Source file src/runtime/proc.go

Documentation: runtime

     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  	"runtime/internal/atomic"
     9  	"runtime/internal/sys"
    10  	"unsafe"
    11  )
    12  
    13  var buildVersion = sys.TheVersion
    14  
    15  // Goroutine scheduler
    16  // The scheduler's job is to distribute ready-to-run goroutines over worker threads.
    17  //
    18  // The main concepts are:
    19  // G - goroutine.
    20  // M - worker thread, or machine.
    21  // P - processor, a resource that is required to execute Go code.
    22  //     M must have an associated P to execute Go code, however it can be
    23  //     blocked or in a syscall w/o an associated P.
    24  //
    25  // Design doc at https://golang.org/s/go11sched.
    26  
    27  // Worker thread parking/unparking.
    28  // We need to balance between keeping enough running worker threads to utilize
    29  // available hardware parallelism and parking excessive running worker threads
    30  // to conserve CPU resources and power. This is not simple for two reasons:
    31  // (1) scheduler state is intentionally distributed (in particular, per-P work
    32  // queues), so it is not possible to compute global predicates on fast paths;
    33  // (2) for optimal thread management we would need to know the future (don't park
    34  // a worker thread when a new goroutine will be readied in near future).
    35  //
    36  // Three rejected approaches that would work badly:
    37  // 1. Centralize all scheduler state (would inhibit scalability).
    38  // 2. Direct goroutine handoff. That is, when we ready a new goroutine and there
    39  //    is a spare P, unpark a thread and handoff it the thread and the goroutine.
    40  //    This would lead to thread state thrashing, as the thread that readied the
    41  //    goroutine can be out of work the very next moment, we will need to park it.
    42  //    Also, it would destroy locality of computation as we want to preserve
    43  //    dependent goroutines on the same thread; and introduce additional latency.
    44  // 3. Unpark an additional thread whenever we ready a goroutine and there is an
    45  //    idle P, but don't do handoff. This would lead to excessive thread parking/
    46  //    unparking as the additional threads will instantly park without discovering
    47  //    any work to do.
    48  //
    49  // The current approach:
    50  // We unpark an additional thread when we ready a goroutine if (1) there is an
    51  // idle P and there are no "spinning" worker threads. A worker thread is considered
    52  // spinning if it is out of local work and did not find work in global run queue/
    53  // netpoller; the spinning state is denoted in m.spinning and in sched.nmspinning.
    54  // Threads unparked this way are also considered spinning; we don't do goroutine
    55  // handoff so such threads are out of work initially. Spinning threads do some
    56  // spinning looking for work in per-P run queues before parking. If a spinning
    57  // thread finds work it takes itself out of the spinning state and proceeds to
    58  // execution. If it does not find work it takes itself out of the spinning state
    59  // and then parks.
    60  // If there is at least one spinning thread (sched.nmspinning>1), we don't unpark
    61  // new threads when readying goroutines. To compensate for that, if the last spinning
    62  // thread finds work and stops spinning, it must unpark a new spinning thread.
    63  // This approach smooths out unjustified spikes of thread unparking,
    64  // but at the same time guarantees eventual maximal CPU parallelism utilization.
    65  //
    66  // The main implementation complication is that we need to be very careful during
    67  // spinning->non-spinning thread transition. This transition can race with submission
    68  // of a new goroutine, and either one part or another needs to unpark another worker
    69  // thread. If they both fail to do that, we can end up with semi-persistent CPU
    70  // underutilization. The general pattern for goroutine readying is: submit a goroutine
    71  // to local work queue, #StoreLoad-style memory barrier, check sched.nmspinning.
    72  // The general pattern for spinning->non-spinning transition is: decrement nmspinning,
    73  // #StoreLoad-style memory barrier, check all per-P work queues for new work.
    74  // Note that all this complexity does not apply to global run queue as we are not
    75  // sloppy about thread unparking when submitting to global queue. Also see comments
    76  // for nmspinning manipulation.
    77  
    78  var (
    79  	m0           m
    80  	g0           g
    81  	raceprocctx0 uintptr
    82  )
    83  
    84  //go:linkname runtime_init runtime.init
    85  func runtime_init()
    86  
    87  //go:linkname main_init main.init
    88  func main_init()
    89  
    90  // main_init_done is a signal used by cgocallbackg that initialization
    91  // has been completed. It is made before _cgo_notify_runtime_init_done,
    92  // so all cgo calls can rely on it existing. When main_init is complete,
    93  // it is closed, meaning cgocallbackg can reliably receive from it.
    94  var main_init_done chan bool
    95  
    96  //go:linkname main_main main.main
    97  func main_main()
    98  
    99  // mainStarted indicates that the main M has started.
   100  var mainStarted bool
   101  
   102  // runtimeInitTime is the nanotime() at which the runtime started.
   103  var runtimeInitTime int64
   104  
   105  // Value to use for signal mask for newly created M's.
   106  var initSigmask sigset
   107  
   108  // The main goroutine.
   109  func main() {
   110  	g := getg()
   111  
   112  	// Racectx of m0->g0 is used only as the parent of the main goroutine.
   113  	// It must not be used for anything else.
   114  	g.m.g0.racectx = 0
   115  
   116  	// Max stack size is 1 GB on 64-bit, 250 MB on 32-bit.
   117  	// Using decimal instead of binary GB and MB because
   118  	// they look nicer in the stack overflow failure message.
   119  	if sys.PtrSize == 8 {
   120  		maxstacksize = 1000000000
   121  	} else {
   122  		maxstacksize = 250000000
   123  	}
   124  
   125  	// Allow newproc to start new Ms.
   126  	mainStarted = true
   127  
   128  	systemstack(func() {
   129  		newm(sysmon, nil)
   130  	})
   131  
   132  	// Lock the main goroutine onto this, the main OS thread,
   133  	// during initialization. Most programs won't care, but a few
   134  	// do require certain calls to be made by the main thread.
   135  	// Those can arrange for main.main to run in the main thread
   136  	// by calling runtime.LockOSThread during initialization
   137  	// to preserve the lock.
   138  	lockOSThread()
   139  
   140  	if g.m != &m0 {
   141  		throw("runtime.main not on m0")
   142  	}
   143  
   144  	runtime_init() // must be before defer
   145  	if nanotime() == 0 {
   146  		throw("nanotime returning zero")
   147  	}
   148  
   149  	// Defer unlock so that runtime.Goexit during init does the unlock too.
   150  	needUnlock := true
   151  	defer func() {
   152  		if needUnlock {
   153  			unlockOSThread()
   154  		}
   155  	}()
   156  
   157  	// Record when the world started. Must be after runtime_init
   158  	// because nanotime on some platforms depends on startNano.
   159  	runtimeInitTime = nanotime()
   160  
   161  	gcenable()
   162  
   163  	main_init_done = make(chan bool)
   164  	if iscgo {
   165  		if _cgo_thread_start == nil {
   166  			throw("_cgo_thread_start missing")
   167  		}
   168  		if GOOS != "windows" {
   169  			if _cgo_setenv == nil {
   170  				throw("_cgo_setenv missing")
   171  			}
   172  			if _cgo_unsetenv == nil {
   173  				throw("_cgo_unsetenv missing")
   174  			}
   175  		}
   176  		if _cgo_notify_runtime_init_done == nil {
   177  			throw("_cgo_notify_runtime_init_done missing")
   178  		}
   179  		// Start the template thread in case we enter Go from
   180  		// a C-created thread and need to create a new thread.
   181  		startTemplateThread()
   182  		cgocall(_cgo_notify_runtime_init_done, nil)
   183  	}
   184  
   185  	fn := main_init // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
   186  	fn()
   187  	close(main_init_done)
   188  
   189  	needUnlock = false
   190  	unlockOSThread()
   191  
   192  	if isarchive || islibrary {
   193  		// A program compiled with -buildmode=c-archive or c-shared
   194  		// has a main, but it is not executed.
   195  		return
   196  	}
   197  	fn = main_main // make an indirect call, as the linker doesn't know the address of the main package when laying down the runtime
   198  	fn()
   199  	if raceenabled {
   200  		racefini()
   201  	}
   202  
   203  	// Make racy client program work: if panicking on
   204  	// another goroutine at the same time as main returns,
   205  	// let the other goroutine finish printing the panic trace.
   206  	// Once it does, it will exit. See issues 3934 and 20018.
   207  	if atomic.Load(&runningPanicDefers) != 0 {
   208  		// Running deferred functions should not take long.
   209  		for c := 0; c < 1000; c++ {
   210  			if atomic.Load(&runningPanicDefers) == 0 {
   211  				break
   212  			}
   213  			Gosched()
   214  		}
   215  	}
   216  	if atomic.Load(&panicking) != 0 {
   217  		gopark(nil, nil, "panicwait", traceEvGoStop, 1)
   218  	}
   219  
   220  	exit(0)
   221  	for {
   222  		var x *int32
   223  		*x = 0
   224  	}
   225  }
   226  
   227  // os_beforeExit is called from os.Exit(0).
   228  //go:linkname os_beforeExit os.runtime_beforeExit
   229  func os_beforeExit() {
   230  	if raceenabled {
   231  		racefini()
   232  	}
   233  }
   234  
   235  // start forcegc helper goroutine
   236  func init() {
   237  	go forcegchelper()
   238  }
   239  
   240  func forcegchelper() {
   241  	forcegc.g = getg()
   242  	for {
   243  		lock(&forcegc.lock)
   244  		if forcegc.idle != 0 {
   245  			throw("forcegc: phase error")
   246  		}
   247  		atomic.Store(&forcegc.idle, 1)
   248  		goparkunlock(&forcegc.lock, "force gc (idle)", traceEvGoBlock, 1)
   249  		// this goroutine is explicitly resumed by sysmon
   250  		if debug.gctrace > 0 {
   251  			println("GC forced")
   252  		}
   253  		// Time-triggered, fully concurrent.
   254  		gcStart(gcBackgroundMode, gcTrigger{kind: gcTriggerTime, now: nanotime()})
   255  	}
   256  }
   257  
   258  //go:nosplit
   259  
   260  // Gosched yields the processor, allowing other goroutines to run. It does not
   261  // suspend the current goroutine, so execution resumes automatically.
   262  func Gosched() {
   263  	mcall(gosched_m)
   264  }
   265  
   266  // goschedguarded yields the processor like gosched, but also checks
   267  // for forbidden states and opts out of the yield in those cases.
   268  //go:nosplit
   269  func goschedguarded() {
   270  	mcall(goschedguarded_m)
   271  }
   272  
   273  // Puts the current goroutine into a waiting state and calls unlockf.
   274  // If unlockf returns false, the goroutine is resumed.
   275  // unlockf must not access this G's stack, as it may be moved between
   276  // the call to gopark and the call to unlockf.
   277  func gopark(unlockf func(*g, unsafe.Pointer) bool, lock unsafe.Pointer, reason string, traceEv byte, traceskip int) {
   278  	mp := acquirem()
   279  	gp := mp.curg
   280  	status := readgstatus(gp)
   281  	if status != _Grunning && status != _Gscanrunning {
   282  		throw("gopark: bad g status")
   283  	}
   284  	mp.waitlock = lock
   285  	mp.waitunlockf = *(*unsafe.Pointer)(unsafe.Pointer(&unlockf))
   286  	gp.waitreason = reason
   287  	mp.waittraceev = traceEv
   288  	mp.waittraceskip = traceskip
   289  	releasem(mp)
   290  	// can't do anything that might move the G between Ms here.
   291  	mcall(park_m)
   292  }
   293  
   294  // Puts the current goroutine into a waiting state and unlocks the lock.
   295  // The goroutine can be made runnable again by calling goready(gp).
   296  func goparkunlock(lock *mutex, reason string, traceEv byte, traceskip int) {
   297  	gopark(parkunlock_c, unsafe.Pointer(lock), reason, traceEv, traceskip)
   298  }
   299  
   300  func goready(gp *g, traceskip int) {
   301  	systemstack(func() {
   302  		ready(gp, traceskip, true)
   303  	})
   304  }
   305  
   306  //go:nosplit
   307  func acquireSudog() *sudog {
   308  	// Delicate dance: the semaphore implementation calls
   309  	// acquireSudog, acquireSudog calls new(sudog),
   310  	// new calls malloc, malloc can call the garbage collector,
   311  	// and the garbage collector calls the semaphore implementation
   312  	// in stopTheWorld.
   313  	// Break the cycle by doing acquirem/releasem around new(sudog).
   314  	// The acquirem/releasem increments m.locks during new(sudog),
   315  	// which keeps the garbage collector from being invoked.
   316  	mp := acquirem()
   317  	pp := mp.p.ptr()
   318  	if len(pp.sudogcache) == 0 {
   319  		lock(&sched.sudoglock)
   320  		// First, try to grab a batch from central cache.
   321  		for len(pp.sudogcache) < cap(pp.sudogcache)/2 && sched.sudogcache != nil {
   322  			s := sched.sudogcache
   323  			sched.sudogcache = s.next
   324  			s.next = nil
   325  			pp.sudogcache = append(pp.sudogcache, s)
   326  		}
   327  		unlock(&sched.sudoglock)
   328  		// If the central cache is empty, allocate a new one.
   329  		if len(pp.sudogcache) == 0 {
   330  			pp.sudogcache = append(pp.sudogcache, new(sudog))
   331  		}
   332  	}
   333  	n := len(pp.sudogcache)
   334  	s := pp.sudogcache[n-1]
   335  	pp.sudogcache[n-1] = nil
   336  	pp.sudogcache = pp.sudogcache[:n-1]
   337  	if s.elem != nil {
   338  		throw("acquireSudog: found s.elem != nil in cache")
   339  	}
   340  	releasem(mp)
   341  	return s
   342  }
   343  
   344  //go:nosplit
   345  func releaseSudog(s *sudog) {
   346  	if s.elem != nil {
   347  		throw("runtime: sudog with non-nil elem")
   348  	}
   349  	if s.isSelect {
   350  		throw("runtime: sudog with non-false isSelect")
   351  	}
   352  	if s.next != nil {
   353  		throw("runtime: sudog with non-nil next")
   354  	}
   355  	if s.prev != nil {
   356  		throw("runtime: sudog with non-nil prev")
   357  	}
   358  	if s.waitlink != nil {
   359  		throw("runtime: sudog with non-nil waitlink")
   360  	}
   361  	if s.c != nil {
   362  		throw("runtime: sudog with non-nil c")
   363  	}
   364  	gp := getg()
   365  	if gp.param != nil {
   366  		throw("runtime: releaseSudog with non-nil gp.param")
   367  	}
   368  	mp := acquirem() // avoid rescheduling to another P
   369  	pp := mp.p.ptr()
   370  	if len(pp.sudogcache) == cap(pp.sudogcache) {
   371  		// Transfer half of local cache to the central cache.
   372  		var first, last *sudog
   373  		for len(pp.sudogcache) > cap(pp.sudogcache)/2 {
   374  			n := len(pp.sudogcache)
   375  			p := pp.sudogcache[n-1]
   376  			pp.sudogcache[n-1] = nil
   377  			pp.sudogcache = pp.sudogcache[:n-1]
   378  			if first == nil {
   379  				first = p
   380  			} else {
   381  				last.next = p
   382  			}
   383  			last = p
   384  		}
   385  		lock(&sched.sudoglock)
   386  		last.next = sched.sudogcache
   387  		sched.sudogcache = first
   388  		unlock(&sched.sudoglock)
   389  	}
   390  	pp.sudogcache = append(pp.sudogcache, s)
   391  	releasem(mp)
   392  }
   393  
   394  // funcPC returns the entry PC of the function f.
   395  // It assumes that f is a func value. Otherwise the behavior is undefined.
   396  //go:nosplit
   397  func funcPC(f interface{}) uintptr {
   398  	return **(**uintptr)(add(unsafe.Pointer(&f), sys.PtrSize))
   399  }
   400  
   401  // called from assembly
   402  func badmcall(fn func(*g)) {
   403  	throw("runtime: mcall called on m->g0 stack")
   404  }
   405  
   406  func badmcall2(fn func(*g)) {
   407  	throw("runtime: mcall function returned")
   408  }
   409  
   410  func badreflectcall() {
   411  	panic(plainError("arg size to reflect.call more than 1GB"))
   412  }
   413  
   414  var badmorestackg0Msg = "fatal: morestack on g0\n"
   415  
   416  //go:nosplit
   417  //go:nowritebarrierrec
   418  func badmorestackg0() {
   419  	sp := stringStructOf(&badmorestackg0Msg)
   420  	write(2, sp.str, int32(sp.len))
   421  }
   422  
   423  var badmorestackgsignalMsg = "fatal: morestack on gsignal\n"
   424  
   425  //go:nosplit
   426  //go:nowritebarrierrec
   427  func badmorestackgsignal() {
   428  	sp := stringStructOf(&badmorestackgsignalMsg)
   429  	write(2, sp.str, int32(sp.len))
   430  }
   431  
   432  //go:nosplit
   433  func badctxt() {
   434  	throw("ctxt != 0")
   435  }
   436  
   437  func lockedOSThread() bool {
   438  	gp := getg()
   439  	return gp.lockedm != 0 && gp.m.lockedg != 0
   440  }
   441  
   442  var (
   443  	allgs    []*g
   444  	allglock mutex
   445  )
   446  
   447  func allgadd(gp *g) {
   448  	if readgstatus(gp) == _Gidle {
   449  		throw("allgadd: bad status Gidle")
   450  	}
   451  
   452  	lock(&allglock)
   453  	allgs = append(allgs, gp)
   454  	allglen = uintptr(len(allgs))
   455  	unlock(&allglock)
   456  }
   457  
   458  const (
   459  	// Number of goroutine ids to grab from sched.goidgen to local per-P cache at once.
   460  	// 16 seems to provide enough amortization, but other than that it's mostly arbitrary number.
   461  	_GoidCacheBatch = 16
   462  )
   463  
   464  // The bootstrap sequence is:
   465  //
   466  //	call osinit
   467  //	call schedinit
   468  //	make & queue new G
   469  //	call runtime·mstart
   470  //
   471  // The new G calls runtime·main.
   472  func schedinit() {
   473  	// raceinit must be the first call to race detector.
   474  	// In particular, it must be done before mallocinit below calls racemapshadow.
   475  	_g_ := getg()
   476  	if raceenabled {
   477  		_g_.racectx, raceprocctx0 = raceinit()
   478  	}
   479  
   480  	sched.maxmcount = 10000
   481  
   482  	tracebackinit()
   483  	moduledataverify()
   484  	stackinit()
   485  	mallocinit()
   486  	mcommoninit(_g_.m)
   487  	alginit()       // maps must not be used before this call
   488  	modulesinit()   // provides activeModules
   489  	typelinksinit() // uses maps, activeModules
   490  	itabsinit()     // uses activeModules
   491  
   492  	msigsave(_g_.m)
   493  	initSigmask = _g_.m.sigmask
   494  
   495  	goargs()
   496  	goenvs()
   497  	parsedebugvars()
   498  	gcinit()
   499  
   500  	sched.lastpoll = uint64(nanotime())
   501  	procs := ncpu
   502  	if n, ok := atoi32(gogetenv("GOMAXPROCS")); ok && n > 0 {
   503  		procs = n
   504  	}
   505  	if procresize(procs) != nil {
   506  		throw("unknown runnable goroutine during bootstrap")
   507  	}
   508  
   509  	// For cgocheck > 1, we turn on the write barrier at all times
   510  	// and check all pointer writes. We can't do this until after
   511  	// procresize because the write barrier needs a P.
   512  	if debug.cgocheck > 1 {
   513  		writeBarrier.cgo = true
   514  		writeBarrier.enabled = true
   515  		for _, p := range allp {
   516  			p.wbBuf.reset()
   517  		}
   518  	}
   519  
   520  	if buildVersion == "" {
   521  		// Condition should never trigger. This code just serves
   522  		// to ensure runtime·buildVersion is kept in the resulting binary.
   523  		buildVersion = "unknown"
   524  	}
   525  }
   526  
   527  func dumpgstatus(gp *g) {
   528  	_g_ := getg()
   529  	print("runtime: gp: gp=", gp, ", goid=", gp.goid, ", gp->atomicstatus=", readgstatus(gp), "\n")
   530  	print("runtime:  g:  g=", _g_, ", goid=", _g_.goid, ",  g->atomicstatus=", readgstatus(_g_), "\n")
   531  }
   532  
   533  func checkmcount() {
   534  	// sched lock is held
   535  	if mcount() > sched.maxmcount {
   536  		print("runtime: program exceeds ", sched.maxmcount, "-thread limit\n")
   537  		throw("thread exhaustion")
   538  	}
   539  }
   540  
   541  func mcommoninit(mp *m) {
   542  	_g_ := getg()
   543  
   544  	// g0 stack won't make sense for user (and is not necessary unwindable).
   545  	if _g_ != _g_.m.g0 {
   546  		callers(1, mp.createstack[:])
   547  	}
   548  
   549  	lock(&sched.lock)
   550  	if sched.mnext+1 < sched.mnext {
   551  		throw("runtime: thread ID overflow")
   552  	}
   553  	mp.id = sched.mnext
   554  	sched.mnext++
   555  	checkmcount()
   556  
   557  	mp.fastrand[0] = 1597334677 * uint32(mp.id)
   558  	mp.fastrand[1] = uint32(cputicks())
   559  	if mp.fastrand[0]|mp.fastrand[1] == 0 {
   560  		mp.fastrand[1] = 1
   561  	}
   562  
   563  	mpreinit(mp)
   564  	if mp.gsignal != nil {
   565  		mp.gsignal.stackguard1 = mp.gsignal.stack.lo + _StackGuard
   566  	}
   567  
   568  	// Add to allm so garbage collector doesn't free g->m
   569  	// when it is just in a register or thread-local storage.
   570  	mp.alllink = allm
   571  
   572  	// NumCgoCall() iterates over allm w/o schedlock,
   573  	// so we need to publish it safely.
   574  	atomicstorep(unsafe.Pointer(&allm), unsafe.Pointer(mp))
   575  	unlock(&sched.lock)
   576  
   577  	// Allocate memory to hold a cgo traceback if the cgo call crashes.
   578  	if iscgo || GOOS == "solaris" || GOOS == "windows" {
   579  		mp.cgoCallers = new(cgoCallers)
   580  	}
   581  }
   582  
   583  // Mark gp ready to run.
   584  func ready(gp *g, traceskip int, next bool) {
   585  	if trace.enabled {
   586  		traceGoUnpark(gp, traceskip)
   587  	}
   588  
   589  	status := readgstatus(gp)
   590  
   591  	// Mark runnable.
   592  	_g_ := getg()
   593  	_g_.m.locks++ // disable preemption because it can be holding p in a local var
   594  	if status&^_Gscan != _Gwaiting {
   595  		dumpgstatus(gp)
   596  		throw("bad g->status in ready")
   597  	}
   598  
   599  	// status is Gwaiting or Gscanwaiting, make Grunnable and put on runq
   600  	casgstatus(gp, _Gwaiting, _Grunnable)
   601  	runqput(_g_.m.p.ptr(), gp, next)
   602  	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
   603  		wakep()
   604  	}
   605  	_g_.m.locks--
   606  	if _g_.m.locks == 0 && _g_.preempt { // restore the preemption request in Case we've cleared it in newstack
   607  		_g_.stackguard0 = stackPreempt
   608  	}
   609  }
   610  
   611  func gcprocs() int32 {
   612  	// Figure out how many CPUs to use during GC.
   613  	// Limited by gomaxprocs, number of actual CPUs, and MaxGcproc.
   614  	lock(&sched.lock)
   615  	n := gomaxprocs
   616  	if n > ncpu {
   617  		n = ncpu
   618  	}
   619  	if n > _MaxGcproc {
   620  		n = _MaxGcproc
   621  	}
   622  	if n > sched.nmidle+1 { // one M is currently running
   623  		n = sched.nmidle + 1
   624  	}
   625  	unlock(&sched.lock)
   626  	return n
   627  }
   628  
   629  func needaddgcproc() bool {
   630  	lock(&sched.lock)
   631  	n := gomaxprocs
   632  	if n > ncpu {
   633  		n = ncpu
   634  	}
   635  	if n > _MaxGcproc {
   636  		n = _MaxGcproc
   637  	}
   638  	n -= sched.nmidle + 1 // one M is currently running
   639  	unlock(&sched.lock)
   640  	return n > 0
   641  }
   642  
   643  func helpgc(nproc int32) {
   644  	_g_ := getg()
   645  	lock(&sched.lock)
   646  	pos := 0
   647  	for n := int32(1); n < nproc; n++ { // one M is currently running
   648  		if allp[pos].mcache == _g_.m.mcache {
   649  			pos++
   650  		}
   651  		mp := mget()
   652  		if mp == nil {
   653  			throw("gcprocs inconsistency")
   654  		}
   655  		mp.helpgc = n
   656  		mp.p.set(allp[pos])
   657  		mp.mcache = allp[pos].mcache
   658  		pos++
   659  		notewakeup(&mp.park)
   660  	}
   661  	unlock(&sched.lock)
   662  }
   663  
   664  // freezeStopWait is a large value that freezetheworld sets
   665  // sched.stopwait to in order to request that all Gs permanently stop.
   666  const freezeStopWait = 0x7fffffff
   667  
   668  // freezing is set to non-zero if the runtime is trying to freeze the
   669  // world.
   670  var freezing uint32
   671  
   672  // Similar to stopTheWorld but best-effort and can be called several times.
   673  // There is no reverse operation, used during crashing.
   674  // This function must not lock any mutexes.
