proc_test.go

     1  // Copyright 2011 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_test
     6  
     7  import (
     8  	"fmt"
     9  	"internal/race"
    10  	"internal/testenv"
    11  	"math"
    12  	"net"
    13  	"runtime"
    14  	"runtime/debug"
    15  	"strings"
    16  	"sync"
    17  	"sync/atomic"
    18  	"syscall"
    19  	"testing"
    20  	"time"
    21  )
    22  
    23  var stop = make(chan bool, 1)
    24  
    25  func perpetuumMobile() {
    26  	select {
    27  	case <-stop:
    28  	default:
    29  		go perpetuumMobile()
    30  	}
    31  }
    32  
    33  func TestStopTheWorldDeadlock(t *testing.T) {
    34  	if runtime.GOARCH == "wasm" {
    35  		t.Skip("no preemption on wasm yet")
    36  	}
    37  	if testing.Short() {
    38  		t.Skip("skipping during short test")
    39  	}
    40  	maxprocs := runtime.GOMAXPROCS(3)
    41  	compl := make(chan bool, 2)
    42  	go func() {
    43  		for i := 0; i != 1000; i += 1 {
    44  			runtime.GC()
    45  		}
    46  		compl <- true
    47  	}()
    48  	go func() {
    49  		for i := 0; i != 1000; i += 1 {
    50  			runtime.GOMAXPROCS(3)
    51  		}
    52  		compl <- true
    53  	}()
    54  	go perpetuumMobile()
    55  	<-compl
    56  	<-compl
    57  	stop <- true
    58  	runtime.GOMAXPROCS(maxprocs)
    59  }
    60  
    61  func TestYieldProgress(t *testing.T) {
    62  	testYieldProgress(false)
    63  }
    64  
    65  func TestYieldLockedProgress(t *testing.T) {
    66  	testYieldProgress(true)
    67  }
    68  
    69  func testYieldProgress(locked bool) {
    70  	c := make(chan bool)
    71  	cack := make(chan bool)
    72  	go func() {
    73  		if locked {
    74  			runtime.LockOSThread()
    75  		}
    76  		for {
    77  			select {
    78  			case <-c:
    79  				cack <- true
    80  				return
    81  			default:
    82  				runtime.Gosched()
    83  			}
    84  		}
    85  	}()
    86  	time.Sleep(10 * time.Millisecond)
    87  	c <- true
    88  	<-cack
    89  }
    90  
    91  func TestYieldLocked(t *testing.T) {
    92  	const N = 10
    93  	c := make(chan bool)
    94  	go func() {
    95  		runtime.LockOSThread()
    96  		for i := 0; i < N; i++ {
    97  			runtime.Gosched()
    98  			time.Sleep(time.Millisecond)
    99  		}
   100  		c <- true
   101  		// runtime.UnlockOSThread() is deliberately omitted
   102  	}()
   103  	<-c
   104  }
   105  
   106  func TestGoroutineParallelism(t *testing.T) {
   107  	if runtime.NumCPU() == 1 {
   108  		// Takes too long, too easy to deadlock, etc.
   109  		t.Skip("skipping on uniprocessor")
   110  	}
   111  	P := 4
   112  	N := 10
   113  	if testing.Short() {
   114  		P = 3
   115  		N = 3
   116  	}
   117  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(P))
   118  	// If runtime triggers a forced GC during this test then it will deadlock,
   119  	// since the goroutines can't be stopped/preempted.
   120  	// Disable GC for this test (see issue #10958).
   121  	defer debug.SetGCPercent(debug.SetGCPercent(-1))
   122  	// SetGCPercent waits until the mark phase is over, but the runtime
   123  	// also preempts at the start of the sweep phase, so make sure that's
   124  	// done too. See #45867.
   125  	runtime.GC()
   126  	for try := 0; try < N; try++ {
   127  		done := make(chan bool)
   128  		x := uint32(0)
   129  		for p := 0; p < P; p++ {
   130  			// Test that all P goroutines are scheduled at the same time
   131  			go func(p int) {
   132  				for i := 0; i < 3; i++ {
   133  					expected := uint32(P*i + p)
   134  					for atomic.LoadUint32(&x) != expected {
   135  					}
   136  					atomic.StoreUint32(&x, expected+1)
   137  				}
   138  				done <- true
   139  			}(p)
   140  		}
   141  		for p := 0; p < P; p++ {
   142  			<-done
   143  		}
   144  	}
   145  }
   146  
   147  // Test that all runnable goroutines are scheduled at the same time.
   148  func TestGoroutineParallelism2(t *testing.T) {
   149  	//testGoroutineParallelism2(t, false, false)
   150  	testGoroutineParallelism2(t, true, false)
   151  	testGoroutineParallelism2(t, false, true)
   152  	testGoroutineParallelism2(t, true, true)
   153  }
   154  
   155  func testGoroutineParallelism2(t *testing.T, load, netpoll bool) {
   156  	if runtime.NumCPU() == 1 {
   157  		// Takes too long, too easy to deadlock, etc.
