Source file src/os/signal/signal_test.go

     1  // Copyright 2009 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  //go:build unix
     6  
     7  package signal
     8  
     9  import (
    10  	"bytes"
    11  	"context"
    12  	"flag"
    13  	"fmt"
    14  	"internal/testenv"
    15  	"os"
    16  	"os/exec"
    17  	"runtime"
    18  	"runtime/trace"
    19  	"strconv"
    20  	"strings"
    21  	"sync"
    22  	"syscall"
    23  	"testing"
    24  	"time"
    25  )
    26  
    27  // settleTime is an upper bound on how long we expect signals to take to be
    28  // delivered. Lower values make the test faster, but also flakier — especially
    29  // on heavily loaded systems.
    30  //
    31  // The current value is set based on flakes observed in the Go builders.
    32  var settleTime = 100 * time.Millisecond
    33  
    34  // fatalWaitingTime is an absurdly long time to wait for signals to be
    35  // delivered but, using it, we (hopefully) eliminate test flakes on the
    36  // build servers. See #46736 for discussion.
    37  var fatalWaitingTime = 30 * time.Second
    38  
    39  func init() {
    40  	if testenv.Builder() == "solaris-amd64-oraclerel" {
    41  		// The solaris-amd64-oraclerel builder has been observed to time out in
    42  		// TestNohup even with a 250ms settle time.
    43  		//
    44  		// Use a much longer settle time on that builder to try to suss out whether
    45  		// the test is flaky due to builder slowness (which may mean we need a
    46  		// longer GO_TEST_TIMEOUT_SCALE) or due to a dropped signal (which may
    47  		// instead need a test-skip and upstream bug filed against the Solaris
    48  		// kernel).
    49  		//
    50  		// See https://golang.org/issue/33174.
    51  		settleTime = 5 * time.Second
    52  	} else if runtime.GOOS == "linux" && strings.HasPrefix(runtime.GOARCH, "ppc64") {
    53  		// Older linux kernels seem to have some hiccups delivering the signal
    54  		// in a timely manner on ppc64 and ppc64le. When running on a
    55  		// ppc64le/ubuntu 16.04/linux 4.4 host the time can vary quite
    56  		// substantially even on an idle system. 5 seconds is twice any value
    57  		// observed when running 10000 tests on such a system.
    58  		settleTime = 5 * time.Second
    59  	} else if s := os.Getenv("GO_TEST_TIMEOUT_SCALE"); s != "" {
    60  		if scale, err := strconv.Atoi(s); err == nil {
    61  			settleTime *= time.Duration(scale)
    62  		}
    63  	}
    64  }
    65  
    66  func waitSig(t *testing.T, c <-chan os.Signal, sig os.Signal) {
    67  	t.Helper()
    68  	waitSig1(t, c, sig, false)
    69  }
    70  func waitSigAll(t *testing.T, c <-chan os.Signal, sig os.Signal) {
    71  	t.Helper()
    72  	waitSig1(t, c, sig, true)
    73  }
    74  
    75  func waitSig1(t *testing.T, c <-chan os.Signal, sig os.Signal, all bool) {
    76  	t.Helper()
    77  
    78  	// Sleep multiple times to give the kernel more tries to
    79  	// deliver the signal.
    80  	start := time.Now()
    81  	timer := time.NewTimer(settleTime / 10)
    82  	defer timer.Stop()
    83  	// If the caller notified for all signals on c, filter out SIGURG,
    84  	// which is used for runtime preemption and can come at unpredictable times.
    85  	// General user code should filter out all unexpected signals instead of just
    86  	// SIGURG, but since os/signal is tightly coupled to the runtime it seems
    87  	// appropriate to be stricter here.
    88  	for time.Since(start) < fatalWaitingTime {
    89  		select {
    90  		case s := <-c:
    91  			if s == sig {
    92  				return
    93  			}
    94  			if !all || s != syscall.SIGURG {
    95  				t.Fatalf("signal was %v, want %v", s, sig)
    96  			}
    97  		case <-timer.C:
    98  			timer.Reset(settleTime / 10)
    99  		}
   100  	}
   101  	t.Fatalf("timeout after %v waiting for %v", fatalWaitingTime, sig)
   102  }
   103  
   104  // quiesce waits until we can be reasonably confident that all pending signals
   105  // have been delivered by the OS.
