
  4537  
  4538  	gp.sched.pc = pc
  4539  	gp.sched.sp = sp
  4540  	gp.sched.lr = 0
  4541  	gp.sched.bp = bp
  4542  	// We need to ensure ctxt is zero, but can't have a write
  4543  	// barrier here. However, it should always already be zero.
  4544  	// Assert that.
  4545  	if gp.sched.ctxt != nil {
  4546  		badctxt()
  4547  	}
  4548  }
  4549  
  4550  // The goroutine g is about to enter a system call.
  4551  // Record that it's not using the cpu anymore.
  4552  // This is called only from the go syscall library and cgocall,
  4553  // not from the low-level system calls used by the runtime.
  4554  //
  4555  // Entersyscall cannot split the stack: the save must
  4556  // make g->sched refer to the caller's stack segment, because
  4557  // entersyscall is going to return immediately after.
  4558  //
  4559  // Nothing entersyscall calls can split the stack either.
  4560  // We cannot safely move the stack during an active call to syscall,
  4561  // because we do not know which of the uintptr arguments are
  4562  // really pointers (back into the stack).
  4563  // In practice, this means that we make the fast path run through
  4564  // entersyscall doing no-split things, and the slow path has to use systemstack
  4565  // to run bigger things on the system stack.
  4566  //
  4567  // reentersyscall is the entry point used by cgo callbacks, where explicitly
  4568  // saved SP and PC are restored. This is needed when exitsyscall will be called
  4569  // from a function further up in the call stack than the parent, as g->syscallsp
  4570  // must always point to a valid stack frame. entersyscall below is the normal
  4571  // entry point for syscalls, which obtains the SP and PC from the caller.
  4572  //
  4573  //go:nosplit
  4574  func reentersyscall(pc, sp, bp uintptr) {
  4575  	trace := traceAcquire()
  4576  	gp := getg()
  4577  
  4578  	// Disable preemption because during this function g is in Gsyscall status,
  4579  	// but can have inconsistent g->sched, do not let GC observe it.
  4580  	gp.m.locks++
  4581  
  4582  	// Entersyscall must not call any function that might split/grow the stack.
  4583  	// (See details in comment above.)
  4584  	// Catch calls that might, by replacing the stack guard with something that
  4585  	// will trip any stack check and leaving a flag to tell newstack to die.
  4586  	gp.stackguard0 = stackPreempt
  4587  	gp.throwsplit = true
  4588  
  4589  	// Leave SP around for GC and traceback.
  4590  	save(pc, sp, bp)
  4591  	gp.syscallsp = sp
  4592  	gp.syscallpc = pc
  4593  	gp.syscallbp = bp
  4594  	casgstatus(gp, _Grunning, _Gsyscall)
  4595  	if staticLockRanking {
  4596  		// When doing static lock ranking casgstatus can call
  4597  		// systemstack which clobbers g.sched.
  4598  		save(pc, sp, bp)
  4599  	}
  4600  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4601  		systemstack(func() {
  4602  			print("entersyscall inconsistent sp ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4603  			throw("entersyscall")
  4604  		})
  4605  	}
  4606  	if gp.syscallbp != 0 && gp.syscallbp < gp.stack.lo || gp.stack.hi < gp.syscallbp {
  4607  		systemstack(func() {
  4608  			print("entersyscall inconsistent bp ", hex(gp.syscallbp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4609  			throw("entersyscall")
  4610  		})
  4611  	}
  4612  
  4613  	if trace.ok() {
  4614  		systemstack(func() {
  4615  			trace.GoSysCall()
  4616  			traceRelease(trace)
  4617  		})
  4618  		// systemstack itself clobbers g.sched.{pc,sp} and we might
  4619  		// need them later when the G is genuinely blocked in a
  4620  		// syscall
  4621  		save(pc, sp, bp)
  4622  	}
  4623  
  4624  	if sched.sysmonwait.Load() {
  4625  		systemstack(entersyscall_sysmon)
  4626  		save(pc, sp, bp)
  4627  	}
  4628  
  4629  	if gp.m.p.ptr().runSafePointFn != 0 {
  4630  		// runSafePointFn may stack split if run on this stack
  4631  		systemstack(runSafePointFn)
  4632  		save(pc, sp, bp)
  4633  	}
  4634  
  4635  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4636  	pp := gp.m.p.ptr()
  4637  	pp.m = 0
  4638  	gp.m.oldp.set(pp)
  4639  	gp.m.p = 0
  4640  	atomic.Store(&pp.status, _Psyscall)
  4641  	if sched.gcwaiting.Load() {
  4642  		systemstack(entersyscall_gcwait)
  4643  		save(pc, sp, bp)
  4644  	}
  4645  
  4646  	gp.m.locks--
  4647  }
  4648  
  4649  // Standard syscall entry used by the go syscall library and normal cgo calls.
  4650  //
  4651  // This is exported via linkname to assembly in the syscall package and x/sys.
  4652  //
  4653  // Other packages should not be accessing entersyscall directly,
  4654  // but widely used packages access it using linkname.
  4655  // Notable members of the hall of shame include:
  4656  //   - gvisor.dev/gvisor
  4657  //
  4658  // Do not remove or change the type signature.
  4659  // See go.dev/issue/67401.
  4660  //
  4661  //go:nosplit
  4662  //go:linkname entersyscall
  4663  func entersyscall() {
  4664  	// N.B. getcallerfp cannot be written directly as argument in the call
  4665  	// to reentersyscall because it forces spilling the other arguments to
  4666  	// the stack. This results in exceeding the nosplit stack requirements
  4667  	// on some platforms.
  4668  	fp := getcallerfp()
  4669  	reentersyscall(sys.GetCallerPC(), sys.GetCallerSP(), fp)
  4670  }
  4671  
  4672  func entersyscall_sysmon() {
  4673  	lock(&sched.lock)
  4674  	if sched.sysmonwait.Load() {
  4675  		sched.sysmonwait.Store(false)
  4676  		notewakeup(&sched.sysmonnote)
  4677  	}
  4678  	unlock(&sched.lock)
  4679  }
  4680  
  4681  func entersyscall_gcwait() {
  4682  	gp := getg()
  4683  	pp := gp.m.oldp.ptr()
  4684  
  4685  	lock(&sched.lock)
  4686  	trace := traceAcquire()
  4687  	if sched.stopwait > 0 && atomic.Cas(&pp.status, _Psyscall, _Pgcstop) {
  4688  		if trace.ok() {
  4689  			// This is a steal in the new tracer. While it's very likely
  4690  			// that we were the ones to put this P into _Psyscall, between
  4691  			// then and now it's totally possible it had been stolen and
  4692  			// then put back into _Psyscall for us to acquire here. In such
  4693  			// case ProcStop would be incorrect.
  4694  			//
  4695  			// TODO(mknyszek): Consider emitting a ProcStop instead when
  4696  			// gp.m.syscalltick == pp.syscalltick, since then we know we never
  4697  			// lost the P.
  4698  			trace.ProcSteal(pp, true)
  4699  			traceRelease(trace)
  4700  		}
  4701  		pp.gcStopTime = nanotime()
  4702  		pp.syscalltick++
  4703  		if sched.stopwait--; sched.stopwait == 0 {
  4704  			notewakeup(&sched.stopnote)
  4705  		}
  4706  	} else if trace.ok() {
  4707  		traceRelease(trace)
  4708  	}
  4709  	unlock(&sched.lock)
  4710  }
  4711  
  4712  // The same as entersyscall(), but with a hint that the syscall is blocking.
  4713  
  4714  // entersyscallblock should be an internal detail,
  4715  // but widely used packages access it using linkname.
  4716  // Notable members of the hall of shame include:
  4717  //   - gvisor.dev/gvisor
  4718  //
  4719  // Do not remove or change the type signature.
  4720  // See go.dev/issue/67401.
  4721  //
  4722  //go:linkname entersyscallblock
  4723  //go:nosplit
  4724  func entersyscallblock() {
  4725  	gp := getg()
  4726  
  4727  	gp.m.locks++ // see comment in entersyscall
  4728  	gp.throwsplit = true
  4729  	gp.stackguard0 = stackPreempt // see comment in entersyscall
  4730  	gp.m.syscalltick = gp.m.p.ptr().syscalltick
  4731  	gp.m.p.ptr().syscalltick++
  4732  
  4733  	// Leave SP around for GC and traceback.
  4734  	pc := sys.GetCallerPC()
  4735  	sp := sys.GetCallerSP()
  4736  	bp := getcallerfp()
  4737  	save(pc, sp, bp)
  4738  	gp.syscallsp = gp.sched.sp
  4739  	gp.syscallpc = gp.sched.pc
  4740  	gp.syscallbp = gp.sched.bp
  4741  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4742  		sp1 := sp
  4743  		sp2 := gp.sched.sp
  4744  		sp3 := gp.syscallsp
  4745  		systemstack(func() {
  4746  			print("entersyscallblock inconsistent sp ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4747  			throw("entersyscallblock")
  4748  		})
  4749  	}
  4750  	casgstatus(gp, _Grunning, _Gsyscall)
  4751  	if gp.syscallsp < gp.stack.lo || gp.stack.hi < gp.syscallsp {
  4752  		systemstack(func() {
  4753  			print("entersyscallblock inconsistent sp ", hex(sp), " ", hex(gp.sched.sp), " ", hex(gp.syscallsp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4754  			throw("entersyscallblock")
  4755  		})
  4756  	}
  4757  	if gp.syscallbp != 0 && gp.syscallbp < gp.stack.lo || gp.stack.hi < gp.syscallbp {
  4758  		systemstack(func() {
  4759  			print("entersyscallblock inconsistent bp ", hex(bp), " ", hex(gp.sched.bp), " ", hex(gp.syscallbp), " [", hex(gp.stack.lo), ",", hex(gp.stack.hi), "]\n")
  4760  			throw("entersyscallblock")
  4761  		})
  4762  	}
  4763  
  4764  	systemstack(entersyscallblock_handoff)
  4765  
  4766  	// Resave for traceback during blocked call.
  4767  	save(sys.GetCallerPC(), sys.GetCallerSP(), getcallerfp())
  4768  
  4769  	gp.m.locks--
  4770  }
  4771  
  4772  func entersyscallblock_handoff() {
  4773  	trace := traceAcquire()
  4774  	if trace.ok() {
  4775  		trace.GoSysCall()
  4776  		traceRelease(trace)
  4777  	}
  4778  	handoffp(releasep())
  4779  }
  4780  
  4781  // The goroutine g exited its system call.
  4782  // Arrange for it to run on a cpu again.
  4783  // This is called only from the go syscall library, not
  4784  // from the low-level system calls used by the runtime.
  4785  //
  4786  // Write barriers are not allowed because our P may have been stolen.
  4787  //
  4788  // This is exported via linkname to assembly in the syscall package.
  4789  //
  4790  // exitsyscall should be an internal detail,
  4791  // but widely used packages access it using linkname.
  4792  // Notable members of the hall of shame include:
  4793  //   - gvisor.dev/gvisor
  4794  //
  4795  // Do not remove or change the type signature.
  4796  // See go.dev/issue/67401.
  4797  //
  4798  //go:nosplit
  4799  //go:nowritebarrierrec
  4800  //go:linkname exitsyscall
  4801  func exitsyscall() {
  4802  	gp := getg()
  4803  
  4804  	gp.m.locks++ // see comment in entersyscall
  4805  	if sys.GetCallerSP() > gp.syscallsp {
  4806  		throw("exitsyscall: syscall frame is no longer valid")
  4807  	}
  4808  
  4809  	gp.waitsince = 0
  4810  	oldp := gp.m.oldp.ptr()
  4811  	gp.m.oldp = 0
  4812  	if exitsyscallfast(oldp) {
  4813  		// When exitsyscallfast returns success, we have a P so can now use
  4814  		// write barriers
  4815  		if goroutineProfile.active {
  4816  			// Make sure that gp has had its stack written out to the goroutine
  4817  			// profile, exactly as it was when the goroutine profiler first
  4818  			// stopped the world.
  4819  			systemstack(func() {
  4820  				tryRecordGoroutineProfileWB(gp)
  4821  			})
  4822  		}
  4823  		trace := traceAcquire()
  4824  		if trace.ok() {
  4825  			lostP := oldp != gp.m.p.ptr() || gp.m.syscalltick != gp.m.p.ptr().syscalltick
  4826  			systemstack(func() {
  4827  				// Write out syscall exit eagerly.
  4828  				//
  4829  				// It's important that we write this *after* we know whether we
  4830  				// lost our P or not (determined by exitsyscallfast).
  4831  				trace.GoSysExit(lostP)
  4832  				if lostP {
  4833  					// We lost the P at some point, even though we got it back here.
  4834  					// Trace that we're starting again, because there was a tracev2.GoSysBlock
  4835  					// call somewhere in exitsyscallfast (indicating that this goroutine
  4836  					// had blocked) and we're about to start running again.
  4837  					trace.GoStart()
  4838  				}
  4839  			})
  4840  		}
  4841  		// There's a cpu for us, so we can run.
  4842  		gp.m.p.ptr().syscalltick++
  4843  		// We need to cas the status and scan before resuming...
  4844  		casgstatus(gp, _Gsyscall, _Grunning)
  4845  		if trace.ok() {
  4846  			traceRelease(trace)
  4847  		}
  4848  
  4849  		// Garbage collector isn't running (since we are),
  4850  		// so okay to clear syscallsp.
  4851  		gp.syscallsp = 0
  4852  		gp.m.locks--
  4853  		if gp.preempt {
  4854  			// restore the preemption request in case we've cleared it in newstack
  4855  			gp.stackguard0 = stackPreempt
  4856  		} else {
  4857  			// otherwise restore the real stackGuard, we've spoiled it in entersyscall/entersyscallblock
  4858  			gp.stackguard0 = gp.stack.lo + stackGuard
  4859  		}
  4860  		gp.throwsplit = false
  4861  
  4862  		if sched.disable.user && !schedEnabled(gp) {
  4863  			// Scheduling of this goroutine is disabled.
  4864  			Gosched()
  4865  		}
  4866  
  4867  		return
  4868  	}
  4869  
  4870  	gp.m.locks--
  4871  
  4872  	// Call the scheduler.
  4873  	mcall(exitsyscall0)
  4874  
  4875  	// Scheduler returned, so we're allowed to run now.
  4876  	// Delete the syscallsp information that we left for
  4877  	// the garbage collector during the system call.
  4878  	// Must wait until now because until gosched returns
  4879  	// we don't know for sure that the garbage collector
  4880  	// is not running.
  4881  	gp.syscallsp = 0
  4882  	gp.m.p.ptr().syscalltick++
  4883  	gp.throwsplit = false
  4884  }
  4885  
  4886  //go:nosplit
  4887  func exitsyscallfast(oldp *p) bool {
  4888  	// Freezetheworld sets stopwait but does not retake P's.
  4889  	if sched.stopwait == freezeStopWait {
  4890  		return false
  4891  	}
  4892  
  4893  	// Try to re-acquire the last P.
  4894  	trace := traceAcquire()
  4895  	if oldp != nil && oldp.status == _Psyscall && atomic.Cas(&oldp.status, _Psyscall, _Pidle) {
  4896  		// There's a cpu for us, so we can run.
  4897  		wirep(oldp)
  4898  		exitsyscallfast_reacquired(trace)
  4899  		if trace.ok() {
  4900  			traceRelease(trace)
  4901  		}
  4902  		return true
  4903  	}
  4904  	if trace.ok() {
  4905  		traceRelease(trace)
  4906  	}
  4907  
  4908  	// Try to get any other idle P.