   158  		t.Skip("skipping on uniprocessor")
   159  	}
   160  	P := 4
   161  	N := 10
   162  	if testing.Short() {
   163  		N = 3
   164  	}
   165  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(P))
   166  	// If runtime triggers a forced GC during this test then it will deadlock,
   167  	// since the goroutines can't be stopped/preempted.
   168  	// Disable GC for this test (see issue #10958).
   169  	defer debug.SetGCPercent(debug.SetGCPercent(-1))
   170  	// SetGCPercent waits until the mark phase is over, but the runtime
   171  	// also preempts at the start of the sweep phase, so make sure that's
   172  	// done too. See #45867.
   173  	runtime.GC()
   174  	for try := 0; try < N; try++ {
   175  		if load {
   176  			// Create P goroutines and wait until they all run.
   177  			// When we run the actual test below, worker threads
   178  			// running the goroutines will start parking.
   179  			done := make(chan bool)
   180  			x := uint32(0)
   181  			for p := 0; p < P; p++ {
   182  				go func() {
   183  					if atomic.AddUint32(&x, 1) == uint32(P) {
   184  						done <- true
   185  						return
   186  					}
   187  					for atomic.LoadUint32(&x) != uint32(P) {
   188  					}
   189  				}()
   190  			}
   191  			<-done
   192  		}
   193  		if netpoll {
   194  			// Enable netpoller, affects schedler behavior.
   195  			laddr := "localhost:0"
   196  			if runtime.GOOS == "android" {
   197  				// On some Android devices, there are no records for localhost,
   198  				// see https://golang.org/issues/14486.
   199  				// Don't use 127.0.0.1 for every case, it won't work on IPv6-only systems.
   200  				laddr = "127.0.0.1:0"
   201  			}
   202  			ln, err := net.Listen("tcp", laddr)
   203  			if err != nil {
   204  				defer ln.Close() // yup, defer in a loop
   205  			}
   206  		}
   207  		done := make(chan bool)
   208  		x := uint32(0)
   209  		// Spawn P goroutines in a nested fashion just to differ from TestGoroutineParallelism.
   210  		for p := 0; p < P/2; p++ {
   211  			go func(p int) {
   212  				for p2 := 0; p2 < 2; p2++ {
   213  					go func(p2 int) {
   214  						for i := 0; i < 3; i++ {
   215  							expected := uint32(P*i + p*2 + p2)
   216  							for atomic.LoadUint32(&x) != expected {
   217  							}
   218  							atomic.StoreUint32(&x, expected+1)
   219  						}
   220  						done <- true
   221  					}(p2)
   222  				}
   223  			}(p)
   224  		}
   225  		for p := 0; p < P; p++ {
   226  			<-done
   227  		}
   228  	}
   229  }
   230  
   231  func TestBlockLocked(t *testing.T) {
   232  	const N = 10
   233  	c := make(chan bool)
   234  	go func() {
   235  		runtime.LockOSThread()
   236  		for i := 0; i < N; i++ {
   237  			c <- true
   238  		}
   239  		runtime.UnlockOSThread()
   240  	}()
   241  	for i := 0; i < N; i++ {
   242  		<-c
   243  	}
   244  }
   245  
   246  func TestTimerFairness(t *testing.T) {
   247  	if runtime.GOARCH == "wasm" {
   248  		t.Skip("no preemption on wasm yet")
   249  	}
   250  
   251  	done := make(chan bool)
   252  	c := make(chan bool)
   253  	for i := 0; i < 2; i++ {
   254  		go func() {
   255  			for {
   256  				select {
   257  				case c <- true:
   258  				case <-done:
   259  					return
   260  				}
   261  			}
   262  		}()
   263  	}
   264  
   265  	timer := time.After(20 * time.Millisecond)
   266  	for {
   267  		select {
   268  		case <-c:
   269  		case <-timer:
   270  			close(done)
   271  			return
   272  		}
   273  	}
   274  }
   275  
   276  func TestTimerFairness2(t *testing.T) {
   277  	if runtime.GOARCH == "wasm" {
   278  		t.Skip("no preemption on wasm yet")
   279  	}
   280  
   281  	done := make(chan bool)
   282  	c := make(chan bool)
   283  	for i := 0; i < 2; i++ {
   284  		go func() {
   285  			timer := time.After(20 * time.Millisecond)
   286  			var buf [1]byte
   287  			for {
   288  				syscall.Read(0, buf[0:0])
   289  				select {
   290  				case c <- true:
   291  				case <-c:
   292  				case <-timer:
   293  					done <- true
   294  					return
   295  				}
   296  			}
   297  		}()
   298  	}
   299  	<-done
   300  	<-done
   301  }
   302  
   303  // The function is used to test preemption at split stack checks.
   304  // Declaring a var avoids inlining at the call site.
   305  var preempt = func() int {
   306  	var a [128]int
   307  	sum := 0
   308  	for _, v := range a {
   309  		sum += v
   310  	}
   311  	return sum
   312  }
   313  
   314  func TestPreemption(t *testing.T) {
   315  	if runtime.GOARCH == "wasm" {
   316  		t.Skip("no preemption on wasm yet")
   317  	}
   318  
   319  	// Test that goroutines are preempted at function calls.