   106  func quiesce() {
   107  	// The kernel will deliver a signal as a thread returns
   108  	// from a syscall. If the only active thread is sleeping,
   109  	// and the system is busy, the kernel may not get around
   110  	// to waking up a thread to catch the signal.
   111  	// We try splitting up the sleep to give the kernel
   112  	// many chances to deliver the signal.
   113  	start := time.Now()
   114  	for time.Since(start) < settleTime {
   115  		time.Sleep(settleTime / 10)
   116  	}
   117  }
   118  
   119  // Test that basic signal handling works.
   120  func TestSignal(t *testing.T) {
   121  	// Ask for SIGHUP
   122  	c := make(chan os.Signal, 1)
   123  	Notify(c, syscall.SIGHUP)
   124  	defer Stop(c)
   125  
   126  	// Send this process a SIGHUP
   127  	t.Logf("sighup...")
   128  	syscall.Kill(syscall.Getpid(), syscall.SIGHUP)
   129  	waitSig(t, c, syscall.SIGHUP)
   130  
   131  	// Ask for everything we can get. The buffer size has to be
   132  	// more than 1, since the runtime might send SIGURG signals.
   133  	// Using 10 is arbitrary.
   134  	c1 := make(chan os.Signal, 10)
   135  	Notify(c1)
   136  	// Stop relaying the SIGURG signals. See #49724
   137  	Reset(syscall.SIGURG)
   138  	defer Stop(c1)
   139  
   140  	// Send this process a SIGWINCH
   141  	t.Logf("sigwinch...")
   142  	syscall.Kill(syscall.Getpid(), syscall.SIGWINCH)
   143  	waitSigAll(t, c1, syscall.SIGWINCH)
   144  
   145  	// Send two more SIGHUPs, to make sure that
   146  	// they get delivered on c1 and that not reading
   147  	// from c does not block everything.
   148  	t.Logf("sighup...")
   149  	syscall.Kill(syscall.Getpid(), syscall.SIGHUP)
   150  	waitSigAll(t, c1, syscall.SIGHUP)
   151  	t.Logf("sighup...")
   152  	syscall.Kill(syscall.Getpid(), syscall.SIGHUP)
   153  	waitSigAll(t, c1, syscall.SIGHUP)
   154  
   155  	// The first SIGHUP should be waiting for us on c.
   156  	waitSig(t, c, syscall.SIGHUP)
   157  }
   158  
   159  func TestStress(t *testing.T) {
   160  	dur := 3 * time.Second
   161  	if testing.Short() {
   162  		dur = 100 * time.Millisecond
   163  	}
   164  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(4))
   165  
   166  	sig := make(chan os.Signal, 1)
   167  	Notify(sig, syscall.SIGUSR1)
   168  
   169  	go func() {
   170  		stop := time.After(dur)
   171  		for {
   172  			select {
   173  			case <-stop:
   174  				// Allow enough time for all signals to be delivered before we stop
   175  				// listening for them.
   176  				quiesce()
   177  				Stop(sig)
   178  				// According to its documentation, “[w]hen Stop returns, it in
   179  				// guaranteed that c will receive no more signals.” So we can safely
   180  				// close sig here: if there is a send-after-close race here, that is a
   181  				// bug in Stop and we would like to detect it.
   182  				close(sig)
   183  				return
   184  
   185  			default:
   186  				syscall.Kill(syscall.Getpid(), syscall.SIGUSR1)
   187  				runtime.Gosched()
   188  			}
   189  		}
   190  	}()
   191  
   192  	for range sig {
   193  		// Receive signals until the sender closes sig.
   194  	}
   195  }
   196  
   197  func testCancel(t *testing.T, ignore bool) {
   198  	// Ask to be notified on c1 when a SIGWINCH is received.