  4909  	if sched.pidle != 0 {
  4910  		var ok bool
  4911  		systemstack(func() {
  4912  			ok = exitsyscallfast_pidle()
  4913  		})
  4914  		if ok {
  4915  			return true
  4916  		}
  4917  	}
  4918  	return false
  4919  }
  4920  
  4921  // exitsyscallfast_reacquired is the exitsyscall path on which this G
  4922  // has successfully reacquired the P it was running on before the
  4923  // syscall.
  4924  //
  4925  //go:nosplit
  4926  func exitsyscallfast_reacquired(trace traceLocker) {
  4927  	gp := getg()
  4928  	if gp.m.syscalltick != gp.m.p.ptr().syscalltick {
  4929  		if trace.ok() {
  4930  			// The p was retaken and then enter into syscall again (since gp.m.syscalltick has changed).
  4931  			// tracev2.GoSysBlock for this syscall was already emitted,
  4932  			// but here we effectively retake the p from the new syscall running on the same p.
  4933  			systemstack(func() {
  4934  				// We're stealing the P. It's treated
  4935  				// as if it temporarily stopped running. Then, start running.
  4936  				trace.ProcSteal(gp.m.p.ptr(), true)
  4937  				trace.ProcStart()
  4938  			})
  4939  		}
  4940  		gp.m.p.ptr().syscalltick++
  4941  	}
  4942  }
  4943  
  4944  func exitsyscallfast_pidle() bool {
  4945  	lock(&sched.lock)
  4946  	pp, _ := pidleget(0)
  4947  	if pp != nil && sched.sysmonwait.Load() {
  4948  		sched.sysmonwait.Store(false)
  4949  		notewakeup(&sched.sysmonnote)
  4950  	}
  4951  	unlock(&sched.lock)
  4952  	if pp != nil {
  4953  		acquirep(pp)
  4954  		return true
  4955  	}
  4956  	return false
  4957  }
  4958  
  4959  // exitsyscall slow path on g0.
  4960  // Failed to acquire P, enqueue gp as runnable.
  4961  //
  4962  // Called via mcall, so gp is the calling g from this M.
  4963  //
  4964  //go:nowritebarrierrec
  4965  func exitsyscall0(gp *g) {
  4966  	var trace traceLocker
  4967  	traceExitingSyscall()
  4968  	trace = traceAcquire()
  4969  	casgstatus(gp, _Gsyscall, _Grunnable)
  4970  	traceExitedSyscall()
  4971  	if trace.ok() {
  4972  		// Write out syscall exit eagerly.
  4973  		//
  4974  		// It's important that we write this *after* we know whether we
  4975  		// lost our P or not (determined by exitsyscallfast).
  4976  		trace.GoSysExit(true)
  4977  		traceRelease(trace)
  4978  	}
  4979  	dropg()
  4980  	lock(&sched.lock)
  4981  	var pp *p
  4982  	if schedEnabled(gp) {
  4983  		pp, _ = pidleget(0)
  4984  	}
  4985  	var locked bool
  4986  	if pp == nil {
  4987  		globrunqput(gp)
  4988  
  4989  		// Below, we stoplockedm if gp is locked. globrunqput releases
  4990  		// ownership of gp, so we must check if gp is locked prior to
  4991  		// committing the release by unlocking sched.lock, otherwise we
  4992  		// could race with another M transitioning gp from unlocked to
  4993  		// locked.
  4994  		locked = gp.lockedm != 0
  4995  	} else if sched.sysmonwait.Load() {
  4996  		sched.sysmonwait.Store(false)
  4997  		notewakeup(&sched.sysmonnote)
  4998  	}
  4999  	unlock(&sched.lock)
  5000  	if pp != nil {
  5001  		acquirep(pp)
  5002  		execute(gp, false) // Never returns.
  5003  	}
  5004  	if locked {
  5005  		// Wait until another thread schedules gp and so m again.
  5006  		//
  5007  		// N.B. lockedm must be this M, as this g was running on this M
  5008  		// before entersyscall.
  5009  		stoplockedm()
  5010  		execute(gp, false) // Never returns.
  5011  	}
  5012  	stopm()
  5013  	schedule() // Never returns.
  5014  }
  5015  
  5016  // Called from syscall package before fork.
  5017  //
  5018  // syscall_runtime_BeforeFork is for package syscall,
  5019  // but widely used packages access it using linkname.
  5020  // Notable members of the hall of shame include:
  5021  //   - gvisor.dev/gvisor
  5022  //
  5023  // Do not remove or change the type signature.
  5024  // See go.dev/issue/67401.
  5025  //
  5026  //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
  5027  //go:nosplit
  5028  func syscall_runtime_BeforeFork() {
  5029  	gp := getg().m.curg
  5030  
  5031  	// Block signals during a fork, so that the child does not run
  5032  	// a signal handler before exec if a signal is sent to the process
  5033  	// group. See issue #18600.
  5034  	gp.m.locks++
  5035  	sigsave(&gp.m.sigmask)
  5036  	sigblock(false)
  5037  
  5038  	// This function is called before fork in syscall package.
  5039  	// Code between fork and exec must not allocate memory nor even try to grow stack.
  5040  	// Here we spoil g.stackguard0 to reliably detect any attempts to grow stack.
  5041  	// runtime_AfterFork will undo this in parent process, but not in child.
  5042  	gp.stackguard0 = stackFork
  5043  }
  5044  
  5045  // Called from syscall package after fork in parent.
  5046  //
  5047  // syscall_runtime_AfterFork is for package syscall,
  5048  // but widely used packages access it using linkname.
  5049  // Notable members of the hall of shame include:
  5050  //   - gvisor.dev/gvisor
  5051  //
  5052  // Do not remove or change the type signature.
  5053  // See go.dev/issue/67401.
  5054  //
  5055  //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
  5056  //go:nosplit
  5057  func syscall_runtime_AfterFork() {
  5058  	gp := getg().m.curg
  5059  
  5060  	// See the comments in beforefork.
  5061  	gp.stackguard0 = gp.stack.lo + stackGuard
  5062  
  5063  	msigrestore(gp.m.sigmask)
  5064  
  5065  	gp.m.locks--
  5066  }
  5067  
  5068  // inForkedChild is true while manipulating signals in the child process.
  5069  // This is used to avoid calling libc functions in case we are using vfork.
  5070  var inForkedChild bool
  5071  
  5072  // Called from syscall package after fork in child.
  5073  // It resets non-sigignored signals to the default handler, and
  5074  // restores the signal mask in preparation for the exec.
  5075  //
  5076  // Because this might be called during a vfork, and therefore may be
  5077  // temporarily sharing address space with the parent process, this must
  5078  // not change any global variables or calling into C code that may do so.
  5079  //
  5080  // syscall_runtime_AfterForkInChild is for package syscall,
  5081  // but widely used packages access it using linkname.
  5082  // Notable members of the hall of shame include:
  5083  //   - gvisor.dev/gvisor
  5084  //
  5085  // Do not remove or change the type signature.
  5086  // See go.dev/issue/67401.
  5087  //
  5088  //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
  5089  //go:nosplit
  5090  //go:nowritebarrierrec
  5091  func syscall_runtime_AfterForkInChild() {
  5092  	// It's OK to change the global variable inForkedChild here
  5093  	// because we are going to change it back. There is no race here,
  5094  	// because if we are sharing address space with the parent process,
  5095  	// then the parent process can not be running concurrently.
  5096  	inForkedChild = true
  5097  
  5098  	clearSignalHandlers()
  5099  
  5100  	// When we are the child we are the only thread running,
  5101  	// so we know that nothing else has changed gp.m.sigmask.
  5102  	msigrestore(getg().m.sigmask)
  5103  
  5104  	inForkedChild = false
  5105  }
  5106  
  5107  // pendingPreemptSignals is the number of preemption signals
  5108  // that have been sent but not received. This is only used on Darwin.
  5109  // For #41702.
  5110  var pendingPreemptSignals atomic.Int32
  5111  
  5112  // Called from syscall package before Exec.
  5113  //
  5114  //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
  5115  func syscall_runtime_BeforeExec() {
  5116  	// Prevent thread creation during exec.
  5117  	execLock.lock()
  5118  
  5119  	// On Darwin, wait for all pending preemption signals to
  5120  	// be received. See issue #41702.
  5121  	if GOOS == "darwin" || GOOS == "ios" {
  5122  		for pendingPreemptSignals.Load() > 0 {
  5123  			osyield()
  5124  		}
  5125  	}
  5126  }
  5127  
  5128  // Called from syscall package after Exec.
  5129  //
  5130  //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
  5131  func syscall_runtime_AfterExec() {
  5132  	execLock.unlock()
  5133  }
  5134  
  5135  // Allocate a new g, with a stack big enough for stacksize bytes.
  5136  func malg(stacksize int32) *g {
  5137  	newg := new(g)
  5138  	if stacksize >= 0 {
  5139  		stacksize = round2(stackSystem + stacksize)
  5140  		systemstack(func() {
  5141  			newg.stack = stackalloc(uint32(stacksize))
  5142  			if valgrindenabled {
  5143  				newg.valgrindStackID = valgrindRegisterStack(unsafe.Pointer(newg.stack.lo), unsafe.Pointer(newg.stack.hi))
  5144  			}
  5145  		})
  5146  		newg.stackguard0 = newg.stack.lo + stackGuard
  5147  		newg.stackguard1 = ^uintptr(0)
  5148  		// Clear the bottom word of the stack. We record g
  5149  		// there on gsignal stack during VDSO on ARM and ARM64.
  5150  		*(*uintptr)(unsafe.Pointer(newg.stack.lo)) = 0
  5151  	}
  5152  	return newg
  5153  }
  5154  
  5155  // Create a new g running fn.
  5156  // Put it on the queue of g's waiting to run.
  5157  // The compiler turns a go statement into a call to this.
  5158  func newproc(fn *funcval) {
  5159  	gp := getg()
  5160  	pc := sys.GetCallerPC()
  5161  	systemstack(func() {
  5162  		newg := newproc1(fn, gp, pc, false, waitReasonZero)
  5163  
  5164  		pp := getg().m.p.ptr()
  5165  		runqput(pp, newg, true)
  5166  
  5167  		if mainStarted {
  5168  			wakep()
  5169  		}
  5170  	})
  5171  }
  5172  
  5173  // Create a new g in state _Grunnable (or _Gwaiting if parked is true), starting at fn.
  5174  // callerpc is the address of the go statement that created this. The caller is responsible
  5175  // for adding the new g to the scheduler. If parked is true, waitreason must be non-zero.
  5176  func newproc1(fn *funcval, callergp *g, callerpc uintptr, parked bool, waitreason waitReason) *g {
  5177  	if fn == nil {
  5178  		fatal("go of nil func value")
  5179  	}
  5180  
  5181  	mp := acquirem() // disable preemption because we hold M and P in local vars.
  5182  	pp := mp.p.ptr()
  5183  	newg := gfget(pp)
  5184  	if newg == nil {
  5185  		newg = malg(stackMin)
  5186  		casgstatus(newg, _Gidle, _Gdead)
  5187  		allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
  5188  	}
  5189  	if newg.stack.hi == 0 {
  5190  		throw("newproc1: newg missing stack")
  5191  	}
  5192  
  5193  	if readgstatus(newg) != _Gdead {
  5194  		throw("newproc1: new g is not Gdead")
  5195  	}
  5196  
  5197  	totalSize := uintptr(4*goarch.PtrSize + sys.MinFrameSize) // extra space in case of reads slightly beyond frame
  5198  	totalSize = alignUp(totalSize, sys.StackAlign)
  5199  	sp := newg.stack.hi - totalSize
  5200  	if usesLR {
  5201  		// caller's LR
  5202  		*(*uintptr)(unsafe.Pointer(sp)) = 0
  5203  		prepGoExitFrame(sp)
  5204  	}
  5205  	if GOARCH == "arm64" {
  5206  		// caller's FP
  5207  		*(*uintptr)(unsafe.Pointer(sp - goarch.PtrSize)) = 0
  5208  	}
  5209  
  5210  	memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
  5211  	newg.sched.sp = sp
  5212  	newg.stktopsp = sp
  5213  	newg.sched.pc = abi.FuncPCABI0(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
  5214  	newg.sched.g = guintptr(unsafe.Pointer(newg))
  5215  	gostartcallfn(&newg.sched, fn)
  5216  	newg.parentGoid = callergp.goid
  5217  	newg.gopc = callerpc
  5218  	newg.ancestors = saveAncestors(callergp)
  5219  	newg.startpc = fn.fn
  5220  	newg.runningCleanups.Store(false)
  5221  	if isSystemGoroutine(newg, false) {
  5222  		sched.ngsys.Add(1)
  5223  	} else {
  5224  		// Only user goroutines inherit synctest groups and pprof labels.
  5225  		newg.bubble = callergp.bubble
  5226  		if mp.curg != nil {
  5227  			newg.labels = mp.curg.labels
  5228  		}
  5229  		if goroutineProfile.active {
  5230  			// A concurrent goroutine profile is running. It should include
  5231  			// exactly the set of goroutines that were alive when the goroutine
  5232  			// profiler first stopped the world. That does not include newg, so
  5233  			// mark it as not needing a profile before transitioning it from
  5234  			// _Gdead.
  5235  			newg.goroutineProfiled.Store(goroutineProfileSatisfied)
  5236  		}
  5237  	}
  5238  	// Track initial transition?
  5239  	newg.trackingSeq = uint8(cheaprand())
  5240  	if newg.trackingSeq%gTrackingPeriod == 0 {
  5241  		newg.tracking = true
  5242  	}
  5243  	gcController.addScannableStack(pp, int64(newg.stack.hi-newg.stack.lo))
  5244  
  5245  	// Get a goid and switch to runnable. Make all this atomic to the tracer.
  5246  	trace := traceAcquire()
  5247  	var status uint32 = _Grunnable
  5248  	if parked {
  5249  		status = _Gwaiting
  5250  		newg.waitreason = waitreason
  5251  	}
  5252  	if pp.goidcache == pp.goidcacheend {
  5253  		// Sched.goidgen is the last allocated id,
  5254  		// this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
  5255  		// At startup sched.goidgen=0, so main goroutine receives goid=1.
  5256  		pp.goidcache = sched.goidgen.Add(_GoidCacheBatch)
  5257  		pp.goidcache -= _GoidCacheBatch - 1
  5258  		pp.goidcacheend = pp.goidcache + _GoidCacheBatch
  5259  	}
  5260  	newg.goid = pp.goidcache
  5261  	casgstatus(newg, _Gdead, status)
  5262  	pp.goidcache++
  5263  	newg.trace.reset()
  5264  	if trace.ok() {
  5265  		trace.GoCreate(newg, newg.startpc, parked)
  5266  		traceRelease(trace)
  5267  	}
  5268  
  5269  	// Set up race context.
  5270  	if raceenabled {
  5271  		newg.racectx = racegostart(callerpc)
  5272  		newg.raceignore = 0
  5273  		if newg.labels != nil {
  5274  			// See note in proflabel.go on labelSync's role in synchronizing
  5275  			// with the reads in the signal handler.
  5276  			racereleasemergeg(newg, unsafe.Pointer(&labelSync))
  5277  		}
  5278  	}
  5279  	releasem(mp)
  5280  
  5281  	return newg
  5282  }
  5283  
  5284  // saveAncestors copies previous ancestors of the given caller g and
  5285  // includes info for the current caller into a new set of tracebacks for
  5286  // a g being created.
  5287  func saveAncestors(callergp *g) *[]ancestorInfo {
  5288  	// Copy all prior info, except for the root goroutine (goid 0).