   320  	N := 5
   321  	if testing.Short() {
   322  		N = 2
   323  	}
   324  	c := make(chan bool)
   325  	var x uint32
   326  	for g := 0; g < 2; g++ {
   327  		go func(g int) {
   328  			for i := 0; i < N; i++ {
   329  				for atomic.LoadUint32(&x) != uint32(g) {
   330  					preempt()
   331  				}
   332  				atomic.StoreUint32(&x, uint32(1-g))
   333  			}
   334  			c <- true
   335  		}(g)
   336  	}
   337  	<-c
   338  	<-c
   339  }
   340  
   341  func TestPreemptionGC(t *testing.T) {
   342  	if runtime.GOARCH == "wasm" {
   343  		t.Skip("no preemption on wasm yet")
   344  	}
   345  
   346  	// Test that pending GC preempts running goroutines.
   347  	P := 5
   348  	N := 10
   349  	if testing.Short() {
   350  		P = 3
   351  		N = 2
   352  	}
   353  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(P + 1))
   354  	var stop uint32
   355  	for i := 0; i < P; i++ {
   356  		go func() {
   357  			for atomic.LoadUint32(&stop) == 0 {
   358  				preempt()
   359  			}
   360  		}()
   361  	}
   362  	for i := 0; i < N; i++ {
   363  		runtime.Gosched()
   364  		runtime.GC()
   365  	}
   366  	atomic.StoreUint32(&stop, 1)
   367  }
   368  
   369  func TestAsyncPreempt(t *testing.T) {
   370  	if !runtime.PreemptMSupported {
   371  		t.Skip("asynchronous preemption not supported on this platform")
   372  	}
   373  	output := runTestProg(t, "testprog", "AsyncPreempt")
   374  	want := "OK\n"
   375  	if output != want {
   376  		t.Fatalf("want %s, got %s\n", want, output)
   377  	}
   378  }
   379  
   380  func TestGCFairness(t *testing.T) {
   381  	output := runTestProg(t, "testprog", "GCFairness")
   382  	want := "OK\n"
   383  	if output != want {
   384  		t.Fatalf("want %s, got %s\n", want, output)
   385  	}
   386  }
   387  
   388  func TestGCFairness2(t *testing.T) {
   389  	output := runTestProg(t, "testprog", "GCFairness2")
   390  	want := "OK\n"
   391  	if output != want {
   392  		t.Fatalf("want %s, got %s\n", want, output)
   393  	}
   394  }
   395  
   396  func TestNumGoroutine(t *testing.T) {
   397  	output := runTestProg(t, "testprog", "NumGoroutine")
   398  	want := "1\n"
   399  	if output != want {
   400  		t.Fatalf("want %q, got %q", want, output)
   401  	}
   402  
   403  	buf := make([]byte, 1<<20)
   404  
   405  	// Try up to 10 times for a match before giving up.
   406  	// This is a fundamentally racy check but it's important
   407  	// to notice if NumGoroutine and Stack are _always_ out of sync.
   408  	for i := 0; ; i++ {
   409  		// Give goroutines about to exit a chance to exit.
   410  		// The NumGoroutine and Stack below need to see
   411  		// the same state of the world, so anything we can do
   412  		// to keep it quiet is good.
   413  		runtime.Gosched()
   414  
   415  		n := runtime.NumGoroutine()
   416  		buf = buf[:runtime.Stack(buf, true)]
   417  
   418  		nstk := strings.Count(string(buf), "goroutine ")
   419  		if n == nstk {
   420  			break
   421  		}
   422  		if i >= 10 {
   423  			t.Fatalf("NumGoroutine=%d, but found %d goroutines in stack dump: %s", n, nstk, buf)
   424  		}
   425  	}
   426  }
   427  
   428  func TestPingPongHog(t *testing.T) {
   429  	if runtime.GOARCH == "wasm" {
   430  		t.Skip("no preemption on wasm yet")
   431  	}
   432  	if testing.Short() {
   433  		t.Skip("skipping in -short mode")
   434  	}
   435  	if race.Enabled {
   436  		// The race detector randomizes the scheduler,
   437  		// which causes this test to fail (#38266).
   438  		t.Skip("skipping in -race mode")
   439  	}
   440  
   441  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(1))
   442  	done := make(chan bool)
   443  	hogChan, lightChan := make(chan bool), make(chan bool)
   444  	hogCount, lightCount := 0, 0
   445  
   446  	run := func(limit int, counter *int, wake chan bool) {
   447  		for {
   448  			select {
   449  			case <-done:
   450  				return
   451  
   452  			case <-wake:
   453  				for i := 0; i < limit; i++ {
   454  					*counter++
   455  				}
   456  				wake <- true
   457  			}
   458  		}
   459  	}
   460  
   461  	// Start two co-scheduled hog goroutines.
   462  	for i := 0; i < 2; i++ {
   463  		go run(1e6, &hogCount, hogChan)
   464  	}
   465  
   466  	// Start two co-scheduled light goroutines.