   199  	c1 := make(chan os.Signal, 1)
   200  	Notify(c1, syscall.SIGWINCH)
   201  	defer Stop(c1)
   202  
   203  	// Ask to be notified on c2 when a SIGHUP is received.
   204  	c2 := make(chan os.Signal, 1)
   205  	Notify(c2, syscall.SIGHUP)
   206  	defer Stop(c2)
   207  
   208  	// Send this process a SIGWINCH and wait for notification on c1.
   209  	syscall.Kill(syscall.Getpid(), syscall.SIGWINCH)
   210  	waitSig(t, c1, syscall.SIGWINCH)
   211  
   212  	// Send this process a SIGHUP and wait for notification on c2.
   213  	syscall.Kill(syscall.Getpid(), syscall.SIGHUP)
   214  	waitSig(t, c2, syscall.SIGHUP)
   215  
   216  	// Ignore, or reset the signal handlers for, SIGWINCH and SIGHUP.
   217  	// Either way, this should undo both calls to Notify above.
   218  	if ignore {
   219  		Ignore(syscall.SIGWINCH, syscall.SIGHUP)
   220  		// Don't bother deferring a call to Reset: it is documented to undo Notify,
   221  		// but its documentation says nothing about Ignore, and (as of the time of
   222  		// writing) it empirically does not undo an Ignore.
   223  	} else {
   224  		Reset(syscall.SIGWINCH, syscall.SIGHUP)
   225  	}
   226  
   227  	// Send this process a SIGWINCH. It should be ignored.
   228  	syscall.Kill(syscall.Getpid(), syscall.SIGWINCH)
   229  
   230  	// If ignoring, Send this process a SIGHUP. It should be ignored.
   231  	if ignore {
   232  		syscall.Kill(syscall.Getpid(), syscall.SIGHUP)
   233  	}
   234  
   235  	quiesce()
   236  
   237  	select {
   238  	case s := <-c1:
   239  		t.Errorf("unexpected signal %v", s)
   240  	default:
   241  		// nothing to read - good
   242  	}
   243  
   244  	select {
   245  	case s := <-c2:
   246  		t.Errorf("unexpected signal %v", s)
   247  	default:
   248  		// nothing to read - good
   249  	}
   250  
   251  	// One or both of the signals may have been blocked for this process
   252  	// by the calling process.
   253  	// Discard any queued signals now to avoid interfering with other tests.
   254  	Notify(c1, syscall.SIGWINCH)
   255  	Notify(c2, syscall.SIGHUP)
   256  	quiesce()
   257  }
   258  
   259  // Test that Reset cancels registration for listed signals on all channels.
   260  func TestReset(t *testing.T) {
   261  	testCancel(t, false)
   262  }
   263  
   264  // Test that Ignore cancels registration for listed signals on all channels.
   265  func TestIgnore(t *testing.T) {
   266  	testCancel(t, true)
   267  }
   268  
   269  // Test that Ignored correctly detects changes to the ignored status of a signal.
   270  func TestIgnored(t *testing.T) {
   271  	// Ask to be notified on SIGWINCH.
   272  	c := make(chan os.Signal, 1)
   273  	Notify(c, syscall.SIGWINCH)
   274  
   275  	// If we're being notified, then the signal should not be ignored.
   276  	if Ignored(syscall.SIGWINCH) {
   277  		t.Errorf("expected SIGWINCH to not be ignored.")
   278  	}
   279  	Stop(c)
   280  	Ignore(syscall.SIGWINCH)
   281  
   282  	// We're no longer paying attention to this signal.
   283  	if !Ignored(syscall.SIGWINCH) {
   284  		t.Errorf("expected SIGWINCH to be ignored when explicitly ignoring it.")
   285  	}
   286  
   287  	Reset()
   288  }
   289  
   290  var checkSighupIgnored = flag.Bool("check_sighup_ignored", false, "if true, TestDetectNohup will fail if SIGHUP is not ignored.")
   291  
   292  // Test that Ignored(SIGHUP) correctly detects whether it is being run under nohup.