   675  func freezetheworld() {
   676  	atomic.Store(&freezing, 1)
   677  	// stopwait and preemption requests can be lost
   678  	// due to races with concurrently executing threads,
   679  	// so try several times
   680  	for i := 0; i < 5; i++ {
   681  		// this should tell the scheduler to not start any new goroutines
   682  		sched.stopwait = freezeStopWait
   683  		atomic.Store(&sched.gcwaiting, 1)
   684  		// this should stop running goroutines
   685  		if !preemptall() {
   686  			break // no running goroutines
   687  		}
   688  		usleep(1000)
   689  	}
   690  	// to be sure
   691  	usleep(1000)
   692  	preemptall()
   693  	usleep(1000)
   694  }
   695  
   696  func isscanstatus(status uint32) bool {
   697  	if status == _Gscan {
   698  		throw("isscanstatus: Bad status Gscan")
   699  	}
   700  	return status&_Gscan == _Gscan
   701  }
   702  
   703  // All reads and writes of g's status go through readgstatus, casgstatus
   704  // castogscanstatus, casfrom_Gscanstatus.
   705  //go:nosplit
   706  func readgstatus(gp *g) uint32 {
   707  	return atomic.Load(&gp.atomicstatus)
   708  }
   709  
   710  // Ownership of gcscanvalid:
   711  //
   712  // If gp is running (meaning status == _Grunning or _Grunning|_Gscan),
   713  // then gp owns gp.gcscanvalid, and other goroutines must not modify it.
   714  //
   715  // Otherwise, a second goroutine can lock the scan state by setting _Gscan
   716  // in the status bit and then modify gcscanvalid, and then unlock the scan state.
   717  //
   718  // Note that the first condition implies an exception to the second:
   719  // if a second goroutine changes gp's status to _Grunning|_Gscan,
   720  // that second goroutine still does not have the right to modify gcscanvalid.
   721  
   722  // The Gscanstatuses are acting like locks and this releases them.
   723  // If it proves to be a performance hit we should be able to make these
   724  // simple atomic stores but for now we are going to throw if
   725  // we see an inconsistent state.
   726  func casfrom_Gscanstatus(gp *g, oldval, newval uint32) {
   727  	success := false
   728  
   729  	// Check that transition is valid.
   730  	switch oldval {
   731  	default:
   732  		print("runtime: casfrom_Gscanstatus bad oldval gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
   733  		dumpgstatus(gp)
   734  		throw("casfrom_Gscanstatus:top gp->status is not in scan state")
   735  	case _Gscanrunnable,
   736  		_Gscanwaiting,
   737  		_Gscanrunning,
   738  		_Gscansyscall:
   739  		if newval == oldval&^_Gscan {
   740  			success = atomic.Cas(&gp.atomicstatus, oldval, newval)
   741  		}
   742  	}
   743  	if !success {
   744  		print("runtime: casfrom_Gscanstatus failed gp=", gp, ", oldval=", hex(oldval), ", newval=", hex(newval), "\n")
   745  		dumpgstatus(gp)
   746  		throw("casfrom_Gscanstatus: gp->status is not in scan state")
   747  	}
   748  }
   749  
   750  // This will return false if the gp is not in the expected status and the cas fails.
   751  // This acts like a lock acquire while the casfromgstatus acts like a lock release.
   752  func castogscanstatus(gp *g, oldval, newval uint32) bool {
   753  	switch oldval {
   754  	case _Grunnable,
   755  		_Grunning,
   756  		_Gwaiting,
   757  		_Gsyscall:
   758  		if newval == oldval|_Gscan {
   759  			return atomic.Cas(&gp.atomicstatus, oldval, newval)
   760  		}
   761  	}
   762  	print("runtime: castogscanstatus oldval=", hex(oldval), " newval=", hex(newval), "\n")
   763  	throw("castogscanstatus")
   764  	panic("not reached")
   765  }
   766  
   767  // If asked to move to or from a Gscanstatus this will throw. Use the castogscanstatus
   768  // and casfrom_Gscanstatus instead.
   769  // casgstatus will loop if the g->atomicstatus is in a Gscan status until the routine that
   770  // put it in the Gscan state is finished.
   771  //go:nosplit
   772  func casgstatus(gp *g, oldval, newval uint32) {
   773  	if (oldval&_Gscan != 0) || (newval&_Gscan != 0) || oldval == newval {
   774  		systemstack(func() {
   775  			print("runtime: casgstatus: oldval=", hex(oldval), " newval=", hex(newval), "\n")
   776  			throw("casgstatus: bad incoming values")
   777  		})
   778  	}
   779  
   780  	if oldval == _Grunning && gp.gcscanvalid {
   781  		// If oldvall == _Grunning, then the actual status must be
   782  		// _Grunning or _Grunning|_Gscan; either way,
   783  		// we own gp.gcscanvalid, so it's safe to read.
   784  		// gp.gcscanvalid must not be true when we are running.
   785  		systemstack(func() {
   786  			print("runtime: casgstatus ", hex(oldval), "->", hex(newval), " gp.status=", hex(gp.atomicstatus), " gp.gcscanvalid=true\n")
   787  			throw("casgstatus")
   788  		})
   789  	}
   790  
   791  	// See http://golang.org/cl/21503 for justification of the yield delay.
   792  	const yieldDelay = 5 * 1000
   793  	var nextYield int64
   794  
   795  	// loop if gp->atomicstatus is in a scan state giving
   796  	// GC time to finish and change the state to oldval.
   797  	for i := 0; !atomic.Cas(&gp.atomicstatus, oldval, newval); i++ {
   798  		if oldval == _Gwaiting && gp.atomicstatus == _Grunnable {
   799  			systemstack(func() {
   800  				throw("casgstatus: waiting for Gwaiting but is Grunnable")
   801  			})
   802  		}
   803  		// Help GC if needed.
   804  		// if gp.preemptscan && !gp.gcworkdone && (oldval == _Grunning || oldval == _Gsyscall) {
   805  		// 	gp.preemptscan = false
   806  		// 	systemstack(func() {
   807  		// 		gcphasework(gp)
   808  		// 	})
   809  		// }
   810  		// But meanwhile just yield.
   811  		if i == 0 {
   812  			nextYield = nanotime() + yieldDelay
   813  		}
   814  		if nanotime() < nextYield {
   815  			for x := 0; x < 10 && gp.atomicstatus != oldval; x++ {
   816  				procyield(1)
   817  			}
   818  		} else {
   819  			osyield()
   820  			nextYield = nanotime() + yieldDelay/2
   821  		}
   822  	}
   823  	if newval == _Grunning {
   824  		gp.gcscanvalid = false
   825  	}
   826  }
   827  
   828  // casgstatus(gp, oldstatus, Gcopystack), assuming oldstatus is Gwaiting or Grunnable.
   829  // Returns old status. Cannot call casgstatus directly, because we are racing with an
   830  // async wakeup that might come in from netpoll. If we see Gwaiting from the readgstatus,
   831  // it might have become Grunnable by the time we get to the cas. If we called casgstatus,
   832  // it would loop waiting for the status to go back to Gwaiting, which it never will.
   833  //go:nosplit
   834  func casgcopystack(gp *g) uint32 {
   835  	for {
   836  		oldstatus := readgstatus(gp) &^ _Gscan
   837  		if oldstatus != _Gwaiting && oldstatus != _Grunnable {
   838  			throw("copystack: bad status, not Gwaiting or Grunnable")
   839  		}
   840  		if atomic.Cas(&gp.atomicstatus, oldstatus, _Gcopystack) {
   841  			return oldstatus
   842  		}
   843  	}
   844  }
   845  
   846  // scang blocks until gp's stack has been scanned.
   847  // It might be scanned by scang or it might be scanned by the goroutine itself.
   848  // Either way, the stack scan has completed when scang returns.
   849  func scang(gp *g, gcw *gcWork) {
   850  	// Invariant; we (the caller, markroot for a specific goroutine) own gp.gcscandone.
   851  	// Nothing is racing with us now, but gcscandone might be set to true left over
   852  	// from an earlier round of stack scanning (we scan twice per GC).
   853  	// We use gcscandone to record whether the scan has been done during this round.
   854  
   855  	gp.gcscandone = false
   856  
   857  	// See http://golang.org/cl/21503 for justification of the yield delay.
   858  	const yieldDelay = 10 * 1000
   859  	var nextYield int64
   860  
   861  	// Endeavor to get gcscandone set to true,
   862  	// either by doing the stack scan ourselves or by coercing gp to scan itself.
   863  	// gp.gcscandone can transition from false to true when we're not looking
   864  	// (if we asked for preemption), so any time we lock the status using
   865  	// castogscanstatus we have to double-check that the scan is still not done.
   866  loop:
   867  	for i := 0; !gp.gcscandone; i++ {
   868  		switch s := readgstatus(gp); s {
   869  		default:
   870  			dumpgstatus(gp)
   871  			throw("stopg: invalid status")
   872  
   873  		case _Gdead:
   874  			// No stack.
   875  			gp.gcscandone = true
   876  			break loop
   877  
   878  		case _Gcopystack:
   879  		// Stack being switched. Go around again.
   880  
   881  		case _Grunnable, _Gsyscall, _Gwaiting:
   882  			// Claim goroutine by setting scan bit.
   883  			// Racing with execution or readying of gp.
   884  			// The scan bit keeps them from running
   885  			// the goroutine until we're done.
   886  			if castogscanstatus(gp, s, s|_Gscan) {
   887  				if !gp.gcscandone {
   888  					scanstack(gp, gcw)
   889  					gp.gcscandone = true
   890  				}
   891  				restartg(gp)
   892  				break loop
   893  			}
   894  
   895  		case _Gscanwaiting:
   896  		// newstack is doing a scan for us right now. Wait.
   897  
   898  		case _Grunning:
   899  			// Goroutine running. Try to preempt execution so it can scan itself.
   900  			// The preemption handler (in newstack) does the actual scan.
   901  
   902  			// Optimization: if there is already a pending preemption request
   903  			// (from the previous loop iteration), don't bother with the atomics.
   904  			if gp.preemptscan && gp.preempt && gp.stackguard0 == stackPreempt {
   905  				break
   906  			}
   907  
   908  			// Ask for preemption and self scan.
   909  			if castogscanstatus(gp, _Grunning, _Gscanrunning) {
   910  				if !gp.gcscandone {
   911  					gp.preemptscan = true
   912  					gp.preempt = true
   913  					gp.stackguard0 = stackPreempt
   914  				}
   915  				casfrom_Gscanstatus(gp, _Gscanrunning, _Grunning)
   916  			}
   917  		}
   918  
   919  		if i == 0 {
   920  			nextYield = nanotime() + yieldDelay
   921  		}
   922  		if nanotime() < nextYield {
   923  			procyield(10)
   924  		} else {
   925  			osyield()
   926  			nextYield = nanotime() + yieldDelay/2
   927  		}
   928  	}
   929  
   930  	gp.preemptscan = false // cancel scan request if no longer needed
   931  }
   932  
   933  // The GC requests that this routine be moved from a scanmumble state to a mumble state.
   934  func restartg(gp *g) {
   935  	s := readgstatus(gp)
   936  	switch s {
   937  	default:
   938  		dumpgstatus(gp)
   939  		throw("restartg: unexpected status")
   940  
   941  	case _Gdead:
   942  	// ok
   943  
   944  	case _Gscanrunnable,
   945  		_Gscanwaiting,
   946  		_Gscansyscall:
   947  		casfrom_Gscanstatus(gp, s, s&^_Gscan)
   948  	}
   949  }
   950  
   951  // stopTheWorld stops all P's from executing goroutines, interrupting
   952  // all goroutines at GC safe points and records reason as the reason
   953  // for the stop. On return, only the current goroutine's P is running.
   954  // stopTheWorld must not be called from a system stack and the caller
   955  // must not hold worldsema. The caller must call startTheWorld when
   956  // other P's should resume execution.
   957  //
   958  // stopTheWorld is safe for multiple goroutines to call at the
   959  // same time. Each will execute its own stop, and the stops will
   960  // be serialized.
   961  //
   962  // This is also used by routines that do stack dumps. If the system is
   963  // in panic or being exited, this may not reliably stop all
   964  // goroutines.
   965  func stopTheWorld(reason string) {
   966  	semacquire(&worldsema)
   967  	getg().m.preemptoff = reason
   968  	systemstack(stopTheWorldWithSema)
   969  }
   970  
   971  // startTheWorld undoes the effects of stopTheWorld.
   972  func startTheWorld() {
   973  	systemstack(func() { startTheWorldWithSema(false) })
   974  	// worldsema must be held over startTheWorldWithSema to ensure
   975  	// gomaxprocs cannot change while worldsema is held.
   976  	semrelease(&worldsema)
   977  	getg().m.preemptoff = ""
   978  }
   979  
   980  // Holding worldsema grants an M the right to try to stop the world
   981  // and prevents gomaxprocs from changing concurrently.
   982  var worldsema uint32 = 1
   983  
   984  // stopTheWorldWithSema is the core implementation of stopTheWorld.
   985  // The caller is responsible for acquiring worldsema and disabling
   986  // preemption first and then should stopTheWorldWithSema on the system
   987  // stack:
   988  //
   989  //	semacquire(&worldsema, 0)
   990  //	m.preemptoff = "reason"
   991  //	systemstack(stopTheWorldWithSema)
   992  //
   993  // When finished, the caller must either call startTheWorld or undo
   994  // these three operations separately:
   995  //
   996  //	m.preemptoff = ""
   997  //	systemstack(startTheWorldWithSema)
   998  //	semrelease(&worldsema)
   999  //
  1000  // It is allowed to acquire worldsema once and then execute multiple
  1001  // startTheWorldWithSema/stopTheWorldWithSema pairs.
  1002  // Other P's are able to execute between successive calls to
  1003  // startTheWorldWithSema and stopTheWorldWithSema.
  1004  // Holding worldsema causes any other goroutines invoking
  1005  // stopTheWorld to block.
  1006  func stopTheWorldWithSema() {
  1007  	_g_ := getg()
  1008  
  1009  	// If we hold a lock, then we won't be able to stop another M
  1010  	// that is blocked trying to acquire the lock.
  1011  	if _g_.m.locks > 0 {
  1012  		throw("stopTheWorld: holding locks")
  1013  	}
  1014  
  1015  	lock(&sched.lock)
  1016  	sched.stopwait = gomaxprocs
  1017  	atomic.Store(&sched.gcwaiting, 1)
  1018  	preemptall()
  1019  	// stop current P
  1020  	_g_.m.p.ptr().status = _Pgcstop // Pgcstop is only diagnostic.
  1021  	sched.stopwait--
  1022  	// try to retake all P's in Psyscall status
  1023  	for _, p := range allp {
  1024  		s := p.status
  1025  		if s == _Psyscall && atomic.Cas(&p.status, s, _Pgcstop) {
  1026  			if trace.enabled {
  1027  				traceGoSysBlock(p)
  1028  				traceProcStop(p)
  1029  			}
  1030  			p.syscalltick++
  1031  			sched.stopwait--
  1032  		}
  1033  	}
  1034  	// stop idle P's
  1035  	for {
  1036  		p := pidleget()
  1037  		if p == nil {
  1038  			break
  1039  		}
  1040  		p.status = _Pgcstop
  1041  		sched.stopwait--
  1042  	}
  1043  	wait := sched.stopwait > 0
  1044  	unlock(&sched.lock)
  1045  
  1046  	// wait for remaining P's to stop voluntarily
  1047  	if wait {
  1048  		for {
  1049  			// wait for 100us, then try to re-preempt in case of any races
  1050  			if notetsleep(&sched.stopnote, 100*1000) {
  1051  				noteclear(&sched.stopnote)
  1052  				break
  1053  			}
  1054  			preemptall()
  1055  		}
  1056  	}
  1057  
  1058  	// sanity checks
  1059  	bad := ""
  1060  	if sched.stopwait != 0 {
  1061  		bad = "stopTheWorld: not stopped (stopwait != 0)"
  1062  	} else {
  1063  		for _, p := range allp {
  1064  			if p.status != _Pgcstop {
  1065  				bad = "stopTheWorld: not stopped (status != _Pgcstop)"
  1066  			}
  1067  		}
  1068  	}
  1069  	if atomic.Load(&freezing) != 0 {
  1070  		// Some other thread is panicking. This can cause the
  1071  		// sanity checks above to fail if the panic happens in
  1072  		// the signal handler on a stopped thread. Either way,
  1073  		// we should halt this thread.
  1074  		lock(&deadlock)
  1075  		lock(&deadlock)
  1076  	}
  1077  	if bad != "" {
  1078  		throw(bad)
  1079  	}
  1080  }
  1081  
  1082  func mhelpgc() {
  1083  	_g_ := getg()
  1084  	_g_.m.helpgc = -1
  1085  }
  1086  
  1087  func startTheWorldWithSema(emitTraceEvent bool) int64 {
  1088  	_g_ := getg()
  1089  
  1090  	_g_.m.locks++ // disable preemption because it can be holding p in a local var
  1091  	if netpollinited() {
  1092  		gp := netpoll(false) // non-blocking
  1093  		injectglist(gp)
  1094  	}
  1095  	add := needaddgcproc()
  1096  	lock(&sched.lock)
  1097  
  1098  	procs := gomaxprocs
  1099  	if newprocs != 0 {
  1100  		procs = newprocs
  1101  		newprocs = 0
  1102  	}
  1103  	p1 := procresize(procs)
  1104  	sched.gcwaiting = 0
  1105  	if sched.sysmonwait != 0 {
  1106  		sched.sysmonwait = 0
  1107  		notewakeup(&sched.sysmonnote)
  1108  	}
  1109  	unlock(&sched.lock)
  1110  
  1111  	for p1 != nil {
  1112  		p := p1
  1113  		p1 = p1.link.ptr()
  1114  		if p.m != 0 {
  1115  			mp := p.m.ptr()
  1116  			p.m = 0
  1117  			if mp.nextp != 0 {
  1118  				throw("startTheWorld: inconsistent mp->nextp")
  1119  			}
  1120  			mp.nextp.set(p)
  1121  			notewakeup(&mp.park)
  1122  		} else {
  1123  			// Start M to run P.  Do not start another M below.
  1124  			newm(nil, p)
  1125  			add = false
  1126  		}
  1127  	}
  1128  
  1129  	// Capture start-the-world time before doing clean-up tasks.
  1130  	startTime := nanotime()
  1131  	if emitTraceEvent {
  1132  		traceGCSTWDone()
  1133  	}
  1134  
  1135  	// Wakeup an additional proc in case we have excessive runnable goroutines
  1136  	// in local queues or in the global queue. If we don't, the proc will park itself.
  1137  	// If we have lots of excessive work, resetspinning will unpark additional procs as necessary.
  1138  	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 {
  1139  		wakep()
  1140  	}
  1141  
  1142  	if add {
  1143  		// If GC could have used another helper proc, start one now,
  1144  		// in the hope that it will be available next time.
  1145  		// It would have been even better to start it before the collection,
  1146  		// but doing so requires allocating memory, so it's tricky to
  1147  		// coordinate. This lazy approach works out in practice:
  1148  		// we don't mind if the first couple gc rounds don't have quite
  1149  		// the maximum number of procs.
  1150  		newm(mhelpgc, nil)
  1151  	}
  1152  	_g_.m.locks--
  1153  	if _g_.m.locks == 0 && _g_.preempt { // restore the preemption request in case we've cleared it in newstack
  1154  		_g_.stackguard0 = stackPreempt
  1155  	}
  1156  
  1157  	return startTime
  1158  }
  1159  
  1160  // Called to start an M.
  1161  //
  1162  // This must not split the stack because we may not even have stack
  1163  // bounds set up yet.
  1164  //
  1165  // May run during STW (because it doesn't have a P yet), so write
  1166  // barriers are not allowed.
  1167  //
  1168  //go:nosplit
  1169  //go:nowritebarrierrec
  1170  func mstart() {
  1171  	_g_ := getg()
  1172  
  1173  	osStack := _g_.stack.lo == 0
  1174  	if osStack {
  1175  		// Initialize stack bounds from system stack.
  1176  		// Cgo may have left stack size in stack.hi.
  1177  		size := _g_.stack.hi
  1178  		if size == 0 {
  1179  			size = 8192 * sys.StackGuardMultiplier
  1180  		}
  1181  		_g_.stack.hi = uintptr(noescape(unsafe.Pointer(&size)))
  1182  		_g_.stack.lo = _g_.stack.hi - size + 1024
  1183  	}
  1184  	// Initialize stack guards so that we can start calling
  1185  	// both Go and C functions with stack growth prologues.
  1186  	_g_.stackguard0 = _g_.stack.lo + _StackGuard
  1187  	_g_.stackguard1 = _g_.stackguard0
  1188  	mstart1(0)
  1189  
  1190  	// Exit this thread.
  1191  	if GOOS == "windows" || GOOS == "solaris" || GOOS == "plan9" {
  1192  		// Window, Solaris and Plan 9 always system-allocate
  1193  		// the stack, but put it in _g_.stack before mstart,
  1194  		// so the logic above hasn't set osStack yet.
  1195  		osStack = true
  1196  	}
  1197  	mexit(osStack)
  1198  }
  1199  
  1200  func mstart1(dummy int32) {
  1201  	_g_ := getg()
  1202  
  1203  	if _g_ != _g_.m.g0 {
  1204  		throw("bad runtime·mstart")
  1205  	}
  1206  
  1207  	// Record the caller for use as the top of stack in mcall and
  1208  	// for terminating the thread.
  1209  	// We're never coming back to mstart1 after we call schedule,
  1210  	// so other calls can reuse the current frame.
  1211  	save(getcallerpc(), getcallersp(unsafe.Pointer(&dummy)))
  1212  	asminit()
  1213  	minit()
  1214  
  1215  	// Install signal handlers; after minit so that minit can
  1216  	// prepare the thread to be able to handle the signals.
  1217  	if _g_.m == &m0 {
  1218  		mstartm0()
  1219  	}
  1220  
  1221  	if fn := _g_.m.mstartfn; fn != nil {
  1222  		fn()
  1223  	}
  1224  
  1225  	if _g_.m.helpgc != 0 {
  1226  		_g_.m.helpgc = 0
  1227  		stopm()
  1228  	} else if _g_.m != &m0 {
  1229  		acquirep(_g_.m.nextp.ptr())
  1230  		_g_.m.nextp = 0
  1231  	}
  1232  	schedule()
  1233  }
  1234  
  1235  // mstartm0 implements part of mstart1 that only runs on the m0.
  1236  //
  1237  // Write barriers are allowed here because we know the GC can't be
  1238  // running yet, so they'll be no-ops.
  1239  //
  1240  //go:yeswritebarrierrec
  1241  func mstartm0() {
  1242  	// Create an extra M for callbacks on threads not created by Go.
  1243  	if iscgo && !cgoHasExtraM {
  1244  		cgoHasExtraM = true
  1245  		newextram()
  1246  	}
  1247  	initsig(false)
  1248  }
  1249  
  1250  // mexit tears down and exits the current thread.
  1251  //
  1252  // Don't call this directly to exit the thread, since it must run at
  1253  // the top of the thread stack. Instead, use gogo(&_g_.m.g0.sched) to
  1254  // unwind the stack to the point that exits the thread.
  1255  //
  1256  // It is entered with m.p != nil, so write barriers are allowed. It
  1257  // will release the P before exiting.
  1258  //
  1259  //go:yeswritebarrierrec
  1260  func mexit(osStack bool) {
  1261  	g := getg()
  1262  	m := g.m
  1263  
  1264  	if m == &m0 {
  1265  		// This is the main thread. Just wedge it.
  1266  		//
  1267  		// On Linux, exiting the main thread puts the process
  1268  		// into a non-waitable zombie state. On Plan 9,
  1269  		// exiting the main thread unblocks wait even though
  1270  		// other threads are still running. On Solaris we can
  1271  		// neither exitThread nor return from mstart. Other
  1272  		// bad things probably happen on other platforms.
  1273  		//
  1274  		// We could try to clean up this M more before wedging
  1275  		// it, but that complicates signal handling.
  1276  		handoffp(releasep())
  1277  		lock(&sched.lock)
  1278  		sched.nmfreed++
  1279  		checkdead()
  1280  		unlock(&sched.lock)
  1281  		notesleep(&m.park)
  1282  		throw("locked m0 woke up")
  1283  	}
  1284  
  1285  	sigblock()
  1286  	unminit()
  1287  
  1288  	// Free the gsignal stack.
  1289  	if m.gsignal != nil {
  1290  		stackfree(m.gsignal.stack)
  1291  	}
  1292  
  1293  	// Remove m from allm.
  1294  	lock(&sched.lock)
  1295  	for pprev := &allm; *pprev != nil; pprev = &(*pprev).alllink {
  1296  		if *pprev == m {
  1297  			*pprev = m.alllink
  1298  			goto found
  1299  		}
  1300  	}
  1301  	throw("m not found in allm")
  1302  found:
  1303  	if !osStack {
  1304  		// Delay reaping m until it's done with the stack.
  1305  		//
  1306  		// If this is using an OS stack, the OS will free it
  1307  		// so there's no need for reaping.
  1308  		atomic.Store(&m.freeWait, 1)
  1309  		// Put m on the free list, though it will not be reaped until
  1310  		// freeWait is 0. Note that the free list must not be linked
  1311  		// through alllink because some functions walk allm without
  1312  		// locking, so may be using alllink.
  1313  		m.freelink = sched.freem
  1314  		sched.freem = m
  1315  	}
  1316  	unlock(&sched.lock)
  1317  
  1318  	// Release the P.
  1319  	handoffp(releasep())
  1320  	// After this point we must not have write barriers.
  1321  
  1322  	// Invoke the deadlock detector. This must happen after
  1323  	// handoffp because it may have started a new M to take our
  1324  	// P's work.