  5289  	if debug.tracebackancestors <= 0 || callergp.goid == 0 {
  5290  		return nil
  5291  	}
  5292  	var callerAncestors []ancestorInfo
  5293  	if callergp.ancestors != nil {
  5294  		callerAncestors = *callergp.ancestors
  5295  	}
  5296  	n := int32(len(callerAncestors)) + 1
  5297  	if n > debug.tracebackancestors {
  5298  		n = debug.tracebackancestors
  5299  	}
  5300  	ancestors := make([]ancestorInfo, n)
  5301  	copy(ancestors[1:], callerAncestors)
  5302  
  5303  	var pcs [tracebackInnerFrames]uintptr
  5304  	npcs := gcallers(callergp, 0, pcs[:])
  5305  	ipcs := make([]uintptr, npcs)
  5306  	copy(ipcs, pcs[:])
  5307  	ancestors[0] = ancestorInfo{
  5308  		pcs:  ipcs,
  5309  		goid: callergp.goid,
  5310  		gopc: callergp.gopc,
  5311  	}
  5312  
  5313  	ancestorsp := new([]ancestorInfo)
  5314  	*ancestorsp = ancestors
  5315  	return ancestorsp
  5316  }
  5317  
  5318  // Put on gfree list.
  5319  // If local list is too long, transfer a batch to the global list.
  5320  func gfput(pp *p, gp *g) {
  5321  	if readgstatus(gp) != _Gdead {
  5322  		throw("gfput: bad status (not Gdead)")
  5323  	}
  5324  
  5325  	stksize := gp.stack.hi - gp.stack.lo
  5326  
  5327  	if stksize != uintptr(startingStackSize) {
  5328  		// non-standard stack size - free it.
  5329  		stackfree(gp.stack)
  5330  		gp.stack.lo = 0
  5331  		gp.stack.hi = 0
  5332  		gp.stackguard0 = 0
  5333  		if valgrindenabled {
  5334  			valgrindDeregisterStack(gp.valgrindStackID)
  5335  			gp.valgrindStackID = 0
  5336  		}
  5337  	}
  5338  
  5339  	pp.gFree.push(gp)
  5340  	if pp.gFree.size >= 64 {
  5341  		var (
  5342  			stackQ   gQueue
  5343  			noStackQ gQueue
  5344  		)
  5345  		for pp.gFree.size >= 32 {
  5346  			gp := pp.gFree.pop()
  5347  			if gp.stack.lo == 0 {
  5348  				noStackQ.push(gp)
  5349  			} else {
  5350  				stackQ.push(gp)
  5351  			}
  5352  		}
  5353  		lock(&sched.gFree.lock)
  5354  		sched.gFree.noStack.pushAll(noStackQ)
  5355  		sched.gFree.stack.pushAll(stackQ)
  5356  		unlock(&sched.gFree.lock)
  5357  	}
  5358  }
  5359  
  5360  // Get from gfree list.
  5361  // If local list is empty, grab a batch from global list.
  5362  func gfget(pp *p) *g {
  5363  retry:
  5364  	if pp.gFree.empty() && (!sched.gFree.stack.empty() || !sched.gFree.noStack.empty()) {
  5365  		lock(&sched.gFree.lock)
  5366  		// Move a batch of free Gs to the P.
  5367  		for pp.gFree.size < 32 {
  5368  			// Prefer Gs with stacks.
  5369  			gp := sched.gFree.stack.pop()
  5370  			if gp == nil {
  5371  				gp = sched.gFree.noStack.pop()
  5372  				if gp == nil {
  5373  					break
  5374  				}
  5375  			}
  5376  			pp.gFree.push(gp)
  5377  		}
  5378  		unlock(&sched.gFree.lock)
  5379  		goto retry
  5380  	}
  5381  	gp := pp.gFree.pop()
  5382  	if gp == nil {
  5383  		return nil
  5384  	}
  5385  	if gp.stack.lo != 0 && gp.stack.hi-gp.stack.lo != uintptr(startingStackSize) {
  5386  		// Deallocate old stack. We kept it in gfput because it was the
  5387  		// right size when the goroutine was put on the free list, but
  5388  		// the right size has changed since then.
  5389  		systemstack(func() {
  5390  			stackfree(gp.stack)
  5391  			gp.stack.lo = 0
  5392  			gp.stack.hi = 0
  5393  			gp.stackguard0 = 0
  5394  			if valgrindenabled {
  5395  				valgrindDeregisterStack(gp.valgrindStackID)
  5396  				gp.valgrindStackID = 0
  5397  			}
  5398  		})
  5399  	}
  5400  	if gp.stack.lo == 0 {
  5401  		// Stack was deallocated in gfput or just above. Allocate a new one.
  5402  		systemstack(func() {
  5403  			gp.stack = stackalloc(startingStackSize)
  5404  			if valgrindenabled {
  5405  				gp.valgrindStackID = valgrindRegisterStack(unsafe.Pointer(gp.stack.lo), unsafe.Pointer(gp.stack.hi))
  5406  			}
  5407  		})
  5408  		gp.stackguard0 = gp.stack.lo + stackGuard
  5409  	} else {
  5410  		if raceenabled {
  5411  			racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5412  		}
  5413  		if msanenabled {
  5414  			msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5415  		}
  5416  		if asanenabled {
  5417  			asanunpoison(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  5418  		}
  5419  	}
  5420  	return gp
  5421  }
  5422  
  5423  // Purge all cached G's from gfree list to the global list.
  5424  func gfpurge(pp *p) {
  5425  	var (
  5426  		stackQ   gQueue
  5427  		noStackQ gQueue
  5428  	)
  5429  	for !pp.gFree.empty() {
  5430  		gp := pp.gFree.pop()
  5431  		if gp.stack.lo == 0 {
  5432  			noStackQ.push(gp)
  5433  		} else {
  5434  			stackQ.push(gp)
  5435  		}
  5436  	}
  5437  	lock(&sched.gFree.lock)
  5438  	sched.gFree.noStack.pushAll(noStackQ)
  5439  	sched.gFree.stack.pushAll(stackQ)
  5440  	unlock(&sched.gFree.lock)
  5441  }
  5442  
  5443  // Breakpoint executes a breakpoint trap.
  5444  func Breakpoint() {
  5445  	breakpoint()
  5446  }
  5447  
  5448  // dolockOSThread is called by LockOSThread and lockOSThread below
  5449  // after they modify m.locked. Do not allow preemption during this call,
  5450  // or else the m might be different in this function than in the caller.
  5451  //
  5452  //go:nosplit
  5453  func dolockOSThread() {
  5454  	if GOARCH == "wasm" {
  5455  		return // no threads on wasm yet
  5456  	}
  5457  	gp := getg()
  5458  	gp.m.lockedg.set(gp)
  5459  	gp.lockedm.set(gp.m)
  5460  }
  5461  
  5462  // LockOSThread wires the calling goroutine to its current operating system thread.
  5463  // The calling goroutine will always execute in that thread,
  5464  // and no other goroutine will execute in it,
  5465  // until the calling goroutine has made as many calls to
  5466  // [UnlockOSThread] as to LockOSThread.
  5467  // If the calling goroutine exits without unlocking the thread,
  5468  // the thread will be terminated.
  5469  //
  5470  // All init functions are run on the startup thread. Calling LockOSThread
  5471  // from an init function will cause the main function to be invoked on
  5472  // that thread.
  5473  //
  5474  // A goroutine should call LockOSThread before calling OS services or
  5475  // non-Go library functions that depend on per-thread state.
  5476  //
  5477  //go:nosplit
  5478  func LockOSThread() {
  5479  	if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
  5480  		// If we need to start a new thread from the locked
  5481  		// thread, we need the template thread. Start it now
  5482  		// while we're in a known-good state.
  5483  		startTemplateThread()
  5484  	}
  5485  	gp := getg()
  5486  	gp.m.lockedExt++
  5487  	if gp.m.lockedExt == 0 {
  5488  		gp.m.lockedExt--
  5489  		panic("LockOSThread nesting overflow")
  5490  	}
  5491  	dolockOSThread()
  5492  }
  5493  
  5494  //go:nosplit
  5495  func lockOSThread() {
  5496  	getg().m.lockedInt++
  5497  	dolockOSThread()
  5498  }
  5499  
  5500  // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
  5501  // after they update m->locked. Do not allow preemption during this call,
  5502  // or else the m might be in different in this function than in the caller.
  5503  //
  5504  //go:nosplit
  5505  func dounlockOSThread() {
  5506  	if GOARCH == "wasm" {
  5507  		return // no threads on wasm yet
  5508  	}
  5509  	gp := getg()
  5510  	if gp.m.lockedInt != 0 || gp.m.lockedExt != 0 {
  5511  		return
  5512  	}
  5513  	gp.m.lockedg = 0
  5514  	gp.lockedm = 0
  5515  }
  5516  
  5517  // UnlockOSThread undoes an earlier call to LockOSThread.
  5518  // If this drops the number of active LockOSThread calls on the
  5519  // calling goroutine to zero, it unwires the calling goroutine from
  5520  // its fixed operating system thread.
  5521  // If there are no active LockOSThread calls, this is a no-op.
  5522  //
  5523  // Before calling UnlockOSThread, the caller must ensure that the OS
  5524  // thread is suitable for running other goroutines. If the caller made
  5525  // any permanent changes to the state of the thread that would affect
  5526  // other goroutines, it should not call this function and thus leave
  5527  // the goroutine locked to the OS thread until the goroutine (and
  5528  // hence the thread) exits.
  5529  //
  5530  //go:nosplit
  5531  func UnlockOSThread() {
  5532  	gp := getg()
  5533  	if gp.m.lockedExt == 0 {
  5534  		return
  5535  	}
  5536  	gp.m.lockedExt--
  5537  	dounlockOSThread()
  5538  }
  5539  
  5540  //go:nosplit
  5541  func unlockOSThread() {
  5542  	gp := getg()
  5543  	if gp.m.lockedInt == 0 {
  5544  		systemstack(badunlockosthread)
  5545  	}
  5546  	gp.m.lockedInt--
  5547  	dounlockOSThread()
  5548  }
  5549  
  5550  func badunlockosthread() {
  5551  	throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
  5552  }
  5553  
  5554  func gcount() int32 {
  5555  	n := int32(atomic.Loaduintptr(&allglen)) - sched.gFree.stack.size - sched.gFree.noStack.size - sched.ngsys.Load()
  5556  	for _, pp := range allp {
  5557  		n -= pp.gFree.size
  5558  	}
  5559  
  5560  	// All these variables can be changed concurrently, so the result can be inconsistent.
  5561  	// But at least the current goroutine is running.
  5562  	if n < 1 {
  5563  		n = 1
  5564  	}
  5565  	return n
  5566  }
  5567  
  5568  func mcount() int32 {
  5569  	return int32(sched.mnext - sched.nmfreed)
  5570  }
  5571  
  5572  var prof struct {
  5573  	signalLock atomic.Uint32
  5574  
  5575  	// Must hold signalLock to write. Reads may be lock-free, but
  5576  	// signalLock should be taken to synchronize with changes.
  5577  	hz atomic.Int32
  5578  }
  5579  
  5580  func _System()                    { _System() }
  5581  func _ExternalCode()              { _ExternalCode() }
  5582  func _LostExternalCode()          { _LostExternalCode() }
  5583  func _GC()                        { _GC() }
  5584  func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
  5585  func _LostContendedRuntimeLock()  { _LostContendedRuntimeLock() }
  5586  func _VDSO()                      { _VDSO() }
  5587  
  5588  // Called if we receive a SIGPROF signal.
  5589  // Called by the signal handler, may run during STW.
  5590  //
  5591  //go:nowritebarrierrec
  5592  func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
  5593  	if prof.hz.Load() == 0 {
  5594  		return
  5595  	}
  5596  
  5597  	// If mp.profilehz is 0, then profiling is not enabled for this thread.
  5598  	// We must check this to avoid a deadlock between setcpuprofilerate
  5599  	// and the call to cpuprof.add, below.
  5600  	if mp != nil && mp.profilehz == 0 {
  5601  		return
  5602  	}
  5603  
  5604  	// On mips{,le}/arm, 64bit atomics are emulated with spinlocks, in
  5605  	// internal/runtime/atomic. If SIGPROF arrives while the program is inside
  5606  	// the critical section, it creates a deadlock (when writing the sample).
  5607  	// As a workaround, create a counter of SIGPROFs while in critical section
  5608  	// to store the count, and pass it to sigprof.add() later when SIGPROF is
  5609  	// received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
  5610  	if GOARCH == "mips" || GOARCH == "mipsle" || GOARCH == "arm" {
  5611  		if f := findfunc(pc); f.valid() {
  5612  			if stringslite.HasPrefix(funcname(f), "internal/runtime/atomic") {
  5613  				cpuprof.lostAtomic++
  5614  				return
  5615  			}
  5616  		}
  5617  		if GOARCH == "arm" && goarm < 7 && GOOS == "linux" && pc&0xffff0000 == 0xffff0000 {
  5618  			// internal/runtime/atomic functions call into kernel
  5619  			// helpers on arm < 7. See
  5620  			// internal/runtime/atomic/sys_linux_arm.s.
  5621  			cpuprof.lostAtomic++
  5622  			return
  5623  		}
  5624  	}
  5625  
  5626  	// Profiling runs concurrently with GC, so it must not allocate.
  5627  	// Set a trap in case the code does allocate.
  5628  	// Note that on windows, one thread takes profiles of all the
  5629  	// other threads, so mp is usually not getg().m.
  5630  	// In fact mp may not even be stopped.
  5631  	// See golang.org/issue/17165.
  5632  	getg().m.mallocing++
  5633  
  5634  	var u unwinder
  5635  	var stk [maxCPUProfStack]uintptr
  5636  	n := 0
  5637  	if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
  5638  		cgoOff := 0
  5639  		// Check cgoCallersUse to make sure that we are not
  5640  		// interrupting other code that is fiddling with
  5641  		// cgoCallers.  We are running in a signal handler
  5642  		// with all signals blocked, so we don't have to worry
  5643  		// about any other code interrupting us.
  5644  		if mp.cgoCallersUse.Load() == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
  5645  			for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
  5646  				cgoOff++
  5647  			}
  5648  			n += copy(stk[:], mp.cgoCallers[:cgoOff])
  5649  			mp.cgoCallers[0] = 0
  5650  		}
  5651  
  5652  		// Collect Go stack that leads to the cgo call.
  5653  		u.initAt(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, unwindSilentErrors)
  5654  	} else if usesLibcall() && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
  5655  		// Libcall, i.e. runtime syscall on windows.
  5656  		// Collect Go stack that leads to the call.
  5657  		u.initAt(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), unwindSilentErrors)
  5658  	} else if mp != nil && mp.vdsoSP != 0 {
  5659  		// VDSO call, e.g. nanotime1 on Linux.
  5660  		// Collect Go stack that leads to the call.
  5661  		u.initAt(mp.vdsoPC, mp.vdsoSP, 0, gp, unwindSilentErrors|unwindJumpStack)
  5662  	} else {
  5663  		u.initAt(pc, sp, lr, gp, unwindSilentErrors|unwindTrap|unwindJumpStack)
  5664  	}
  5665  	n += tracebackPCs(&u, 0, stk[n:])
  5666  
  5667  	if n <= 0 {
  5668  		// Normal traceback is impossible or has failed.
  5669  		// Account it against abstract "System" or "GC".
  5670  		n = 2
  5671  		if inVDSOPage(pc) {
  5672  			pc = abi.FuncPCABIInternal(_VDSO) + sys.PCQuantum
  5673  		} else if pc > firstmoduledata.etext {
  5674  			// "ExternalCode" is better than "etext".