   293  func TestDetectNohup(t *testing.T) {
   294  	if *checkSighupIgnored {
   295  		if !Ignored(syscall.SIGHUP) {
   296  			t.Fatal("SIGHUP is not ignored.")
   297  		} else {
   298  			t.Log("SIGHUP is ignored.")
   299  		}
   300  	} else {
   301  		defer Reset()
   302  		// Ugly: ask for SIGHUP so that child will not have no-hup set
   303  		// even if test is running under nohup environment.
   304  		// We have no intention of reading from c.
   305  		c := make(chan os.Signal, 1)
   306  		Notify(c, syscall.SIGHUP)
   307  		if out, err := testenv.Command(t, os.Args[0], "-test.run=^TestDetectNohup$", "-check_sighup_ignored").CombinedOutput(); err == nil {
   308  			t.Errorf("ran test with -check_sighup_ignored and it succeeded: expected failure.\nOutput:\n%s", out)
   309  		}
   310  		Stop(c)
   311  
   312  		// Again, this time with nohup, assuming we can find it.
   313  		_, err := os.Stat("/usr/bin/nohup")
   314  		if err != nil {
   315  			t.Skip("cannot find nohup; skipping second half of test")
   316  		}
   317  		Ignore(syscall.SIGHUP)
   318  		os.Remove("nohup.out")
   319  		out, err := testenv.Command(t, "/usr/bin/nohup", os.Args[0], "-test.run=^TestDetectNohup$", "-check_sighup_ignored").CombinedOutput()
   320  
   321  		data, _ := os.ReadFile("nohup.out")
   322  		os.Remove("nohup.out")
   323  		if err != nil {
   324  			// nohup doesn't work on new LUCI darwin builders due to the
   325  			// type of launchd service the test run under. See
   326  			// https://go.dev/issue/63875.
   327  			if runtime.GOOS == "darwin" && strings.Contains(string(out), "nohup: can't detach from console: Inappropriate ioctl for device") {
   328  				t.Skip("Skipping nohup test due to darwin builder limitation. See https://go.dev/issue/63875.")
   329  			}
   330  
   331  			t.Errorf("ran test with -check_sighup_ignored under nohup and it failed: expected success.\nError: %v\nOutput:\n%s%s", err, out, data)
   332  		}
   333  	}
   334  }
   335  
   336  var (
   337  	sendUncaughtSighup = flag.Int("send_uncaught_sighup", 0, "send uncaught SIGHUP during TestStop")
   338  	dieFromSighup      = flag.Bool("die_from_sighup", false, "wait to die from uncaught SIGHUP")
   339  )
   340  
   341  // Test that Stop cancels the channel's registrations.
   342  func TestStop(t *testing.T) {
   343  	sigs := []syscall.Signal{
   344  		syscall.SIGWINCH,
   345  		syscall.SIGHUP,
   346  		syscall.SIGUSR1,
   347  	}
   348  
   349  	for _, sig := range sigs {
   350  		sig := sig
   351  		t.Run(fmt.Sprint(sig), func(t *testing.T) {
   352  			// When calling Notify with a specific signal,
   353  			// independent signals should not interfere with each other,
   354  			// and we end up needing to wait for signals to quiesce a lot.
   355  			// Test the three different signals concurrently.
   356  			t.Parallel()
   357  
   358  			// If the signal is not ignored, send the signal before registering a
   359  			// channel to verify the behavior of the default Go handler.
   360  			// If it's SIGWINCH or SIGUSR1 we should not see it.
   361  			// If it's SIGHUP, maybe we'll die. Let the flag tell us what to do.
   362  			mayHaveBlockedSignal := false
   363  			if !Ignored(sig) && (sig != syscall.SIGHUP || *sendUncaughtSighup == 1) {
   364  				syscall.Kill(syscall.Getpid(), sig)
   365  				quiesce()
   366  
   367  				// We don't know whether sig is blocked for this process; see
   368  				// https://golang.org/issue/38165. Assume that it could be.
   369  				mayHaveBlockedSignal = true
   370  			}
   371  
   372  			// Ask for signal
   373  			c := make(chan os.Signal, 1)
   374  			Notify(c, sig)
   375  
   376  			// Send this process the signal again.