   467  	for i := 0; i < 2; i++ {
   468  		go run(1e3, &lightCount, lightChan)
   469  	}
   470  
   471  	// Start goroutine pairs and wait for a few preemption rounds.
   472  	hogChan <- true
   473  	lightChan <- true
   474  	time.Sleep(100 * time.Millisecond)
   475  	close(done)
   476  	<-hogChan
   477  	<-lightChan
   478  
   479  	// Check that hogCount and lightCount are within a factor of
   480  	// 20, which indicates that both pairs of goroutines handed off
   481  	// the P within a time-slice to their buddy. We can use a
   482  	// fairly large factor here to make this robust: if the
   483  	// scheduler isn't working right, the gap should be ~1000X
   484  	// (was 5, increased to 20, see issue 52207).
   485  	const factor = 20
   486  	if hogCount/factor > lightCount || lightCount/factor > hogCount {
   487  		t.Fatalf("want hogCount/lightCount in [%v, %v]; got %d/%d = %g", 1.0/factor, factor, hogCount, lightCount, float64(hogCount)/float64(lightCount))
   488  	}
   489  }
   490  
   491  func BenchmarkPingPongHog(b *testing.B) {
   492  	if b.N == 0 {
   493  		return
   494  	}
   495  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(1))
   496  
   497  	// Create a CPU hog
   498  	stop, done := make(chan bool), make(chan bool)
   499  	go func() {
   500  		for {
   501  			select {
   502  			case <-stop:
   503  				done <- true
   504  				return
   505  			default:
   506  			}
   507  		}
   508  	}()
   509  
   510  	// Ping-pong b.N times
   511  	ping, pong := make(chan bool), make(chan bool)
   512  	go func() {
   513  		for j := 0; j < b.N; j++ {
   514  			pong <- <-ping
   515  		}
   516  		close(stop)
   517  		done <- true
   518  	}()
   519  	go func() {
   520  		for i := 0; i < b.N; i++ {
   521  			ping <- <-pong
   522  		}
   523  		done <- true
   524  	}()
   525  	b.ResetTimer()
   526  	ping <- true // Start ping-pong
   527  	<-stop
   528  	b.StopTimer()
   529  	<-ping // Let last ponger exit
   530  	<-done // Make sure goroutines exit
   531  	<-done
   532  	<-done
   533  }
   534  
   535  var padData [128]uint64
   536  
   537  func stackGrowthRecursive(i int) {
   538  	var pad [128]uint64
   539  	pad = padData
   540  	for j := range pad {
   541  		if pad[j] != 0 {
   542  			return
   543  		}
   544  	}
   545  	if i != 0 {
   546  		stackGrowthRecursive(i - 1)
   547  	}
   548  }
   549  
   550  func TestPreemptSplitBig(t *testing.T) {
   551  	if testing.Short() {
   552  		t.Skip("skipping in -short mode")
   553  	}
   554  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(2))
   555  	stop := make(chan int)
   556  	go big(stop)
   557  	for i := 0; i < 3; i++ {
   558  		time.Sleep(10 * time.Microsecond) // let big start running
   559  		runtime.GC()
   560  	}
   561  	close(stop)
   562  }
   563  
   564  func big(stop chan int) int {
   565  	n := 0
   566  	for {
   567  		// delay so that gc is sure to have asked for a preemption
   568  		for i := 0; i < 1e9; i++ {
   569  			n++
   570  		}
   571  
   572  		// call bigframe, which used to miss the preemption in its prologue.
   573  		bigframe(stop)
   574  
   575  		// check if we've been asked to stop.
   576  		select {
   577  		case <-stop:
   578  			return n
   579  		}
   580  	}
   581  }
   582  
   583  func bigframe(stop chan int) int {
   584  	// not splitting the stack will overflow.
   585  	// small will notice that it needs a stack split and will
   586  	// catch the overflow.
   587  	var x [8192]byte
   588  	return small(stop, &x)
   589  }
   590  
   591  func small(stop chan int, x *[8192]byte) int {
   592  	for i := range x {
   593  		x[i] = byte(i)
   594  	}
   595  	sum := 0
   596  	for i := range x {
   597  		sum += int(x[i])
   598  	}
   599  
   600  	// keep small from being a leaf function, which might
   601  	// make it not do any stack check at all.
   602  	nonleaf(stop)
   603  
   604  	return sum
   605  }
   606  
   607  func nonleaf(stop chan int) bool {
   608  	// do something that won't be inlined:
   609  	select {
   610  	case <-stop:
   611  		return true
   612  	default:
   613  		return false
   614  	}
   615  }
   616  
   617  func TestSchedLocalQueue(t *testing.T) {
   618  	runtime.RunSchedLocalQueueTest()
   619  }
   620  
   621  func TestSchedLocalQueueSteal(t *testing.T) {
   622  	runtime.RunSchedLocalQueueStealTest()
   623  }
   624  
   625  func TestSchedLocalQueueEmpty(t *testing.T) {
   626  	if runtime.NumCPU() == 1 {
   627  		// Takes too long and does not trigger the race.
   628  		t.Skip("skipping on uniprocessor")
   629  	}
   630  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(4))
   631  
   632  	// If runtime triggers a forced GC during this test then it will deadlock,
   633  	// since the goroutines can't be stopped/preempted during spin wait.