  5675  			pc = abi.FuncPCABIInternal(_ExternalCode) + sys.PCQuantum
  5676  		}
  5677  		stk[0] = pc
  5678  		if mp.preemptoff != "" {
  5679  			stk[1] = abi.FuncPCABIInternal(_GC) + sys.PCQuantum
  5680  		} else {
  5681  			stk[1] = abi.FuncPCABIInternal(_System) + sys.PCQuantum
  5682  		}
  5683  	}
  5684  
  5685  	if prof.hz.Load() != 0 {
  5686  		// Note: it can happen on Windows that we interrupted a system thread
  5687  		// with no g, so gp could nil. The other nil checks are done out of
  5688  		// caution, but not expected to be nil in practice.
  5689  		var tagPtr *unsafe.Pointer
  5690  		if gp != nil && gp.m != nil && gp.m.curg != nil {
  5691  			tagPtr = &gp.m.curg.labels
  5692  		}
  5693  		cpuprof.add(tagPtr, stk[:n])
  5694  
  5695  		gprof := gp
  5696  		var mp *m
  5697  		var pp *p
  5698  		if gp != nil && gp.m != nil {
  5699  			if gp.m.curg != nil {
  5700  				gprof = gp.m.curg
  5701  			}
  5702  			mp = gp.m
  5703  			pp = gp.m.p.ptr()
  5704  		}
  5705  		traceCPUSample(gprof, mp, pp, stk[:n])
  5706  	}
  5707  	getg().m.mallocing--
  5708  }
  5709  
  5710  // setcpuprofilerate sets the CPU profiling rate to hz times per second.
  5711  // If hz <= 0, setcpuprofilerate turns off CPU profiling.
  5712  func setcpuprofilerate(hz int32) {
  5713  	// Force sane arguments.
  5714  	if hz < 0 {
  5715  		hz = 0
  5716  	}
  5717  
  5718  	// Disable preemption, otherwise we can be rescheduled to another thread
  5719  	// that has profiling enabled.
  5720  	gp := getg()
  5721  	gp.m.locks++
  5722  
  5723  	// Stop profiler on this thread so that it is safe to lock prof.
  5724  	// if a profiling signal came in while we had prof locked,
  5725  	// it would deadlock.
  5726  	setThreadCPUProfiler(0)
  5727  
  5728  	for !prof.signalLock.CompareAndSwap(0, 1) {
  5729  		osyield()
  5730  	}
  5731  	if prof.hz.Load() != hz {
  5732  		setProcessCPUProfiler(hz)
  5733  		prof.hz.Store(hz)
  5734  	}
  5735  	prof.signalLock.Store(0)
  5736  
  5737  	lock(&sched.lock)
  5738  	sched.profilehz = hz
  5739  	unlock(&sched.lock)
  5740  
  5741  	if hz != 0 {
  5742  		setThreadCPUProfiler(hz)
  5743  	}
  5744  
  5745  	gp.m.locks--
  5746  }
  5747  
  5748  // init initializes pp, which may be a freshly allocated p or a
  5749  // previously destroyed p, and transitions it to status _Pgcstop.
  5750  func (pp *p) init(id int32) {
  5751  	pp.id = id
  5752  	pp.status = _Pgcstop
  5753  	pp.sudogcache = pp.sudogbuf[:0]
  5754  	pp.deferpool = pp.deferpoolbuf[:0]
  5755  	pp.wbBuf.reset()
  5756  	if pp.mcache == nil {
  5757  		if id == 0 {
  5758  			if mcache0 == nil {
  5759  				throw("missing mcache?")
  5760  			}
  5761  			// Use the bootstrap mcache0. Only one P will get
  5762  			// mcache0: the one with ID 0.
  5763  			pp.mcache = mcache0
  5764  		} else {
  5765  			pp.mcache = allocmcache()
  5766  		}
  5767  	}
  5768  	if raceenabled && pp.raceprocctx == 0 {
  5769  		if id == 0 {
  5770  			pp.raceprocctx = raceprocctx0
  5771  			raceprocctx0 = 0 // bootstrap
  5772  		} else {
  5773  			pp.raceprocctx = raceproccreate()
  5774  		}
  5775  	}
  5776  	lockInit(&pp.timers.mu, lockRankTimers)
  5777  
  5778  	// This P may get timers when it starts running. Set the mask here
  5779  	// since the P may not go through pidleget (notably P 0 on startup).
  5780  	timerpMask.set(id)
  5781  	// Similarly, we may not go through pidleget before this P starts
  5782  	// running if it is P 0 on startup.
  5783  	idlepMask.clear(id)
  5784  }
  5785  
  5786  // destroy releases all of the resources associated with pp and
  5787  // transitions it to status _Pdead.
  5788  //
  5789  // sched.lock must be held and the world must be stopped.
  5790  func (pp *p) destroy() {
  5791  	assertLockHeld(&sched.lock)
  5792  	assertWorldStopped()
  5793  
  5794  	// Move all runnable goroutines to the global queue
  5795  	for pp.runqhead != pp.runqtail {
  5796  		// Pop from tail of local queue
  5797  		pp.runqtail--
  5798  		gp := pp.runq[pp.runqtail%uint32(len(pp.runq))].ptr()
  5799  		// Push onto head of global queue
  5800  		globrunqputhead(gp)
  5801  	}
  5802  	if pp.runnext != 0 {
  5803  		globrunqputhead(pp.runnext.ptr())
  5804  		pp.runnext = 0
  5805  	}
  5806  
  5807  	// Move all timers to the local P.
  5808  	getg().m.p.ptr().timers.take(&pp.timers)
  5809  
  5810  	// Flush p's write barrier buffer.
  5811  	if gcphase != _GCoff {
  5812  		wbBufFlush1(pp)
  5813  		pp.gcw.dispose()
  5814  	}
  5815  	clear(pp.sudogbuf[:])
  5816  	pp.sudogcache = pp.sudogbuf[:0]
  5817  	pp.pinnerCache = nil
  5818  	clear(pp.deferpoolbuf[:])
  5819  	pp.deferpool = pp.deferpoolbuf[:0]
  5820  	systemstack(func() {
  5821  		for i := 0; i < pp.mspancache.len; i++ {
  5822  			// Safe to call since the world is stopped.
  5823  			mheap_.spanalloc.free(unsafe.Pointer(pp.mspancache.buf[i]))
  5824  		}
  5825  		pp.mspancache.len = 0
  5826  		lock(&mheap_.lock)
  5827  		pp.pcache.flush(&mheap_.pages)
  5828  		unlock(&mheap_.lock)
  5829  	})
  5830  	freemcache(pp.mcache)
  5831  	pp.mcache = nil
  5832  	gfpurge(pp)
  5833  	if raceenabled {
  5834  		if pp.timers.raceCtx != 0 {
  5835  			// The race detector code uses a callback to fetch
  5836  			// the proc context, so arrange for that callback
  5837  			// to see the right thing.
  5838  			// This hack only works because we are the only
  5839  			// thread running.
  5840  			mp := getg().m
  5841  			phold := mp.p.ptr()
  5842  			mp.p.set(pp)
  5843  
  5844  			racectxend(pp.timers.raceCtx)
  5845  			pp.timers.raceCtx = 0
  5846  
  5847  			mp.p.set(phold)
  5848  		}
  5849  		raceprocdestroy(pp.raceprocctx)
  5850  		pp.raceprocctx = 0
  5851  	}
  5852  	pp.gcAssistTime = 0
  5853  	gcCleanups.queued += pp.cleanupsQueued
  5854  	pp.cleanupsQueued = 0
  5855  	pp.status = _Pdead
  5856  }
  5857  
  5858  // Change number of processors.
  5859  //
  5860  // sched.lock must be held, and the world must be stopped.
  5861  //
  5862  // gcworkbufs must not be being modified by either the GC or the write barrier
  5863  // code, so the GC must not be running if the number of Ps actually changes.
  5864  //
  5865  // Returns list of Ps with local work, they need to be scheduled by the caller.
  5866  func procresize(nprocs int32) *p {
  5867  	assertLockHeld(&sched.lock)
  5868  	assertWorldStopped()
  5869  
  5870  	old := gomaxprocs
  5871  	if old < 0 || nprocs <= 0 {
  5872  		throw("procresize: invalid arg")
  5873  	}
  5874  	trace := traceAcquire()
  5875  	if trace.ok() {
  5876  		trace.Gomaxprocs(nprocs)
  5877  		traceRelease(trace)
  5878  	}
  5879  
  5880  	// update statistics
  5881  	now := nanotime()
  5882  	if sched.procresizetime != 0 {
  5883  		sched.totaltime += int64(old) * (now - sched.procresizetime)
  5884  	}
  5885  	sched.procresizetime = now
  5886  
  5887  	maskWords := (nprocs + 31) / 32
  5888  
  5889  	// Grow allp if necessary.
  5890  	if nprocs > int32(len(allp)) {
  5891  		// Synchronize with retake, which could be running
  5892  		// concurrently since it doesn't run on a P.
  5893  		lock(&allpLock)
  5894  		if nprocs <= int32(cap(allp)) {
  5895  			allp = allp[:nprocs]
  5896  		} else {
  5897  			nallp := make([]*p, nprocs)
  5898  			// Copy everything up to allp's cap so we
  5899  			// never lose old allocated Ps.
  5900  			copy(nallp, allp[:cap(allp)])
  5901  			allp = nallp
  5902  		}
  5903  
  5904  		if maskWords <= int32(cap(idlepMask)) {
  5905  			idlepMask = idlepMask[:maskWords]
  5906  			timerpMask = timerpMask[:maskWords]
  5907  		} else {
  5908  			nidlepMask := make([]uint32, maskWords)
  5909  			// No need to copy beyond len, old Ps are irrelevant.
  5910  			copy(nidlepMask, idlepMask)
  5911  			idlepMask = nidlepMask
  5912  
  5913  			ntimerpMask := make([]uint32, maskWords)
  5914  			copy(ntimerpMask, timerpMask)
  5915  			timerpMask = ntimerpMask
  5916  		}
  5917  		unlock(&allpLock)
  5918  	}
  5919  
  5920  	// initialize new P's
  5921  	for i := old; i < nprocs; i++ {
  5922  		pp := allp[i]
  5923  		if pp == nil {
  5924  			pp = new(p)
  5925  		}
  5926  		pp.init(i)
  5927  		atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
  5928  	}
  5929  
  5930  	gp := getg()
  5931  	if gp.m.p != 0 && gp.m.p.ptr().id < nprocs {
  5932  		// continue to use the current P
  5933  		gp.m.p.ptr().status = _Prunning
  5934  		gp.m.p.ptr().mcache.prepareForSweep()
  5935  	} else {
  5936  		// release the current P and acquire allp[0].
  5937  		//
  5938  		// We must do this before destroying our current P
  5939  		// because p.destroy itself has write barriers, so we
  5940  		// need to do that from a valid P.
  5941  		if gp.m.p != 0 {
  5942  			trace := traceAcquire()
  5943  			if trace.ok() {
  5944  				// Pretend that we were descheduled
  5945  				// and then scheduled again to keep
  5946  				// the trace consistent.
  5947  				trace.GoSched()
  5948  				trace.ProcStop(gp.m.p.ptr())
  5949  				traceRelease(trace)
  5950  			}
  5951  			gp.m.p.ptr().m = 0
  5952  		}
  5953  		gp.m.p = 0
  5954  		pp := allp[0]
  5955  		pp.m = 0
  5956  		pp.status = _Pidle
  5957  		acquirep(pp)
  5958  		trace := traceAcquire()
  5959  		if trace.ok() {
  5960  			trace.GoStart()
  5961  			traceRelease(trace)
  5962  		}
  5963  	}
  5964  
  5965  	// g.m.p is now set, so we no longer need mcache0 for bootstrapping.
  5966  	mcache0 = nil
  5967  
  5968  	// release resources from unused P's
  5969  	for i := nprocs; i < old; i++ {
  5970  		pp := allp[i]
  5971  		pp.destroy()
  5972  		// can't free P itself because it can be referenced by an M in syscall
  5973  	}
  5974  
  5975  	// Trim allp.
  5976  	if int32(len(allp)) != nprocs {
  5977  		lock(&allpLock)
  5978  		allp = allp[:nprocs]
  5979  		idlepMask = idlepMask[:maskWords]
  5980  		timerpMask = timerpMask[:maskWords]
  5981  		unlock(&allpLock)
  5982  	}
  5983  
  5984  	var runnablePs *p
  5985  	for i := nprocs - 1; i >= 0; i-- {
  5986  		pp := allp[i]
  5987  		if gp.m.p.ptr() == pp {
  5988  			continue
  5989  		}
  5990  		pp.status = _Pidle
  5991  		if runqempty(pp) {
  5992  			pidleput(pp, now)
  5993  		} else {
  5994  			pp.m.set(mget())
  5995  			pp.link.set(runnablePs)
  5996  			runnablePs = pp
  5997  		}
  5998  	}
  5999  	stealOrder.reset(uint32(nprocs))
  6000  	var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
  6001  	atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
  6002  	if old != nprocs {
  6003  		// Notify the limiter that the amount of procs has changed.
  6004  		gcCPULimiter.resetCapacity(now, nprocs)
  6005  	}
  6006  	return runnablePs
  6007  }
  6008  
  6009  // Associate p and the current m.
  6010  //
  6011  // This function is allowed to have write barriers even if the caller
  6012  // isn't because it immediately acquires pp.
  6013  //
  6014  //go:yeswritebarrierrec
  6015  func acquirep(pp *p) {
  6016  	// Do the part that isn't allowed to have write barriers.
  6017  	wirep(pp)
  6018  
  6019  	// Have p; write barriers now allowed.
  6020  
  6021  	// Perform deferred mcache flush before this P can allocate
  6022  	// from a potentially stale mcache.
  6023  	pp.mcache.prepareForSweep()
  6024  
  6025  	trace := traceAcquire()
  6026  	if trace.ok() {
  6027  		trace.ProcStart()
  6028  		traceRelease(trace)
  6029  	}
  6030  }
  6031  
  6032  // wirep is the first step of acquirep, which actually associates the
  6033  // current M to pp. This is broken out so we can disallow write
  6034  // barriers for this part, since we don't yet have a P.
  6035  //
  6036  //go:nowritebarrierrec
  6037  //go:nosplit
  6038  func wirep(pp *p) {
  6039  	gp := getg()
  6040  
  6041  	if gp.m.p != 0 {
  6042  		// Call on the systemstack to avoid a nosplit overflow build failure
  6043  		// on some platforms when built with -N -l. See #64113.
  6044  		systemstack(func() {
  6045  			throw("wirep: already in go")
  6046  		})
  6047  	}
  6048  	if pp.m != 0 || pp.status != _Pidle {
  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  			id := int64(0)
  6053  			if pp.m != 0 {
  6054  				id = pp.m.ptr().id
  6055  			}
  6056  			print("wirep: p->m=", pp.m, "(", id, ") p->status=", pp.status, "\n")
  6057  			throw("wirep: invalid p state")
  6058  		})
  6059  	}
  6060  	gp.m.p.set(pp)
  6061  	pp.m.set(gp.m)
  6062  	pp.status = _Prunning
  6063  }
  6064  
  6065  // Disassociate p and the current m.
  6066  func releasep() *p {
  6067  	trace := traceAcquire()
  6068  	if trace.ok() {
  6069  		trace.ProcStop(getg().m.p.ptr())
  6070  		traceRelease(trace)
  6071  	}
  6072  	return releasepNoTrace()
  6073  }
  6074  
  6075  // Disassociate p and the current m without tracing an event.