   377  			syscall.Kill(syscall.Getpid(), sig)
   378  			waitSig(t, c, sig)
   379  
   380  			if mayHaveBlockedSignal {
   381  				// We may have received a queued initial signal in addition to the one
   382  				// that we sent after Notify. If so, waitSig may have observed that
   383  				// initial signal instead of the second one, and we may need to wait for
   384  				// the second signal to clear. Do that now.
   385  				quiesce()
   386  				select {
   387  				case <-c:
   388  				default:
   389  				}
   390  			}
   391  
   392  			// Stop watching for the signal and send it again.
   393  			// If it's SIGHUP, maybe we'll die. Let the flag tell us what to do.
   394  			Stop(c)
   395  			if sig != syscall.SIGHUP || *sendUncaughtSighup == 2 {
   396  				syscall.Kill(syscall.Getpid(), sig)
   397  				quiesce()
   398  
   399  				select {
   400  				case s := <-c:
   401  					t.Errorf("unexpected signal %v", s)
   402  				default:
   403  					// nothing to read - good
   404  				}
   405  
   406  				// If we're going to receive a signal, it has almost certainly been
   407  				// received by now. However, it may have been blocked for this process —
   408  				// we don't know. Explicitly unblock it and wait for it to clear now.
   409  				Notify(c, sig)
   410  				quiesce()
   411  				Stop(c)
   412  			}
   413  		})
   414  	}
   415  }
   416  
   417  // Test that when run under nohup, an uncaught SIGHUP does not kill the program.
   418  func TestNohup(t *testing.T) {
   419  	// When run without nohup, the test should crash on an uncaught SIGHUP.
   420  	// When run under nohup, the test should ignore uncaught SIGHUPs,
   421  	// because the runtime is not supposed to be listening for them.
   422  	// Either way, TestStop should still be able to catch them when it wants them
   423  	// and then when it stops wanting them, the original behavior should resume.
   424  	//
   425  	// send_uncaught_sighup=1 sends the SIGHUP before starting to listen for SIGHUPs.
   426  	// send_uncaught_sighup=2 sends the SIGHUP after no longer listening for SIGHUPs.
   427  	//
   428  	// Both should fail without nohup and succeed with nohup.
   429  
   430  	t.Run("uncaught", func(t *testing.T) {
   431  		// Ugly: ask for SIGHUP so that child will not have no-hup set
   432  		// even if test is running under nohup environment.
   433  		// We have no intention of reading from c.
   434  		c := make(chan os.Signal, 1)
   435  		Notify(c, syscall.SIGHUP)
   436  		t.Cleanup(func() { Stop(c) })
   437  
   438  		var subTimeout time.Duration
   439  		if deadline, ok := t.Deadline(); ok {
   440  			subTimeout = time.Until(deadline)
   441  			subTimeout -= subTimeout / 10 // Leave 10% headroom for propagating output.
   442  		}
   443  		for i := 1; i <= 2; i++ {
   444  			i := i
   445  			t.Run(fmt.Sprintf("%d", i), func(t *testing.T) {
   446  				t.Parallel()
   447  
   448  				args := []string{
   449  					"-test.v",
   450  					"-test.run=^TestStop$",
   451  					"-send_uncaught_sighup=" + strconv.Itoa(i),
   452  					"-die_from_sighup",
   453  				}
   454  				if subTimeout != 0 {
   455  					args = append(args, fmt.Sprintf("-test.timeout=%v", subTimeout))
   456  				}
   457  				out, err := testenv.Command(t, os.Args[0], args...).CombinedOutput()
   458  
   459  				if err == nil {
   460  					t.Errorf("ran test with -send_uncaught_sighup=%d and it succeeded: expected failure.\nOutput:\n%s", i, out)
   461  				} else {
   462  					t.Logf("test with -send_uncaught_sighup=%d failed as expected.\nError: %v\nOutput:\n%s", i, err, out)
   463  				}
   464  			})
   465  		}
   466  	})
   467  
   468  	t.Run("nohup", func(t *testing.T) {
   469  		// Skip the nohup test below when running in tmux on darwin, since nohup
   470  		// doesn't work correctly there. See issue #5135.