   634  	defer debug.SetGCPercent(debug.SetGCPercent(-1))
   635  	// SetGCPercent waits until the mark phase is over, but the runtime
   636  	// also preempts at the start of the sweep phase, so make sure that's
   637  	// done too. See #45867.
   638  	runtime.GC()
   639  
   640  	iters := int(1e5)
   641  	if testing.Short() {
   642  		iters = 1e2
   643  	}
   644  	runtime.RunSchedLocalQueueEmptyTest(iters)
   645  }
   646  
   647  func benchmarkStackGrowth(b *testing.B, rec int) {
   648  	b.RunParallel(func(pb *testing.PB) {
   649  		for pb.Next() {
   650  			stackGrowthRecursive(rec)
   651  		}
   652  	})
   653  }
   654  
   655  func BenchmarkStackGrowth(b *testing.B) {
   656  	benchmarkStackGrowth(b, 10)
   657  }
   658  
   659  func BenchmarkStackGrowthDeep(b *testing.B) {
   660  	benchmarkStackGrowth(b, 1024)
   661  }
   662  
   663  func BenchmarkCreateGoroutines(b *testing.B) {
   664  	benchmarkCreateGoroutines(b, 1)
   665  }
   666  
   667  func BenchmarkCreateGoroutinesParallel(b *testing.B) {
   668  	benchmarkCreateGoroutines(b, runtime.GOMAXPROCS(-1))
   669  }
   670  
   671  func benchmarkCreateGoroutines(b *testing.B, procs int) {
   672  	c := make(chan bool)
   673  	var f func(n int)
   674  	f = func(n int) {
   675  		if n == 0 {
   676  			c <- true
   677  			return
   678  		}
   679  		go f(n - 1)
   680  	}
   681  	for i := 0; i < procs; i++ {
   682  		go f(b.N / procs)
   683  	}
   684  	for i := 0; i < procs; i++ {
   685  		<-c
   686  	}
   687  }
   688  
   689  func BenchmarkCreateGoroutinesCapture(b *testing.B) {
   690  	b.ReportAllocs()
   691  	for i := 0; i < b.N; i++ {
   692  		const N = 4
   693  		var wg sync.WaitGroup
   694  		wg.Add(N)
   695  		for i := 0; i < N; i++ {
   696  			i := i
   697  			go func() {
   698  				if i >= N {
   699  					b.Logf("bad") // just to capture b
   700  				}
   701  				wg.Done()
   702  			}()
   703  		}
   704  		wg.Wait()
   705  	}
   706  }
   707  
   708  // warmupScheduler ensures the scheduler has at least targetThreadCount threads
   709  // in its thread pool.
   710  func warmupScheduler(targetThreadCount int) {
   711  	var wg sync.WaitGroup
   712  	var count int32
   713  	for i := 0; i < targetThreadCount; i++ {
   714  		wg.Add(1)
   715  		go func() {
   716  			atomic.AddInt32(&count, 1)
   717  			for atomic.LoadInt32(&count) < int32(targetThreadCount) {
   718  				// spin until all threads started
   719  			}
   720  
   721  			// spin a bit more to ensure they are all running on separate CPUs.
   722  			doWork(time.Millisecond)
   723  			wg.Done()
   724  		}()
   725  	}
   726  	wg.Wait()
   727  }
   728  
   729  func doWork(dur time.Duration) {
   730  	start := time.Now()
   731  	for time.Since(start) < dur {
   732  	}
   733  }
   734  
   735  // BenchmarkCreateGoroutinesSingle creates many goroutines, all from a single
   736  // producer (the main benchmark goroutine).
   737  //
   738  // Compared to BenchmarkCreateGoroutines, this causes different behavior in the
   739  // scheduler because Ms are much more likely to need to steal work from the
   740  // main P rather than having work in the local run queue.
   741  func BenchmarkCreateGoroutinesSingle(b *testing.B) {
   742  	// Since we are interested in stealing behavior, warm the scheduler to
   743  	// get all the Ps running first.
   744  	warmupScheduler(runtime.GOMAXPROCS(0))
   745  	b.ResetTimer()
   746  
   747  	var wg sync.WaitGroup
   748  	wg.Add(b.N)
   749  	for i := 0; i < b.N; i++ {
   750  		go func() {
   751  			wg.Done()
   752  		}()
   753  	}
   754  	wg.Wait()
   755  }
   756  
   757  func BenchmarkClosureCall(b *testing.B) {
   758  	sum := 0
   759  	off1 := 1
   760  	for i := 0; i < b.N; i++ {
   761  		off2 := 2
   762  		func() {
   763  			sum += i + off1 + off2
   764  		}()
   765  	}
   766  	_ = sum
   767  }
   768  
   769  func benchmarkWakeupParallel(b *testing.B, spin func(time.Duration)) {
   770  	if runtime.GOMAXPROCS(0) == 1 {
   771  		b.Skip("skipping: GOMAXPROCS=1")
   772  	}
   773  
   774  	wakeDelay := 5 * time.Microsecond
   775  	for _, delay := range []time.Duration{
   776  		0,
   777  		1 * time.Microsecond,
   778  		2 * time.Microsecond,
   779  		5 * time.Microsecond,
   780  		10 * time.Microsecond,
   781  		20 * time.Microsecond,
   782  		50 * time.Microsecond,
   783  		100 * time.Microsecond,
   784  	} {
   785  		b.Run(delay.String(), func(b *testing.B) {
   786  			if b.N == 0 {
   787  				return
   788  			}
   789  			// Start two goroutines, which alternate between being
   790  			// sender and receiver in the following protocol:
   791  			//
   792  			// - The receiver spins for `delay` and then does a
   793  			// blocking receive on a channel.