  6076  func releasepNoTrace() *p {
  6077  	gp := getg()
  6078  
  6079  	if gp.m.p == 0 {
  6080  		throw("releasep: invalid arg")
  6081  	}
  6082  	pp := gp.m.p.ptr()
  6083  	if pp.m.ptr() != gp.m || pp.status != _Prunning {
  6084  		print("releasep: m=", gp.m, " m->p=", gp.m.p.ptr(), " p->m=", hex(pp.m), " p->status=", pp.status, "\n")
  6085  		throw("releasep: invalid p state")
  6086  	}
  6087  	gp.m.p = 0
  6088  	pp.m = 0
  6089  	pp.status = _Pidle
  6090  	return pp
  6091  }
  6092  
  6093  func incidlelocked(v int32) {
  6094  	lock(&sched.lock)
  6095  	sched.nmidlelocked += v
  6096  	if v > 0 {
  6097  		checkdead()
  6098  	}
  6099  	unlock(&sched.lock)
  6100  }
  6101  
  6102  // Check for deadlock situation.
  6103  // The check is based on number of running M's, if 0 -> deadlock.
  6104  // sched.lock must be held.
  6105  func checkdead() {
  6106  	assertLockHeld(&sched.lock)
  6107  
  6108  	// For -buildmode=c-shared or -buildmode=c-archive it's OK if
  6109  	// there are no running goroutines. The calling program is
  6110  	// assumed to be running.
  6111  	// One exception is Wasm, which is single-threaded. If we are
  6112  	// in Go and all goroutines are blocked, it deadlocks.
  6113  	if (islibrary || isarchive) && GOARCH != "wasm" {
  6114  		return
  6115  	}
  6116  
  6117  	// If we are dying because of a signal caught on an already idle thread,
  6118  	// freezetheworld will cause all running threads to block.
  6119  	// And runtime will essentially enter into deadlock state,
  6120  	// except that there is a thread that will call exit soon.
  6121  	if panicking.Load() > 0 {
  6122  		return
  6123  	}
  6124  
  6125  	// If we are not running under cgo, but we have an extra M then account
  6126  	// for it. (It is possible to have an extra M on Windows without cgo to
  6127  	// accommodate callbacks created by syscall.NewCallback. See issue #6751
  6128  	// for details.)
  6129  	var run0 int32
  6130  	if !iscgo && cgoHasExtraM && extraMLength.Load() > 0 {
  6131  		run0 = 1
  6132  	}
  6133  
  6134  	run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
  6135  	if run > run0 {
  6136  		return
  6137  	}
  6138  	if run < 0 {
  6139  		print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
  6140  		unlock(&sched.lock)
  6141  		throw("checkdead: inconsistent counts")
  6142  	}
  6143  
  6144  	grunning := 0
  6145  	forEachG(func(gp *g) {
  6146  		if isSystemGoroutine(gp, false) {
  6147  			return
  6148  		}
  6149  		s := readgstatus(gp)
  6150  		switch s &^ _Gscan {
  6151  		case _Gwaiting,
  6152  			_Gpreempted:
  6153  			grunning++
  6154  		case _Grunnable,
  6155  			_Grunning,
  6156  			_Gsyscall:
  6157  			print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
  6158  			unlock(&sched.lock)
  6159  			throw("checkdead: runnable g")
  6160  		}
  6161  	})
  6162  	if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
  6163  		unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  6164  		fatal("no goroutines (main called runtime.Goexit) - deadlock!")
  6165  	}
  6166  
  6167  	// Maybe jump time forward for playground.
  6168  	if faketime != 0 {
  6169  		if when := timeSleepUntil(); when < maxWhen {
  6170  			faketime = when
  6171  
  6172  			// Start an M to steal the timer.
  6173  			pp, _ := pidleget(faketime)
  6174  			if pp == nil {
  6175  				// There should always be a free P since
  6176  				// nothing is running.
  6177  				unlock(&sched.lock)
  6178  				throw("checkdead: no p for timer")
  6179  			}
  6180  			mp := mget()
  6181  			if mp == nil {
  6182  				// There should always be a free M since
  6183  				// nothing is running.
  6184  				unlock(&sched.lock)
  6185  				throw("checkdead: no m for timer")
  6186  			}
  6187  			// M must be spinning to steal. We set this to be
  6188  			// explicit, but since this is the only M it would
  6189  			// become spinning on its own anyways.
  6190  			sched.nmspinning.Add(1)
  6191  			mp.spinning = true
  6192  			mp.nextp.set(pp)
  6193  			notewakeup(&mp.park)
  6194  			return
  6195  		}
  6196  	}
  6197  
  6198  	// There are no goroutines running, so we can look at the P's.
  6199  	for _, pp := range allp {
  6200  		if len(pp.timers.heap) > 0 {
  6201  			return
  6202  		}
  6203  	}
  6204  
  6205  	unlock(&sched.lock) // unlock so that GODEBUG=scheddetail=1 doesn't hang
  6206  	fatal("all goroutines are asleep - deadlock!")
  6207  }
  6208  
  6209  // forcegcperiod is the maximum time in nanoseconds between garbage
  6210  // collections. If we go this long without a garbage collection, one
  6211  // is forced to run.
  6212  //
  6213  // This is a variable for testing purposes. It normally doesn't change.
  6214  var forcegcperiod int64 = 2 * 60 * 1e9
  6215  
  6216  // needSysmonWorkaround is true if the workaround for
  6217  // golang.org/issue/42515 is needed on NetBSD.
  6218  var needSysmonWorkaround bool = false
  6219  
  6220  // haveSysmon indicates whether there is sysmon thread support.
  6221  //
  6222  // No threads on wasm yet, so no sysmon.
  6223  const haveSysmon = GOARCH != "wasm"
  6224  
  6225  // Always runs without a P, so write barriers are not allowed.
  6226  //
  6227  //go:nowritebarrierrec
  6228  func sysmon() {
  6229  	lock(&sched.lock)
  6230  	sched.nmsys++
  6231  	checkdead()
  6232  	unlock(&sched.lock)
  6233  
  6234  	lastgomaxprocs := int64(0)
  6235  	lasttrace := int64(0)
  6236  	idle := 0 // how many cycles in succession we had not wokeup somebody
  6237  	delay := uint32(0)
  6238  
  6239  	for {
  6240  		if idle == 0 { // start with 20us sleep...
  6241  			delay = 20
  6242  		} else if idle > 50 { // start doubling the sleep after 1ms...
  6243  			delay *= 2
  6244  		}
  6245  		if delay > 10*1000 { // up to 10ms
  6246  			delay = 10 * 1000
  6247  		}
  6248  		usleep(delay)
  6249  
  6250  		// sysmon should not enter deep sleep if schedtrace is enabled so that
  6251  		// it can print that information at the right time.
  6252  		//
  6253  		// It should also not enter deep sleep if there are any active P's so
  6254  		// that it can retake P's from syscalls, preempt long running G's, and
  6255  		// poll the network if all P's are busy for long stretches.
  6256  		//
  6257  		// It should wakeup from deep sleep if any P's become active either due
  6258  		// to exiting a syscall or waking up due to a timer expiring so that it
  6259  		// can resume performing those duties. If it wakes from a syscall it
  6260  		// resets idle and delay as a bet that since it had retaken a P from a
  6261  		// syscall before, it may need to do it again shortly after the
  6262  		// application starts work again. It does not reset idle when waking
  6263  		// from a timer to avoid adding system load to applications that spend
  6264  		// most of their time sleeping.
  6265  		now := nanotime()
  6266  		if debug.schedtrace <= 0 && (sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs) {
  6267  			lock(&sched.lock)
  6268  			if sched.gcwaiting.Load() || sched.npidle.Load() == gomaxprocs {
  6269  				syscallWake := false
  6270  				next := timeSleepUntil()
  6271  				if next > now {
  6272  					sched.sysmonwait.Store(true)
  6273  					unlock(&sched.lock)
  6274  					// Make wake-up period small enough
  6275  					// for the sampling to be correct.
  6276  					sleep := forcegcperiod / 2
  6277  					if next-now < sleep {
  6278  						sleep = next - now
  6279  					}
  6280  					shouldRelax := sleep >= osRelaxMinNS
  6281  					if shouldRelax {
  6282  						osRelax(true)
  6283  					}
  6284  					syscallWake = notetsleep(&sched.sysmonnote, sleep)
  6285  					if shouldRelax {
  6286  						osRelax(false)
  6287  					}
  6288  					lock(&sched.lock)
  6289  					sched.sysmonwait.Store(false)
  6290  					noteclear(&sched.sysmonnote)
  6291  				}
  6292  				if syscallWake {
  6293  					idle = 0
  6294  					delay = 20
  6295  				}
  6296  			}
  6297  			unlock(&sched.lock)
  6298  		}
  6299  
  6300  		lock(&sched.sysmonlock)
  6301  		// Update now in case we blocked on sysmonnote or spent a long time
  6302  		// blocked on schedlock or sysmonlock above.
  6303  		now = nanotime()
  6304  
  6305  		// trigger libc interceptors if needed
  6306  		if *cgo_yield != nil {
  6307  			asmcgocall(*cgo_yield, nil)
  6308  		}
  6309  		// poll network if not polled for more than 10ms
  6310  		lastpoll := sched.lastpoll.Load()
  6311  		if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
  6312  			sched.lastpoll.CompareAndSwap(lastpoll, now)
  6313  			list, delta := netpoll(0) // non-blocking - returns list of goroutines
  6314  			if !list.empty() {
  6315  				// Need to decrement number of idle locked M's
  6316  				// (pretending that one more is running) before injectglist.
  6317  				// Otherwise it can lead to the following situation:
  6318  				// injectglist grabs all P's but before it starts M's to run the P's,
  6319  				// another M returns from syscall, finishes running its G,
  6320  				// observes that there is no work to do and no other running M's
  6321  				// and reports deadlock.
  6322  				incidlelocked(-1)
  6323  				injectglist(&list)
  6324  				incidlelocked(1)
  6325  				netpollAdjustWaiters(delta)
  6326  			}
  6327  		}
  6328  		if GOOS == "netbsd" && needSysmonWorkaround {
  6329  			// netpoll is responsible for waiting for timer
  6330  			// expiration, so we typically don't have to worry
  6331  			// about starting an M to service timers. (Note that
  6332  			// sleep for timeSleepUntil above simply ensures sysmon
  6333  			// starts running again when that timer expiration may
  6334  			// cause Go code to run again).
  6335  			//
  6336  			// However, netbsd has a kernel bug that sometimes
  6337  			// misses netpollBreak wake-ups, which can lead to
  6338  			// unbounded delays servicing timers. If we detect this
  6339  			// overrun, then startm to get something to handle the
  6340  			// timer.
  6341  			//
  6342  			// See issue 42515 and
  6343  			// https://gnats.netbsd.org/cgi-bin/query-pr-single.pl?number=50094.
  6344  			if next := timeSleepUntil(); next < now {
  6345  				startm(nil, false, false)
  6346  			}
  6347  		}
  6348  		// Check if we need to update GOMAXPROCS at most once per second.
  6349  		if debug.updatemaxprocs != 0 && lastgomaxprocs+1e9 <= now {
  6350  			sysmonUpdateGOMAXPROCS()
  6351  			lastgomaxprocs = now
  6352  		}
  6353  		if scavenger.sysmonWake.Load() != 0 {
  6354  			// Kick the scavenger awake if someone requested it.
  6355  			scavenger.wake()
  6356  		}
  6357  		// retake P's blocked in syscalls
  6358  		// and preempt long running G's
  6359  		if retake(now) != 0 {
  6360  			idle = 0
  6361  		} else {
  6362  			idle++
  6363  		}
  6364  		// check if we need to force a GC
  6365  		if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && forcegc.idle.Load() {
  6366  			lock(&forcegc.lock)
  6367  			forcegc.idle.Store(false)
  6368  			var list gList
  6369  			list.push(forcegc.g)
  6370  			injectglist(&list)
  6371  			unlock(&forcegc.lock)
  6372  		}
  6373  		if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
  6374  			lasttrace = now
  6375  			schedtrace(debug.scheddetail > 0)
  6376  		}
  6377  		unlock(&sched.sysmonlock)
  6378  	}
  6379  }
  6380  
  6381  type sysmontick struct {
  6382  	schedtick   uint32
  6383  	syscalltick uint32
  6384  	schedwhen   int64
  6385  	syscallwhen int64
  6386  }
  6387  
  6388  // forcePreemptNS is the time slice given to a G before it is
  6389  // preempted.
  6390  const forcePreemptNS = 10 * 1000 * 1000 // 10ms
  6391  
  6392  func retake(now int64) uint32 {
  6393  	n := 0
  6394  	// Prevent allp slice changes. This lock will be completely
  6395  	// uncontended unless we're already stopping the world.
  6396  	lock(&allpLock)
  6397  	// We can't use a range loop over allp because we may
  6398  	// temporarily drop the allpLock. Hence, we need to re-fetch
  6399  	// allp each time around the loop.
  6400  	for i := 0; i < len(allp); i++ {
  6401  		pp := allp[i]
  6402  		if pp == nil {
  6403  			// This can happen if procresize has grown
  6404  			// allp but not yet created new Ps.
  6405  			continue
  6406  		}
  6407  		pd := &pp.sysmontick
  6408  		s := pp.status
  6409  		sysretake := false
  6410  		if s == _Prunning || s == _Psyscall {
  6411  			// Preempt G if it's running on the same schedtick for
  6412  			// too long. This could be from a single long-running
  6413  			// goroutine or a sequence of goroutines run via
  6414  			// runnext, which share a single schedtick time slice.
  6415  			t := int64(pp.schedtick)
  6416  			if int64(pd.schedtick) != t {
  6417  				pd.schedtick = uint32(t)
  6418  				pd.schedwhen = now
  6419  			} else if pd.schedwhen+forcePreemptNS <= now {
  6420  				preemptone(pp)
  6421  				// In case of syscall, preemptone() doesn't
  6422  				// work, because there is no M wired to P.
  6423  				sysretake = true
  6424  			}
  6425  		}
  6426  		if s == _Psyscall {
  6427  			// Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
  6428  			t := int64(pp.syscalltick)
  6429  			if !sysretake && int64(pd.syscalltick) != t {
  6430  				pd.syscalltick = uint32(t)
  6431  				pd.syscallwhen = now
  6432  				continue
  6433  			}
  6434  			// On the one hand we don't want to retake Ps if there is no other work to do,
  6435  			// but on the other hand we want to retake them eventually
  6436  			// because they can prevent the sysmon thread from deep sleep.
  6437  			if runqempty(pp) && sched.nmspinning.Load()+sched.npidle.Load() > 0 && pd.syscallwhen+10*1000*1000 > now {
  6438  				continue
  6439  			}
  6440  			// Drop allpLock so we can take sched.lock.
  6441  			unlock(&allpLock)
  6442  			// Need to decrement number of idle locked M's
  6443  			// (pretending that one more is running) before the CAS.
  6444  			// Otherwise the M from which we retake can exit the syscall,
  6445  			// increment nmidle and report deadlock.
  6446  			incidlelocked(-1)
  6447  			trace := traceAcquire()
  6448  			if atomic.Cas(&pp.status, s, _Pidle) {
  6449  				if trace.ok() {
  6450  					trace.ProcSteal(pp, false)
  6451  					traceRelease(trace)
  6452  				}
  6453  				n++
  6454  				pp.syscalltick++
  6455  				handoffp(pp)
  6456  			} else if trace.ok() {
  6457  				traceRelease(trace)
  6458  			}
  6459  			incidlelocked(1)
  6460  			lock(&allpLock)
  6461  		}
  6462  	}
  6463  	unlock(&allpLock)
  6464  	return uint32(n)
  6465  }
  6466  
  6467  // Tell all goroutines that they have been preempted and they should stop.
  6468  // This function is purely best-effort. It can fail to inform a goroutine if a
  6469  // processor just started running it.
  6470  // No locks need to be held.
  6471  // Returns true if preemption request was issued to at least one goroutine.