   471  		if runtime.GOOS == "darwin" && os.Getenv("TMUX") != "" {
   472  			t.Skip("Skipping nohup test due to running in tmux on darwin")
   473  		}
   474  
   475  		// Again, this time with nohup, assuming we can find it.
   476  		_, err := exec.LookPath("nohup")
   477  		if err != nil {
   478  			t.Skip("cannot find nohup; skipping second half of test")
   479  		}
   480  
   481  		var subTimeout time.Duration
   482  		if deadline, ok := t.Deadline(); ok {
   483  			subTimeout = time.Until(deadline)
   484  			subTimeout -= subTimeout / 10 // Leave 10% headroom for propagating output.
   485  		}
   486  		for i := 1; i <= 2; i++ {
   487  			i := i
   488  			t.Run(fmt.Sprintf("%d", i), func(t *testing.T) {
   489  				t.Parallel()
   490  
   491  				// POSIX specifies that nohup writes to a file named nohup.out if standard
   492  				// output is a terminal. However, for an exec.Cmd, standard output is
   493  				// not a terminal — so we don't need to read or remove that file (and,
   494  				// indeed, cannot even create it if the current user is unable to write to
   495  				// GOROOT/src, such as when GOROOT is installed and owned by root).
   496  
   497  				args := []string{
   498  					os.Args[0],
   499  					"-test.v",
   500  					"-test.run=^TestStop$",
   501  					"-send_uncaught_sighup=" + strconv.Itoa(i),
   502  				}
   503  				if subTimeout != 0 {
   504  					args = append(args, fmt.Sprintf("-test.timeout=%v", subTimeout))
   505  				}
   506  				out, err := testenv.Command(t, "nohup", args...).CombinedOutput()
   507  
   508  				if err != nil {
   509  					// nohup doesn't work on new LUCI darwin builders due to the
   510  					// type of launchd service the test run under. See
   511  					// https://go.dev/issue/63875.
   512  					if runtime.GOOS == "darwin" && strings.Contains(string(out), "nohup: can't detach from console: Inappropriate ioctl for device") {
   513  						// TODO(go.dev/issue/63799): A false-positive in vet reports a
   514  						// t.Skip here as invalid. Switch back to t.Skip once fixed.
   515  						t.Logf("Skipping nohup test due to darwin builder limitation. See https://go.dev/issue/63875.")
   516  						return
   517  					}
   518  
   519  					t.Errorf("ran test with -send_uncaught_sighup=%d under nohup and it failed: expected success.\nError: %v\nOutput:\n%s", i, err, out)
   520  				} else {
   521  					t.Logf("ran test with -send_uncaught_sighup=%d under nohup.\nOutput:\n%s", i, out)
   522  				}
   523  			})
   524  		}
   525  	})
   526  }
   527  
   528  // Test that SIGCONT works (issue 8953).
   529  func TestSIGCONT(t *testing.T) {
   530  	c := make(chan os.Signal, 1)
   531  	Notify(c, syscall.SIGCONT)
   532  	defer Stop(c)
   533  	syscall.Kill(syscall.Getpid(), syscall.SIGCONT)
   534  	waitSig(t, c, syscall.SIGCONT)
   535  }
   536  
   537  // Test race between stopping and receiving a signal (issue 14571).
   538  func TestAtomicStop(t *testing.T) {
   539  	if os.Getenv("GO_TEST_ATOMIC_STOP") != "" {
   540  		atomicStopTestProgram(t)
   541  		t.Fatal("atomicStopTestProgram returned")
   542  	}
   543  
   544  	testenv.MustHaveExec(t)
   545  
   546  	// Call Notify for SIGINT before starting the child process.
   547  	// That ensures that SIGINT is not ignored for the child.