  1325  	lock(&sched.lock)
  1326  	sched.nmfreed++
  1327  	checkdead()
  1328  	unlock(&sched.lock)
  1329  
  1330  	if osStack {
  1331  		// Return from mstart and let the system thread
  1332  		// library free the g0 stack and terminate the thread.
  1333  		return
  1334  	}
  1335  
  1336  	// mstart is the thread's entry point, so there's nothing to
  1337  	// return to. Exit the thread directly. exitThread will clear
  1338  	// m.freeWait when it's done with the stack and the m can be
  1339  	// reaped.
  1340  	exitThread(&m.freeWait)
  1341  }
  1342  
  1343  // forEachP calls fn(p) for every P p when p reaches a GC safe point.
  1344  // If a P is currently executing code, this will bring the P to a GC
  1345  // safe point and execute fn on that P. If the P is not executing code
  1346  // (it is idle or in a syscall), this will call fn(p) directly while
  1347  // preventing the P from exiting its state. This does not ensure that
  1348  // fn will run on every CPU executing Go code, but it acts as a global
  1349  // memory barrier. GC uses this as a "ragged barrier."
  1350  //
  1351  // The caller must hold worldsema.
  1352  //
  1353  //go:systemstack
  1354  func forEachP(fn func(*p)) {
  1355  	mp := acquirem()
  1356  	_p_ := getg().m.p.ptr()
  1357  
  1358  	lock(&sched.lock)
  1359  	if sched.safePointWait != 0 {
  1360  		throw("forEachP: sched.safePointWait != 0")
  1361  	}
  1362  	sched.safePointWait = gomaxprocs - 1
  1363  	sched.safePointFn = fn
  1364  
  1365  	// Ask all Ps to run the safe point function.
  1366  	for _, p := range allp {
  1367  		if p != _p_ {
  1368  			atomic.Store(&p.runSafePointFn, 1)
  1369  		}
  1370  	}
  1371  	preemptall()
  1372  
  1373  	// Any P entering _Pidle or _Psyscall from now on will observe
  1374  	// p.runSafePointFn == 1 and will call runSafePointFn when
  1375  	// changing its status to _Pidle/_Psyscall.
  1376  
  1377  	// Run safe point function for all idle Ps. sched.pidle will
  1378  	// not change because we hold sched.lock.
  1379  	for p := sched.pidle.ptr(); p != nil; p = p.link.ptr() {
  1380  		if atomic.Cas(&p.runSafePointFn, 1, 0) {
  1381  			fn(p)
  1382  			sched.safePointWait--
  1383  		}
  1384  	}
  1385  
  1386  	wait := sched.safePointWait > 0
  1387  	unlock(&sched.lock)
  1388  
  1389  	// Run fn for the current P.
  1390  	fn(_p_)
  1391  
  1392  	// Force Ps currently in _Psyscall into _Pidle and hand them
  1393  	// off to induce safe point function execution.
  1394  	for _, p := range allp {
  1395  		s := p.status
  1396  		if s == _Psyscall && p.runSafePointFn == 1 && atomic.Cas(&p.status, s, _Pidle) {
  1397  			if trace.enabled {
  1398  				traceGoSysBlock(p)
  1399  				traceProcStop(p)
  1400  			}
  1401  			p.syscalltick++
  1402  			handoffp(p)
  1403  		}
  1404  	}
  1405  
  1406  	// Wait for remaining Ps to run fn.
  1407  	if wait {
  1408  		for {
  1409  			// Wait for 100us, then try to re-preempt in
  1410  			// case of any races.
  1411  			//
  1412  			// Requires system stack.
  1413  			if notetsleep(&sched.safePointNote, 100*1000) {
  1414  				noteclear(&sched.safePointNote)
  1415  				break
  1416  			}
  1417  			preemptall()
  1418  		}
  1419  	}
  1420  	if sched.safePointWait != 0 {
  1421  		throw("forEachP: not done")
  1422  	}
  1423  	for _, p := range allp {
  1424  		if p.runSafePointFn != 0 {
  1425  			throw("forEachP: P did not run fn")
  1426  		}
  1427  	}
  1428  
  1429  	lock(&sched.lock)
  1430  	sched.safePointFn = nil
  1431  	unlock(&sched.lock)
  1432  	releasem(mp)
  1433  }
  1434  
  1435  // runSafePointFn runs the safe point function, if any, for this P.
  1436  // This should be called like
  1437  //
  1438  //     if getg().m.p.runSafePointFn != 0 {
  1439  //         runSafePointFn()
  1440  //     }
  1441  //
  1442  // runSafePointFn must be checked on any transition in to _Pidle or
  1443  // _Psyscall to avoid a race where forEachP sees that the P is running
  1444  // just before the P goes into _Pidle/_Psyscall and neither forEachP
  1445  // nor the P run the safe-point function.
  1446  func runSafePointFn() {
  1447  	p := getg().m.p.ptr()
  1448  	// Resolve the race between forEachP running the safe-point
  1449  	// function on this P's behalf and this P running the
  1450  	// safe-point function directly.
  1451  	if !atomic.Cas(&p.runSafePointFn, 1, 0) {
  1452  		return
  1453  	}
  1454  	sched.safePointFn(p)
  1455  	lock(&sched.lock)
  1456  	sched.safePointWait--
  1457  	if sched.safePointWait == 0 {
  1458  		notewakeup(&sched.safePointNote)
  1459  	}
  1460  	unlock(&sched.lock)
  1461  }
  1462  
  1463  // When running with cgo, we call _cgo_thread_start
  1464  // to start threads for us so that we can play nicely with
  1465  // foreign code.
  1466  var cgoThreadStart unsafe.Pointer
  1467  
  1468  type cgothreadstart struct {
  1469  	g   guintptr
  1470  	tls *uint64
  1471  	fn  unsafe.Pointer
  1472  }
  1473  
  1474  // Allocate a new m unassociated with any thread.
  1475  // Can use p for allocation context if needed.
  1476  // fn is recorded as the new m's m.mstartfn.
  1477  //
  1478  // This function is allowed to have write barriers even if the caller
  1479  // isn't because it borrows _p_.
  1480  //
  1481  //go:yeswritebarrierrec
  1482  func allocm(_p_ *p, fn func()) *m {
  1483  	_g_ := getg()
  1484  	_g_.m.locks++ // disable GC because it can be called from sysmon
  1485  	if _g_.m.p == 0 {
  1486  		acquirep(_p_) // temporarily borrow p for mallocs in this function
  1487  	}
  1488  
  1489  	// Release the free M list. We need to do this somewhere and
  1490  	// this may free up a stack we can use.
  1491  	if sched.freem != nil {
  1492  		lock(&sched.lock)
  1493  		var newList *m
  1494  		for freem := sched.freem; freem != nil; {
  1495  			if freem.freeWait != 0 {
  1496  				next := freem.freelink
  1497  				freem.freelink = newList
  1498  				newList = freem
  1499  				freem = next
  1500  				continue
  1501  			}
  1502  			stackfree(freem.g0.stack)
  1503  			freem = freem.freelink
  1504  		}
  1505  		sched.freem = newList
  1506  		unlock(&sched.lock)
  1507  	}
  1508  
  1509  	mp := new(m)
  1510  	mp.mstartfn = fn
  1511  	mcommoninit(mp)
  1512  
  1513  	// In case of cgo or Solaris, pthread_create will make us a stack.
  1514  	// Windows and Plan 9 will layout sched stack on OS stack.
  1515  	if iscgo || GOOS == "solaris" || GOOS == "windows" || GOOS == "plan9" {
  1516  		mp.g0 = malg(-1)
  1517  	} else {
  1518  		mp.g0 = malg(8192 * sys.StackGuardMultiplier)
  1519  	}
  1520  	mp.g0.m = mp
  1521  
  1522  	if _p_ == _g_.m.p.ptr() {
  1523  		releasep()
  1524  	}
  1525  	_g_.m.locks--
  1526  	if _g_.m.locks == 0 && _g_.preempt { // restore the preemption request in case we've cleared it in newstack
  1527  		_g_.stackguard0 = stackPreempt
  1528  	}
  1529  
  1530  	return mp
  1531  }
  1532  
  1533  // needm is called when a cgo callback happens on a
  1534  // thread without an m (a thread not created by Go).
  1535  // In this case, needm is expected to find an m to use
  1536  // and return with m, g initialized correctly.
  1537  // Since m and g are not set now (likely nil, but see below)
  1538  // needm is limited in what routines it can call. In particular
  1539  // it can only call nosplit functions (textflag 7) and cannot
  1540  // do any scheduling that requires an m.
  1541  //
  1542  // In order to avoid needing heavy lifting here, we adopt
  1543  // the following strategy: there is a stack of available m's
  1544  // that can be stolen. Using compare-and-swap
  1545  // to pop from the stack has ABA races, so we simulate
  1546  // a lock by doing an exchange (via casp) to steal the stack
  1547  // head and replace the top pointer with MLOCKED (1).
  1548  // This serves as a simple spin lock that we can use even
  1549  // without an m. The thread that locks the stack in this way
  1550  // unlocks the stack by storing a valid stack head pointer.
  1551  //
  1552  // In order to make sure that there is always an m structure
  1553  // available to be stolen, we maintain the invariant that there
  1554  // is always one more than needed. At the beginning of the
  1555  // program (if cgo is in use) the list is seeded with a single m.
  1556  // If needm finds that it has taken the last m off the list, its job
  1557  // is - once it has installed its own m so that it can do things like
  1558  // allocate memory - to create a spare m and put it on the list.
  1559  //
  1560  // Each of these extra m's also has a g0 and a curg that are
  1561  // pressed into service as the scheduling stack and current
  1562  // goroutine for the duration of the cgo callback.
  1563  //
  1564  // When the callback is done with the m, it calls dropm to
  1565  // put the m back on the list.
  1566  //go:nosplit
  1567  func needm(x byte) {
  1568  	if iscgo && !cgoHasExtraM {
  1569  		// Can happen if C/C++ code calls Go from a global ctor.
  1570  		// Can not throw, because scheduler is not initialized yet.
  1571  		write(2, unsafe.Pointer(&earlycgocallback[0]), int32(len(earlycgocallback)))
  1572  		exit(1)
  1573  	}
  1574  
  1575  	// Lock extra list, take head, unlock popped list.
  1576  	// nilokay=false is safe here because of the invariant above,
  1577  	// that the extra list always contains or will soon contain
  1578  	// at least one m.
  1579  	mp := lockextra(false)
  1580  
  1581  	// Set needextram when we've just emptied the list,
  1582  	// so that the eventual call into cgocallbackg will
  1583  	// allocate a new m for the extra list. We delay the
  1584  	// allocation until then so that it can be done
  1585  	// after exitsyscall makes sure it is okay to be
  1586  	// running at all (that is, there's no garbage collection
  1587  	// running right now).
  1588  	mp.needextram = mp.schedlink == 0
  1589  	extraMCount--
  1590  	unlockextra(mp.schedlink.ptr())
  1591  
  1592  	// Save and block signals before installing g.
  1593  	// Once g is installed, any incoming signals will try to execute,
  1594  	// but we won't have the sigaltstack settings and other data
  1595  	// set up appropriately until the end of minit, which will
  1596  	// unblock the signals. This is the same dance as when
  1597  	// starting a new m to run Go code via newosproc.
  1598  	msigsave(mp)
  1599  	sigblock()
  1600  
  1601  	// Install g (= m->g0) and set the stack bounds
  1602  	// to match the current stack. We don't actually know
  1603  	// how big the stack is, like we don't know how big any
  1604  	// scheduling stack is, but we assume there's at least 32 kB,
  1605  	// which is more than enough for us.
  1606  	setg(mp.g0)
  1607  	_g_ := getg()
  1608  	_g_.stack.hi = uintptr(noescape(unsafe.Pointer(&x))) + 1024
  1609  	_g_.stack.lo = uintptr(noescape(unsafe.Pointer(&x))) - 32*1024
  1610  	_g_.stackguard0 = _g_.stack.lo + _StackGuard
  1611  
  1612  	// Initialize this thread to use the m.
  1613  	asminit()
  1614  	minit()
  1615  
  1616  	// mp.curg is now a real goroutine.
  1617  	casgstatus(mp.curg, _Gdead, _Gsyscall)
  1618  	atomic.Xadd(&sched.ngsys, -1)
  1619  }
  1620  
  1621  var earlycgocallback = []byte("fatal error: cgo callback before cgo call\n")
  1622  
  1623  // newextram allocates m's and puts them on the extra list.
  1624  // It is called with a working local m, so that it can do things
  1625  // like call schedlock and allocate.
  1626  func newextram() {
  1627  	c := atomic.Xchg(&extraMWaiters, 0)
  1628  	if c > 0 {
  1629  		for i := uint32(0); i < c; i++ {
  1630  			oneNewExtraM()
  1631  		}
  1632  	} else {
  1633  		// Make sure there is at least one extra M.
  1634  		mp := lockextra(true)
  1635  		unlockextra(mp)
  1636  		if mp == nil {
  1637  			oneNewExtraM()
  1638  		}
  1639  	}
  1640  }
  1641  
  1642  // oneNewExtraM allocates an m and puts it on the extra list.
  1643  func oneNewExtraM() {
  1644  	// Create extra goroutine locked to extra m.
  1645  	// The goroutine is the context in which the cgo callback will run.
  1646  	// The sched.pc will never be returned to, but setting it to
  1647  	// goexit makes clear to the traceback routines where
  1648  	// the goroutine stack ends.
  1649  	mp := allocm(nil, nil)
  1650  	gp := malg(4096)
  1651  	gp.sched.pc = funcPC(goexit) + sys.PCQuantum
  1652  	gp.sched.sp = gp.stack.hi
  1653  	gp.sched.sp -= 4 * sys.RegSize // extra space in case of reads slightly beyond frame
  1654  	gp.sched.lr = 0
  1655  	gp.sched.g = guintptr(unsafe.Pointer(gp))
  1656  	gp.syscallpc = gp.sched.pc
  1657  	gp.syscallsp = gp.sched.sp
  1658  	gp.stktopsp = gp.sched.sp
  1659  	gp.gcscanvalid = true
  1660  	gp.gcscandone = true
  1661  	// malg returns status as _Gidle. Change to _Gdead before
  1662  	// adding to allg where GC can see it. We use _Gdead to hide
  1663  	// this from tracebacks and stack scans since it isn't a
  1664  	// "real" goroutine until needm grabs it.
  1665  	casgstatus(gp, _Gidle, _Gdead)
  1666  	gp.m = mp
  1667  	mp.curg = gp
  1668  	mp.lockedInt++
  1669  	mp.lockedg.set(gp)
  1670  	gp.lockedm.set(mp)
  1671  	gp.goid = int64(atomic.Xadd64(&sched.goidgen, 1))
  1672  	if raceenabled {
  1673  		gp.racectx = racegostart(funcPC(newextram) + sys.PCQuantum)
  1674  	}
  1675  	// put on allg for garbage collector
  1676  	allgadd(gp)
  1677  
  1678  	// gp is now on the allg list, but we don't want it to be
  1679  	// counted by gcount. It would be more "proper" to increment
  1680  	// sched.ngfree, but that requires locking. Incrementing ngsys
  1681  	// has the same effect.
  1682  	atomic.Xadd(&sched.ngsys, +1)
  1683  
  1684  	// Add m to the extra list.
  1685  	mnext := lockextra(true)
  1686  	mp.schedlink.set(mnext)
  1687  	extraMCount++
  1688  	unlockextra(mp)
  1689  }
  1690  
  1691  // dropm is called when a cgo callback has called needm but is now
  1692  // done with the callback and returning back into the non-Go thread.
  1693  // It puts the current m back onto the extra list.
  1694  //
  1695  // The main expense here is the call to signalstack to release the
  1696  // m's signal stack, and then the call to needm on the next callback
  1697  // from this thread. It is tempting to try to save the m for next time,
  1698  // which would eliminate both these costs, but there might not be
  1699  // a next time: the current thread (which Go does not control) might exit.
  1700  // If we saved the m for that thread, there would be an m leak each time
  1701  // such a thread exited. Instead, we acquire and release an m on each
  1702  // call. These should typically not be scheduling operations, just a few
  1703  // atomics, so the cost should be small.
  1704  //
  1705  // TODO(rsc): An alternative would be to allocate a dummy pthread per-thread
  1706  // variable using pthread_key_create. Unlike the pthread keys we already use
  1707  // on OS X, this dummy key would never be read by Go code. It would exist
  1708  // only so that we could register at thread-exit-time destructor.
  1709  // That destructor would put the m back onto the extra list.
  1710  // This is purely a performance optimization. The current version,
  1711  // in which dropm happens on each cgo call, is still correct too.
  1712  // We may have to keep the current version on systems with cgo
  1713  // but without pthreads, like Windows.
  1714  func dropm() {
  1715  	// Clear m and g, and return m to the extra list.
  1716  	// After the call to setg we can only call nosplit functions
  1717  	// with no pointer manipulation.
  1718  	mp := getg().m
  1719  
  1720  	// Return mp.curg to dead state.
  1721  	casgstatus(mp.curg, _Gsyscall, _Gdead)
  1722  	atomic.Xadd(&sched.ngsys, +1)
  1723  
  1724  	// Block signals before unminit.
  1725  	// Unminit unregisters the signal handling stack (but needs g on some systems).
  1726  	// Setg(nil) clears g, which is the signal handler's cue not to run Go handlers.
  1727  	// It's important not to try to handle a signal between those two steps.
  1728  	sigmask := mp.sigmask
  1729  	sigblock()
  1730  	unminit()
  1731  
  1732  	mnext := lockextra(true)
  1733  	extraMCount++
  1734  	mp.schedlink.set(mnext)
  1735  
  1736  	setg(nil)
  1737  
  1738  	// Commit the release of mp.
  1739  	unlockextra(mp)
  1740  
  1741  	msigrestore(sigmask)
  1742  }
  1743  
  1744  // A helper function for EnsureDropM.
  1745  func getm() uintptr {
  1746  	return uintptr(unsafe.Pointer(getg().m))
  1747  }
  1748  
  1749  var extram uintptr
  1750  var extraMCount uint32 // Protected by lockextra
  1751  var extraMWaiters uint32
  1752  
  1753  // lockextra locks the extra list and returns the list head.
  1754  // The caller must unlock the list by storing a new list head
  1755  // to extram. If nilokay is true, then lockextra will
  1756  // return a nil list head if that's what it finds. If nilokay is false,
  1757  // lockextra will keep waiting until the list head is no longer nil.
  1758  //go:nosplit
  1759  func lockextra(nilokay bool) *m {
  1760  	const locked = 1
  1761  
  1762  	incr := false
  1763  	for {
  1764  		old := atomic.Loaduintptr(&extram)
  1765  		if old == locked {
  1766  			yield := osyield
  1767  			yield()
  1768  			continue
  1769  		}
  1770  		if old == 0 && !nilokay {
  1771  			if !incr {
  1772  				// Add 1 to the number of threads
  1773  				// waiting for an M.
  1774  				// This is cleared by newextram.
  1775  				atomic.Xadd(&extraMWaiters, 1)
  1776  				incr = true
  1777  			}
  1778  			usleep(1)
  1779  			continue
  1780  		}
  1781  		if atomic.Casuintptr(&extram, old, locked) {
  1782  			return (*m)(unsafe.Pointer(old))
  1783  		}
  1784  		yield := osyield
  1785  		yield()
  1786  		continue
  1787  	}
  1788  }
  1789  
  1790  //go:nosplit
  1791  func unlockextra(mp *m) {
  1792  	atomic.Storeuintptr(&extram, uintptr(unsafe.Pointer(mp)))
  1793  }
  1794  
  1795  // execLock serializes exec and clone to avoid bugs or unspecified behaviour
  1796  // around exec'ing while creating/destroying threads.  See issue #19546.
  1797  var execLock rwmutex
  1798  
  1799  // newmHandoff contains a list of m structures that need new OS threads.
  1800  // This is used by newm in situations where newm itself can't safely
  1801  // start an OS thread.
  1802  var newmHandoff struct {
  1803  	lock mutex
  1804  
  1805  	// newm points to a list of M structures that need new OS
  1806  	// threads. The list is linked through m.schedlink.
  1807  	newm muintptr
  1808  
  1809  	// waiting indicates that wake needs to be notified when an m
  1810  	// is put on the list.
  1811  	waiting bool
  1812  	wake    note
  1813  
  1814  	// haveTemplateThread indicates that the templateThread has
  1815  	// been started. This is not protected by lock. Use cas to set
  1816  	// to 1.
  1817  	haveTemplateThread uint32
  1818  }
  1819  
  1820  // Create a new m. It will start off with a call to fn, or else the scheduler.
  1821  // fn needs to be static and not a heap allocated closure.
  1822  // May run with m.p==nil, so write barriers are not allowed.
  1823  //go:nowritebarrierrec
  1824  func newm(fn func(), _p_ *p) {
  1825  	mp := allocm(_p_, fn)
  1826  	mp.nextp.set(_p_)
  1827  	mp.sigmask = initSigmask
  1828  	if gp := getg(); gp != nil && gp.m != nil && (gp.m.lockedExt != 0 || gp.m.incgo) && GOOS != "plan9" {
  1829  		// We're on a locked M or a thread that may have been
  1830  		// started by C. The kernel state of this thread may
  1831  		// be strange (the user may have locked it for that
  1832  		// purpose). We don't want to clone that into another
  1833  		// thread. Instead, ask a known-good thread to create
  1834  		// the thread for us.
  1835  		//
  1836  		// This is disabled on Plan 9. See golang.org/issue/22227.
  1837  		//
  1838  		// TODO: This may be unnecessary on Windows, which
  1839  		// doesn't model thread creation off fork.
  1840  		lock(&newmHandoff.lock)
  1841  		if newmHandoff.haveTemplateThread == 0 {
  1842  			throw("on a locked thread with no template thread")
  1843  		}
  1844  		mp.schedlink = newmHandoff.newm
  1845  		newmHandoff.newm.set(mp)
  1846  		if newmHandoff.waiting {
  1847  			newmHandoff.waiting = false
  1848  			notewakeup(&newmHandoff.wake)
  1849  		}
  1850  		unlock(&newmHandoff.lock)
  1851  		return
  1852  	}
  1853  	newm1(mp)
  1854  }
  1855  
  1856  func newm1(mp *m) {
  1857  	if iscgo {
  1858  		var ts cgothreadstart
  1859  		if _cgo_thread_start == nil {
  1860  			throw("_cgo_thread_start missing")
  1861  		}
  1862  		ts.g.set(mp.g0)
  1863  		ts.tls = (*uint64)(unsafe.Pointer(&mp.tls[0]))
  1864  		ts.fn = unsafe.Pointer(funcPC(mstart))
  1865  		if msanenabled {
  1866  			msanwrite(unsafe.Pointer(&ts), unsafe.Sizeof(ts))
  1867  		}
  1868  		execLock.rlock() // Prevent process clone.
  1869  		asmcgocall(_cgo_thread_start, unsafe.Pointer(&ts))
  1870  		execLock.runlock()
  1871  		return
  1872  	}
  1873  	execLock.rlock() // Prevent process clone.
  1874  	newosproc(mp, unsafe.Pointer(mp.g0.stack.hi))
  1875  	execLock.runlock()
  1876  }
  1877  
  1878  // startTemplateThread starts the template thread if it is not already
  1879  // running.
  1880  //
  1881  // The calling thread must itself be in a known-good state.
  1882  func startTemplateThread() {
  1883  	if !atomic.Cas(&newmHandoff.haveTemplateThread, 0, 1) {
  1884  		return
  1885  	}
  1886  	newm(templateThread, nil)
  1887  }
  1888  
  1889  // tmeplateThread is a thread in a known-good state that exists solely
  1890  // to start new threads in known-good states when the calling thread
  1891  // may not be a a good state.
  1892  //
  1893  // Many programs never need this, so templateThread is started lazily
  1894  // when we first enter a state that might lead to running on a thread
  1895  // in an unknown state.
  1896  //
  1897  // templateThread runs on an M without a P, so it must not have write
  1898  // barriers.
  1899  //
  1900  //go:nowritebarrierrec
  1901  func templateThread() {
  1902  	lock(&sched.lock)
  1903  	sched.nmsys++
  1904  	checkdead()
  1905  	unlock(&sched.lock)
  1906  
  1907  	for {
  1908  		lock(&newmHandoff.lock)
  1909  		for newmHandoff.newm != 0 {
  1910  			newm := newmHandoff.newm.ptr()
  1911  			newmHandoff.newm = 0
  1912  			unlock(&newmHandoff.lock)
  1913  			for newm != nil {
  1914  				next := newm.schedlink.ptr()
  1915  				newm.schedlink = 0
  1916  				newm1(newm)
  1917  				newm = next
  1918  			}
  1919  			lock(&newmHandoff.lock)
  1920  		}
  1921  		newmHandoff.waiting = true
  1922  		noteclear(&newmHandoff.wake)
  1923  		unlock(&newmHandoff.lock)
  1924  		notesleep(&newmHandoff.wake)
  1925  	}
  1926  }
  1927  
  1928  // Stops execution of the current m until new work is available.
  1929  // Returns with acquired P.
  1930  func stopm() {
  1931  	_g_ := getg()
  1932  
  1933  	if _g_.m.locks != 0 {
  1934  		throw("stopm holding locks")
  1935  	}
  1936  	if _g_.m.p != 0 {
  1937  		throw("stopm holding p")
  1938  	}
  1939  	if _g_.m.spinning {
  1940  		throw("stopm spinning")
  1941  	}
  1942  
  1943  retry:
  1944  	lock(&sched.lock)
  1945  	mput(_g_.m)
  1946  	unlock(&sched.lock)
  1947  	notesleep(&_g_.m.park)
  1948  	noteclear(&_g_.m.park)
  1949  	if _g_.m.helpgc != 0 {
  1950  		// helpgc() set _g_.m.p and _g_.m.mcache, so we have a P.