  6472  func preemptall() bool {
  6473  	res := false
  6474  	for _, pp := range allp {
  6475  		if pp.status != _Prunning {
  6476  			continue
  6477  		}
  6478  		if preemptone(pp) {
  6479  			res = true
  6480  		}
  6481  	}
  6482  	return res
  6483  }
  6484  
  6485  // Tell the goroutine running on processor P to stop.
  6486  // This function is purely best-effort. It can incorrectly fail to inform the
  6487  // goroutine. It can inform the wrong goroutine. Even if it informs the
  6488  // correct goroutine, that goroutine might ignore the request if it is
  6489  // simultaneously executing newstack.
  6490  // No lock needs to be held.
  6491  // Returns true if preemption request was issued.
  6492  // The actual preemption will happen at some point in the future
  6493  // and will be indicated by the gp->status no longer being
  6494  // Grunning
  6495  func preemptone(pp *p) bool {
  6496  	mp := pp.m.ptr()
  6497  	if mp == nil || mp == getg().m {
  6498  		return false
  6499  	}
  6500  	gp := mp.curg
  6501  	if gp == nil || gp == mp.g0 {
  6502  		return false
  6503  	}
  6504  
  6505  	gp.preempt = true
  6506  
  6507  	// Every call in a goroutine checks for stack overflow by
  6508  	// comparing the current stack pointer to gp->stackguard0.
  6509  	// Setting gp->stackguard0 to StackPreempt folds
  6510  	// preemption into the normal stack overflow check.
  6511  	gp.stackguard0 = stackPreempt
  6512  
  6513  	// Request an async preemption of this P.
  6514  	if preemptMSupported && debug.asyncpreemptoff == 0 {
  6515  		pp.preempt = true
  6516  		preemptM(mp)
  6517  	}
  6518  
  6519  	return true
  6520  }
  6521  
  6522  var starttime int64
  6523  
  6524  func schedtrace(detailed bool) {
  6525  	now := nanotime()
  6526  	if starttime == 0 {
  6527  		starttime = now
  6528  	}
  6529  
  6530  	lock(&sched.lock)
  6531  	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)
  6532  	if detailed {
  6533  		print(" gcwaiting=", sched.gcwaiting.Load(), " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait.Load(), "\n")
  6534  	}
  6535  	// We must be careful while reading data from P's, M's and G's.
  6536  	// Even if we hold schedlock, most data can be changed concurrently.
  6537  	// E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
  6538  	for i, pp := range allp {
  6539  		h := atomic.Load(&pp.runqhead)
  6540  		t := atomic.Load(&pp.runqtail)
  6541  		if detailed {
  6542  			print("  P", i, ": status=", pp.status, " schedtick=", pp.schedtick, " syscalltick=", pp.syscalltick, " m=")
  6543  			mp := pp.m.ptr()
  6544  			if mp != nil {
  6545  				print(mp.id)
  6546  			} else {
  6547  				print("nil")
  6548  			}
  6549  			print(" runqsize=", t-h, " gfreecnt=", pp.gFree.size, " timerslen=", len(pp.timers.heap), "\n")
  6550  		} else {
  6551  			// In non-detailed mode format lengths of per-P run queues as:
  6552  			// [ len1 len2 len3 len4 ]
  6553  			print(" ")
  6554  			if i == 0 {
  6555  				print("[ ")
  6556  			}
  6557  			print(t - h)
  6558  			if i == len(allp)-1 {
  6559  				print(" ]")
  6560  			}
  6561  		}
  6562  	}
  6563  
  6564  	if !detailed {
  6565  		// Format per-P schedticks as: schedticks=[ tick1 tick2 tick3 tick4 ].
  6566  		print(" schedticks=[ ")
  6567  		for _, pp := range allp {
  6568  			print(pp.schedtick)
  6569  			print(" ")
  6570  		}
  6571  		print("]\n")
  6572  	}
  6573  
  6574  	if !detailed {
  6575  		unlock(&sched.lock)
  6576  		return
  6577  	}
  6578  
  6579  	for mp := allm; mp != nil; mp = mp.alllink {
  6580  		pp := mp.p.ptr()
  6581  		print("  M", mp.id, ": p=")
  6582  		if pp != nil {
  6583  			print(pp.id)
  6584  		} else {
  6585  			print("nil")
  6586  		}
  6587  		print(" curg=")
  6588  		if mp.curg != nil {
  6589  			print(mp.curg.goid)
  6590  		} else {
  6591  			print("nil")
  6592  		}
  6593  		print(" mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, " locks=", mp.locks, " dying=", mp.dying, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=")
  6594  		if lockedg := mp.lockedg.ptr(); lockedg != nil {
  6595  			print(lockedg.goid)
  6596  		} else {
  6597  			print("nil")
  6598  		}
  6599  		print("\n")
  6600  	}
  6601  
  6602  	forEachG(func(gp *g) {
  6603  		print("  G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason.String(), ") m=")
  6604  		if gp.m != nil {
  6605  			print(gp.m.id)
  6606  		} else {
  6607  			print("nil")
  6608  		}
  6609  		print(" lockedm=")
  6610  		if lockedm := gp.lockedm.ptr(); lockedm != nil {
  6611  			print(lockedm.id)
  6612  		} else {
  6613  			print("nil")
  6614  		}
  6615  		print("\n")
  6616  	})
  6617  	unlock(&sched.lock)
  6618  }
  6619  
  6620  type updateMaxProcsGState struct {
  6621  	lock mutex
  6622  	g    *g
  6623  	idle atomic.Bool
  6624  
  6625  	// Readable when idle == false, writable when idle == true.
  6626  	procs int32 // new GOMAXPROCS value
  6627  }
  6628  
  6629  var (
  6630  	// GOMAXPROCS update godebug metric. Incremented if automatic
  6631  	// GOMAXPROCS updates actually change the value of GOMAXPROCS.
  6632  	updatemaxprocs = &godebugInc{name: "updatemaxprocs"}
  6633  
  6634  	// Synchronization and state between updateMaxProcsGoroutine and
  6635  	// sysmon.
  6636  	updateMaxProcsG updateMaxProcsGState
  6637  
  6638  	// Synchronization between GOMAXPROCS and sysmon.
  6639  	//
  6640  	// Setting GOMAXPROCS via a call to GOMAXPROCS disables automatic
  6641  	// GOMAXPROCS updates.
  6642  	//
  6643  	// We want to make two guarantees to callers of GOMAXPROCS. After
  6644  	// GOMAXPROCS returns:
  6645  	//
  6646  	// 1. The runtime will not make any automatic changes to GOMAXPROCS.
  6647  	//
  6648  	// 2. The runtime will not perform any of the system calls used to
  6649  	//    determine the appropriate value of GOMAXPROCS (i.e., it won't
  6650  	//    call defaultGOMAXPROCS).
  6651  	//
  6652  	// (1) is the baseline guarantee that everyone needs. The GOMAXPROCS
  6653  	// API isn't useful to anyone if automatic updates may occur after it
  6654  	// returns. This is easily achieved by double-checking the state under
  6655  	// STW before committing an automatic GOMAXPROCS update.
  6656  	//
  6657  	// (2) doesn't matter to most users, as it is isn't observable as long
  6658  	// as (1) holds. However, it can be important to users sandboxing Go.
  6659  	// They want disable these system calls and need some way to know when
  6660  	// they are guaranteed the calls will stop.
  6661  	//
  6662  	// This would be simple to achieve if we simply called
  6663  	// defaultGOMAXPROCS under STW in updateMaxProcsGoroutine below.
  6664  	// However, we would like to avoid scheduling this goroutine every
  6665  	// second when it will almost never do anything. Instead, sysmon calls
  6666  	// defaultGOMAXPROCS to decide whether to schedule
  6667  	// updateMaxProcsGoroutine. Thus we need to synchronize between sysmon
  6668  	// and GOMAXPROCS calls.
  6669  	//
  6670  	// GOMAXPROCS can't hold a runtime mutex across STW. It could hold a
  6671  	// semaphore, but sysmon cannot take semaphores. Instead, we have a
  6672  	// more complex scheme:
  6673  	//
  6674  	// * sysmon holds computeMaxProcsLock while calling defaultGOMAXPROCS.
  6675  	// * sysmon skips the current update if sched.customGOMAXPROCS is
  6676  	//   set.
  6677  	// * GOMAXPROCS sets sched.customGOMAXPROCS once it is committed to
  6678  	//   changing GOMAXPROCS.
  6679  	// * GOMAXPROCS takes computeMaxProcsLock to wait for outstanding
  6680  	//   defaultGOMAXPROCS calls to complete.
  6681  	//
  6682  	// N.B. computeMaxProcsLock could simply be sched.lock, but we want to
  6683  	// avoid holding that lock during the potentially slow
  6684  	// defaultGOMAXPROCS.
  6685  	computeMaxProcsLock mutex
  6686  )
  6687  
  6688  // Start GOMAXPROCS update helper goroutine.
  6689  //
  6690  // This is based on forcegchelper.
  6691  func defaultGOMAXPROCSUpdateEnable() {
  6692  	if debug.updatemaxprocs == 0 {
  6693  		// Unconditionally increment the metric when updates are disabled.
  6694  		//
  6695  		// It would be more descriptive if we did a dry run of the
  6696  		// complete update, determining the appropriate value of
  6697  		// GOMAXPROCS and the bailing out and just incrementing the
  6698  		// metric if a change would occur.
  6699  		//
  6700  		// Not only is that a lot of ongoing work for a disabled
  6701  		// feature, but some users need to be able to completely
  6702  		// disable the update system calls (such as sandboxes).
  6703  		// Currently, updatemaxprocs=0 serves that purpose.
  6704  		updatemaxprocs.IncNonDefault()
  6705  		return
  6706  	}
  6707  
  6708  	go updateMaxProcsGoroutine()
  6709  }
  6710  
  6711  func updateMaxProcsGoroutine() {
  6712  	updateMaxProcsG.g = getg()
  6713  	lockInit(&updateMaxProcsG.lock, lockRankUpdateMaxProcsG)
  6714  	for {
  6715  		lock(&updateMaxProcsG.lock)
  6716  		if updateMaxProcsG.idle.Load() {
  6717  			throw("updateMaxProcsGoroutine: phase error")
  6718  		}
  6719  		updateMaxProcsG.idle.Store(true)
  6720  		goparkunlock(&updateMaxProcsG.lock, waitReasonUpdateGOMAXPROCSIdle, traceBlockSystemGoroutine, 1)
  6721  		// This goroutine is explicitly resumed by sysmon.
  6722  
  6723  		stw := stopTheWorldGC(stwGOMAXPROCS)
  6724  
  6725  		// Still OK to update?
  6726  		lock(&sched.lock)
  6727  		custom := sched.customGOMAXPROCS
  6728  		unlock(&sched.lock)
  6729  		if custom {
  6730  			startTheWorldGC(stw)
  6731  			return
  6732  		}
  6733  
  6734  		// newprocs will be processed by startTheWorld
  6735  		//
  6736  		// TODO(prattmic): this could use a nicer API. Perhaps add it to the
  6737  		// stw parameter?
  6738  		newprocs = updateMaxProcsG.procs
  6739  		lock(&sched.lock)
  6740  		sched.customGOMAXPROCS = false
  6741  		unlock(&sched.lock)
  6742  
  6743  		startTheWorldGC(stw)
  6744  	}
  6745  }
  6746  
  6747  func sysmonUpdateGOMAXPROCS() {
  6748  	// Synchronize with GOMAXPROCS. See comment on computeMaxProcsLock.
  6749  	lock(&computeMaxProcsLock)
  6750  
  6751  	// No update if GOMAXPROCS was set manually.
  6752  	lock(&sched.lock)
  6753  	custom := sched.customGOMAXPROCS
  6754  	curr := gomaxprocs
  6755  	unlock(&sched.lock)
  6756  	if custom {
  6757  		unlock(&computeMaxProcsLock)
  6758  		return
  6759  	}
  6760  
  6761  	// Don't hold sched.lock while we read the filesystem.
  6762  	procs := defaultGOMAXPROCS(0)
  6763  	unlock(&computeMaxProcsLock)
  6764  	if procs == curr {
  6765  		// Nothing to do.
  6766  		return
  6767  	}
  6768  
  6769  	// Sysmon can't directly stop the world. Run the helper to do so on our
  6770  	// behalf. If updateGOMAXPROCS.idle is false, then a previous update is
  6771  	// still pending.
  6772  	if updateMaxProcsG.idle.Load() {
  6773  		lock(&updateMaxProcsG.lock)
  6774  		updateMaxProcsG.procs = procs
  6775  		updateMaxProcsG.idle.Store(false)
  6776  		var list gList
  6777  		list.push(updateMaxProcsG.g)
  6778  		injectglist(&list)
  6779  		unlock(&updateMaxProcsG.lock)
  6780  	}
  6781  }
  6782  
  6783  // schedEnableUser enables or disables the scheduling of user
  6784  // goroutines.
  6785  //
  6786  // This does not stop already running user goroutines, so the caller
  6787  // should first stop the world when disabling user goroutines.
  6788  func schedEnableUser(enable bool) {
  6789  	lock(&sched.lock)
  6790  	if sched.disable.user == !enable {
  6791  		unlock(&sched.lock)
  6792  		return
  6793  	}
  6794  	sched.disable.user = !enable
  6795  	if enable {
  6796  		n := sched.disable.runnable.size
  6797  		globrunqputbatch(&sched.disable.runnable)
  6798  		unlock(&sched.lock)
  6799  		for ; n != 0 && sched.npidle.Load() != 0; n-- {
  6800  			startm(nil, false, false)
  6801  		}
  6802  	} else {
  6803  		unlock(&sched.lock)
  6804  	}
  6805  }
  6806  
  6807  // schedEnabled reports whether gp should be scheduled. It returns
  6808  // false is scheduling of gp is disabled.
  6809  //
  6810  // sched.lock must be held.
  6811  func schedEnabled(gp *g) bool {
  6812  	assertLockHeld(&sched.lock)
  6813  
  6814  	if sched.disable.user {
  6815  		return isSystemGoroutine(gp, true)
  6816  	}
  6817  	return true
  6818  }
  6819  
  6820  // Put mp on midle list.
  6821  // sched.lock must be held.
  6822  // May run during STW, so write barriers are not allowed.
  6823  //
  6824  //go:nowritebarrierrec
  6825  func mput(mp *m) {
  6826  	assertLockHeld(&sched.lock)
  6827  
  6828  	mp.schedlink = sched.midle
  6829  	sched.midle.set(mp)
  6830  	sched.nmidle++
  6831  	checkdead()
  6832  }
  6833  
  6834  // Try to get an m from midle list.
  6835  // sched.lock must be held.
  6836  // May run during STW, so write barriers are not allowed.
  6837  //
  6838  //go:nowritebarrierrec
  6839  func mget() *m {
  6840  	assertLockHeld(&sched.lock)
  6841  
  6842  	mp := sched.midle.ptr()
  6843  	if mp != nil {
  6844  		sched.midle = mp.schedlink
  6845  		sched.nmidle--
  6846  	}
  6847  	return mp
  6848  }
  6849  
  6850  // Put gp on the global runnable queue.
  6851  // sched.lock must be held.
  6852  // May run during STW, so write barriers are not allowed.
  6853  //
  6854  //go:nowritebarrierrec
  6855  func globrunqput(gp *g) {
  6856  	assertLockHeld(&sched.lock)
  6857  
  6858  	sched.runq.pushBack(gp)
  6859  }
  6860  
  6861  // Put gp at the head of the global runnable queue.
  6862  // sched.lock must be held.
  6863  // May run during STW, so write barriers are not allowed.