   548  	// This is necessary because if SIGINT is ignored when a
   549  	// Go program starts, then it remains ignored, and closing
   550  	// the last notification channel for SIGINT will switch it
   551  	// back to being ignored. In that case the assumption of
   552  	// atomicStopTestProgram, that it will either die from SIGINT
   553  	// or have it be reported, breaks down, as there is a third
   554  	// option: SIGINT might be ignored.
   555  	cs := make(chan os.Signal, 1)
   556  	Notify(cs, syscall.SIGINT)
   557  	defer Stop(cs)
   558  
   559  	const execs = 10
   560  	for i := 0; i < execs; i++ {
   561  		timeout := "0"
   562  		if deadline, ok := t.Deadline(); ok {
   563  			timeout = time.Until(deadline).String()
   564  		}
   565  		cmd := testenv.Command(t, os.Args[0], "-test.run=^TestAtomicStop$", "-test.timeout="+timeout)
   566  		cmd.Env = append(os.Environ(), "GO_TEST_ATOMIC_STOP=1")
   567  		out, err := cmd.CombinedOutput()
   568  		if err == nil {
   569  			if len(out) > 0 {
   570  				t.Logf("iteration %d: output %s", i, out)
   571  			}
   572  		} else {
   573  			t.Logf("iteration %d: exit status %q: output: %s", i, err, out)
   574  		}
   575  
   576  		lost := bytes.Contains(out, []byte("lost signal"))
   577  		if lost {
   578  			t.Errorf("iteration %d: lost signal", i)
   579  		}
   580  
   581  		// The program should either die due to SIGINT,
   582  		// or exit with success without printing "lost signal".
   583  		if err == nil {
   584  			if len(out) > 0 && !lost {
   585  				t.Errorf("iteration %d: unexpected output", i)
   586  			}
   587  		} else {
   588  			if ee, ok := err.(*exec.ExitError); !ok {
   589  				t.Errorf("iteration %d: error (%v) has type %T; expected exec.ExitError", i, err, err)
   590  			} else if ws, ok := ee.Sys().(syscall.WaitStatus); !ok {
   591  				t.Errorf("iteration %d: error.Sys (%v) has type %T; expected syscall.WaitStatus", i, ee.Sys(), ee.Sys())
   592  			} else if !ws.Signaled() || ws.Signal() != syscall.SIGINT {
   593  				t.Errorf("iteration %d: got exit status %v; expected SIGINT", i, ee)
   594  			}
   595  		}
   596  	}
   597  }
   598  
   599  // atomicStopTestProgram is run in a subprocess by TestAtomicStop.
   600  // It tries to trigger a signal delivery race. This function should
   601  // either catch a signal or die from it.
   602  func atomicStopTestProgram(t *testing.T) {
   603  	// This test won't work if SIGINT is ignored here.
   604  	if Ignored(syscall.SIGINT) {
   605  		fmt.Println("SIGINT is ignored")
   606  		os.Exit(1)
   607  	}
   608  
   609  	const tries = 10
   610  
   611  	timeout := 2 * time.Second
   612  	if deadline, ok := t.Deadline(); ok {
   613  		// Give each try an equal slice of the deadline, with one slice to spare for
   614  		// cleanup.
   615  		timeout = time.Until(deadline) / (tries + 1)
   616  	}
   617  
   618  	pid := syscall.Getpid()
   619  	printed := false
   620  	for i := 0; i < tries; i++ {
   621  		cs := make(chan os.Signal, 1)
   622  		Notify(cs, syscall.SIGINT)
   623  
   624  		var wg sync.WaitGroup
   625  		wg.Add(1)
   626  		go func() {
   627  			defer wg.Done()
   628  			Stop(cs)
   629  		}()
   630  
   631  		syscall.Kill(pid, syscall.SIGINT)
   632  
   633  		// At this point we should either die from SIGINT or
   634  		// get a notification on cs. If neither happens, we
   635  		// dropped the signal. It is given 2 seconds to
   636  		// deliver, as needed for gccgo on some loaded test systems.