  1951  		gchelper()
  1952  		// Undo the effects of helpgc().
  1953  		_g_.m.helpgc = 0
  1954  		_g_.m.mcache = nil
  1955  		_g_.m.p = 0
  1956  		goto retry
  1957  	}
  1958  	acquirep(_g_.m.nextp.ptr())
  1959  	_g_.m.nextp = 0
  1960  }
  1961  
  1962  func mspinning() {
  1963  	// startm's caller incremented nmspinning. Set the new M's spinning.
  1964  	getg().m.spinning = true
  1965  }
  1966  
  1967  // Schedules some M to run the p (creates an M if necessary).
  1968  // If p==nil, tries to get an idle P, if no idle P's does nothing.
  1969  // May run with m.p==nil, so write barriers are not allowed.
  1970  // If spinning is set, the caller has incremented nmspinning and startm will
  1971  // either decrement nmspinning or set m.spinning in the newly started M.
  1972  //go:nowritebarrierrec
  1973  func startm(_p_ *p, spinning bool) {
  1974  	lock(&sched.lock)
  1975  	if _p_ == nil {
  1976  		_p_ = pidleget()
  1977  		if _p_ == nil {
  1978  			unlock(&sched.lock)
  1979  			if spinning {
  1980  				// The caller incremented nmspinning, but there are no idle Ps,
  1981  				// so it's okay to just undo the increment and give up.
  1982  				if int32(atomic.Xadd(&sched.nmspinning, -1)) < 0 {
  1983  					throw("startm: negative nmspinning")
  1984  				}
  1985  			}
  1986  			return
  1987  		}
  1988  	}
  1989  	mp := mget()
  1990  	unlock(&sched.lock)
  1991  	if mp == nil {
  1992  		var fn func()
  1993  		if spinning {
  1994  			// The caller incremented nmspinning, so set m.spinning in the new M.
  1995  			fn = mspinning
  1996  		}
  1997  		newm(fn, _p_)
  1998  		return
  1999  	}
  2000  	if mp.spinning {
  2001  		throw("startm: m is spinning")
  2002  	}
  2003  	if mp.nextp != 0 {
  2004  		throw("startm: m has p")
  2005  	}
  2006  	if spinning && !runqempty(_p_) {
  2007  		throw("startm: p has runnable gs")
  2008  	}
  2009  	// The caller incremented nmspinning, so set m.spinning in the new M.
  2010  	mp.spinning = spinning
  2011  	mp.nextp.set(_p_)
  2012  	notewakeup(&mp.park)
  2013  }
  2014  
  2015  // Hands off P from syscall or locked M.
  2016  // Always runs without a P, so write barriers are not allowed.
  2017  //go:nowritebarrierrec
  2018  func handoffp(_p_ *p) {
  2019  	// handoffp must start an M in any situation where
  2020  	// findrunnable would return a G to run on _p_.
  2021  
  2022  	// if it has local work, start it straight away
  2023  	if !runqempty(_p_) || sched.runqsize != 0 {
  2024  		startm(_p_, false)
  2025  		return
  2026  	}
  2027  	// if it has GC work, start it straight away
  2028  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(_p_) {
  2029  		startm(_p_, false)
  2030  		return
  2031  	}
  2032  	// no local work, check that there are no spinning/idle M's,
  2033  	// otherwise our help is not required
  2034  	if atomic.Load(&sched.nmspinning)+atomic.Load(&sched.npidle) == 0 && atomic.Cas(&sched.nmspinning, 0, 1) { // TODO: fast atomic
  2035  		startm(_p_, true)
  2036  		return
  2037  	}
  2038  	lock(&sched.lock)
  2039  	if sched.gcwaiting != 0 {
  2040  		_p_.status = _Pgcstop
  2041  		sched.stopwait--
  2042  		if sched.stopwait == 0 {
  2043  			notewakeup(&sched.stopnote)
  2044  		}
  2045  		unlock(&sched.lock)
  2046  		return
  2047  	}
  2048  	if _p_.runSafePointFn != 0 && atomic.Cas(&_p_.runSafePointFn, 1, 0) {
  2049  		sched.safePointFn(_p_)
  2050  		sched.safePointWait--
  2051  		if sched.safePointWait == 0 {
  2052  			notewakeup(&sched.safePointNote)
  2053  		}
  2054  	}
  2055  	if sched.runqsize != 0 {
  2056  		unlock(&sched.lock)
  2057  		startm(_p_, false)
  2058  		return
  2059  	}
  2060  	// If this is the last running P and nobody is polling network,
  2061  	// need to wakeup another M to poll network.
  2062  	if sched.npidle == uint32(gomaxprocs-1) && atomic.Load64(&sched.lastpoll) != 0 {
  2063  		unlock(&sched.lock)
  2064  		startm(_p_, false)
  2065  		return
  2066  	}
  2067  	pidleput(_p_)
  2068  	unlock(&sched.lock)
  2069  }
  2070  
  2071  // Tries to add one more P to execute G's.
  2072  // Called when a G is made runnable (newproc, ready).
  2073  func wakep() {
  2074  	// be conservative about spinning threads
  2075  	if !atomic.Cas(&sched.nmspinning, 0, 1) {
  2076  		return
  2077  	}
  2078  	startm(nil, true)
  2079  }
  2080  
  2081  // Stops execution of the current m that is locked to a g until the g is runnable again.
  2082  // Returns with acquired P.
  2083  func stoplockedm() {
  2084  	_g_ := getg()
  2085  
  2086  	if _g_.m.lockedg == 0 || _g_.m.lockedg.ptr().lockedm.ptr() != _g_.m {
  2087  		throw("stoplockedm: inconsistent locking")
  2088  	}
  2089  	if _g_.m.p != 0 {
  2090  		// Schedule another M to run this p.
  2091  		_p_ := releasep()
  2092  		handoffp(_p_)
  2093  	}
  2094  	incidlelocked(1)
  2095  	// Wait until another thread schedules lockedg again.
  2096  	notesleep(&_g_.m.park)
  2097  	noteclear(&_g_.m.park)
  2098  	status := readgstatus(_g_.m.lockedg.ptr())
  2099  	if status&^_Gscan != _Grunnable {
  2100  		print("runtime:stoplockedm: g is not Grunnable or Gscanrunnable\n")
  2101  		dumpgstatus(_g_)
  2102  		throw("stoplockedm: not runnable")
  2103  	}
  2104  	acquirep(_g_.m.nextp.ptr())
  2105  	_g_.m.nextp = 0
  2106  }
  2107  
  2108  // Schedules the locked m to run the locked gp.
  2109  // May run during STW, so write barriers are not allowed.
  2110  //go:nowritebarrierrec
  2111  func startlockedm(gp *g) {
  2112  	_g_ := getg()
  2113  
  2114  	mp := gp.lockedm.ptr()
  2115  	if mp == _g_.m {
  2116  		throw("startlockedm: locked to me")
  2117  	}
  2118  	if mp.nextp != 0 {
  2119  		throw("startlockedm: m has p")
  2120  	}
  2121  	// directly handoff current P to the locked m
  2122  	incidlelocked(-1)
  2123  	_p_ := releasep()
  2124  	mp.nextp.set(_p_)
  2125  	notewakeup(&mp.park)
  2126  	stopm()
  2127  }
  2128  
  2129  // Stops the current m for stopTheWorld.
  2130  // Returns when the world is restarted.
  2131  func gcstopm() {
  2132  	_g_ := getg()
  2133  
  2134  	if sched.gcwaiting == 0 {
  2135  		throw("gcstopm: not waiting for gc")
  2136  	}
  2137  	if _g_.m.spinning {
  2138  		_g_.m.spinning = false
  2139  		// OK to just drop nmspinning here,
  2140  		// startTheWorld will unpark threads as necessary.
  2141  		if int32(atomic.Xadd(&sched.nmspinning, -1)) < 0 {
  2142  			throw("gcstopm: negative nmspinning")
  2143  		}
  2144  	}
  2145  	_p_ := releasep()
  2146  	lock(&sched.lock)
  2147  	_p_.status = _Pgcstop
  2148  	sched.stopwait--
  2149  	if sched.stopwait == 0 {
  2150  		notewakeup(&sched.stopnote)
  2151  	}
  2152  	unlock(&sched.lock)
  2153  	stopm()
  2154  }
  2155  
  2156  // Schedules gp to run on the current M.
  2157  // If inheritTime is true, gp inherits the remaining time in the
  2158  // current time slice. Otherwise, it starts a new time slice.
  2159  // Never returns.
  2160  //
  2161  // Write barriers are allowed because this is called immediately after
  2162  // acquiring a P in several places.
  2163  //
  2164  //go:yeswritebarrierrec
  2165  func execute(gp *g, inheritTime bool) {
  2166  	_g_ := getg()
  2167  
  2168  	casgstatus(gp, _Grunnable, _Grunning)
  2169  	gp.waitsince = 0
  2170  	gp.preempt = false
  2171  	gp.stackguard0 = gp.stack.lo + _StackGuard
  2172  	if !inheritTime {
  2173  		_g_.m.p.ptr().schedtick++
  2174  	}
  2175  	_g_.m.curg = gp
  2176  	gp.m = _g_.m
  2177  
  2178  	// Check whether the profiler needs to be turned on or off.
  2179  	hz := sched.profilehz
  2180  	if _g_.m.profilehz != hz {
  2181  		setThreadCPUProfiler(hz)
  2182  	}
  2183  
  2184  	if trace.enabled {
  2185  		// GoSysExit has to happen when we have a P, but before GoStart.
  2186  		// So we emit it here.
  2187  		if gp.syscallsp != 0 && gp.sysblocktraced {
  2188  			traceGoSysExit(gp.sysexitticks)
  2189  		}
  2190  		traceGoStart()
  2191  	}
  2192  
  2193  	gogo(&gp.sched)
  2194  }
  2195  
  2196  // Finds a runnable goroutine to execute.
  2197  // Tries to steal from other P's, get g from global queue, poll network.
  2198  func findrunnable() (gp *g, inheritTime bool) {
  2199  	_g_ := getg()
  2200  
  2201  	// The conditions here and in handoffp must agree: if
  2202  	// findrunnable would return a G to run, handoffp must start
  2203  	// an M.
  2204  
  2205  top:
  2206  	_p_ := _g_.m.p.ptr()
  2207  	if sched.gcwaiting != 0 {
  2208  		gcstopm()
  2209  		goto top
  2210  	}
  2211  	if _p_.runSafePointFn != 0 {
  2212  		runSafePointFn()
  2213  	}
  2214  	if fingwait && fingwake {
  2215  		if gp := wakefing(); gp != nil {
  2216  			ready(gp, 0, true)
  2217  		}
  2218  	}
  2219  	if *cgo_yield != nil {
  2220  		asmcgocall(*cgo_yield, nil)
  2221  	}
  2222  
  2223  	// local runq
  2224  	if gp, inheritTime := runqget(_p_); gp != nil {
  2225  		return gp, inheritTime
  2226  	}
  2227  
  2228  	// global runq
  2229  	if sched.runqsize != 0 {
  2230  		lock(&sched.lock)
  2231  		gp := globrunqget(_p_, 0)
  2232  		unlock(&sched.lock)
  2233  		if gp != nil {
  2234  			return gp, false
  2235  		}
  2236  	}
  2237  
  2238  	// Poll network.
  2239  	// This netpoll is only an optimization before we resort to stealing.
  2240  	// We can safely skip it if there are no waiters or a thread is blocked
  2241  	// in netpoll already. If there is any kind of logical race with that
  2242  	// blocked thread (e.g. it has already returned from netpoll, but does
  2243  	// not set lastpoll yet), this thread will do blocking netpoll below
  2244  	// anyway.
  2245  	if netpollinited() && atomic.Load(&netpollWaiters) > 0 && atomic.Load64(&sched.lastpoll) != 0 {
  2246  		if gp := netpoll(false); gp != nil { // non-blocking
  2247  			// netpoll returns list of goroutines linked by schedlink.
  2248  			injectglist(gp.schedlink.ptr())
  2249  			casgstatus(gp, _Gwaiting, _Grunnable)
  2250  			if trace.enabled {
  2251  				traceGoUnpark(gp, 0)
  2252  			}
  2253  			return gp, false
  2254  		}
  2255  	}
  2256  
  2257  	// Steal work from other P's.
  2258  	procs := uint32(gomaxprocs)
  2259  	if atomic.Load(&sched.npidle) == procs-1 {
  2260  		// Either GOMAXPROCS=1 or everybody, except for us, is idle already.
  2261  		// New work can appear from returning syscall/cgocall, network or timers.
  2262  		// Neither of that submits to local run queues, so no point in stealing.
  2263  		goto stop
  2264  	}
  2265  	// If number of spinning M's >= number of busy P's, block.
  2266  	// This is necessary to prevent excessive CPU consumption
  2267  	// when GOMAXPROCS>>1 but the program parallelism is low.
  2268  	if !_g_.m.spinning && 2*atomic.Load(&sched.nmspinning) >= procs-atomic.Load(&sched.npidle) {
  2269  		goto stop
  2270  	}
  2271  	if !_g_.m.spinning {
  2272  		_g_.m.spinning = true
  2273  		atomic.Xadd(&sched.nmspinning, 1)
  2274  	}
  2275  	for i := 0; i < 4; i++ {
  2276  		for enum := stealOrder.start(fastrand()); !enum.done(); enum.next() {
  2277  			if sched.gcwaiting != 0 {
  2278  				goto top
  2279  			}
  2280  			stealRunNextG := i > 2 // first look for ready queues with more than 1 g
  2281  			if gp := runqsteal(_p_, allp[enum.position()], stealRunNextG); gp != nil {
  2282  				return gp, false
  2283  			}
  2284  		}
  2285  	}
  2286  
  2287  stop:
  2288  
  2289  	// We have nothing to do. If we're in the GC mark phase, can
  2290  	// safely scan and blacken objects, and have work to do, run
  2291  	// idle-time marking rather than give up the P.
  2292  	if gcBlackenEnabled != 0 && _p_.gcBgMarkWorker != 0 && gcMarkWorkAvailable(_p_) {
  2293  		_p_.gcMarkWorkerMode = gcMarkWorkerIdleMode
  2294  		gp := _p_.gcBgMarkWorker.ptr()
  2295  		casgstatus(gp, _Gwaiting, _Grunnable)
  2296  		if trace.enabled {
  2297  			traceGoUnpark(gp, 0)
  2298  		}
  2299  		return gp, false
  2300  	}
  2301  
  2302  	// Before we drop our P, make a snapshot of the allp slice,
  2303  	// which can change underfoot once we no longer block
  2304  	// safe-points. We don't need to snapshot the contents because
  2305  	// everything up to cap(allp) is immutable.
  2306  	allpSnapshot := allp
  2307  
  2308  	// return P and block
  2309  	lock(&sched.lock)
  2310  	if sched.gcwaiting != 0 || _p_.runSafePointFn != 0 {
  2311  		unlock(&sched.lock)
  2312  		goto top
  2313  	}
  2314  	if sched.runqsize != 0 {
  2315  		gp := globrunqget(_p_, 0)
  2316  		unlock(&sched.lock)
  2317  		return gp, false
  2318  	}
  2319  	if releasep() != _p_ {
  2320  		throw("findrunnable: wrong p")
  2321  	}
  2322  	pidleput(_p_)
  2323  	unlock(&sched.lock)
  2324  
  2325  	// Delicate dance: thread transitions from spinning to non-spinning state,
  2326  	// potentially concurrently with submission of new goroutines. We must
  2327  	// drop nmspinning first and then check all per-P queues again (with
  2328  	// #StoreLoad memory barrier in between). If we do it the other way around,
  2329  	// another thread can submit a goroutine after we've checked all run queues
  2330  	// but before we drop nmspinning; as the result nobody will unpark a thread
  2331  	// to run the goroutine.
  2332  	// If we discover new work below, we need to restore m.spinning as a signal
  2333  	// for resetspinning to unpark a new worker thread (because there can be more
  2334  	// than one starving goroutine). However, if after discovering new work
  2335  	// we also observe no idle Ps, it is OK to just park the current thread:
  2336  	// the system is fully loaded so no spinning threads are required.
  2337  	// Also see "Worker thread parking/unparking" comment at the top of the file.
  2338  	wasSpinning := _g_.m.spinning
  2339  	if _g_.m.spinning {
  2340  		_g_.m.spinning = false
  2341  		if int32(atomic.Xadd(&sched.nmspinning, -1)) < 0 {
  2342  			throw("findrunnable: negative nmspinning")
  2343  		}
  2344  	}
  2345  
  2346  	// check all runqueues once again
  2347  	for _, _p_ := range allpSnapshot {
  2348  		if !runqempty(_p_) {
  2349  			lock(&sched.lock)
  2350  			_p_ = pidleget()
  2351  			unlock(&sched.lock)
  2352  			if _p_ != nil {
  2353  				acquirep(_p_)
  2354  				if wasSpinning {
  2355  					_g_.m.spinning = true
  2356  					atomic.Xadd(&sched.nmspinning, 1)
  2357  				}
  2358  				goto top
  2359  			}
  2360  			break
  2361  		}
  2362  	}
  2363  
  2364  	// Check for idle-priority GC work again.
  2365  	if gcBlackenEnabled != 0 && gcMarkWorkAvailable(nil) {
  2366  		lock(&sched.lock)
  2367  		_p_ = pidleget()
  2368  		if _p_ != nil && _p_.gcBgMarkWorker == 0 {
  2369  			pidleput(_p_)
  2370  			_p_ = nil
  2371  		}
  2372  		unlock(&sched.lock)
  2373  		if _p_ != nil {
  2374  			acquirep(_p_)
  2375  			if wasSpinning {
  2376  				_g_.m.spinning = true
  2377  				atomic.Xadd(&sched.nmspinning, 1)
  2378  			}
  2379  			// Go back to idle GC check.
  2380  			goto stop
  2381  		}
  2382  	}
  2383  
  2384  	// poll network
  2385  	if netpollinited() && atomic.Load(&netpollWaiters) > 0 && atomic.Xchg64(&sched.lastpoll, 0) != 0 {
  2386  		if _g_.m.p != 0 {
  2387  			throw("findrunnable: netpoll with p")
  2388  		}
  2389  		if _g_.m.spinning {
  2390  			throw("findrunnable: netpoll with spinning")
  2391  		}
  2392  		gp := netpoll(true) // block until new work is available
  2393  		atomic.Store64(&sched.lastpoll, uint64(nanotime()))
  2394  		if gp != nil {
  2395  			lock(&sched.lock)
  2396  			_p_ = pidleget()
  2397  			unlock(&sched.lock)
  2398  			if _p_ != nil {
  2399  				acquirep(_p_)
  2400  				injectglist(gp.schedlink.ptr())
  2401  				casgstatus(gp, _Gwaiting, _Grunnable)
  2402  				if trace.enabled {
  2403  					traceGoUnpark(gp, 0)
  2404  				}
  2405  				return gp, false
  2406  			}
  2407  			injectglist(gp)
  2408  		}
  2409  	}
  2410  	stopm()
  2411  	goto top
  2412  }
  2413  
  2414  // pollWork returns true if there is non-background work this P could
  2415  // be doing. This is a fairly lightweight check to be used for
  2416  // background work loops, like idle GC. It checks a subset of the
  2417  // conditions checked by the actual scheduler.
  2418  func pollWork() bool {
  2419  	if sched.runqsize != 0 {
  2420  		return true
  2421  	}
  2422  	p := getg().m.p.ptr()
  2423  	if !runqempty(p) {
  2424  		return true
  2425  	}
  2426  	if netpollinited() && atomic.Load(&netpollWaiters) > 0 && sched.lastpoll != 0 {
  2427  		if gp := netpoll(false); gp != nil {
  2428  			injectglist(gp)
  2429  			return true
  2430  		}
  2431  	}
  2432  	return false
  2433  }
  2434  
  2435  func resetspinning() {
  2436  	_g_ := getg()
  2437  	if !_g_.m.spinning {
  2438  		throw("resetspinning: not a spinning m")
  2439  	}
  2440  	_g_.m.spinning = false
  2441  	nmspinning := atomic.Xadd(&sched.nmspinning, -1)
  2442  	if int32(nmspinning) < 0 {
  2443  		throw("findrunnable: negative nmspinning")
  2444  	}
  2445  	// M wakeup policy is deliberately somewhat conservative, so check if we
  2446  	// need to wakeup another P here. See "Worker thread parking/unparking"
  2447  	// comment at the top of the file for details.
  2448  	if nmspinning == 0 && atomic.Load(&sched.npidle) > 0 {
  2449  		wakep()
  2450  	}
  2451  }
  2452  
  2453  // Injects the list of runnable G's into the scheduler.
  2454  // Can run concurrently with GC.
  2455  func injectglist(glist *g) {
  2456  	if glist == nil {
  2457  		return
  2458  	}
  2459  	if trace.enabled {
  2460  		for gp := glist; gp != nil; gp = gp.schedlink.ptr() {
  2461  			traceGoUnpark(gp, 0)
  2462  		}
  2463  	}
  2464  	lock(&sched.lock)
  2465  	var n int
  2466  	for n = 0; glist != nil; n++ {
  2467  		gp := glist
  2468  		glist = gp.schedlink.ptr()
  2469  		casgstatus(gp, _Gwaiting, _Grunnable)
  2470  		globrunqput(gp)
  2471  	}
  2472  	unlock(&sched.lock)
  2473  	for ; n != 0 && sched.npidle != 0; n-- {
  2474  		startm(nil, false)
  2475  	}
  2476  }
  2477  
  2478  // One round of scheduler: find a runnable goroutine and execute it.
  2479  // Never returns.
  2480  func schedule() {
  2481  	_g_ := getg()
  2482  
  2483  	if _g_.m.locks != 0 {
  2484  		throw("schedule: holding locks")
  2485  	}
  2486  
  2487  	if _g_.m.lockedg != 0 {
  2488  		stoplockedm()
  2489  		execute(_g_.m.lockedg.ptr(), false) // Never returns.
  2490  	}
  2491  
  2492  	// We should not schedule away from a g that is executing a cgo call,
  2493  	// since the cgo call is using the m's g0 stack.
  2494  	if _g_.m.incgo {
  2495  		throw("schedule: in cgo")
  2496  	}
  2497  
  2498  top:
  2499  	if sched.gcwaiting != 0 {
  2500  		gcstopm()
  2501  		goto top
  2502  	}
  2503  	if _g_.m.p.ptr().runSafePointFn != 0 {
  2504  		runSafePointFn()
  2505  	}
  2506  
  2507  	var gp *g
  2508  	var inheritTime bool
  2509  	if trace.enabled || trace.shutdown {
  2510  		gp = traceReader()
  2511  		if gp != nil {
  2512  			casgstatus(gp, _Gwaiting, _Grunnable)
  2513  			traceGoUnpark(gp, 0)
  2514  		}
  2515  	}
  2516  	if gp == nil && gcBlackenEnabled != 0 {
  2517  		gp = gcController.findRunnableGCWorker(_g_.m.p.ptr())
  2518  	}
  2519  	if gp == nil {
  2520  		// Check the global runnable queue once in a while to ensure fairness.
  2521  		// Otherwise two goroutines can completely occupy the local runqueue
  2522  		// by constantly respawning each other.
  2523  		if _g_.m.p.ptr().schedtick%61 == 0 && sched.runqsize > 0 {
  2524  			lock(&sched.lock)
  2525  			gp = globrunqget(_g_.m.p.ptr(), 1)
  2526  			unlock(&sched.lock)
  2527  		}
  2528  	}
  2529  	if gp == nil {
  2530  		gp, inheritTime = runqget(_g_.m.p.ptr())
  2531  		if gp != nil && _g_.m.spinning {
  2532  			throw("schedule: spinning with local work")
  2533  		}
  2534  	}
  2535  	if gp == nil {
  2536  		gp, inheritTime = findrunnable() // blocks until work is available
  2537  	}
  2538  
  2539  	// This thread is going to run a goroutine and is not spinning anymore,
  2540  	// so if it was marked as spinning we need to reset it now and potentially
  2541  	// start a new spinning M.
  2542  	if _g_.m.spinning {
  2543  		resetspinning()
  2544  	}
  2545  
  2546  	if gp.lockedm != 0 {
  2547  		// Hands off own p to the locked m,
  2548  		// then blocks waiting for a new p.
  2549  		startlockedm(gp)
  2550  		goto top
  2551  	}
  2552  
  2553  	execute(gp, inheritTime)
  2554  }
  2555  
  2556  // dropg removes the association between m and the current goroutine m->curg (gp for short).
  2557  // Typically a caller sets gp's status away from Grunning and then
  2558  // immediately calls dropg to finish the job. The caller is also responsible
  2559  // for arranging that gp will be restarted using ready at an
  2560  // appropriate time. After calling dropg and arranging for gp to be
  2561  // readied later, the caller can do other work but eventually should
  2562  // call schedule to restart the scheduling of goroutines on this m.
  2563  func dropg() {
  2564  	_g_ := getg()
  2565  
  2566  	setMNoWB(&_g_.m.curg.m, nil)
  2567  	setGNoWB(&_g_.m.curg, nil)
  2568  }
  2569  
  2570  func parkunlock_c(gp *g, lock unsafe.Pointer) bool {
  2571  	unlock((*mutex)(lock))
  2572  	return true
  2573  }
  2574  
  2575  // park continuation on g0.