  6864  //
  6865  //go:nowritebarrierrec
  6866  func globrunqputhead(gp *g) {
  6867  	assertLockHeld(&sched.lock)
  6868  
  6869  	sched.runq.push(gp)
  6870  }
  6871  
  6872  // Put a batch of runnable goroutines on the global runnable queue.
  6873  // This clears *batch.
  6874  // sched.lock must be held.
  6875  // May run during STW, so write barriers are not allowed.
  6876  //
  6877  //go:nowritebarrierrec
  6878  func globrunqputbatch(batch *gQueue) {
  6879  	assertLockHeld(&sched.lock)
  6880  
  6881  	sched.runq.pushBackAll(*batch)
  6882  	*batch = gQueue{}
  6883  }
  6884  
  6885  // Try get a single G from the global runnable queue.
  6886  // sched.lock must be held.
  6887  func globrunqget() *g {
  6888  	assertLockHeld(&sched.lock)
  6889  
  6890  	if sched.runq.size == 0 {
  6891  		return nil
  6892  	}
  6893  
  6894  	return sched.runq.pop()
  6895  }
  6896  
  6897  // Try get a batch of G's from the global runnable queue.
  6898  // sched.lock must be held.
  6899  func globrunqgetbatch(n int32) (gp *g, q gQueue) {
  6900  	assertLockHeld(&sched.lock)
  6901  
  6902  	if sched.runq.size == 0 {
  6903  		return
  6904  	}
  6905  
  6906  	n = min(n, sched.runq.size, sched.runq.size/gomaxprocs+1)
  6907  
  6908  	gp = sched.runq.pop()
  6909  	n--
  6910  
  6911  	for ; n > 0; n-- {
  6912  		gp1 := sched.runq.pop()
  6913  		q.pushBack(gp1)
  6914  	}
  6915  	return
  6916  }
  6917  
  6918  // pMask is an atomic bitstring with one bit per P.
  6919  type pMask []uint32
  6920  
  6921  // read returns true if P id's bit is set.
  6922  func (p pMask) read(id uint32) bool {
  6923  	word := id / 32
  6924  	mask := uint32(1) << (id % 32)
  6925  	return (atomic.Load(&p[word]) & mask) != 0
  6926  }
  6927  
  6928  // set sets P id's bit.
  6929  func (p pMask) set(id int32) {
  6930  	word := id / 32
  6931  	mask := uint32(1) << (id % 32)
  6932  	atomic.Or(&p[word], mask)
  6933  }
  6934  
  6935  // clear clears P id's bit.
  6936  func (p pMask) clear(id int32) {
  6937  	word := id / 32
  6938  	mask := uint32(1) << (id % 32)
  6939  	atomic.And(&p[word], ^mask)
  6940  }
  6941  
  6942  // pidleput puts p on the _Pidle list. now must be a relatively recent call
  6943  // to nanotime or zero. Returns now or the current time if now was zero.
  6944  //
  6945  // This releases ownership of p. Once sched.lock is released it is no longer
  6946  // safe to use p.
  6947  //
  6948  // sched.lock must be held.
  6949  //
  6950  // May run during STW, so write barriers are not allowed.
  6951  //
  6952  //go:nowritebarrierrec
  6953  func pidleput(pp *p, now int64) int64 {
  6954  	assertLockHeld(&sched.lock)
  6955  
  6956  	if !runqempty(pp) {
  6957  		throw("pidleput: P has non-empty run queue")
  6958  	}
  6959  	if now == 0 {
  6960  		now = nanotime()
  6961  	}
  6962  	if pp.timers.len.Load() == 0 {
  6963  		timerpMask.clear(pp.id)
  6964  	}
  6965  	idlepMask.set(pp.id)
  6966  	pp.link = sched.pidle
  6967  	sched.pidle.set(pp)
  6968  	sched.npidle.Add(1)
  6969  	if !pp.limiterEvent.start(limiterEventIdle, now) {
  6970  		throw("must be able to track idle limiter event")
  6971  	}
  6972  	return now
  6973  }
  6974  
  6975  // pidleget tries to get a p from the _Pidle list, acquiring ownership.
  6976  //
  6977  // sched.lock must be held.
  6978  //
  6979  // May run during STW, so write barriers are not allowed.
  6980  //
  6981  //go:nowritebarrierrec
  6982  func pidleget(now int64) (*p, int64) {
  6983  	assertLockHeld(&sched.lock)
  6984  
  6985  	pp := sched.pidle.ptr()
  6986  	if pp != nil {
  6987  		// Timer may get added at any time now.
  6988  		if now == 0 {
  6989  			now = nanotime()
  6990  		}
  6991  		timerpMask.set(pp.id)
  6992  		idlepMask.clear(pp.id)
  6993  		sched.pidle = pp.link
  6994  		sched.npidle.Add(-1)
  6995  		pp.limiterEvent.stop(limiterEventIdle, now)
  6996  	}
  6997  	return pp, now
  6998  }
  6999  
  7000  // pidlegetSpinning tries to get a p from the _Pidle list, acquiring ownership.
  7001  // This is called by spinning Ms (or callers than need a spinning M) that have
  7002  // found work. If no P is available, this must synchronized with non-spinning
  7003  // Ms that may be preparing to drop their P without discovering this work.
  7004  //
  7005  // sched.lock must be held.
  7006  //
  7007  // May run during STW, so write barriers are not allowed.
  7008  //
  7009  //go:nowritebarrierrec
  7010  func pidlegetSpinning(now int64) (*p, int64) {
  7011  	assertLockHeld(&sched.lock)
  7012  
  7013  	pp, now := pidleget(now)
  7014  	if pp == nil {
  7015  		// See "Delicate dance" comment in findrunnable. We found work
  7016  		// that we cannot take, we must synchronize with non-spinning
  7017  		// Ms that may be preparing to drop their P.
  7018  		sched.needspinning.Store(1)
  7019  		return nil, now
  7020  	}
  7021  
  7022  	return pp, now
  7023  }
  7024  
  7025  // runqempty reports whether pp has no Gs on its local run queue.
  7026  // It never returns true spuriously.
  7027  func runqempty(pp *p) bool {
  7028  	// Defend against a race where 1) pp has G1 in runqnext but runqhead == runqtail,
  7029  	// 2) runqput on pp kicks G1 to the runq, 3) runqget on pp empties runqnext.
  7030  	// Simply observing that runqhead == runqtail and then observing that runqnext == nil
  7031  	// does not mean the queue is empty.
  7032  	for {
  7033  		head := atomic.Load(&pp.runqhead)
  7034  		tail := atomic.Load(&pp.runqtail)
  7035  		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&pp.runnext)))
  7036  		if tail == atomic.Load(&pp.runqtail) {
  7037  			return head == tail && runnext == 0
  7038  		}
  7039  	}
  7040  }
  7041  
  7042  // To shake out latent assumptions about scheduling order,
  7043  // we introduce some randomness into scheduling decisions
  7044  // when running with the race detector.
  7045  // The need for this was made obvious by changing the
  7046  // (deterministic) scheduling order in Go 1.5 and breaking
  7047  // many poorly-written tests.
  7048  // With the randomness here, as long as the tests pass
  7049  // consistently with -race, they shouldn't have latent scheduling
  7050  // assumptions.
  7051  const randomizeScheduler = raceenabled
  7052  
  7053  // runqput tries to put g on the local runnable queue.
  7054  // If next is false, runqput adds g to the tail of the runnable queue.
  7055  // If next is true, runqput puts g in the pp.runnext slot.
  7056  // If the run queue is full, runnext puts g on the global queue.
  7057  // Executed only by the owner P.
  7058  func runqput(pp *p, gp *g, next bool) {
  7059  	if !haveSysmon && next {
  7060  		// A runnext goroutine shares the same time slice as the
  7061  		// current goroutine (inheritTime from runqget). To prevent a
  7062  		// ping-pong pair of goroutines from starving all others, we
  7063  		// depend on sysmon to preempt "long-running goroutines". That
  7064  		// is, any set of goroutines sharing the same time slice.
  7065  		//
  7066  		// If there is no sysmon, we must avoid runnext entirely or
  7067  		// risk starvation.
  7068  		next = false
  7069  	}
  7070  	if randomizeScheduler && next && randn(2) == 0 {
  7071  		next = false
  7072  	}
  7073  
  7074  	if next {
  7075  	retryNext:
  7076  		oldnext := pp.runnext
  7077  		if !pp.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
  7078  			goto retryNext
  7079  		}
  7080  		if oldnext == 0 {
  7081  			return
  7082  		}
  7083  		// Kick the old runnext out to the regular run queue.
  7084  		gp = oldnext.ptr()
  7085  	}
  7086  
  7087  retry:
  7088  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  7089  	t := pp.runqtail
  7090  	if t-h < uint32(len(pp.runq)) {
  7091  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  7092  		atomic.StoreRel(&pp.runqtail, t+1) // store-release, makes the item available for consumption
  7093  		return
  7094  	}
  7095  	if runqputslow(pp, gp, h, t) {
  7096  		return
  7097  	}
  7098  	// the queue is not full, now the put above must succeed
  7099  	goto retry
  7100  }
  7101  
  7102  // Put g and a batch of work from local runnable queue on global queue.
  7103  // Executed only by the owner P.
  7104  func runqputslow(pp *p, gp *g, h, t uint32) bool {
  7105  	var batch [len(pp.runq)/2 + 1]*g
  7106  
  7107  	// First, grab a batch from local queue.
  7108  	n := t - h
  7109  	n = n / 2
  7110  	if n != uint32(len(pp.runq)/2) {
  7111  		throw("runqputslow: queue is not full")
  7112  	}
  7113  	for i := uint32(0); i < n; i++ {
  7114  		batch[i] = pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  7115  	}
  7116  	if !atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  7117  		return false
  7118  	}
  7119  	batch[n] = gp
  7120  
  7121  	if randomizeScheduler {
  7122  		for i := uint32(1); i <= n; i++ {
  7123  			j := cheaprandn(i + 1)
  7124  			batch[i], batch[j] = batch[j], batch[i]
  7125  		}
  7126  	}
  7127  
  7128  	// Link the goroutines.
  7129  	for i := uint32(0); i < n; i++ {
  7130  		batch[i].schedlink.set(batch[i+1])
  7131  	}
  7132  
  7133  	q := gQueue{batch[0].guintptr(), batch[n].guintptr(), int32(n + 1)}
  7134  
  7135  	// Now put the batch on global queue.
  7136  	lock(&sched.lock)
  7137  	globrunqputbatch(&q)
  7138  	unlock(&sched.lock)
  7139  	return true
  7140  }
  7141  
  7142  // runqputbatch tries to put all the G's on q on the local runnable queue.
  7143  // If the local runq is full the input queue still contains unqueued Gs.
  7144  // Executed only by the owner P.
  7145  func runqputbatch(pp *p, q *gQueue) {
  7146  	if q.empty() {
  7147  		return
  7148  	}
  7149  	h := atomic.LoadAcq(&pp.runqhead)
  7150  	t := pp.runqtail
  7151  	n := uint32(0)
  7152  	for !q.empty() && t-h < uint32(len(pp.runq)) {
  7153  		gp := q.pop()
  7154  		pp.runq[t%uint32(len(pp.runq))].set(gp)
  7155  		t++
  7156  		n++
  7157  	}
  7158  
  7159  	if randomizeScheduler {
  7160  		off := func(o uint32) uint32 {
  7161  			return (pp.runqtail + o) % uint32(len(pp.runq))
  7162  		}
  7163  		for i := uint32(1); i < n; i++ {
  7164  			j := cheaprandn(i + 1)
  7165  			pp.runq[off(i)], pp.runq[off(j)] = pp.runq[off(j)], pp.runq[off(i)]
  7166  		}
  7167  	}
  7168  
  7169  	atomic.StoreRel(&pp.runqtail, t)
  7170  
  7171  	return
  7172  }
  7173  
  7174  // Get g from local runnable queue.
  7175  // If inheritTime is true, gp should inherit the remaining time in the
  7176  // current time slice. Otherwise, it should start a new time slice.
  7177  // Executed only by the owner P.
  7178  func runqget(pp *p) (gp *g, inheritTime bool) {
  7179  	// If there's a runnext, it's the next G to run.
  7180  	next := pp.runnext
  7181  	// If the runnext is non-0 and the CAS fails, it could only have been stolen by another P,
  7182  	// because other Ps can race to set runnext to 0, but only the current P can set it to non-0.
  7183  	// Hence, there's no need to retry this CAS if it fails.
  7184  	if next != 0 && pp.runnext.cas(next, 0) {
  7185  		return next.ptr(), true
  7186  	}
  7187  
  7188  	for {
  7189  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7190  		t := pp.runqtail
  7191  		if t == h {
  7192  			return nil, false
  7193  		}
  7194  		gp := pp.runq[h%uint32(len(pp.runq))].ptr()
  7195  		if atomic.CasRel(&pp.runqhead, h, h+1) { // cas-release, commits consume
  7196  			return gp, false
  7197  		}
  7198  	}
  7199  }
  7200  
  7201  // runqdrain drains the local runnable queue of pp and returns all goroutines in it.
  7202  // Executed only by the owner P.
  7203  func runqdrain(pp *p) (drainQ gQueue) {
  7204  	oldNext := pp.runnext
  7205  	if oldNext != 0 && pp.runnext.cas(oldNext, 0) {
  7206  		drainQ.pushBack(oldNext.ptr())
  7207  	}
  7208  
  7209  retry:
  7210  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7211  	t := pp.runqtail
  7212  	qn := t - h
  7213  	if qn == 0 {
  7214  		return
  7215  	}
  7216  	if qn > uint32(len(pp.runq)) { // read inconsistent h and t
  7217  		goto retry
  7218  	}
  7219  
  7220  	if !atomic.CasRel(&pp.runqhead, h, h+qn) { // cas-release, commits consume
  7221  		goto retry
  7222  	}
  7223  
  7224  	// We've inverted the order in which it gets G's from the local P's runnable queue
  7225  	// and then advances the head pointer because we don't want to mess up the statuses of G's
  7226  	// while runqdrain() and runqsteal() are running in parallel.
  7227  	// Thus we should advance the head pointer before draining the local P into a gQueue,
  7228  	// so that we can update any gp.schedlink only after we take the full ownership of G,
  7229  	// meanwhile, other P's can't access to all G's in local P's runnable queue and steal them.
  7230  	// See https://groups.google.com/g/golang-dev/c/0pTKxEKhHSc/m/6Q85QjdVBQAJ for more details.
  7231  	for i := uint32(0); i < qn; i++ {
  7232  		gp := pp.runq[(h+i)%uint32(len(pp.runq))].ptr()
  7233  		drainQ.pushBack(gp)
  7234  	}
  7235  	return
  7236  }
  7237  
  7238  // Grabs a batch of goroutines from pp's runnable queue into batch.
  7239  // Batch is a ring buffer starting at batchHead.
  7240  // Returns number of grabbed goroutines.
  7241  // Can be executed by any P.
  7242  func runqgrab(pp *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
  7243  	for {
  7244  		h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with other consumers
  7245  		t := atomic.LoadAcq(&pp.runqtail) // load-acquire, synchronize with the producer
  7246  		n := t - h
  7247  		n = n - n/2
  7248  		if n == 0 {
  7249  			if stealRunNextG {
  7250  				// Try to steal from pp.runnext.
  7251  				if next := pp.runnext; next != 0 {
  7252  					if pp.status == _Prunning {
  7253  						// Sleep to ensure that pp isn't about to run the g
  7254  						// we are about to steal.
  7255  						// The important use case here is when the g running
  7256  						// on pp ready()s another g and then almost
  7257  						// immediately blocks. Instead of stealing runnext
  7258  						// in this window, back off to give pp a chance to
  7259  						// schedule runnext. This will avoid thrashing gs
  7260  						// between different Ps.