   794  			//
   795  			// - The sender spins for `delay+wakeDelay` and then
   796  			// sends to the same channel. (The addition of
   797  			// `wakeDelay` improves the probability that the
   798  			// receiver will be blocking when the send occurs when
   799  			// the goroutines execute in parallel.)
   800  			//
   801  			// In each iteration of the benchmark, each goroutine
   802  			// acts once as sender and once as receiver, so each
   803  			// goroutine spins for delay twice.
   804  			//
   805  			// BenchmarkWakeupParallel is used to estimate how
   806  			// efficiently the scheduler parallelizes goroutines in
   807  			// the presence of blocking:
   808  			//
   809  			// - If both goroutines are executed on the same core,
   810  			// an increase in delay by N will increase the time per
   811  			// iteration by 4*N, because all 4 delays are
   812  			// serialized.
   813  			//
   814  			// - Otherwise, an increase in delay by N will increase
   815  			// the time per iteration by 2*N, and the time per
   816  			// iteration is 2 * (runtime overhead + chan
   817  			// send/receive pair + delay + wakeDelay). This allows
   818  			// the runtime overhead, including the time it takes
   819  			// for the unblocked goroutine to be scheduled, to be
   820  			// estimated.
   821  			ping, pong := make(chan struct{}), make(chan struct{})
   822  			start := make(chan struct{})
   823  			done := make(chan struct{})
   824  			go func() {
   825  				<-start
   826  				for i := 0; i < b.N; i++ {
   827  					// sender
   828  					spin(delay + wakeDelay)
   829  					ping <- struct{}{}
   830  					// receiver
   831  					spin(delay)
   832  					<-pong
   833  				}
   834  				done <- struct{}{}
   835  			}()
   836  			go func() {
   837  				for i := 0; i < b.N; i++ {
   838  					// receiver
   839  					spin(delay)
   840  					<-ping
   841  					// sender
   842  					spin(delay + wakeDelay)
   843  					pong <- struct{}{}
   844  				}
   845  				done <- struct{}{}
   846  			}()
   847  			b.ResetTimer()
   848  			start <- struct{}{}
   849  			<-done
   850  			<-done
   851  		})
   852  	}
   853  }
   854  
   855  func BenchmarkWakeupParallelSpinning(b *testing.B) {
   856  	benchmarkWakeupParallel(b, func(d time.Duration) {
   857  		end := time.Now().Add(d)
   858  		for time.Now().Before(end) {
   859  			// do nothing
   860  		}
   861  	})
   862  }
   863  
   864  // sysNanosleep is defined by OS-specific files (such as runtime_linux_test.go)
   865  // to sleep for the given duration. If nil, dependent tests are skipped.
   866  // The implementation should invoke a blocking system call and not
   867  // call time.Sleep, which would deschedule the goroutine.
   868  var sysNanosleep func(d time.Duration)
   869  
   870  func BenchmarkWakeupParallelSyscall(b *testing.B) {
   871  	if sysNanosleep == nil {
   872  		b.Skipf("skipping on %v; sysNanosleep not defined", runtime.GOOS)
   873  	}
   874  	benchmarkWakeupParallel(b, func(d time.Duration) {
   875  		sysNanosleep(d)
   876  	})
   877  }
   878  
   879  type Matrix [][]float64
   880  
   881  func BenchmarkMatmult(b *testing.B) {
   882  	b.StopTimer()
   883  	// matmult is O(N**3) but testing expects O(b.N),
   884  	// so we need to take cube root of b.N
   885  	n := int(math.Cbrt(float64(b.N))) + 1
   886  	A := makeMatrix(n)
   887  	B := makeMatrix(n)
   888  	C := makeMatrix(n)
   889  	b.StartTimer()
   890  	matmult(nil, A, B, C, 0, n, 0, n, 0, n, 8)
   891  }
   892  
   893  func makeMatrix(n int) Matrix {
   894  	m := make(Matrix, n)
   895  	for i := 0; i < n; i++ {
   896  		m[i] = make([]float64, n)
   897  		for j := 0; j < n; j++ {
   898  			m[i][j] = float64(i*n + j)
   899  		}
   900  	}
   901  	return m
   902  }
   903  
   904  func matmult(done chan<- struct{}, A, B, C Matrix, i0, i1, j0, j1, k0, k1, threshold int) {
   905  	di := i1 - i0
   906  	dj := j1 - j0