  2576  func park_m(gp *g) {
  2577  	_g_ := getg()
  2578  
  2579  	if trace.enabled {
  2580  		traceGoPark(_g_.m.waittraceev, _g_.m.waittraceskip)
  2581  	}
  2582  
  2583  	casgstatus(gp, _Grunning, _Gwaiting)
  2584  	dropg()
  2585  
  2586  	if _g_.m.waitunlockf != nil {
  2587  		fn := *(*func(*g, unsafe.Pointer) bool)(unsafe.Pointer(&_g_.m.waitunlockf))
  2588  		ok := fn(gp, _g_.m.waitlock)
  2589  		_g_.m.waitunlockf = nil
  2590  		_g_.m.waitlock = nil
  2591  		if !ok {
  2592  			if trace.enabled {
  2593  				traceGoUnpark(gp, 2)
  2594  			}
  2595  			casgstatus(gp, _Gwaiting, _Grunnable)
  2596  			execute(gp, true) // Schedule it back, never returns.
  2597  		}
  2598  	}
  2599  	schedule()
  2600  }
  2601  
  2602  func goschedImpl(gp *g) {
  2603  	status := readgstatus(gp)
  2604  	if status&^_Gscan != _Grunning {
  2605  		dumpgstatus(gp)
  2606  		throw("bad g status")
  2607  	}
  2608  	casgstatus(gp, _Grunning, _Grunnable)
  2609  	dropg()
  2610  	lock(&sched.lock)
  2611  	globrunqput(gp)
  2612  	unlock(&sched.lock)
  2613  
  2614  	schedule()
  2615  }
  2616  
  2617  // Gosched continuation on g0.
  2618  func gosched_m(gp *g) {
  2619  	if trace.enabled {
  2620  		traceGoSched()
  2621  	}
  2622  	goschedImpl(gp)
  2623  }
  2624  
  2625  // goschedguarded is a forbidden-states-avoided version of gosched_m
  2626  func goschedguarded_m(gp *g) {
  2627  
  2628  	if gp.m.locks != 0 || gp.m.mallocing != 0 || gp.m.preemptoff != "" || gp.m.p.ptr().status != _Prunning {
  2629  		gogo(&gp.sched) // never return
  2630  	}
  2631  
  2632  	if trace.enabled {
  2633  		traceGoSched()
  2634  	}
  2635  	goschedImpl(gp)
  2636  }
  2637  
  2638  func gopreempt_m(gp *g) {
  2639  	if trace.enabled {
  2640  		traceGoPreempt()
  2641  	}
  2642  	goschedImpl(gp)
  2643  }
  2644  
  2645  // Finishes execution of the current goroutine.
  2646  func goexit1() {
  2647  	if raceenabled {
  2648  		racegoend()
  2649  	}
  2650  	if trace.enabled {
  2651  		traceGoEnd()
  2652  	}
  2653  	mcall(goexit0)
  2654  }
  2655  
  2656  // goexit continuation on g0.
  2657  func goexit0(gp *g) {
  2658  	_g_ := getg()
  2659  
  2660  	casgstatus(gp, _Grunning, _Gdead)
  2661  	if isSystemGoroutine(gp) {
  2662  		atomic.Xadd(&sched.ngsys, -1)
  2663  	}
  2664  	gp.m = nil
  2665  	locked := gp.lockedm != 0
  2666  	gp.lockedm = 0
  2667  	_g_.m.lockedg = 0
  2668  	gp.paniconfault = false
  2669  	gp._defer = nil // should be true already but just in case.
  2670  	gp._panic = nil // non-nil for Goexit during panic. points at stack-allocated data.
  2671  	gp.writebuf = nil
  2672  	gp.waitreason = ""
  2673  	gp.param = nil
  2674  	gp.labels = nil
  2675  	gp.timer = nil
  2676  
  2677  	if gcBlackenEnabled != 0 && gp.gcAssistBytes > 0 {
  2678  		// Flush assist credit to the global pool. This gives
  2679  		// better information to pacing if the application is
  2680  		// rapidly creating an exiting goroutines.
  2681  		scanCredit := int64(gcController.assistWorkPerByte * float64(gp.gcAssistBytes))
  2682  		atomic.Xaddint64(&gcController.bgScanCredit, scanCredit)
  2683  		gp.gcAssistBytes = 0
  2684  	}
  2685  
  2686  	// Note that gp's stack scan is now "valid" because it has no
  2687  	// stack.
  2688  	gp.gcscanvalid = true
  2689  	dropg()
  2690  
  2691  	if _g_.m.lockedInt != 0 {
  2692  		print("invalid m->lockedInt = ", _g_.m.lockedInt, "\n")
  2693  		throw("internal lockOSThread error")
  2694  	}
  2695  	_g_.m.lockedExt = 0
  2696  	gfput(_g_.m.p.ptr(), gp)
  2697  	if locked {
  2698  		// The goroutine may have locked this thread because
  2699  		// it put it in an unusual kernel state. Kill it
  2700  		// rather than returning it to the thread pool.
  2701  
  2702  		// Return to mstart, which will release the P and exit
  2703  		// the thread.
  2704  		if GOOS != "plan9" { // See golang.org/issue/22227.
  2705  			gogo(&_g_.m.g0.sched)
  2706  		}
  2707  	}
  2708  	schedule()
  2709  }
  2710  
  2711  // save updates getg().sched to refer to pc and sp so that a following
  2712  // gogo will restore pc and sp.
  2713  //
  2714  // save must not have write barriers because invoking a write barrier
  2715  // can clobber getg().sched.
  2716  //
  2717  //go:nosplit
  2718  //go:nowritebarrierrec
  2719  func save(pc, sp uintptr) {
  2720  	_g_ := getg()
  2721  
  2722  	_g_.sched.pc = pc
  2723  	_g_.sched.sp = sp
  2724  	_g_.sched.lr = 0
  2725  	_g_.sched.ret = 0
  2726  	_g_.sched.g = guintptr(unsafe.Pointer(_g_))
  2727  	// We need to ensure ctxt is zero, but can't have a write
  2728  	// barrier here. However, it should always already be zero.
  2729  	// Assert that.
  2730  	if _g_.sched.ctxt != nil {
  2731  		badctxt()
  2732  	}
  2733  }
  2734  
  2735  // The goroutine g is about to enter a system call.
  2736  // Record that it's not using the cpu anymore.
  2737  // This is called only from the go syscall library and cgocall,
  2738  // not from the low-level system calls used by the runtime.
  2739  //
  2740  // Entersyscall cannot split the stack: the gosave must
  2741  // make g->sched refer to the caller's stack segment, because
  2742  // entersyscall is going to return immediately after.
  2743  //
  2744  // Nothing entersyscall calls can split the stack either.
  2745  // We cannot safely move the stack during an active call to syscall,
  2746  // because we do not know which of the uintptr arguments are
  2747  // really pointers (back into the stack).
  2748  // In practice, this means that we make the fast path run through
  2749  // entersyscall doing no-split things, and the slow path has to use systemstack
  2750  // to run bigger things on the system stack.
  2751  //
  2752  // reentersyscall is the entry point used by cgo callbacks, where explicitly
  2753  // saved SP and PC are restored. This is needed when exitsyscall will be called
  2754  // from a function further up in the call stack than the parent, as g->syscallsp
  2755  // must always point to a valid stack frame. entersyscall below is the normal
  2756  // entry point for syscalls, which obtains the SP and PC from the caller.
  2757  //
  2758  // Syscall tracing:
  2759  // At the start of a syscall we emit traceGoSysCall to capture the stack trace.
  2760  // If the syscall does not block, that is it, we do not emit any other events.
  2761  // If the syscall blocks (that is, P is retaken), retaker emits traceGoSysBlock;
  2762  // when syscall returns we emit traceGoSysExit and when the goroutine starts running
  2763  // (potentially instantly, if exitsyscallfast returns true) we emit traceGoStart.
  2764  // To ensure that traceGoSysExit is emitted strictly after traceGoSysBlock,
  2765  // we remember current value of syscalltick in m (_g_.m.syscalltick = _g_.m.p.ptr().syscalltick),
  2766  // whoever emits traceGoSysBlock increments p.syscalltick afterwards;
  2767  // and we wait for the increment before emitting traceGoSysExit.
  2768  // Note that the increment is done even if tracing is not enabled,
  2769  // because tracing can be enabled in the middle of syscall. We don't want the wait to hang.
  2770  //
  2771  //go:nosplit
  2772  func reentersyscall(pc, sp uintptr) {
  2773  	_g_ := getg()
  2774  
  2775  	// Disable preemption because during this function g is in Gsyscall status,
  2776  	// but can have inconsistent g->sched, do not let GC observe it.
  2777  	_g_.m.locks++
  2778  
  2779  	// Entersyscall must not call any function that might split/grow the stack.
  2780  	// (See details in comment above.)
  2781  	// Catch calls that might, by replacing the stack guard with something that
  2782  	// will trip any stack check and leaving a flag to tell newstack to die.
  2783  	_g_.stackguard0 = stackPreempt
  2784  	_g_.throwsplit = true
  2785  
  2786  	// Leave SP around for GC and traceback.
  2787  	save(pc, sp)
  2788  	_g_.syscallsp = sp
  2789  	_g_.syscallpc = pc
  2790  	casgstatus(_g_, _Grunning, _Gsyscall)
  2791  	if _g_.syscallsp < _g_.stack.lo || _g_.stack.hi < _g_.syscallsp {
  2792  		systemstack(func() {
  2793  			print("entersyscall inconsistent ", hex(_g_.syscallsp), " [", hex(_g_.stack.lo), ",", hex(_g_.stack.hi), "]\n")
  2794  			throw("entersyscall")
  2795  		})
  2796  	}
  2797  
  2798  	if trace.enabled {
  2799  		systemstack(traceGoSysCall)
  2800  		// systemstack itself clobbers g.sched.{pc,sp} and we might
  2801  		// need them later when the G is genuinely blocked in a
  2802  		// syscall
  2803  		save(pc, sp)
  2804  	}
  2805  
  2806  	if atomic.Load(&sched.sysmonwait) != 0 {
  2807  		systemstack(entersyscall_sysmon)
  2808  		save(pc, sp)
  2809  	}
  2810  
  2811  	if _g_.m.p.ptr().runSafePointFn != 0 {
  2812  		// runSafePointFn may stack split if run on this stack
  2813  		systemstack(runSafePointFn)
  2814  		save(pc, sp)
  2815  	}
  2816  
  2817  	_g_.m.syscalltick = _g_.m.p.ptr().syscalltick
  2818  	_g_.sysblocktraced = true
  2819  	_g_.m.mcache = nil
  2820  	_g_.m.p.ptr().m = 0
  2821  	atomic.Store(&_g_.m.p.ptr().status, _Psyscall)
  2822  	if sched.gcwaiting != 0 {
  2823  		systemstack(entersyscall_gcwait)
  2824  		save(pc, sp)
  2825  	}
  2826  
  2827  	// Goroutines must not split stacks in Gsyscall status (it would corrupt g->sched).
  2828  	// We set _StackGuard to StackPreempt so that first split stack check calls morestack.
  2829  	// Morestack detects this case and throws.
  2830  	_g_.stackguard0 = stackPreempt
  2831  	_g_.m.locks--
  2832  }
  2833  
  2834  // Standard syscall entry used by the go syscall library and normal cgo calls.
  2835  //go:nosplit
  2836  func entersyscall(dummy int32) {
  2837  	reentersyscall(getcallerpc(), getcallersp(unsafe.Pointer(&dummy)))
  2838  }
  2839  
  2840  func entersyscall_sysmon() {
  2841  	lock(&sched.lock)
  2842  	if atomic.Load(&sched.sysmonwait) != 0 {
  2843  		atomic.Store(&sched.sysmonwait, 0)
  2844  		notewakeup(&sched.sysmonnote)
  2845  	}
  2846  	unlock(&sched.lock)
  2847  }
  2848  
  2849  func entersyscall_gcwait() {
  2850  	_g_ := getg()
  2851  	_p_ := _g_.m.p.ptr()
  2852  
  2853  	lock(&sched.lock)
  2854  	if sched.stopwait > 0 && atomic.Cas(&_p_.status, _Psyscall, _Pgcstop) {
  2855  		if trace.enabled {
  2856  			traceGoSysBlock(_p_)
  2857  			traceProcStop(_p_)
  2858  		}
  2859  		_p_.syscalltick++
  2860  		if sched.stopwait--; sched.stopwait == 0 {
  2861  			notewakeup(&sched.stopnote)
  2862  		}
  2863  	}
  2864  	unlock(&sched.lock)
  2865  }
  2866  
  2867  // The same as entersyscall(), but with a hint that the syscall is blocking.
  2868  //go:nosplit
  2869  func entersyscallblock(dummy int32) {
  2870  	_g_ := getg()
  2871  
  2872  	_g_.m.locks++ // see comment in entersyscall
  2873  	_g_.throwsplit = true
  2874  	_g_.stackguard0 = stackPreempt // see comment in entersyscall
  2875  	_g_.m.syscalltick = _g_.m.p.ptr().syscalltick
  2876  	_g_.sysblocktraced = true
  2877  	_g_.m.p.ptr().syscalltick++
  2878  
  2879  	// Leave SP around for GC and traceback.
  2880  	pc := getcallerpc()
  2881  	sp := getcallersp(unsafe.Pointer(&dummy))
  2882  	save(pc, sp)
  2883  	_g_.syscallsp = _g_.sched.sp
  2884  	_g_.syscallpc = _g_.sched.pc
  2885  	if _g_.syscallsp < _g_.stack.lo || _g_.stack.hi < _g_.syscallsp {
  2886  		sp1 := sp
  2887  		sp2 := _g_.sched.sp
  2888  		sp3 := _g_.syscallsp
  2889  		systemstack(func() {
  2890  			print("entersyscallblock inconsistent ", hex(sp1), " ", hex(sp2), " ", hex(sp3), " [", hex(_g_.stack.lo), ",", hex(_g_.stack.hi), "]\n")
  2891  			throw("entersyscallblock")
  2892  		})
  2893  	}
  2894  	casgstatus(_g_, _Grunning, _Gsyscall)
  2895  	if _g_.syscallsp < _g_.stack.lo || _g_.stack.hi < _g_.syscallsp {
  2896  		systemstack(func() {
  2897  			print("entersyscallblock inconsistent ", hex(sp), " ", hex(_g_.sched.sp), " ", hex(_g_.syscallsp), " [", hex(_g_.stack.lo), ",", hex(_g_.stack.hi), "]\n")
  2898  			throw("entersyscallblock")
  2899  		})
  2900  	}
  2901  
  2902  	systemstack(entersyscallblock_handoff)
  2903  
  2904  	// Resave for traceback during blocked call.
  2905  	save(getcallerpc(), getcallersp(unsafe.Pointer(&dummy)))
  2906  
  2907  	_g_.m.locks--
  2908  }
  2909  
  2910  func entersyscallblock_handoff() {
  2911  	if trace.enabled {
  2912  		traceGoSysCall()
  2913  		traceGoSysBlock(getg().m.p.ptr())
  2914  	}
  2915  	handoffp(releasep())
  2916  }
  2917  
  2918  // The goroutine g exited its system call.
  2919  // Arrange for it to run on a cpu again.
  2920  // This is called only from the go syscall library, not
  2921  // from the low-level system calls used by the runtime.
  2922  //
  2923  // Write barriers are not allowed because our P may have been stolen.
  2924  //
  2925  //go:nosplit
  2926  //go:nowritebarrierrec
  2927  func exitsyscall(dummy int32) {
  2928  	_g_ := getg()
  2929  
  2930  	_g_.m.locks++ // see comment in entersyscall
  2931  	if getcallersp(unsafe.Pointer(&dummy)) > _g_.syscallsp {
  2932  		// throw calls print which may try to grow the stack,
  2933  		// but throwsplit == true so the stack can not be grown;
  2934  		// use systemstack to avoid that possible problem.
  2935  		systemstack(func() {
  2936  			throw("exitsyscall: syscall frame is no longer valid")
  2937  		})
  2938  	}
  2939  
  2940  	_g_.waitsince = 0
  2941  	oldp := _g_.m.p.ptr()
  2942  	if exitsyscallfast() {
  2943  		if _g_.m.mcache == nil {
  2944  			systemstack(func() {
  2945  				throw("lost mcache")
  2946  			})
  2947  		}
  2948  		if trace.enabled {
  2949  			if oldp != _g_.m.p.ptr() || _g_.m.syscalltick != _g_.m.p.ptr().syscalltick {
  2950  				systemstack(traceGoStart)
  2951  			}
  2952  		}
  2953  		// There's a cpu for us, so we can run.
  2954  		_g_.m.p.ptr().syscalltick++
  2955  		// We need to cas the status and scan before resuming...
  2956  		casgstatus(_g_, _Gsyscall, _Grunning)
  2957  
  2958  		// Garbage collector isn't running (since we are),
  2959  		// so okay to clear syscallsp.
  2960  		_g_.syscallsp = 0
  2961  		_g_.m.locks--
  2962  		if _g_.preempt {
  2963  			// restore the preemption request in case we've cleared it in newstack
  2964  			_g_.stackguard0 = stackPreempt
  2965  		} else {
  2966  			// otherwise restore the real _StackGuard, we've spoiled it in entersyscall/entersyscallblock
  2967  			_g_.stackguard0 = _g_.stack.lo + _StackGuard
  2968  		}
  2969  		_g_.throwsplit = false
  2970  		return
  2971  	}
  2972  
  2973  	_g_.sysexitticks = 0
  2974  	if trace.enabled {
  2975  		// Wait till traceGoSysBlock event is emitted.
  2976  		// This ensures consistency of the trace (the goroutine is started after it is blocked).
  2977  		for oldp != nil && oldp.syscalltick == _g_.m.syscalltick {
  2978  			osyield()
  2979  		}
  2980  		// We can't trace syscall exit right now because we don't have a P.
  2981  		// Tracing code can invoke write barriers that cannot run without a P.
  2982  		// So instead we remember the syscall exit time and emit the event
  2983  		// in execute when we have a P.
  2984  		_g_.sysexitticks = cputicks()
  2985  	}
  2986  
  2987  	_g_.m.locks--
  2988  
  2989  	// Call the scheduler.
  2990  	mcall(exitsyscall0)
  2991  
  2992  	if _g_.m.mcache == nil {
  2993  		systemstack(func() {
  2994  			throw("lost mcache")
  2995  		})
  2996  	}
  2997  
  2998  	// Scheduler returned, so we're allowed to run now.
  2999  	// Delete the syscallsp information that we left for
  3000  	// the garbage collector during the system call.
  3001  	// Must wait until now because until gosched returns
  3002  	// we don't know for sure that the garbage collector
  3003  	// is not running.
  3004  	_g_.syscallsp = 0
  3005  	_g_.m.p.ptr().syscalltick++
  3006  	_g_.throwsplit = false
  3007  }
  3008  
  3009  //go:nosplit
  3010  func exitsyscallfast() bool {
  3011  	_g_ := getg()
  3012  
  3013  	// Freezetheworld sets stopwait but does not retake P's.
  3014  	if sched.stopwait == freezeStopWait {
  3015  		_g_.m.mcache = nil
  3016  		_g_.m.p = 0
  3017  		return false
  3018  	}
  3019  
  3020  	// Try to re-acquire the last P.
  3021  	if _g_.m.p != 0 && _g_.m.p.ptr().status == _Psyscall && atomic.Cas(&_g_.m.p.ptr().status, _Psyscall, _Prunning) {
  3022  		// There's a cpu for us, so we can run.
  3023  		exitsyscallfast_reacquired()
  3024  		return true
  3025  	}
  3026  
  3027  	// Try to get any other idle P.
  3028  	oldp := _g_.m.p.ptr()
  3029  	_g_.m.mcache = nil
  3030  	_g_.m.p = 0
  3031  	if sched.pidle != 0 {
  3032  		var ok bool
  3033  		systemstack(func() {
  3034  			ok = exitsyscallfast_pidle()
  3035  			if ok && trace.enabled {
  3036  				if oldp != nil {
  3037  					// Wait till traceGoSysBlock event is emitted.
  3038  					// This ensures consistency of the trace (the goroutine is started after it is blocked).
  3039  					for oldp.syscalltick == _g_.m.syscalltick {
  3040  						osyield()
  3041  					}
  3042  				}
  3043  				traceGoSysExit(0)
  3044  			}
  3045  		})
  3046  		if ok {
  3047  			return true
  3048  		}
  3049  	}
  3050  	return false
  3051  }
  3052  
  3053  // exitsyscallfast_reacquired is the exitsyscall path on which this G
  3054  // has successfully reacquired the P it was running on before the
  3055  // syscall.
  3056  //
  3057  // This function is allowed to have write barriers because exitsyscall
  3058  // has acquired a P at this point.
  3059  //
  3060  //go:yeswritebarrierrec
  3061  //go:nosplit
  3062  func exitsyscallfast_reacquired() {
  3063  	_g_ := getg()
  3064  	_g_.m.mcache = _g_.m.p.ptr().mcache
  3065  	_g_.m.p.ptr().m.set(_g_.m)
  3066  	if _g_.m.syscalltick != _g_.m.p.ptr().syscalltick {
  3067  		if trace.enabled {
  3068  			// The p was retaken and then enter into syscall again (since _g_.m.syscalltick has changed).
  3069  			// traceGoSysBlock for this syscall was already emitted,
  3070  			// but here we effectively retake the p from the new syscall running on the same p.
  3071  			systemstack(func() {
  3072  				// Denote blocking of the new syscall.
  3073  				traceGoSysBlock(_g_.m.p.ptr())
  3074  				// Denote completion of the current syscall.
  3075  				traceGoSysExit(0)
  3076  			})
  3077  		}
  3078  		_g_.m.p.ptr().syscalltick++
  3079  	}
  3080  }
  3081  
  3082  func exitsyscallfast_pidle() bool {
  3083  	lock(&sched.lock)
  3084  	_p_ := pidleget()
  3085  	if _p_ != nil && atomic.Load(&sched.sysmonwait) != 0 {
  3086  		atomic.Store(&sched.sysmonwait, 0)
  3087  		notewakeup(&sched.sysmonnote)
  3088  	}
  3089  	unlock(&sched.lock)
  3090  	if _p_ != nil {
  3091  		acquirep(_p_)
  3092  		return true
  3093  	}
  3094  	return false
  3095  }
  3096  
  3097  // exitsyscall slow path on g0.
  3098  // Failed to acquire P, enqueue gp as runnable.
  3099  //
  3100  //go:nowritebarrierrec
  3101  func exitsyscall0(gp *g) {
  3102  	_g_ := getg()
  3103  
  3104  	casgstatus(gp, _Gsyscall, _Grunnable)
  3105  	dropg()
  3106  	lock(&sched.lock)
  3107  	_p_ := pidleget()
  3108  	if _p_ == nil {
  3109  		globrunqput(gp)
  3110  	} else if atomic.Load(&sched.sysmonwait) != 0 {
  3111  		atomic.Store(&sched.sysmonwait, 0)
  3112  		notewakeup(&sched.sysmonnote)
  3113  	}
  3114  	unlock(&sched.lock)
  3115  	if _p_ != nil {
  3116  		acquirep(_p_)
  3117  		execute(gp, false) // Never returns.
  3118  	}
  3119  	if _g_.m.lockedg != 0 {
  3120  		// Wait until another thread schedules gp and so m again.
  3121  		stoplockedm()
  3122  		execute(gp, false) // Never returns.
  3123  	}
  3124  	stopm()
  3125  	schedule() // Never returns.
  3126  }
  3127  
  3128  func beforefork() {
  3129  	gp := getg().m.curg
  3130  
  3131  	// Block signals during a fork, so that the child does not run
  3132  	// a signal handler before exec if a signal is sent to the process
  3133  	// group. See issue #18600.
  3134  	gp.m.locks++
  3135  	msigsave(gp.m)
  3136  	sigblock()
  3137  
  3138  	// This function is called before fork in syscall package.
  3139  	// Code between fork and exec must not allocate memory nor even try to grow stack.
  3140  	// Here we spoil g->_StackGuard to reliably detect any attempts to grow stack.
  3141  	// runtime_AfterFork will undo this in parent process, but not in child.
  3142  	gp.stackguard0 = stackFork
  3143  }
  3144  
  3145  // Called from syscall package before fork.
  3146  //go:linkname syscall_runtime_BeforeFork syscall.runtime_BeforeFork
  3147  //go:nosplit
  3148  func syscall_runtime_BeforeFork() {
  3149  	systemstack(beforefork)
  3150  }
  3151  
  3152  func afterfork() {
  3153  	gp := getg().m.curg
  3154  
  3155  	// See the comments in beforefork.
  3156  	gp.stackguard0 = gp.stack.lo + _StackGuard
  3157  
  3158  	msigrestore(gp.m.sigmask)
  3159  
  3160  	gp.m.locks--
  3161  }
  3162  
  3163  // Called from syscall package after fork in parent.
  3164  //go:linkname syscall_runtime_AfterFork syscall.runtime_AfterFork
  3165  //go:nosplit
  3166  func syscall_runtime_AfterFork() {
  3167  	systemstack(afterfork)
  3168  }
  3169  
  3170  // inForkedChild is true while manipulating signals in the child process.
  3171  // This is used to avoid calling libc functions in case we are using vfork.
  3172  var inForkedChild bool
  3173  
  3174  // Called from syscall package after fork in child.
  3175  // It resets non-sigignored signals to the default handler, and
  3176  // restores the signal mask in preparation for the exec.
  3177  //
  3178  // Because this might be called during a vfork, and therefore may be
  3179  // temporarily sharing address space with the parent process, this must
  3180  // not change any global variables or calling into C code that may do so.