  7261  						// A sync chan send/recv takes ~50ns as of time of
  7262  						// writing, so 3us gives ~50x overshoot.
  7263  						if !osHasLowResTimer {
  7264  							usleep(3)
  7265  						} else {
  7266  							// On some platforms system timer granularity is
  7267  							// 1-15ms, which is way too much for this
  7268  							// optimization. So just yield.
  7269  							osyield()
  7270  						}
  7271  					}
  7272  					if !pp.runnext.cas(next, 0) {
  7273  						continue
  7274  					}
  7275  					batch[batchHead%uint32(len(batch))] = next
  7276  					return 1
  7277  				}
  7278  			}
  7279  			return 0
  7280  		}
  7281  		if n > uint32(len(pp.runq)/2) { // read inconsistent h and t
  7282  			continue
  7283  		}
  7284  		for i := uint32(0); i < n; i++ {
  7285  			g := pp.runq[(h+i)%uint32(len(pp.runq))]
  7286  			batch[(batchHead+i)%uint32(len(batch))] = g
  7287  		}
  7288  		if atomic.CasRel(&pp.runqhead, h, h+n) { // cas-release, commits consume
  7289  			return n
  7290  		}
  7291  	}
  7292  }
  7293  
  7294  // Steal half of elements from local runnable queue of p2
  7295  // and put onto local runnable queue of p.
  7296  // Returns one of the stolen elements (or nil if failed).
  7297  func runqsteal(pp, p2 *p, stealRunNextG bool) *g {
  7298  	t := pp.runqtail
  7299  	n := runqgrab(p2, &pp.runq, t, stealRunNextG)
  7300  	if n == 0 {
  7301  		return nil
  7302  	}
  7303  	n--
  7304  	gp := pp.runq[(t+n)%uint32(len(pp.runq))].ptr()
  7305  	if n == 0 {
  7306  		return gp
  7307  	}
  7308  	h := atomic.LoadAcq(&pp.runqhead) // load-acquire, synchronize with consumers
  7309  	if t-h+n >= uint32(len(pp.runq)) {
  7310  		throw("runqsteal: runq overflow")
  7311  	}
  7312  	atomic.StoreRel(&pp.runqtail, t+n) // store-release, makes the item available for consumption
  7313  	return gp
  7314  }
  7315  
  7316  // A gQueue is a dequeue of Gs linked through g.schedlink. A G can only
  7317  // be on one gQueue or gList at a time.
  7318  type gQueue struct {
  7319  	head guintptr
  7320  	tail guintptr
  7321  	size int32
  7322  }
  7323  
  7324  // empty reports whether q is empty.
  7325  func (q *gQueue) empty() bool {
  7326  	return q.head == 0
  7327  }
  7328  
  7329  // push adds gp to the head of q.
  7330  func (q *gQueue) push(gp *g) {
  7331  	gp.schedlink = q.head
  7332  	q.head.set(gp)
  7333  	if q.tail == 0 {
  7334  		q.tail.set(gp)
  7335  	}
  7336  	q.size++
  7337  }
  7338  
  7339  // pushBack adds gp to the tail of q.
  7340  func (q *gQueue) pushBack(gp *g) {
  7341  	gp.schedlink = 0
  7342  	if q.tail != 0 {
  7343  		q.tail.ptr().schedlink.set(gp)
  7344  	} else {
  7345  		q.head.set(gp)
  7346  	}
  7347  	q.tail.set(gp)
  7348  	q.size++
  7349  }
  7350  
  7351  // pushBackAll adds all Gs in q2 to the tail of q. After this q2 must
  7352  // not be used.
  7353  func (q *gQueue) pushBackAll(q2 gQueue) {
  7354  	if q2.tail == 0 {
  7355  		return
  7356  	}
  7357  	q2.tail.ptr().schedlink = 0
  7358  	if q.tail != 0 {
  7359  		q.tail.ptr().schedlink = q2.head
  7360  	} else {
  7361  		q.head = q2.head
  7362  	}
  7363  	q.tail = q2.tail
  7364  	q.size += q2.size
  7365  }
  7366  
  7367  // pop removes and returns the head of queue q. It returns nil if
  7368  // q is empty.
  7369  func (q *gQueue) pop() *g {
  7370  	gp := q.head.ptr()
  7371  	if gp != nil {
  7372  		q.head = gp.schedlink
  7373  		if q.head == 0 {
  7374  			q.tail = 0
  7375  		}
  7376  		q.size--
  7377  	}
  7378  	return gp
  7379  }
  7380  
  7381  // popList takes all Gs in q and returns them as a gList.
  7382  func (q *gQueue) popList() gList {
  7383  	stack := gList{q.head, q.size}
  7384  	*q = gQueue{}
  7385  	return stack
  7386  }
  7387  
  7388  // A gList is a list of Gs linked through g.schedlink. A G can only be
  7389  // on one gQueue or gList at a time.
  7390  type gList struct {
  7391  	head guintptr
  7392  	size int32
  7393  }
  7394  
  7395  // empty reports whether l is empty.
  7396  func (l *gList) empty() bool {
  7397  	return l.head == 0
  7398  }
  7399  
  7400  // push adds gp to the head of l.
  7401  func (l *gList) push(gp *g) {
  7402  	gp.schedlink = l.head
  7403  	l.head.set(gp)
  7404  	l.size++
  7405  }
  7406  
  7407  // pushAll prepends all Gs in q to l. After this q must not be used.
  7408  func (l *gList) pushAll(q gQueue) {
  7409  	if !q.empty() {
  7410  		q.tail.ptr().schedlink = l.head
  7411  		l.head = q.head
  7412  		l.size += q.size
  7413  	}
  7414  }
  7415  
  7416  // pop removes and returns the head of l. If l is empty, it returns nil.
  7417  func (l *gList) pop() *g {
  7418  	gp := l.head.ptr()
  7419  	if gp != nil {
  7420  		l.head = gp.schedlink
  7421  		l.size--
  7422  	}
  7423  	return gp
  7424  }
  7425  
  7426  //go:linkname setMaxThreads runtime/debug.setMaxThreads
  7427  func setMaxThreads(in int) (out int) {
  7428  	lock(&sched.lock)
  7429  	out = int(sched.maxmcount)
  7430  	if in > 0x7fffffff { // MaxInt32
  7431  		sched.maxmcount = 0x7fffffff
  7432  	} else {
  7433  		sched.maxmcount = int32(in)
  7434  	}
  7435  	checkmcount()
  7436  	unlock(&sched.lock)
  7437  	return
  7438  }
  7439  
  7440  // procPin should be an internal detail,
  7441  // but widely used packages access it using linkname.
  7442  // Notable members of the hall of shame include:
  7443  //   - github.com/bytedance/gopkg
  7444  //   - github.com/choleraehyq/pid
  7445  //   - github.com/songzhibin97/gkit
  7446  //
  7447  // Do not remove or change the type signature.
  7448  // See go.dev/issue/67401.
  7449  //
  7450  //go:linkname procPin
  7451  //go:nosplit
  7452  func procPin() int {
  7453  	gp := getg()
  7454  	mp := gp.m
  7455  
  7456  	mp.locks++
  7457  	return int(mp.p.ptr().id)
  7458  }
  7459  
  7460  // procUnpin should be an internal detail,
  7461  // but widely used packages access it using linkname.
  7462  // Notable members of the hall of shame include:
  7463  //   - github.com/bytedance/gopkg
  7464  //   - github.com/choleraehyq/pid
  7465  //   - github.com/songzhibin97/gkit
  7466  //
  7467  // Do not remove or change the type signature.
  7468  // See go.dev/issue/67401.
  7469  //
  7470  //go:linkname procUnpin
  7471  //go:nosplit
  7472  func procUnpin() {
  7473  	gp := getg()
  7474  	gp.m.locks--
  7475  }
  7476  
  7477  //go:linkname sync_runtime_procPin sync.runtime_procPin
  7478  //go:nosplit
  7479  func sync_runtime_procPin() int {
  7480  	return procPin()
  7481  }
  7482  
  7483  //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
  7484  //go:nosplit
  7485  func sync_runtime_procUnpin() {
  7486  	procUnpin()
  7487  }
  7488  
  7489  //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
  7490  //go:nosplit
  7491  func sync_atomic_runtime_procPin() int {
  7492  	return procPin()
  7493  }
  7494  
  7495  //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
  7496  //go:nosplit
  7497  func sync_atomic_runtime_procUnpin() {
  7498  	procUnpin()
  7499  }
  7500  
  7501  // Active spinning for sync.Mutex.
  7502  //
  7503  //go:linkname internal_sync_runtime_canSpin internal/sync.runtime_canSpin
  7504  //go:nosplit
  7505  func internal_sync_runtime_canSpin(i int) bool {
  7506  	// sync.Mutex is cooperative, so we are conservative with spinning.
  7507  	// Spin only few times and only if running on a multicore machine and
  7508  	// GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
  7509  	// As opposed to runtime mutex we don't do passive spinning here,
  7510  	// because there can be work on global runq or on other Ps.
  7511  	if i >= active_spin || numCPUStartup <= 1 || gomaxprocs <= sched.npidle.Load()+sched.nmspinning.Load()+1 {
  7512  		return false
  7513  	}
  7514  	if p := getg().m.p.ptr(); !runqempty(p) {
  7515  		return false
  7516  	}
  7517  	return true
  7518  }
  7519  
  7520  //go:linkname internal_sync_runtime_doSpin internal/sync.runtime_doSpin
  7521  //go:nosplit
  7522  func internal_sync_runtime_doSpin() {
  7523  	procyield(active_spin_cnt)
  7524  }
  7525  
  7526  // Active spinning for sync.Mutex.
  7527  //
  7528  // sync_runtime_canSpin should be an internal detail,
  7529  // but widely used packages access it using linkname.
  7530  // Notable members of the hall of shame include:
  7531  //   - github.com/livekit/protocol
  7532  //   - github.com/sagernet/gvisor
  7533  //   - gvisor.dev/gvisor
  7534  //
  7535  // Do not remove or change the type signature.
  7536  // See go.dev/issue/67401.
  7537  //
  7538  //go:linkname sync_runtime_canSpin sync.runtime_canSpin
  7539  //go:nosplit
  7540  func sync_runtime_canSpin(i int) bool {
  7541  	return internal_sync_runtime_canSpin(i)
  7542  }
  7543  
  7544  // sync_runtime_doSpin should be an internal detail,
  7545  // but widely used packages access it using linkname.
  7546  // Notable members of the hall of shame include:
  7547  //   - github.com/livekit/protocol
  7548  //   - github.com/sagernet/gvisor
  7549  //   - gvisor.dev/gvisor
  7550  //
  7551  // Do not remove or change the type signature.
  7552  // See go.dev/issue/67401.
  7553  //
  7554  //go:linkname sync_runtime_doSpin sync.runtime_doSpin
  7555  //go:nosplit
  7556  func sync_runtime_doSpin() {
  7557  	internal_sync_runtime_doSpin()
  7558  }
  7559  
  7560  var stealOrder randomOrder
  7561  
  7562  // randomOrder/randomEnum are helper types for randomized work stealing.
  7563  // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
  7564  // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
  7565  // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
  7566  type randomOrder struct {
  7567  	count    uint32
  7568  	coprimes []uint32
  7569  }
  7570  
  7571  type randomEnum struct {
  7572  	i     uint32
  7573  	count uint32
  7574  	pos   uint32
  7575  	inc   uint32
  7576  }
  7577  
  7578  func (ord *randomOrder) reset(count uint32) {
  7579  	ord.count = count
  7580  	ord.coprimes = ord.coprimes[:0]
  7581  	for i := uint32(1); i <= count; i++ {
  7582  		if gcd(i, count) == 1 {
  7583  			ord.coprimes = append(ord.coprimes, i)
  7584  		}
  7585  	}
  7586  }
  7587  
  7588  func (ord *randomOrder) start(i uint32) randomEnum {
  7589  	return randomEnum{
  7590  		count: ord.count,
  7591  		pos:   i % ord.count,
  7592  		inc:   ord.coprimes[i/ord.count%uint32(len(ord.coprimes))],
  7593  	}
  7594  }
  7595  
  7596  func (enum *randomEnum) done() bool {
  7597  	return enum.i == enum.count
  7598  }
  7599  
  7600  func (enum *randomEnum) next() {
  7601  	enum.i++
  7602  	enum.pos = (enum.pos + enum.inc) % enum.count
  7603  }
  7604  
  7605  func (enum *randomEnum) position() uint32 {
  7606  	return enum.pos
  7607  }
  7608  
  7609  func gcd(a, b uint32) uint32 {
  7610  	for b != 0 {
  7611  		a, b = b, a%b
  7612  	}
  7613  	return a
  7614  }
  7615  
  7616  // An initTask represents the set of initializations that need to be done for a package.
  7617  // Keep in sync with ../../test/noinit.go:initTask
  7618  type initTask struct {
  7619  	state uint32 // 0 = uninitialized, 1 = in progress, 2 = done
  7620  	nfns  uint32
  7621  	// followed by nfns pcs, uintptr sized, one per init function to run
  7622  }
  7623  
  7624  // inittrace stores statistics for init functions which are
  7625  // updated by malloc and newproc when active is true.
  7626  var inittrace tracestat
  7627  
  7628  type tracestat struct {
  7629  	active bool   // init tracing activation status
  7630  	id     uint64 // init goroutine id
  7631  	allocs uint64 // heap allocations
  7632  	bytes  uint64 // heap allocated bytes
  7633  }
  7634  
  7635  func doInit(ts []*initTask) {
  7636  	for _, t := range ts {
  7637  		doInit1(t)
  7638  	}
  7639  }
  7640  
  7641  func doInit1(t *initTask) {
  7642  	switch t.state {
  7643  	case 2: // fully initialized
  7644  		return
  7645  	case 1: // initialization in progress
  7646  		throw("recursive call during initialization - linker skew")
  7647  	default: // not initialized yet
  7648  		t.state = 1 // initialization in progress
  7649  
  7650  		var (
  7651  			start  int64
  7652  			before tracestat
  7653  		)
  7654  
  7655  		if inittrace.active {
  7656  			start = nanotime()
  7657  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  7658  			before = inittrace
  7659  		}
  7660  
  7661  		if t.nfns == 0 {
  7662  			// We should have pruned all of these in the linker.
  7663  			throw("inittask with no functions")
  7664  		}
  7665  
  7666  		firstFunc := add(unsafe.Pointer(t), 8)
  7667  		for i := uint32(0); i < t.nfns; i++ {
  7668  			p := add(firstFunc, uintptr(i)*goarch.PtrSize)
  7669  			f := *(*func())(unsafe.Pointer(&p))
  7670  			f()
  7671  		}
  7672  
  7673  		if inittrace.active {
  7674  			end := nanotime()
  7675  			// Load stats non-atomically since tracinit is updated only by this init goroutine.
  7676  			after := inittrace
  7677  
  7678  			f := *(*func())(unsafe.Pointer(&firstFunc))
  7679  			pkg := funcpkgpath(findfunc(abi.FuncPCABIInternal(f)))
  7680  
  7681  			var sbuf [24]byte
  7682  			print("init ", pkg, " @")
  7683  			print(string(fmtNSAsMS(sbuf[:], uint64(start-runtimeInitTime))), " ms, ")
  7684  			print(string(fmtNSAsMS(sbuf[:], uint64(end-start))), " ms clock, ")
  7685  			print(string(itoa(sbuf[:], after.bytes-before.bytes)), " bytes, ")
  7686  			print(string(itoa(sbuf[:], after.allocs-before.allocs)), " allocs")
  7687  			print("\n")
  7688  		}
  7689  
  7690  		t.state = 2 // initialization done
  7691  	}
  7692  }
  7693