   907  	dk := k1 - k0
   908  	if di >= dj && di >= dk && di >= threshold {
   909  		// divide in two by y axis
   910  		mi := i0 + di/2
   911  		done1 := make(chan struct{}, 1)
   912  		go matmult(done1, A, B, C, i0, mi, j0, j1, k0, k1, threshold)
   913  		matmult(nil, A, B, C, mi, i1, j0, j1, k0, k1, threshold)
   914  		<-done1
   915  	} else if dj >= dk && dj >= threshold {
   916  		// divide in two by x axis
   917  		mj := j0 + dj/2
   918  		done1 := make(chan struct{}, 1)
   919  		go matmult(done1, A, B, C, i0, i1, j0, mj, k0, k1, threshold)
   920  		matmult(nil, A, B, C, i0, i1, mj, j1, k0, k1, threshold)
   921  		<-done1
   922  	} else if dk >= threshold {
   923  		// divide in two by "k" axis
   924  		// deliberately not parallel because of data races
   925  		mk := k0 + dk/2
   926  		matmult(nil, A, B, C, i0, i1, j0, j1, k0, mk, threshold)
   927  		matmult(nil, A, B, C, i0, i1, j0, j1, mk, k1, threshold)
   928  	} else {
   929  		// the matrices are small enough, compute directly
   930  		for i := i0; i < i1; i++ {
   931  			for j := j0; j < j1; j++ {
   932  				for k := k0; k < k1; k++ {
   933  					C[i][j] += A[i][k] * B[k][j]
   934  				}
   935  			}
   936  		}
   937  	}
   938  	if done != nil {
   939  		done <- struct{}{}
   940  	}
   941  }
   942  
   943  func TestStealOrder(t *testing.T) {
   944  	runtime.RunStealOrderTest()
   945  }
   946  
   947  func TestLockOSThreadNesting(t *testing.T) {
   948  	if runtime.GOARCH == "wasm" {
   949  		t.Skip("no threads on wasm yet")
   950  	}
   951  
   952  	go func() {
   953  		e, i := runtime.LockOSCounts()
   954  		if e != 0 || i != 0 {
   955  			t.Errorf("want locked counts 0, 0; got %d, %d", e, i)
   956  			return
   957  		}
   958  		runtime.LockOSThread()
   959  		runtime.LockOSThread()
   960  		runtime.UnlockOSThread()
   961  		e, i = runtime.LockOSCounts()
   962  		if e != 1 || i != 0 {
   963  			t.Errorf("want locked counts 1, 0; got %d, %d", e, i)
   964  			return
   965  		}
   966  		runtime.UnlockOSThread()
   967  		e, i = runtime.LockOSCounts()
   968  		if e != 0 || i != 0 {
   969  			t.Errorf("want locked counts 0, 0; got %d, %d", e, i)
   970  			return
   971  		}
   972  	}()
   973  }
   974  
   975  func TestLockOSThreadExit(t *testing.T) {
   976  	testLockOSThreadExit(t, "testprog")
   977  }
   978  
   979  func testLockOSThreadExit(t *testing.T, prog string) {
   980  	output := runTestProg(t, prog, "LockOSThreadMain", "GOMAXPROCS=1")
   981  	want := "OK\n"
   982  	if output != want {
   983  		t.Errorf("want %q, got %q", want, output)
   984  	}
   985  
   986  	output = runTestProg(t, prog, "LockOSThreadAlt")
   987  	if output != want {
   988  		t.Errorf("want %q, got %q", want, output)
   989  	}
   990  }
   991  
   992  func TestLockOSThreadAvoidsStatePropagation(t *testing.T) {
   993  	want := "OK\n"
   994  	skip := "unshare not permitted\n"
   995  	output := runTestProg(t, "testprog", "LockOSThreadAvoidsStatePropagation", "GOMAXPROCS=1")
   996  	if output == skip {
   997  		t.Skip("unshare syscall not permitted on this system")
   998  	} else if output != want {
   999  		t.Errorf("want %q, got %q", want, output)
  1000  	}
  1001  }
  1002  
  1003  func TestLockOSThreadTemplateThreadRace(t *testing.T) {
  1004  	testenv.MustHaveGoRun(t)
  1005  
  1006  	exe, err := buildTestProg(t, "testprog")
  1007  	if err != nil {
  1008  		t.Fatal(err)
  1009  	}
  1010  
  1011  	iterations := 100
  1012  	if testing.Short() {
  1013  		// Reduce run time to ~100ms, with much lower probability of
  1014  		// catching issues.
  1015  		iterations = 5
  1016  	}
  1017  	for i := 0; i < iterations; i++ {
  1018  		want := "OK\n"
  1019  		output := runBuiltTestProg(t, exe, "LockOSThreadTemplateThreadRace")
  1020  		if output != want {
  1021  			t.Fatalf("run %d: want %q, got %q", i, want, output)
  1022  		}
  1023  	}
  1024  }
  1025  
  1026  // fakeSyscall emulates a system call.