  3181  //
  3182  //go:linkname syscall_runtime_AfterForkInChild syscall.runtime_AfterForkInChild
  3183  //go:nosplit
  3184  //go:nowritebarrierrec
  3185  func syscall_runtime_AfterForkInChild() {
  3186  	// It's OK to change the global variable inForkedChild here
  3187  	// because we are going to change it back. There is no race here,
  3188  	// because if we are sharing address space with the parent process,
  3189  	// then the parent process can not be running concurrently.
  3190  	inForkedChild = true
  3191  
  3192  	clearSignalHandlers()
  3193  
  3194  	// When we are the child we are the only thread running,
  3195  	// so we know that nothing else has changed gp.m.sigmask.
  3196  	msigrestore(getg().m.sigmask)
  3197  
  3198  	inForkedChild = false
  3199  }
  3200  
  3201  // Called from syscall package before Exec.
  3202  //go:linkname syscall_runtime_BeforeExec syscall.runtime_BeforeExec
  3203  func syscall_runtime_BeforeExec() {
  3204  	// Prevent thread creation during exec.
  3205  	execLock.lock()
  3206  }
  3207  
  3208  // Called from syscall package after Exec.
  3209  //go:linkname syscall_runtime_AfterExec syscall.runtime_AfterExec
  3210  func syscall_runtime_AfterExec() {
  3211  	execLock.unlock()
  3212  }
  3213  
  3214  // Allocate a new g, with a stack big enough for stacksize bytes.
  3215  func malg(stacksize int32) *g {
  3216  	newg := new(g)
  3217  	if stacksize >= 0 {
  3218  		stacksize = round2(_StackSystem + stacksize)
  3219  		systemstack(func() {
  3220  			newg.stack = stackalloc(uint32(stacksize))
  3221  		})
  3222  		newg.stackguard0 = newg.stack.lo + _StackGuard
  3223  		newg.stackguard1 = ^uintptr(0)
  3224  	}
  3225  	return newg
  3226  }
  3227  
  3228  // Create a new g running fn with siz bytes of arguments.
  3229  // Put it on the queue of g's waiting to run.
  3230  // The compiler turns a go statement into a call to this.
  3231  // Cannot split the stack because it assumes that the arguments
  3232  // are available sequentially after &fn; they would not be
  3233  // copied if a stack split occurred.
  3234  //go:nosplit
  3235  func newproc(siz int32, fn *funcval) {
  3236  	argp := add(unsafe.Pointer(&fn), sys.PtrSize)
  3237  	pc := getcallerpc()
  3238  	systemstack(func() {
  3239  		newproc1(fn, (*uint8)(argp), siz, pc)
  3240  	})
  3241  }
  3242  
  3243  // Create a new g running fn with narg bytes of arguments starting
  3244  // at argp. callerpc is the address of the go statement that created
  3245  // this. The new g is put on the queue of g's waiting to run.
  3246  func newproc1(fn *funcval, argp *uint8, narg int32, callerpc uintptr) {
  3247  	_g_ := getg()
  3248  
  3249  	if fn == nil {
  3250  		_g_.m.throwing = -1 // do not dump full stacks
  3251  		throw("go of nil func value")
  3252  	}
  3253  	_g_.m.locks++ // disable preemption because it can be holding p in a local var
  3254  	siz := narg
  3255  	siz = (siz + 7) &^ 7
  3256  
  3257  	// We could allocate a larger initial stack if necessary.
  3258  	// Not worth it: this is almost always an error.
  3259  	// 4*sizeof(uintreg): extra space added below
  3260  	// sizeof(uintreg): caller's LR (arm) or return address (x86, in gostartcall).
  3261  	if siz >= _StackMin-4*sys.RegSize-sys.RegSize {
  3262  		throw("newproc: function arguments too large for new goroutine")
  3263  	}
  3264  
  3265  	_p_ := _g_.m.p.ptr()
  3266  	newg := gfget(_p_)
  3267  	if newg == nil {
  3268  		newg = malg(_StackMin)
  3269  		casgstatus(newg, _Gidle, _Gdead)
  3270  		allgadd(newg) // publishes with a g->status of Gdead so GC scanner doesn't look at uninitialized stack.
  3271  	}
  3272  	if newg.stack.hi == 0 {
  3273  		throw("newproc1: newg missing stack")
  3274  	}
  3275  
  3276  	if readgstatus(newg) != _Gdead {
  3277  		throw("newproc1: new g is not Gdead")
  3278  	}
  3279  
  3280  	totalSize := 4*sys.RegSize + uintptr(siz) + sys.MinFrameSize // extra space in case of reads slightly beyond frame
  3281  	totalSize += -totalSize & (sys.SpAlign - 1)                  // align to spAlign
  3282  	sp := newg.stack.hi - totalSize
  3283  	spArg := sp
  3284  	if usesLR {
  3285  		// caller's LR
  3286  		*(*uintptr)(unsafe.Pointer(sp)) = 0
  3287  		prepGoExitFrame(sp)
  3288  		spArg += sys.MinFrameSize
  3289  	}
  3290  	if narg > 0 {
  3291  		memmove(unsafe.Pointer(spArg), unsafe.Pointer(argp), uintptr(narg))
  3292  		// This is a stack-to-stack copy. If write barriers
  3293  		// are enabled and the source stack is grey (the
  3294  		// destination is always black), then perform a
  3295  		// barrier copy. We do this *after* the memmove
  3296  		// because the destination stack may have garbage on
  3297  		// it.
  3298  		if writeBarrier.needed && !_g_.m.curg.gcscandone {
  3299  			f := findfunc(fn.fn)
  3300  			stkmap := (*stackmap)(funcdata(f, _FUNCDATA_ArgsPointerMaps))
  3301  			// We're in the prologue, so it's always stack map index 0.
  3302  			bv := stackmapdata(stkmap, 0)
  3303  			bulkBarrierBitmap(spArg, spArg, uintptr(narg), 0, bv.bytedata)
  3304  		}
  3305  	}
  3306  
  3307  	memclrNoHeapPointers(unsafe.Pointer(&newg.sched), unsafe.Sizeof(newg.sched))
  3308  	newg.sched.sp = sp
  3309  	newg.stktopsp = sp
  3310  	newg.sched.pc = funcPC(goexit) + sys.PCQuantum // +PCQuantum so that previous instruction is in same function
  3311  	newg.sched.g = guintptr(unsafe.Pointer(newg))
  3312  	gostartcallfn(&newg.sched, fn)
  3313  	newg.gopc = callerpc
  3314  	newg.startpc = fn.fn
  3315  	if _g_.m.curg != nil {
  3316  		newg.labels = _g_.m.curg.labels
  3317  	}
  3318  	if isSystemGoroutine(newg) {
  3319  		atomic.Xadd(&sched.ngsys, +1)
  3320  	}
  3321  	newg.gcscanvalid = false
  3322  	casgstatus(newg, _Gdead, _Grunnable)
  3323  
  3324  	if _p_.goidcache == _p_.goidcacheend {
  3325  		// Sched.goidgen is the last allocated id,
  3326  		// this batch must be [sched.goidgen+1, sched.goidgen+GoidCacheBatch].
  3327  		// At startup sched.goidgen=0, so main goroutine receives goid=1.
  3328  		_p_.goidcache = atomic.Xadd64(&sched.goidgen, _GoidCacheBatch)
  3329  		_p_.goidcache -= _GoidCacheBatch - 1
  3330  		_p_.goidcacheend = _p_.goidcache + _GoidCacheBatch
  3331  	}
  3332  	newg.goid = int64(_p_.goidcache)
  3333  	_p_.goidcache++
  3334  	if raceenabled {
  3335  		newg.racectx = racegostart(callerpc)
  3336  	}
  3337  	if trace.enabled {
  3338  		traceGoCreate(newg, newg.startpc)
  3339  	}
  3340  	runqput(_p_, newg, true)
  3341  
  3342  	if atomic.Load(&sched.npidle) != 0 && atomic.Load(&sched.nmspinning) == 0 && mainStarted {
  3343  		wakep()
  3344  	}
  3345  	_g_.m.locks--
  3346  	if _g_.m.locks == 0 && _g_.preempt { // restore the preemption request in case we've cleared it in newstack
  3347  		_g_.stackguard0 = stackPreempt
  3348  	}
  3349  }
  3350  
  3351  // Put on gfree list.
  3352  // If local list is too long, transfer a batch to the global list.
  3353  func gfput(_p_ *p, gp *g) {
  3354  	if readgstatus(gp) != _Gdead {
  3355  		throw("gfput: bad status (not Gdead)")
  3356  	}
  3357  
  3358  	stksize := gp.stack.hi - gp.stack.lo
  3359  
  3360  	if stksize != _FixedStack {
  3361  		// non-standard stack size - free it.
  3362  		stackfree(gp.stack)
  3363  		gp.stack.lo = 0
  3364  		gp.stack.hi = 0
  3365  		gp.stackguard0 = 0
  3366  	}
  3367  
  3368  	gp.schedlink.set(_p_.gfree)
  3369  	_p_.gfree = gp
  3370  	_p_.gfreecnt++
  3371  	if _p_.gfreecnt >= 64 {
  3372  		lock(&sched.gflock)
  3373  		for _p_.gfreecnt >= 32 {
  3374  			_p_.gfreecnt--
  3375  			gp = _p_.gfree
  3376  			_p_.gfree = gp.schedlink.ptr()
  3377  			if gp.stack.lo == 0 {
  3378  				gp.schedlink.set(sched.gfreeNoStack)
  3379  				sched.gfreeNoStack = gp
  3380  			} else {
  3381  				gp.schedlink.set(sched.gfreeStack)
  3382  				sched.gfreeStack = gp
  3383  			}
  3384  			sched.ngfree++
  3385  		}
  3386  		unlock(&sched.gflock)
  3387  	}
  3388  }
  3389  
  3390  // Get from gfree list.
  3391  // If local list is empty, grab a batch from global list.
  3392  func gfget(_p_ *p) *g {
  3393  retry:
  3394  	gp := _p_.gfree
  3395  	if gp == nil && (sched.gfreeStack != nil || sched.gfreeNoStack != nil) {
  3396  		lock(&sched.gflock)
  3397  		for _p_.gfreecnt < 32 {
  3398  			if sched.gfreeStack != nil {
  3399  				// Prefer Gs with stacks.
  3400  				gp = sched.gfreeStack
  3401  				sched.gfreeStack = gp.schedlink.ptr()
  3402  			} else if sched.gfreeNoStack != nil {
  3403  				gp = sched.gfreeNoStack
  3404  				sched.gfreeNoStack = gp.schedlink.ptr()
  3405  			} else {
  3406  				break
  3407  			}
  3408  			_p_.gfreecnt++
  3409  			sched.ngfree--
  3410  			gp.schedlink.set(_p_.gfree)
  3411  			_p_.gfree = gp
  3412  		}
  3413  		unlock(&sched.gflock)
  3414  		goto retry
  3415  	}
  3416  	if gp != nil {
  3417  		_p_.gfree = gp.schedlink.ptr()
  3418  		_p_.gfreecnt--
  3419  		if gp.stack.lo == 0 {
  3420  			// Stack was deallocated in gfput. Allocate a new one.
  3421  			systemstack(func() {
  3422  				gp.stack = stackalloc(_FixedStack)
  3423  			})
  3424  			gp.stackguard0 = gp.stack.lo + _StackGuard
  3425  		} else {
  3426  			if raceenabled {
  3427  				racemalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  3428  			}
  3429  			if msanenabled {
  3430  				msanmalloc(unsafe.Pointer(gp.stack.lo), gp.stack.hi-gp.stack.lo)
  3431  			}
  3432  		}
  3433  	}
  3434  	return gp
  3435  }
  3436  
  3437  // Purge all cached G's from gfree list to the global list.
  3438  func gfpurge(_p_ *p) {
  3439  	lock(&sched.gflock)
  3440  	for _p_.gfreecnt != 0 {
  3441  		_p_.gfreecnt--
  3442  		gp := _p_.gfree
  3443  		_p_.gfree = gp.schedlink.ptr()
  3444  		if gp.stack.lo == 0 {
  3445  			gp.schedlink.set(sched.gfreeNoStack)
  3446  			sched.gfreeNoStack = gp
  3447  		} else {
  3448  			gp.schedlink.set(sched.gfreeStack)
  3449  			sched.gfreeStack = gp
  3450  		}
  3451  		sched.ngfree++
  3452  	}
  3453  	unlock(&sched.gflock)
  3454  }
  3455  
  3456  // Breakpoint executes a breakpoint trap.
  3457  func Breakpoint() {
  3458  	breakpoint()
  3459  }
  3460  
  3461  // dolockOSThread is called by LockOSThread and lockOSThread below
  3462  // after they modify m.locked. Do not allow preemption during this call,
  3463  // or else the m might be different in this function than in the caller.
  3464  //go:nosplit
  3465  func dolockOSThread() {
  3466  	_g_ := getg()
  3467  	_g_.m.lockedg.set(_g_)
  3468  	_g_.lockedm.set(_g_.m)
  3469  }
  3470  
  3471  //go:nosplit
  3472  
  3473  // LockOSThread wires the calling goroutine to its current operating system thread.
  3474  // The calling goroutine will always execute in that thread,
  3475  // and no other goroutine will execute in it,
  3476  // until the calling goroutine has made as many calls to
  3477  // UnlockOSThread as to LockOSThread.
  3478  // If the calling goroutine exits without unlocking the thread,
  3479  // the thread will be terminated.
  3480  //
  3481  // A goroutine should call LockOSThread before calling OS services or
  3482  // non-Go library functions that depend on per-thread state.
  3483  func LockOSThread() {
  3484  	if atomic.Load(&newmHandoff.haveTemplateThread) == 0 && GOOS != "plan9" {
  3485  		// If we need to start a new thread from the locked
  3486  		// thread, we need the template thread. Start it now
  3487  		// while we're in a known-good state.
  3488  		startTemplateThread()
  3489  	}
  3490  	_g_ := getg()
  3491  	_g_.m.lockedExt++
  3492  	if _g_.m.lockedExt == 0 {
  3493  		_g_.m.lockedExt--
  3494  		panic("LockOSThread nesting overflow")
  3495  	}
  3496  	dolockOSThread()
  3497  }
  3498  
  3499  //go:nosplit
  3500  func lockOSThread() {
  3501  	getg().m.lockedInt++
  3502  	dolockOSThread()
  3503  }
  3504  
  3505  // dounlockOSThread is called by UnlockOSThread and unlockOSThread below
  3506  // after they update m->locked. Do not allow preemption during this call,
  3507  // or else the m might be in different in this function than in the caller.
  3508  //go:nosplit
  3509  func dounlockOSThread() {
  3510  	_g_ := getg()
  3511  	if _g_.m.lockedInt != 0 || _g_.m.lockedExt != 0 {
  3512  		return
  3513  	}
  3514  	_g_.m.lockedg = 0
  3515  	_g_.lockedm = 0
  3516  }
  3517  
  3518  //go:nosplit
  3519  
  3520  // UnlockOSThread undoes an earlier call to LockOSThread.
  3521  // If this drops the number of active LockOSThread calls on the
  3522  // calling goroutine to zero, it unwires the calling goroutine from
  3523  // its fixed operating system thread.
  3524  // If there are no active LockOSThread calls, this is a no-op.
  3525  //
  3526  // Before calling UnlockOSThread, the caller must ensure that the OS
  3527  // thread is suitable for running other goroutines. If the caller made
  3528  // any permanent changes to the state of the thread that would affect
  3529  // other goroutines, it should not call this function and thus leave
  3530  // the goroutine locked to the OS thread until the goroutine (and
  3531  // hence the thread) exits.
  3532  func UnlockOSThread() {
  3533  	_g_ := getg()
  3534  	if _g_.m.lockedExt == 0 {
  3535  		return
  3536  	}
  3537  	_g_.m.lockedExt--
  3538  	dounlockOSThread()
  3539  }
  3540  
  3541  //go:nosplit
  3542  func unlockOSThread() {
  3543  	_g_ := getg()
  3544  	if _g_.m.lockedInt == 0 {
  3545  		systemstack(badunlockosthread)
  3546  	}
  3547  	_g_.m.lockedInt--
  3548  	dounlockOSThread()
  3549  }
  3550  
  3551  func badunlockosthread() {
  3552  	throw("runtime: internal error: misuse of lockOSThread/unlockOSThread")
  3553  }
  3554  
  3555  func gcount() int32 {
  3556  	n := int32(allglen) - sched.ngfree - int32(atomic.Load(&sched.ngsys))
  3557  	for _, _p_ := range allp {
  3558  		n -= _p_.gfreecnt
  3559  	}
  3560  
  3561  	// All these variables can be changed concurrently, so the result can be inconsistent.
  3562  	// But at least the current goroutine is running.
  3563  	if n < 1 {
  3564  		n = 1
  3565  	}
  3566  	return n
  3567  }
  3568  
  3569  func mcount() int32 {
  3570  	return int32(sched.mnext - sched.nmfreed)
  3571  }
  3572  
  3573  var prof struct {
  3574  	signalLock uint32
  3575  	hz         int32
  3576  }
  3577  
  3578  func _System()                    { _System() }
  3579  func _ExternalCode()              { _ExternalCode() }
  3580  func _LostExternalCode()          { _LostExternalCode() }
  3581  func _GC()                        { _GC() }
  3582  func _LostSIGPROFDuringAtomic64() { _LostSIGPROFDuringAtomic64() }
  3583  
  3584  // Counts SIGPROFs received while in atomic64 critical section, on mips{,le}
  3585  var lostAtomic64Count uint64
  3586  
  3587  // Called if we receive a SIGPROF signal.
  3588  // Called by the signal handler, may run during STW.
  3589  //go:nowritebarrierrec
  3590  func sigprof(pc, sp, lr uintptr, gp *g, mp *m) {
  3591  	if prof.hz == 0 {
  3592  		return
  3593  	}
  3594  
  3595  	// On mips{,le}, 64bit atomics are emulated with spinlocks, in
  3596  	// runtime/internal/atomic. If SIGPROF arrives while the program is inside
  3597  	// the critical section, it creates a deadlock (when writing the sample).
  3598  	// As a workaround, create a counter of SIGPROFs while in critical section
  3599  	// to store the count, and pass it to sigprof.add() later when SIGPROF is
  3600  	// received from somewhere else (with _LostSIGPROFDuringAtomic64 as pc).
  3601  	if GOARCH == "mips" || GOARCH == "mipsle" {
  3602  		if f := findfunc(pc); f.valid() {
  3603  			if hasprefix(funcname(f), "runtime/internal/atomic") {
  3604  				lostAtomic64Count++
  3605  				return
  3606  			}
  3607  		}
  3608  	}
  3609  
  3610  	// Profiling runs concurrently with GC, so it must not allocate.
  3611  	// Set a trap in case the code does allocate.
  3612  	// Note that on windows, one thread takes profiles of all the
  3613  	// other threads, so mp is usually not getg().m.
  3614  	// In fact mp may not even be stopped.
  3615  	// See golang.org/issue/17165.
  3616  	getg().m.mallocing++
  3617  
  3618  	// Define that a "user g" is a user-created goroutine, and a "system g"
  3619  	// is one that is m->g0 or m->gsignal.
  3620  	//
  3621  	// We might be interrupted for profiling halfway through a
  3622  	// goroutine switch. The switch involves updating three (or four) values:
  3623  	// g, PC, SP, and (on arm) LR. The PC must be the last to be updated,
  3624  	// because once it gets updated the new g is running.
  3625  	//
  3626  	// When switching from a user g to a system g, LR is not considered live,
  3627  	// so the update only affects g, SP, and PC. Since PC must be last, there
  3628  	// the possible partial transitions in ordinary execution are (1) g alone is updated,
  3629  	// (2) both g and SP are updated, and (3) SP alone is updated.
  3630  	// If SP or g alone is updated, we can detect the partial transition by checking
  3631  	// whether the SP is within g's stack bounds. (We could also require that SP
  3632  	// be changed only after g, but the stack bounds check is needed by other
  3633  	// cases, so there is no need to impose an additional requirement.)
  3634  	//
  3635  	// There is one exceptional transition to a system g, not in ordinary execution.
  3636  	// When a signal arrives, the operating system starts the signal handler running
  3637  	// with an updated PC and SP. The g is updated last, at the beginning of the
  3638  	// handler. There are two reasons this is okay. First, until g is updated the
  3639  	// g and SP do not match, so the stack bounds check detects the partial transition.
  3640  	// Second, signal handlers currently run with signals disabled, so a profiling
  3641  	// signal cannot arrive during the handler.
  3642  	//
  3643  	// When switching from a system g to a user g, there are three possibilities.
  3644  	//
  3645  	// First, it may be that the g switch has no PC update, because the SP
  3646  	// either corresponds to a user g throughout (as in asmcgocall)
  3647  	// or because it has been arranged to look like a user g frame
  3648  	// (as in cgocallback_gofunc). In this case, since the entire
  3649  	// transition is a g+SP update, a partial transition updating just one of
  3650  	// those will be detected by the stack bounds check.
  3651  	//
  3652  	// Second, when returning from a signal handler, the PC and SP updates
  3653  	// are performed by the operating system in an atomic update, so the g
  3654  	// update must be done before them. The stack bounds check detects
  3655  	// the partial transition here, and (again) signal handlers run with signals
  3656  	// disabled, so a profiling signal cannot arrive then anyway.
  3657  	//
  3658  	// Third, the common case: it may be that the switch updates g, SP, and PC
  3659  	// separately. If the PC is within any of the functions that does this,
  3660  	// we don't ask for a traceback. C.F. the function setsSP for more about this.
  3661  	//
  3662  	// There is another apparently viable approach, recorded here in case
  3663  	// the "PC within setsSP function" check turns out not to be usable.
  3664  	// It would be possible to delay the update of either g or SP until immediately
  3665  	// before the PC update instruction. Then, because of the stack bounds check,
  3666  	// the only problematic interrupt point is just before that PC update instruction,
  3667  	// and the sigprof handler can detect that instruction and simulate stepping past
  3668  	// it in order to reach a consistent state. On ARM, the update of g must be made
  3669  	// in two places (in R10 and also in a TLS slot), so the delayed update would
  3670  	// need to be the SP update. The sigprof handler must read the instruction at
  3671  	// the current PC and if it was the known instruction (for example, JMP BX or
  3672  	// MOV R2, PC), use that other register in place of the PC value.
  3673  	// The biggest drawback to this solution is that it requires that we can tell
  3674  	// whether it's safe to read from the memory pointed at by PC.
  3675  	// In a correct program, we can test PC == nil and otherwise read,
  3676  	// but if a profiling signal happens at the instant that a program executes
  3677  	// a bad jump (before the program manages to handle the resulting fault)
  3678  	// the profiling handler could fault trying to read nonexistent memory.
  3679  	//
  3680  	// To recap, there are no constraints on the assembly being used for the
  3681  	// transition. We simply require that g and SP match and that the PC is not
  3682  	// in gogo.
  3683  	traceback := true
  3684  	if gp == nil || sp < gp.stack.lo || gp.stack.hi < sp || setsSP(pc) {
  3685  		traceback = false
  3686  	}
  3687  	var stk [maxCPUProfStack]uintptr
  3688  	n := 0
  3689  	if mp.ncgo > 0 && mp.curg != nil && mp.curg.syscallpc != 0 && mp.curg.syscallsp != 0 {
  3690  		cgoOff := 0
  3691  		// Check cgoCallersUse to make sure that we are not
  3692  		// interrupting other code that is fiddling with
  3693  		// cgoCallers.  We are running in a signal handler
  3694  		// with all signals blocked, so we don't have to worry
  3695  		// about any other code interrupting us.
  3696  		if atomic.Load(&mp.cgoCallersUse) == 0 && mp.cgoCallers != nil && mp.cgoCallers[0] != 0 {
  3697  			for cgoOff < len(mp.cgoCallers) && mp.cgoCallers[cgoOff] != 0 {
  3698  				cgoOff++
  3699  			}
  3700  			copy(stk[:], mp.cgoCallers[:cgoOff])
  3701  			mp.cgoCallers[0] = 0
  3702  		}
  3703  
  3704  		// Collect Go stack that leads to the cgo call.
  3705  		n = gentraceback(mp.curg.syscallpc, mp.curg.syscallsp, 0, mp.curg, 0, &stk[cgoOff], len(stk)-cgoOff, nil, nil, 0)
  3706  	} else if traceback {
  3707  		n = gentraceback(pc, sp, lr, gp, 0, &stk[0], len(stk), nil, nil, _TraceTrap|_TraceJumpStack)
  3708  	}
  3709  
  3710  	if n <= 0 {
  3711  		// Normal traceback is impossible or has failed.
  3712  		// See if it falls into several common cases.
  3713  		n = 0
  3714  		if GOOS == "windows" && mp.libcallg != 0 && mp.libcallpc != 0 && mp.libcallsp != 0 {
  3715  			// Libcall, i.e. runtime syscall on windows.
  3716  			// Collect Go stack that leads to the call.
  3717  			n = gentraceback(mp.libcallpc, mp.libcallsp, 0, mp.libcallg.ptr(), 0, &stk[0], len(stk), nil, nil, 0)
  3718  		}
  3719  		if n == 0 {
  3720  			// If all of the above has failed, account it against abstract "System" or "GC".
  3721  			n = 2
  3722  			// "ExternalCode" is better than "etext".
  3723  			if pc > firstmoduledata.etext {
  3724  				pc = funcPC(_ExternalCode) + sys.PCQuantum
  3725  			}
  3726  			stk[0] = pc
  3727  			if mp.preemptoff != "" || mp.helpgc != 0 {
  3728  				stk[1] = funcPC(_GC) + sys.PCQuantum
  3729  			} else {
  3730  				stk[1] = funcPC(_System) + sys.PCQuantum
  3731  			}
  3732  		}
  3733  	}
  3734  
  3735  	if prof.hz != 0 {
  3736  		if (GOARCH == "mips" || GOARCH == "mipsle") && lostAtomic64Count > 0 {
  3737  			cpuprof.addLostAtomic64(lostAtomic64Count)
  3738  			lostAtomic64Count = 0
  3739  		}
  3740  		cpuprof.add(gp, stk[:n])
  3741  	}
  3742  	getg().m.mallocing--
  3743  }
  3744  
  3745  // If the signal handler receives a SIGPROF signal on a non-Go thread,
  3746  // it tries to collect a traceback into sigprofCallers.
  3747  // sigprofCallersUse is set to non-zero while sigprofCallers holds a traceback.
  3748  var sigprofCallers cgoCallers
  3749  var sigprofCallersUse uint32
  3750  
  3751  // sigprofNonGo is called if we receive a SIGPROF signal on a non-Go thread,
  3752  // and the signal handler collected a stack trace in sigprofCallers.
  3753  // When this is called, sigprofCallersUse will be non-zero.
  3754  // g is nil, and what we can do is very limited.
  3755  //go:nosplit
  3756  //go:nowritebarrierrec
  3757  func sigprofNonGo() {
  3758  	if prof.hz != 0 {
  3759  		n := 0
  3760  		for n < len(sigprofCallers) && sigprofCallers[n] != 0 {
  3761  			n++
  3762  		}
  3763  		cpuprof.addNonGo(sigprofCallers[:n])
  3764  	}
  3765  
  3766  	atomic.Store(&sigprofCallersUse, 0)
  3767  }
  3768  
  3769  // sigprofNonGoPC is called when a profiling signal arrived on a
  3770  // non-Go thread and we have a single PC value, not a stack trace.
  3771  // g is nil, and what we can do is very limited.
  3772  //go:nosplit
  3773  //go:nowritebarrierrec
  3774  func sigprofNonGoPC(pc uintptr) {
  3775  	if prof.hz != 0 {
  3776  		stk := []uintptr{
  3777  			pc,
  3778  			funcPC(_ExternalCode) + sys.PCQuantum,
  3779  		}
  3780  		cpuprof.addNonGo(stk)
  3781  	}
  3782  }
  3783  
  3784  // Reports whether a function will set the SP
  3785  // to an absolute value. Important that
  3786  // we don't traceback when these are at the bottom
  3787  // of the stack since we can't be sure that we will
  3788  // find the caller.
  3789  //
  3790  // If the function is not on the bottom of the stack
  3791  // we assume that it will have set it up so that traceback will be consistent,
  3792  // either by being a traceback terminating function
  3793  // or putting one on the stack at the right offset.
  3794  func setsSP(pc uintptr) bool {
  3795  	f := findfunc(pc)
  3796  	if !f.valid() {
  3797  		// couldn't find the function for this PC,
  3798  		// so assume the worst and stop traceback
  3799  		return true
  3800  	}
  3801  	switch f.entry {
  3802  	case gogoPC, systemstackPC, mcallPC, morestackPC:
  3803  		return true
  3804  	}
  3805  	return false
  3806  }
  3807  
  3808  // setcpuprofilerate sets the CPU profiling rate to hz times per second.
  3809  // If hz <= 0, setcpuprofilerate turns off CPU profiling.
  3810  func setcpuprofilerate(hz int32) {
  3811  	// Force sane arguments.
  3812  	if hz < 0 {
  3813  		hz = 0
  3814  	}
  3815  
  3816  	// Disable preemption, otherwise we can be rescheduled to another thread
  3817  	// that has profiling enabled.
  3818  	_g_ := getg()
  3819  	_g_.m.locks++
  3820  
  3821  	// Stop profiler on this thread so that it is safe to lock prof.
  3822  	// if a profiling signal came in while we had prof locked,
  3823  	// it would deadlock.
  3824  	setThreadCPUProfiler(0)
  3825  
  3826  	for !atomic.Cas(&prof.signalLock, 0, 1) {
  3827  		osyield()
  3828  	}
  3829  	if prof.hz != hz {
  3830  		setProcessCPUProfiler(hz)
  3831  		prof.hz = hz
  3832  	}
  3833  	atomic.Store(&prof.signalLock, 0)
  3834  
  3835  	lock(&sched.lock)
  3836  	sched.profilehz = hz
  3837  	unlock(&sched.lock)
  3838  
  3839  	if hz != 0 {
  3840  		setThreadCPUProfiler(hz)
  3841  	}
  3842  
  3843  	_g_.m.locks--
  3844  }
  3845  
  3846  // Change number of processors. The world is stopped, sched is locked.
  3847  // gcworkbufs are not being modified by either the GC or
  3848  // the write barrier code.
  3849  // Returns list of Ps with local work, they need to be scheduled by the caller.
  3850  func procresize(nprocs int32) *p {
  3851  	old := gomaxprocs
  3852  	if old < 0 || nprocs <= 0 {
  3853  		throw("procresize: invalid arg")
  3854  	}
  3855  	if trace.enabled {
  3856  		traceGomaxprocs(nprocs)
  3857  	}
  3858  
  3859  	// update statistics
  3860  	now := nanotime()
  3861  	if sched.procresizetime != 0 {
  3862  		sched.totaltime += int64(old) * (now - sched.procresizetime)
  3863  	}
  3864  	sched.procresizetime = now
  3865  
  3866  	// Grow allp if necessary.
  3867  	if nprocs > int32(len(allp)) {
  3868  		// Synchronize with retake, which could be running
  3869  		// concurrently since it doesn't run on a P.
  3870  		lock(&allpLock)
  3871  		if nprocs <= int32(cap(allp)) {
  3872  			allp = allp[:nprocs]
  3873  		} else {
  3874  			nallp := make([]*p, nprocs)
  3875  			// Copy everything up to allp's cap so we
  3876  			// never lose old allocated Ps.
  3877  			copy(nallp, allp[:cap(allp)])
  3878  			allp = nallp
  3879  		}
  3880  		unlock(&allpLock)
  3881  	}
  3882  
  3883  	// initialize new P's
  3884  	for i := int32(0); i < nprocs; i++ {
  3885  		pp := allp[i]
  3886  		if pp == nil {
  3887  			pp = new(p)
  3888  			pp.id = i
  3889  			pp.status = _Pgcstop
  3890  			pp.sudogcache = pp.sudogbuf[:0]
  3891  			for i := range pp.deferpool {
  3892  				pp.deferpool[i] = pp.deferpoolbuf[i][:0]
  3893  			}
  3894  			pp.wbBuf.reset()
  3895  			atomicstorep(unsafe.Pointer(&allp[i]), unsafe.Pointer(pp))
  3896  		}
  3897  		if pp.mcache == nil {
  3898  			if old == 0 && i == 0 {
  3899  				if getg().m.mcache == nil {
  3900  					throw("missing mcache?")
  3901  				}
  3902  				pp.mcache = getg().m.mcache // bootstrap
  3903  			} else {
  3904  				pp.mcache = allocmcache()
  3905  			}
  3906  		}
  3907  		if raceenabled && pp.racectx == 0 {
  3908  			if old == 0 && i == 0 {
  3909  				pp.racectx = raceprocctx0
  3910  				raceprocctx0 = 0 // bootstrap
  3911  			} else {
  3912  				pp.racectx = raceproccreate()
  3913  			}
  3914  		}
  3915  	}
  3916  
  3917  	// free unused P's
  3918  	for i := nprocs; i < old; i++ {
  3919  		p := allp[i]
  3920  		if trace.enabled && p == getg().m.p.ptr() {
  3921  			// moving to p[0], pretend that we were descheduled
  3922  			// and then scheduled again to keep the trace sane.
  3923  			traceGoSched()
  3924  			traceProcStop(p)
  3925  		}
  3926  		// move all runnable goroutines to the global queue
  3927  		for p.runqhead != p.runqtail {
  3928  			// pop from tail of local queue
  3929  			p.runqtail--
  3930  			gp := p.runq[p.runqtail%uint32(len(p.runq))].ptr()
  3931  			// push onto head of global queue
  3932  			globrunqputhead(gp)
  3933  		}
  3934  		if p.runnext != 0 {
  3935  			globrunqputhead(p.runnext.ptr())
  3936  			p.runnext = 0
  3937  		}
  3938  		// if there's a background worker, make it runnable and put
  3939  		// it on the global queue so it can clean itself up
  3940  		if gp := p.gcBgMarkWorker.ptr(); gp != nil {
  3941  			casgstatus(gp, _Gwaiting, _Grunnable)
  3942  			if trace.enabled {
  3943  				traceGoUnpark(gp, 0)
  3944  			}
  3945  			globrunqput(gp)
  3946  			// This assignment doesn't race because the
  3947  			// world is stopped.
  3948  			p.gcBgMarkWorker.set(nil)
  3949  		}
  3950  		// Flush p's write barrier buffer.
  3951  		if gcphase != _GCoff {
  3952  			wbBufFlush1(p)
  3953  			p.gcw.dispose()
  3954  		}
  3955  		for i := range p.sudogbuf {
  3956  			p.sudogbuf[i] = nil
  3957  		}
  3958  		p.sudogcache = p.sudogbuf[:0]
  3959  		for i := range p.deferpool {
  3960  			for j := range p.deferpoolbuf[i] {
  3961  				p.deferpoolbuf[i][j] = nil
  3962  			}
  3963  			p.deferpool[i] = p.deferpoolbuf[i][:0]
  3964  		}
  3965  		freemcache(p.mcache)
  3966  		p.mcache = nil
  3967  		gfpurge(p)
  3968  		traceProcFree(p)
  3969  		if raceenabled {
  3970  			raceprocdestroy(p.racectx)
  3971  			p.racectx = 0
  3972  		}
  3973  		p.gcAssistTime = 0
  3974  		p.status = _Pdead
  3975  		// can't free P itself because it can be referenced by an M in syscall
  3976  	}
  3977  
  3978  	// Trim allp.
  3979  	if int32(len(allp)) != nprocs {
  3980  		lock(&allpLock)
  3981  		allp = allp[:nprocs]
  3982  		unlock(&allpLock)
  3983  	}
  3984  
  3985  	_g_ := getg()
  3986  	if _g_.m.p != 0 && _g_.m.p.ptr().id < nprocs {
  3987  		// continue to use the current P
  3988  		_g_.m.p.ptr().status = _Prunning
  3989  	} else {
  3990  		// release the current P and acquire allp[0]
  3991  		if _g_.m.p != 0 {
  3992  			_g_.m.p.ptr().m = 0
  3993  		}
  3994  		_g_.m.p = 0
  3995  		_g_.m.mcache = nil
  3996  		p := allp[0]
  3997  		p.m = 0
  3998  		p.status = _Pidle
  3999  		acquirep(p)
  4000  		if trace.enabled {
  4001  			traceGoStart()
  4002  		}
  4003  	}
  4004  	var runnablePs *p
  4005  	for i := nprocs - 1; i >= 0; i-- {
  4006  		p := allp[i]
  4007  		if _g_.m.p.ptr() == p {
  4008  			continue
  4009  		}
  4010  		p.status = _Pidle
  4011  		if runqempty(p) {
  4012  			pidleput(p)
  4013  		} else {
  4014  			p.m.set(mget())
  4015  			p.link.set(runnablePs)
  4016  			runnablePs = p
  4017  		}
  4018  	}
  4019  	stealOrder.reset(uint32(nprocs))
  4020  	var int32p *int32 = &gomaxprocs // make compiler check that gomaxprocs is an int32
  4021  	atomic.Store((*uint32)(unsafe.Pointer(int32p)), uint32(nprocs))
  4022  	return runnablePs
  4023  }
  4024  
  4025  // Associate p and the current m.
  4026  //
  4027  // This function is allowed to have write barriers even if the caller
  4028  // isn't because it immediately acquires _p_.
  4029  //
  4030  //go:yeswritebarrierrec
  4031  func acquirep(_p_ *p) {
  4032  	// Do the part that isn't allowed to have write barriers.
  4033  	acquirep1(_p_)
  4034  
  4035  	// have p; write barriers now allowed
  4036  	_g_ := getg()
  4037  	_g_.m.mcache = _p_.mcache
  4038  
  4039  	if trace.enabled {
  4040  		traceProcStart()
  4041  	}
  4042  }
  4043  
  4044  // acquirep1 is the first step of acquirep, which actually acquires
  4045  // _p_. This is broken out so we can disallow write barriers for this
  4046  // part, since we don't yet have a P.
  4047  //
  4048  //go:nowritebarrierrec
  4049  func acquirep1(_p_ *p) {
  4050  	_g_ := getg()
  4051  
  4052  	if _g_.m.p != 0 || _g_.m.mcache != nil {
  4053  		throw("acquirep: already in go")
  4054  	}
  4055  	if _p_.m != 0 || _p_.status != _Pidle {
  4056  		id := int64(0)
  4057  		if _p_.m != 0 {
  4058  			id = _p_.m.ptr().id
  4059  		}
  4060  		print("acquirep: p->m=", _p_.m, "(", id, ") p->status=", _p_.status, "\n")
  4061  		throw("acquirep: invalid p state")
  4062  	}
  4063  	_g_.m.p.set(_p_)
  4064  	_p_.m.set(_g_.m)
  4065  	_p_.status = _Prunning
  4066  }
  4067  
  4068  // Disassociate p and the current m.
  4069  func releasep() *p {
  4070  	_g_ := getg()
  4071  
  4072  	if _g_.m.p == 0 || _g_.m.mcache == nil {
  4073  		throw("releasep: invalid arg")
  4074  	}
  4075  	_p_ := _g_.m.p.ptr()
  4076  	if _p_.m.ptr() != _g_.m || _p_.mcache != _g_.m.mcache || _p_.status != _Prunning {
  4077  		print("releasep: m=", _g_.m, " m->p=", _g_.m.p.ptr(), " p->m=", _p_.m, " m->mcache=", _g_.m.mcache, " p->mcache=", _p_.mcache, " p->status=", _p_.status, "\n")
  4078  		throw("releasep: invalid p state")
  4079  	}
  4080  	if trace.enabled {
  4081  		traceProcStop(_g_.m.p.ptr())
  4082  	}
  4083  	_g_.m.p = 0
  4084  	_g_.m.mcache = nil
  4085  	_p_.m = 0
  4086  	_p_.status = _Pidle
  4087  	return _p_
  4088  }
  4089  
  4090  func incidlelocked(v int32) {
  4091  	lock(&sched.lock)
  4092  	sched.nmidlelocked += v
  4093  	if v > 0 {
  4094  		checkdead()
  4095  	}
  4096  	unlock(&sched.lock)
  4097  }
  4098  
  4099  // Check for deadlock situation.
  4100  // The check is based on number of running M's, if 0 -> deadlock.
  4101  // sched.lock must be held.
  4102  func checkdead() {
  4103  	// For -buildmode=c-shared or -buildmode=c-archive it's OK if
  4104  	// there are no running goroutines. The calling program is
  4105  	// assumed to be running.
  4106  	if islibrary || isarchive {
  4107  		return
  4108  	}
  4109  
  4110  	// If we are dying because of a signal caught on an already idle thread,
  4111  	// freezetheworld will cause all running threads to block.
  4112  	// And runtime will essentially enter into deadlock state,
  4113  	// except that there is a thread that will call exit soon.
  4114  	if panicking > 0 {
  4115  		return
  4116  	}
  4117  
  4118  	run := mcount() - sched.nmidle - sched.nmidlelocked - sched.nmsys
  4119  	if run > 0 {
  4120  		return
  4121  	}
  4122  	if run < 0 {
  4123  		print("runtime: checkdead: nmidle=", sched.nmidle, " nmidlelocked=", sched.nmidlelocked, " mcount=", mcount(), " nmsys=", sched.nmsys, "\n")
  4124  		throw("checkdead: inconsistent counts")
  4125  	}
  4126  
  4127  	grunning := 0
  4128  	lock(&allglock)
  4129  	for i := 0; i < len(allgs); i++ {
  4130  		gp := allgs[i]
  4131  		if isSystemGoroutine(gp) {
  4132  			continue
  4133  		}
  4134  		s := readgstatus(gp)
  4135  		switch s &^ _Gscan {
  4136  		case _Gwaiting:
  4137  			grunning++
  4138  		case _Grunnable,
  4139  			_Grunning,
  4140  			_Gsyscall:
  4141  			unlock(&allglock)
  4142  			print("runtime: checkdead: find g ", gp.goid, " in status ", s, "\n")
  4143  			throw("checkdead: runnable g")
  4144  		}
  4145  	}
  4146  	unlock(&allglock)
  4147  	if grunning == 0 { // possible if main goroutine calls runtime·Goexit()
  4148  		throw("no goroutines (main called runtime.Goexit) - deadlock!")
  4149  	}
  4150  
  4151  	// Maybe jump time forward for playground.
  4152  	gp := timejump()
  4153  	if gp != nil {
  4154  		casgstatus(gp, _Gwaiting, _Grunnable)
  4155  		globrunqput(gp)
  4156  		_p_ := pidleget()
  4157  		if _p_ == nil {
  4158  			throw("checkdead: no p for timer")
  4159  		}
  4160  		mp := mget()
  4161  		if mp == nil {
  4162  			// There should always be a free M since
  4163  			// nothing is running.
  4164  			throw("checkdead: no m for timer")
  4165  		}
  4166  		mp.nextp.set(_p_)
  4167  		notewakeup(&mp.park)
  4168  		return
  4169  	}
  4170  
  4171  	getg().m.throwing = -1 // do not dump full stacks
  4172  	throw("all goroutines are asleep - deadlock!")
  4173  }
  4174  
  4175  // forcegcperiod is the maximum time in nanoseconds between garbage
  4176  // collections. If we go this long without a garbage collection, one
  4177  // is forced to run.
  4178  //
  4179  // This is a variable for testing purposes. It normally doesn't change.
  4180  var forcegcperiod int64 = 2 * 60 * 1e9
  4181  
  4182  // Always runs without a P, so write barriers are not allowed.
  4183  //
  4184  //go:nowritebarrierrec
  4185  func sysmon() {
  4186  	lock(&sched.lock)
  4187  	sched.nmsys++
  4188  	checkdead()
  4189  	unlock(&sched.lock)
  4190  
  4191  	// If a heap span goes unused for 5 minutes after a garbage collection,
  4192  	// we hand it back to the operating system.
  4193  	scavengelimit := int64(5 * 60 * 1e9)
  4194  
  4195  	if debug.scavenge > 0 {
  4196  		// Scavenge-a-lot for testing.
  4197  		forcegcperiod = 10 * 1e6
  4198  		scavengelimit = 20 * 1e6
  4199  	}
  4200  
  4201  	lastscavenge := nanotime()
  4202  	nscavenge := 0
  4203  
  4204  	lasttrace := int64(0)
  4205  	idle := 0 // how many cycles in succession we had not wokeup somebody
  4206  	delay := uint32(0)
  4207  	for {
  4208  		if idle == 0 { // start with 20us sleep...
  4209  			delay = 20
  4210  		} else if idle > 50 { // start doubling the sleep after 1ms...
  4211  			delay *= 2
  4212  		}
  4213  		if delay > 10*1000 { // up to 10ms
  4214  			delay = 10 * 1000
  4215  		}
  4216  		usleep(delay)
  4217  		if debug.schedtrace <= 0 && (sched.gcwaiting != 0 || atomic.Load(&sched.npidle) == uint32(gomaxprocs)) {
  4218  			lock(&sched.lock)
  4219  			if atomic.Load(&sched.gcwaiting) != 0 || atomic.Load(&sched.npidle) == uint32(gomaxprocs) {
  4220  				atomic.Store(&sched.sysmonwait, 1)
  4221  				unlock(&sched.lock)
  4222  				// Make wake-up period small enough
  4223  				// for the sampling to be correct.
  4224  				maxsleep := forcegcperiod / 2
  4225  				if scavengelimit < forcegcperiod {
  4226  					maxsleep = scavengelimit / 2
  4227  				}
  4228  				shouldRelax := true
  4229  				if osRelaxMinNS > 0 {
  4230  					next := timeSleepUntil()
  4231  					now := nanotime()
  4232  					if next-now < osRelaxMinNS {
  4233  						shouldRelax = false
  4234  					}
  4235  				}
  4236  				if shouldRelax {
  4237  					osRelax(true)
  4238  				}
  4239  				notetsleep(&sched.sysmonnote, maxsleep)
  4240  				if shouldRelax {
  4241  					osRelax(false)
  4242  				}
  4243  				lock(&sched.lock)
  4244  				atomic.Store(&sched.sysmonwait, 0)
  4245  				noteclear(&sched.sysmonnote)
  4246  				idle = 0
  4247  				delay = 20
  4248  			}
  4249  			unlock(&sched.lock)
  4250  		}
  4251  		// trigger libc interceptors if needed
  4252  		if *cgo_yield != nil {
  4253  			asmcgocall(*cgo_yield, nil)
  4254  		}
  4255  		// poll network if not polled for more than 10ms
  4256  		lastpoll := int64(atomic.Load64(&sched.lastpoll))
  4257  		now := nanotime()
  4258  		if netpollinited() && lastpoll != 0 && lastpoll+10*1000*1000 < now {
  4259  			atomic.Cas64(&sched.lastpoll, uint64(lastpoll), uint64(now))
  4260  			gp := netpoll(false) // non-blocking - returns list of goroutines
  4261  			if gp != nil {
  4262  				// Need to decrement number of idle locked M's
  4263  				// (pretending that one more is running) before injectglist.
  4264  				// Otherwise it can lead to the following situation:
  4265  				// injectglist grabs all P's but before it starts M's to run the P's,
  4266  				// another M returns from syscall, finishes running its G,
  4267  				// observes that there is no work to do and no other running M's
  4268  				// and reports deadlock.
  4269  				incidlelocked(-1)
  4270  				injectglist(gp)
  4271  				incidlelocked(1)
  4272  			}
  4273  		}
  4274  		// retake P's blocked in syscalls
  4275  		// and preempt long running G's
  4276  		if retake(now) != 0 {
  4277  			idle = 0
  4278  		} else {
  4279  			idle++
  4280  		}
  4281  		// check if we need to force a GC
  4282  		if t := (gcTrigger{kind: gcTriggerTime, now: now}); t.test() && atomic.Load(&forcegc.idle) != 0 {
  4283  			lock(&forcegc.lock)
  4284  			forcegc.idle = 0
  4285  			forcegc.g.schedlink = 0
  4286  			injectglist(forcegc.g)
  4287  			unlock(&forcegc.lock)
  4288  		}
  4289  		// scavenge heap once in a while
  4290  		if lastscavenge+scavengelimit/2 < now {
  4291  			mheap_.scavenge(int32(nscavenge), uint64(now), uint64(scavengelimit))
  4292  			lastscavenge = now
  4293  			nscavenge++
  4294  		}
  4295  		if debug.schedtrace > 0 && lasttrace+int64(debug.schedtrace)*1000000 <= now {
  4296  			lasttrace = now
  4297  			schedtrace(debug.scheddetail > 0)
  4298  		}
  4299  	}
  4300  }
  4301  
  4302  type sysmontick struct {
  4303  	schedtick   uint32
  4304  	schedwhen   int64
  4305  	syscalltick uint32
  4306  	syscallwhen int64
  4307  }
  4308  
  4309  // forcePreemptNS is the time slice given to a G before it is
  4310  // preempted.
  4311  const forcePreemptNS = 10 * 1000 * 1000 // 10ms
  4312  
  4313  func retake(now int64) uint32 {
  4314  	n := 0
  4315  	// Prevent allp slice changes. This lock will be completely
  4316  	// uncontended unless we're already stopping the world.
  4317  	lock(&allpLock)
  4318  	// We can't use a range loop over allp because we may
  4319  	// temporarily drop the allpLock. Hence, we need to re-fetch
  4320  	// allp each time around the loop.
  4321  	for i := 0; i < len(allp); i++ {
  4322  		_p_ := allp[i]
  4323  		if _p_ == nil {
  4324  			// This can happen if procresize has grown
  4325  			// allp but not yet created new Ps.
  4326  			continue
  4327  		}
  4328  		pd := &_p_.sysmontick
  4329  		s := _p_.status
  4330  		if s == _Psyscall {
  4331  			// Retake P from syscall if it's there for more than 1 sysmon tick (at least 20us).
  4332  			t := int64(_p_.syscalltick)
  4333  			if int64(pd.syscalltick) != t {
  4334  				pd.syscalltick = uint32(t)
  4335  				pd.syscallwhen = now
  4336  				continue
  4337  			}
  4338  			// On the one hand we don't want to retake Ps if there is no other work to do,
  4339  			// but on the other hand we want to retake them eventually
  4340  			// because they can prevent the sysmon thread from deep sleep.
  4341  			if runqempty(_p_) && atomic.Load(&sched.nmspinning)+atomic.Load(&sched.npidle) > 0 && pd.syscallwhen+10*1000*1000 > now {
  4342  				continue
  4343  			}
  4344  			// Drop allpLock so we can take sched.lock.
  4345  			unlock(&allpLock)
  4346  			// Need to decrement number of idle locked M's
  4347  			// (pretending that one more is running) before the CAS.
  4348  			// Otherwise the M from which we retake can exit the syscall,
  4349  			// increment nmidle and report deadlock.
  4350  			incidlelocked(-1)
  4351  			if atomic.Cas(&_p_.status, s, _Pidle) {
  4352  				if trace.enabled {
  4353  					traceGoSysBlock(_p_)
  4354  					traceProcStop(_p_)
  4355  				}
  4356  				n++
  4357  				_p_.syscalltick++
  4358  				handoffp(_p_)
  4359  			}
  4360  			incidlelocked(1)
  4361  			lock(&allpLock)
  4362  		} else if s == _Prunning {
  4363  			// Preempt G if it's running for too long.