  1027  //
  1028  //go:nosplit
  1029  func fakeSyscall(duration time.Duration) {
  1030  	runtime.Entersyscall()
  1031  	for start := runtime.Nanotime(); runtime.Nanotime()-start < int64(duration); {
  1032  	}
  1033  	runtime.Exitsyscall()
  1034  }
  1035  
  1036  // Check that a goroutine will be preempted if it is calling short system calls.
  1037  func testPreemptionAfterSyscall(t *testing.T, syscallDuration time.Duration) {
  1038  	if runtime.GOARCH == "wasm" {
  1039  		t.Skip("no preemption on wasm yet")
  1040  	}
  1041  
  1042  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(2))
  1043  
  1044  	interations := 10
  1045  	if testing.Short() {
  1046  		interations = 1
  1047  	}
  1048  	const (
  1049  		maxDuration = 5 * time.Second
  1050  		nroutines   = 8
  1051  	)
  1052  
  1053  	for i := 0; i < interations; i++ {
  1054  		c := make(chan bool, nroutines)
  1055  		stop := uint32(0)
  1056  
  1057  		start := time.Now()
  1058  		for g := 0; g < nroutines; g++ {
  1059  			go func(stop *uint32) {
  1060  				c <- true
  1061  				for atomic.LoadUint32(stop) == 0 {
  1062  					fakeSyscall(syscallDuration)
  1063  				}
  1064  				c <- true
  1065  			}(&stop)
  1066  		}
  1067  		// wait until all goroutines have started.
  1068  		for g := 0; g < nroutines; g++ {
  1069  			<-c
  1070  		}
  1071  		atomic.StoreUint32(&stop, 1)
  1072  		// wait until all goroutines have finished.
  1073  		for g := 0; g < nroutines; g++ {
  1074  			<-c
  1075  		}
  1076  		duration := time.Since(start)
  1077  
  1078  		if duration > maxDuration {
  1079  			t.Errorf("timeout exceeded: %v (%v)", duration, maxDuration)
  1080  		}
  1081  	}
  1082  }
  1083  
  1084  func TestPreemptionAfterSyscall(t *testing.T) {
  1085  	if runtime.GOOS == "plan9" {
  1086  		testenv.SkipFlaky(t, 41015)
  1087  	}
  1088  
  1089  	for _, i := range []time.Duration{10, 100, 1000} {
  1090  		d := i * time.Microsecond
  1091  		t.Run(fmt.Sprint(d), func(t *testing.T) {
  1092  			testPreemptionAfterSyscall(t, d)
  1093  		})
  1094  	}
  1095  }
  1096  
  1097  func TestGetgThreadSwitch(t *testing.T) {
  1098  	runtime.RunGetgThreadSwitchTest()
  1099  }
  1100  
  1101  // TestNetpollBreak tests that netpollBreak can break a netpoll.
  1102  // This test is not particularly safe since the call to netpoll
  1103  // will pick up any stray files that are ready, but it should work
  1104  // OK as long it is not run in parallel.
  1105  func TestNetpollBreak(t *testing.T) {
  1106  	if runtime.GOMAXPROCS(0) == 1 {
  1107  		t.Skip("skipping: GOMAXPROCS=1")
  1108  	}
  1109  
  1110  	// Make sure that netpoll is initialized.
  1111  	runtime.NetpollGenericInit()
  1112  
  1113  	start := time.Now()
  1114  	c := make(chan bool, 2)
  1115  	go func() {
  1116  		c <- true
  1117  		runtime.Netpoll(10 * time.Second.Nanoseconds())
  1118  		c <- true
  1119  	}()
  1120  	<-c
  1121  	// Loop because the break might get eaten by the scheduler.
  1122  	// Break twice to break both the netpoll we started and the
  1123  	// scheduler netpoll.
  1124  loop:
  1125  	for {
  1126  		runtime.Usleep(100)
  1127  		runtime.NetpollBreak()
  1128  		runtime.NetpollBreak()
  1129  		select {
  1130  		case <-c:
  1131  			break loop
  1132  		default:
  1133  		}
  1134  	}
  1135  	if dur := time.Since(start); dur > 5*time.Second {
  1136  		t.Errorf("netpollBreak did not interrupt netpoll: slept for: %v", dur)
  1137  	}
  1138  }
  1139  
  1140  // TestBigGOMAXPROCS tests that setting GOMAXPROCS to a large value
  1141  // doesn't cause a crash at startup. See issue 38474.
  1142  func TestBigGOMAXPROCS(t *testing.T) {
  1143  	t.Parallel()
  1144  	output := runTestProg(t, "testprog", "NonexistentTest", "GOMAXPROCS=1024")
  1145  	// Ignore error conditions on small machines.
  1146  	for _, errstr := range []string{
  1147  		"failed to create new OS thread",
  1148  		"cannot allocate memory",
  1149  	} {
  1150  		if strings.Contains(output, errstr) {
  1151  			t.Skipf("failed to create 1024 threads")
  1152  		}
  1153  	}
  1154  	if !strings.Contains(output, "unknown function: NonexistentTest") {
  1155  		t.Errorf("output:\n%s\nwanted:\nunknown function: NonexistentTest", output)
  1156  	}
  1157  }
  1158
View as plain text