  4364  			t := int64(_p_.schedtick)
  4365  			if int64(pd.schedtick) != t {
  4366  				pd.schedtick = uint32(t)
  4367  				pd.schedwhen = now
  4368  				continue
  4369  			}
  4370  			if pd.schedwhen+forcePreemptNS > now {
  4371  				continue
  4372  			}
  4373  			preemptone(_p_)
  4374  		}
  4375  	}
  4376  	unlock(&allpLock)
  4377  	return uint32(n)
  4378  }
  4379  
  4380  // Tell all goroutines that they have been preempted and they should stop.
  4381  // This function is purely best-effort. It can fail to inform a goroutine if a
  4382  // processor just started running it.
  4383  // No locks need to be held.
  4384  // Returns true if preemption request was issued to at least one goroutine.
  4385  func preemptall() bool {
  4386  	res := false
  4387  	for _, _p_ := range allp {
  4388  		if _p_.status != _Prunning {
  4389  			continue
  4390  		}
  4391  		if preemptone(_p_) {
  4392  			res = true
  4393  		}
  4394  	}
  4395  	return res
  4396  }
  4397  
  4398  // Tell the goroutine running on processor P to stop.
  4399  // This function is purely best-effort. It can incorrectly fail to inform the
  4400  // goroutine. It can send inform the wrong goroutine. Even if it informs the
  4401  // correct goroutine, that goroutine might ignore the request if it is
  4402  // simultaneously executing newstack.
  4403  // No lock needs to be held.
  4404  // Returns true if preemption request was issued.
  4405  // The actual preemption will happen at some point in the future
  4406  // and will be indicated by the gp->status no longer being
  4407  // Grunning
  4408  func preemptone(_p_ *p) bool {
  4409  	mp := _p_.m.ptr()
  4410  	if mp == nil || mp == getg().m {
  4411  		return false
  4412  	}
  4413  	gp := mp.curg
  4414  	if gp == nil || gp == mp.g0 {
  4415  		return false
  4416  	}
  4417  
  4418  	gp.preempt = true
  4419  
  4420  	// Every call in a go routine checks for stack overflow by
  4421  	// comparing the current stack pointer to gp->stackguard0.
  4422  	// Setting gp->stackguard0 to StackPreempt folds
  4423  	// preemption into the normal stack overflow check.
  4424  	gp.stackguard0 = stackPreempt
  4425  	return true
  4426  }
  4427  
  4428  var starttime int64
  4429  
  4430  func schedtrace(detailed bool) {
  4431  	now := nanotime()
  4432  	if starttime == 0 {
  4433  		starttime = now
  4434  	}
  4435  
  4436  	lock(&sched.lock)
  4437  	print("SCHED ", (now-starttime)/1e6, "ms: gomaxprocs=", gomaxprocs, " idleprocs=", sched.npidle, " threads=", mcount(), " spinningthreads=", sched.nmspinning, " idlethreads=", sched.nmidle, " runqueue=", sched.runqsize)
  4438  	if detailed {
  4439  		print(" gcwaiting=", sched.gcwaiting, " nmidlelocked=", sched.nmidlelocked, " stopwait=", sched.stopwait, " sysmonwait=", sched.sysmonwait, "\n")
  4440  	}
  4441  	// We must be careful while reading data from P's, M's and G's.
  4442  	// Even if we hold schedlock, most data can be changed concurrently.
  4443  	// E.g. (p->m ? p->m->id : -1) can crash if p->m changes from non-nil to nil.
  4444  	for i, _p_ := range allp {
  4445  		mp := _p_.m.ptr()
  4446  		h := atomic.Load(&_p_.runqhead)
  4447  		t := atomic.Load(&_p_.runqtail)
  4448  		if detailed {
  4449  			id := int64(-1)
  4450  			if mp != nil {
  4451  				id = mp.id
  4452  			}
  4453  			print("  P", i, ": status=", _p_.status, " schedtick=", _p_.schedtick, " syscalltick=", _p_.syscalltick, " m=", id, " runqsize=", t-h, " gfreecnt=", _p_.gfreecnt, "\n")
  4454  		} else {
  4455  			// In non-detailed mode format lengths of per-P run queues as:
  4456  			// [len1 len2 len3 len4]
  4457  			print(" ")
  4458  			if i == 0 {
  4459  				print("[")
  4460  			}
  4461  			print(t - h)
  4462  			if i == len(allp)-1 {
  4463  				print("]\n")
  4464  			}
  4465  		}
  4466  	}
  4467  
  4468  	if !detailed {
  4469  		unlock(&sched.lock)
  4470  		return
  4471  	}
  4472  
  4473  	for mp := allm; mp != nil; mp = mp.alllink {
  4474  		_p_ := mp.p.ptr()
  4475  		gp := mp.curg
  4476  		lockedg := mp.lockedg.ptr()
  4477  		id1 := int32(-1)
  4478  		if _p_ != nil {
  4479  			id1 = _p_.id
  4480  		}
  4481  		id2 := int64(-1)
  4482  		if gp != nil {
  4483  			id2 = gp.goid
  4484  		}
  4485  		id3 := int64(-1)
  4486  		if lockedg != nil {
  4487  			id3 = lockedg.goid
  4488  		}
  4489  		print("  M", mp.id, ": p=", id1, " curg=", id2, " mallocing=", mp.mallocing, " throwing=", mp.throwing, " preemptoff=", mp.preemptoff, ""+" locks=", mp.locks, " dying=", mp.dying, " helpgc=", mp.helpgc, " spinning=", mp.spinning, " blocked=", mp.blocked, " lockedg=", id3, "\n")
  4490  	}
  4491  
  4492  	lock(&allglock)
  4493  	for gi := 0; gi < len(allgs); gi++ {
  4494  		gp := allgs[gi]
  4495  		mp := gp.m
  4496  		lockedm := gp.lockedm.ptr()
  4497  		id1 := int64(-1)
  4498  		if mp != nil {
  4499  			id1 = mp.id
  4500  		}
  4501  		id2 := int64(-1)
  4502  		if lockedm != nil {
  4503  			id2 = lockedm.id
  4504  		}
  4505  		print("  G", gp.goid, ": status=", readgstatus(gp), "(", gp.waitreason, ") m=", id1, " lockedm=", id2, "\n")
  4506  	}
  4507  	unlock(&allglock)
  4508  	unlock(&sched.lock)
  4509  }
  4510  
  4511  // Put mp on midle list.
  4512  // Sched must be locked.
  4513  // May run during STW, so write barriers are not allowed.
  4514  //go:nowritebarrierrec
  4515  func mput(mp *m) {
  4516  	mp.schedlink = sched.midle
  4517  	sched.midle.set(mp)
  4518  	sched.nmidle++
  4519  	checkdead()
  4520  }
  4521  
  4522  // Try to get an m from midle list.
  4523  // Sched must be locked.
  4524  // May run during STW, so write barriers are not allowed.
  4525  //go:nowritebarrierrec
  4526  func mget() *m {
  4527  	mp := sched.midle.ptr()
  4528  	if mp != nil {
  4529  		sched.midle = mp.schedlink
  4530  		sched.nmidle--
  4531  	}
  4532  	return mp
  4533  }
  4534  
  4535  // Put gp on the global runnable queue.
  4536  // Sched must be locked.
  4537  // May run during STW, so write barriers are not allowed.
  4538  //go:nowritebarrierrec
  4539  func globrunqput(gp *g) {
  4540  	gp.schedlink = 0
  4541  	if sched.runqtail != 0 {
  4542  		sched.runqtail.ptr().schedlink.set(gp)
  4543  	} else {
  4544  		sched.runqhead.set(gp)
  4545  	}
  4546  	sched.runqtail.set(gp)
  4547  	sched.runqsize++
  4548  }
  4549  
  4550  // Put gp at the head of the global runnable queue.
  4551  // Sched must be locked.
  4552  // May run during STW, so write barriers are not allowed.
  4553  //go:nowritebarrierrec
  4554  func globrunqputhead(gp *g) {
  4555  	gp.schedlink = sched.runqhead
  4556  	sched.runqhead.set(gp)
  4557  	if sched.runqtail == 0 {
  4558  		sched.runqtail.set(gp)
  4559  	}
  4560  	sched.runqsize++
  4561  }
  4562  
  4563  // Put a batch of runnable goroutines on the global runnable queue.
  4564  // Sched must be locked.
  4565  func globrunqputbatch(ghead *g, gtail *g, n int32) {
  4566  	gtail.schedlink = 0
  4567  	if sched.runqtail != 0 {
  4568  		sched.runqtail.ptr().schedlink.set(ghead)
  4569  	} else {
  4570  		sched.runqhead.set(ghead)
  4571  	}
  4572  	sched.runqtail.set(gtail)
  4573  	sched.runqsize += n
  4574  }
  4575  
  4576  // Try get a batch of G's from the global runnable queue.
  4577  // Sched must be locked.
  4578  func globrunqget(_p_ *p, max int32) *g {
  4579  	if sched.runqsize == 0 {
  4580  		return nil
  4581  	}
  4582  
  4583  	n := sched.runqsize/gomaxprocs + 1
  4584  	if n > sched.runqsize {
  4585  		n = sched.runqsize
  4586  	}
  4587  	if max > 0 && n > max {
  4588  		n = max
  4589  	}
  4590  	if n > int32(len(_p_.runq))/2 {
  4591  		n = int32(len(_p_.runq)) / 2
  4592  	}
  4593  
  4594  	sched.runqsize -= n
  4595  	if sched.runqsize == 0 {
  4596  		sched.runqtail = 0
  4597  	}
  4598  
  4599  	gp := sched.runqhead.ptr()
  4600  	sched.runqhead = gp.schedlink
  4601  	n--
  4602  	for ; n > 0; n-- {
  4603  		gp1 := sched.runqhead.ptr()
  4604  		sched.runqhead = gp1.schedlink
  4605  		runqput(_p_, gp1, false)
  4606  	}
  4607  	return gp
  4608  }
  4609  
  4610  // Put p to on _Pidle list.
  4611  // Sched must be locked.
  4612  // May run during STW, so write barriers are not allowed.
  4613  //go:nowritebarrierrec
  4614  func pidleput(_p_ *p) {
  4615  	if !runqempty(_p_) {
  4616  		throw("pidleput: P has non-empty run queue")
  4617  	}
  4618  	_p_.link = sched.pidle
  4619  	sched.pidle.set(_p_)
  4620  	atomic.Xadd(&sched.npidle, 1) // TODO: fast atomic
  4621  }
  4622  
  4623  // Try get a p from _Pidle list.
  4624  // Sched must be locked.
  4625  // May run during STW, so write barriers are not allowed.
  4626  //go:nowritebarrierrec
  4627  func pidleget() *p {
  4628  	_p_ := sched.pidle.ptr()
  4629  	if _p_ != nil {
  4630  		sched.pidle = _p_.link
  4631  		atomic.Xadd(&sched.npidle, -1) // TODO: fast atomic
  4632  	}
  4633  	return _p_
  4634  }
  4635  
  4636  // runqempty returns true if _p_ has no Gs on its local run queue.
  4637  // It never returns true spuriously.
  4638  func runqempty(_p_ *p) bool {
  4639  	// Defend against a race where 1) _p_ has G1 in runqnext but runqhead == runqtail,
  4640  	// 2) runqput on _p_ kicks G1 to the runq, 3) runqget on _p_ empties runqnext.
  4641  	// Simply observing that runqhead == runqtail and then observing that runqnext == nil
  4642  	// does not mean the queue is empty.
  4643  	for {
  4644  		head := atomic.Load(&_p_.runqhead)
  4645  		tail := atomic.Load(&_p_.runqtail)
  4646  		runnext := atomic.Loaduintptr((*uintptr)(unsafe.Pointer(&_p_.runnext)))
  4647  		if tail == atomic.Load(&_p_.runqtail) {
  4648  			return head == tail && runnext == 0
  4649  		}
  4650  	}
  4651  }
  4652  
  4653  // To shake out latent assumptions about scheduling order,
  4654  // we introduce some randomness into scheduling decisions
  4655  // when running with the race detector.
  4656  // The need for this was made obvious by changing the
  4657  // (deterministic) scheduling order in Go 1.5 and breaking
  4658  // many poorly-written tests.
  4659  // With the randomness here, as long as the tests pass
  4660  // consistently with -race, they shouldn't have latent scheduling
  4661  // assumptions.
  4662  const randomizeScheduler = raceenabled
  4663  
  4664  // runqput tries to put g on the local runnable queue.
  4665  // If next if false, runqput adds g to the tail of the runnable queue.
  4666  // If next is true, runqput puts g in the _p_.runnext slot.
  4667  // If the run queue is full, runnext puts g on the global queue.
  4668  // Executed only by the owner P.
  4669  func runqput(_p_ *p, gp *g, next bool) {
  4670  	if randomizeScheduler && next && fastrand()%2 == 0 {
  4671  		next = false
  4672  	}
  4673  
  4674  	if next {
  4675  	retryNext:
  4676  		oldnext := _p_.runnext
  4677  		if !_p_.runnext.cas(oldnext, guintptr(unsafe.Pointer(gp))) {
  4678  			goto retryNext
  4679  		}
  4680  		if oldnext == 0 {
  4681  			return
  4682  		}
  4683  		// Kick the old runnext out to the regular run queue.
  4684  		gp = oldnext.ptr()
  4685  	}
  4686  
  4687  retry:
  4688  	h := atomic.Load(&_p_.runqhead) // load-acquire, synchronize with consumers
  4689  	t := _p_.runqtail
  4690  	if t-h < uint32(len(_p_.runq)) {
  4691  		_p_.runq[t%uint32(len(_p_.runq))].set(gp)
  4692  		atomic.Store(&_p_.runqtail, t+1) // store-release, makes the item available for consumption
  4693  		return
  4694  	}
  4695  	if runqputslow(_p_, gp, h, t) {
  4696  		return
  4697  	}
  4698  	// the queue is not full, now the put above must succeed
  4699  	goto retry
  4700  }
  4701  
  4702  // Put g and a batch of work from local runnable queue on global queue.
  4703  // Executed only by the owner P.
  4704  func runqputslow(_p_ *p, gp *g, h, t uint32) bool {
  4705  	var batch [len(_p_.runq)/2 + 1]*g
  4706  
  4707  	// First, grab a batch from local queue.
  4708  	n := t - h
  4709  	n = n / 2
  4710  	if n != uint32(len(_p_.runq)/2) {
  4711  		throw("runqputslow: queue is not full")
  4712  	}
  4713  	for i := uint32(0); i < n; i++ {
  4714  		batch[i] = _p_.runq[(h+i)%uint32(len(_p_.runq))].ptr()
  4715  	}
  4716  	if !atomic.Cas(&_p_.runqhead, h, h+n) { // cas-release, commits consume
  4717  		return false
  4718  	}
  4719  	batch[n] = gp
  4720  
  4721  	if randomizeScheduler {
  4722  		for i := uint32(1); i <= n; i++ {
  4723  			j := fastrandn(i + 1)
  4724  			batch[i], batch[j] = batch[j], batch[i]
  4725  		}
  4726  	}
  4727  
  4728  	// Link the goroutines.
  4729  	for i := uint32(0); i < n; i++ {
  4730  		batch[i].schedlink.set(batch[i+1])
  4731  	}
  4732  
  4733  	// Now put the batch on global queue.
  4734  	lock(&sched.lock)
  4735  	globrunqputbatch(batch[0], batch[n], int32(n+1))
  4736  	unlock(&sched.lock)
  4737  	return true
  4738  }
  4739  
  4740  // Get g from local runnable queue.
  4741  // If inheritTime is true, gp should inherit the remaining time in the
  4742  // current time slice. Otherwise, it should start a new time slice.
  4743  // Executed only by the owner P.
  4744  func runqget(_p_ *p) (gp *g, inheritTime bool) {
  4745  	// If there's a runnext, it's the next G to run.
  4746  	for {
  4747  		next := _p_.runnext
  4748  		if next == 0 {
  4749  			break
  4750  		}
  4751  		if _p_.runnext.cas(next, 0) {
  4752  			return next.ptr(), true
  4753  		}
  4754  	}
  4755  
  4756  	for {
  4757  		h := atomic.Load(&_p_.runqhead) // load-acquire, synchronize with other consumers
  4758  		t := _p_.runqtail
  4759  		if t == h {
  4760  			return nil, false
  4761  		}
  4762  		gp := _p_.runq[h%uint32(len(_p_.runq))].ptr()
  4763  		if atomic.Cas(&_p_.runqhead, h, h+1) { // cas-release, commits consume
  4764  			return gp, false
  4765  		}
  4766  	}
  4767  }
  4768  
  4769  // Grabs a batch of goroutines from _p_'s runnable queue into batch.
  4770  // Batch is a ring buffer starting at batchHead.
  4771  // Returns number of grabbed goroutines.
  4772  // Can be executed by any P.
  4773  func runqgrab(_p_ *p, batch *[256]guintptr, batchHead uint32, stealRunNextG bool) uint32 {
  4774  	for {
  4775  		h := atomic.Load(&_p_.runqhead) // load-acquire, synchronize with other consumers
  4776  		t := atomic.Load(&_p_.runqtail) // load-acquire, synchronize with the producer
  4777  		n := t - h
  4778  		n = n - n/2
  4779  		if n == 0 {
  4780  			if stealRunNextG {
  4781  				// Try to steal from _p_.runnext.
  4782  				if next := _p_.runnext; next != 0 {
  4783  					if _p_.status == _Prunning {
  4784  						// Sleep to ensure that _p_ isn't about to run the g
  4785  						// we are about to steal.
  4786  						// The important use case here is when the g running
  4787  						// on _p_ ready()s another g and then almost
  4788  						// immediately blocks. Instead of stealing runnext
  4789  						// in this window, back off to give _p_ a chance to
  4790  						// schedule runnext. This will avoid thrashing gs
  4791  						// between different Ps.
  4792  						// A sync chan send/recv takes ~50ns as of time of
  4793  						// writing, so 3us gives ~50x overshoot.
  4794  						if GOOS != "windows" {
  4795  							usleep(3)
  4796  						} else {
  4797  							// On windows system timer granularity is
  4798  							// 1-15ms, which is way too much for this
  4799  							// optimization. So just yield.
  4800  							osyield()
  4801  						}
  4802  					}
  4803  					if !_p_.runnext.cas(next, 0) {
  4804  						continue
  4805  					}
  4806  					batch[batchHead%uint32(len(batch))] = next
  4807  					return 1
  4808  				}
  4809  			}
  4810  			return 0
  4811  		}
  4812  		if n > uint32(len(_p_.runq)/2) { // read inconsistent h and t
  4813  			continue
  4814  		}
  4815  		for i := uint32(0); i < n; i++ {
  4816  			g := _p_.runq[(h+i)%uint32(len(_p_.runq))]
  4817  			batch[(batchHead+i)%uint32(len(batch))] = g
  4818  		}
  4819  		if atomic.Cas(&_p_.runqhead, h, h+n) { // cas-release, commits consume
  4820  			return n
  4821  		}
  4822  	}
  4823  }
  4824  
  4825  // Steal half of elements from local runnable queue of p2
  4826  // and put onto local runnable queue of p.
  4827  // Returns one of the stolen elements (or nil if failed).
  4828  func runqsteal(_p_, p2 *p, stealRunNextG bool) *g {
  4829  	t := _p_.runqtail
  4830  	n := runqgrab(p2, &_p_.runq, t, stealRunNextG)
  4831  	if n == 0 {
  4832  		return nil
  4833  	}
  4834  	n--
  4835  	gp := _p_.runq[(t+n)%uint32(len(_p_.runq))].ptr()
  4836  	if n == 0 {
  4837  		return gp
  4838  	}
  4839  	h := atomic.Load(&_p_.runqhead) // load-acquire, synchronize with consumers
  4840  	if t-h+n >= uint32(len(_p_.runq)) {
  4841  		throw("runqsteal: runq overflow")
  4842  	}
  4843  	atomic.Store(&_p_.runqtail, t+n) // store-release, makes the item available for consumption
  4844  	return gp
  4845  }
  4846  
  4847  //go:linkname setMaxThreads runtime/debug.setMaxThreads
  4848  func setMaxThreads(in int) (out int) {
  4849  	lock(&sched.lock)
  4850  	out = int(sched.maxmcount)
  4851  	if in > 0x7fffffff { // MaxInt32
  4852  		sched.maxmcount = 0x7fffffff
  4853  	} else {
  4854  		sched.maxmcount = int32(in)
  4855  	}
  4856  	checkmcount()
  4857  	unlock(&sched.lock)
  4858  	return
  4859  }
  4860  
  4861  func haveexperiment(name string) bool {
  4862  	if name == "framepointer" {
  4863  		return framepointer_enabled // set by linker
  4864  	}
  4865  	x := sys.Goexperiment
  4866  	for x != "" {
  4867  		xname := ""
  4868  		i := index(x, ",")
  4869  		if i < 0 {
  4870  			xname, x = x, ""
  4871  		} else {
  4872  			xname, x = x[:i], x[i+1:]
  4873  		}
  4874  		if xname == name {
  4875  			return true
  4876  		}
  4877  		if len(xname) > 2 && xname[:2] == "no" && xname[2:] == name {
  4878  			return false
  4879  		}
  4880  	}
  4881  	return false
  4882  }
  4883  
  4884  //go:nosplit
  4885  func procPin() int {
  4886  	_g_ := getg()
  4887  	mp := _g_.m
  4888  
  4889  	mp.locks++
  4890  	return int(mp.p.ptr().id)
  4891  }
  4892  
  4893  //go:nosplit
  4894  func procUnpin() {
  4895  	_g_ := getg()
  4896  	_g_.m.locks--
  4897  }
  4898  
  4899  //go:linkname sync_runtime_procPin sync.runtime_procPin
  4900  //go:nosplit
  4901  func sync_runtime_procPin() int {
  4902  	return procPin()
  4903  }
  4904  
  4905  //go:linkname sync_runtime_procUnpin sync.runtime_procUnpin
  4906  //go:nosplit
  4907  func sync_runtime_procUnpin() {
  4908  	procUnpin()
  4909  }
  4910  
  4911  //go:linkname sync_atomic_runtime_procPin sync/atomic.runtime_procPin
  4912  //go:nosplit
  4913  func sync_atomic_runtime_procPin() int {
  4914  	return procPin()
  4915  }
  4916  
  4917  //go:linkname sync_atomic_runtime_procUnpin sync/atomic.runtime_procUnpin
  4918  //go:nosplit
  4919  func sync_atomic_runtime_procUnpin() {
  4920  	procUnpin()
  4921  }
  4922  
  4923  // Active spinning for sync.Mutex.
  4924  //go:linkname sync_runtime_canSpin sync.runtime_canSpin
  4925  //go:nosplit
  4926  func sync_runtime_canSpin(i int) bool {
  4927  	// sync.Mutex is cooperative, so we are conservative with spinning.
  4928  	// Spin only few times and only if running on a multicore machine and
  4929  	// GOMAXPROCS>1 and there is at least one other running P and local runq is empty.
  4930  	// As opposed to runtime mutex we don't do passive spinning here,
  4931  	// because there can be work on global runq on on other Ps.
  4932  	if i >= active_spin || ncpu <= 1 || gomaxprocs <= int32(sched.npidle+sched.nmspinning)+1 {
  4933  		return false
  4934  	}
  4935  	if p := getg().m.p.ptr(); !runqempty(p) {
  4936  		return false
  4937  	}
  4938  	return true
  4939  }
  4940  
  4941  //go:linkname sync_runtime_doSpin sync.runtime_doSpin
  4942  //go:nosplit
  4943  func sync_runtime_doSpin() {
  4944  	procyield(active_spin_cnt)
  4945  }
  4946  
  4947  var stealOrder randomOrder
  4948  
  4949  // randomOrder/randomEnum are helper types for randomized work stealing.
  4950  // They allow to enumerate all Ps in different pseudo-random orders without repetitions.
  4951  // The algorithm is based on the fact that if we have X such that X and GOMAXPROCS
  4952  // are coprime, then a sequences of (i + X) % GOMAXPROCS gives the required enumeration.
  4953  type randomOrder struct {
  4954  	count    uint32
  4955  	coprimes []uint32
  4956  }
  4957  
  4958  type randomEnum struct {
  4959  	i     uint32
  4960  	count uint32
  4961  	pos   uint32
  4962  	inc   uint32
  4963  }
  4964  
  4965  func (ord *randomOrder) reset(count uint32) {
  4966  	ord.count = count
  4967  	ord.coprimes = ord.coprimes[:0]
  4968  	for i := uint32(1); i <= count; i++ {
  4969  		if gcd(i, count) == 1 {
  4970  			ord.coprimes = append(ord.coprimes, i)
  4971  		}
  4972  	}
  4973  }
  4974  
  4975  func (ord *randomOrder) start(i uint32) randomEnum {
  4976  	return randomEnum{
  4977  		count: ord.count,
  4978  		pos:   i % ord.count,
  4979  		inc:   ord.coprimes[i%uint32(len(ord.coprimes))],
  4980  	}
  4981  }
  4982  
  4983  func (enum *randomEnum) done() bool {
  4984  	return enum.i == enum.count
  4985  }
  4986  
  4987  func (enum *randomEnum) next() {
  4988  	enum.i++
  4989  	enum.pos = (enum.pos + enum.inc) % enum.count
  4990  }
  4991  
  4992  func (enum *randomEnum) position() uint32 {
  4993  	return enum.pos
  4994  }
  4995  
  4996  func gcd(a, b uint32) uint32 {
  4997  	for b != 0 {
  4998  		a, b = b, a%b
  4999  	}
  5000  	return a
  5001  }
  5002  

View as plain text