   637  
   638  		select {
   639  		case <-cs:
   640  		case <-time.After(timeout):
   641  			if !printed {
   642  				fmt.Print("lost signal on tries:")
   643  				printed = true
   644  			}
   645  			fmt.Printf(" %d", i)
   646  		}
   647  
   648  		wg.Wait()
   649  	}
   650  	if printed {
   651  		fmt.Print("\n")
   652  	}
   653  
   654  	os.Exit(0)
   655  }
   656  
   657  func TestTime(t *testing.T) {
   658  	// Test that signal works fine when we are in a call to get time,
   659  	// which on some platforms is using VDSO. See issue #34391.
   660  	dur := 3 * time.Second
   661  	if testing.Short() {
   662  		dur = 100 * time.Millisecond
   663  	}
   664  	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(4))
   665  
   666  	sig := make(chan os.Signal, 1)
   667  	Notify(sig, syscall.SIGUSR1)
   668  
   669  	stop := make(chan struct{})
   670  	go func() {
   671  		for {
   672  			select {
   673  			case <-stop:
   674  				// Allow enough time for all signals to be delivered before we stop
   675  				// listening for them.
   676  				quiesce()
   677  				Stop(sig)
   678  				// According to its documentation, “[w]hen Stop returns, it in
   679  				// guaranteed that c will receive no more signals.” So we can safely
   680  				// close sig here: if there is a send-after-close race, that is a bug in
   681  				// Stop and we would like to detect it.
   682  				close(sig)
   683  				return
   684  
   685  			default:
   686  				syscall.Kill(syscall.Getpid(), syscall.SIGUSR1)
   687  				runtime.Gosched()
   688  			}
   689  		}
   690  	}()
   691  
   692  	done := make(chan struct{})
   693  	go func() {
   694  		for range sig {
   695  			// Receive signals until the sender closes sig.
   696  		}
   697  		close(done)
   698  	}()
   699  
   700  	t0 := time.Now()
   701  	for t1 := t0; t1.Sub(t0) < dur; t1 = time.Now() {
   702  	} // hammering on getting time
   703  
   704  	close(stop)
   705  	<-done
   706  }
   707  
   708  var (
   709  	checkNotifyContext = flag.Bool("check_notify_ctx", false, "if true, TestNotifyContext will fail if SIGINT is not received.")
   710  	ctxNotifyTimes     = flag.Int("ctx_notify_times", 1, "number of times a SIGINT signal should be received")
   711  )
   712  
   713  func TestNotifyContextNotifications(t *testing.T) {
   714  	if *checkNotifyContext {
   715  		ctx, _ := NotifyContext(context.Background(), syscall.SIGINT)
   716  		// We want to make sure not to be calling Stop() internally on NotifyContext() when processing a received signal.
   717  		// Being able to wait for a number of received system signals allows us to do so.
   718  		var wg sync.WaitGroup
   719  		n := *ctxNotifyTimes
   720  		wg.Add(n)
   721  		for i := 0; i < n; i++ {
   722  			go func() {
   723  				syscall.Kill(syscall.Getpid(), syscall.SIGINT)
   724  				wg.Done()
   725  			}()
   726  		}
   727  		wg.Wait()
   728  		<-ctx.Done()
   729  		fmt.Println("received SIGINT")
   730  		// Sleep to give time to simultaneous signals to reach the process.
   731  		// These signals must be ignored given stop() is not called on this code.
   732  		// We want to guarantee a SIGINT doesn't cause a premature termination of the program.
   733  		time.Sleep(settleTime)
   734  		return
   735  	}
   736  
   737  	t.Parallel()
   738  	testCases := []struct {
   739  		name string
   740  		n    int // number of times a SIGINT should be notified.
   741  	}{
   742  		{"once", 1},
   743  		{"multiple", 10},
   744  	}
   745  	for _, tc := range testCases {
   746  		tc := tc
   747  		t.Run(tc.name, func(t *testing.T) {
   748  			t.Parallel()
   749  
   750  			var subTimeout time.Duration
   751  			if deadline, ok := t.Deadline(); ok {
   752  				timeout := time.Until(deadline)
   753  				if timeout < 2*settleTime {
   754  					t.Fatalf("starting test with less than %v remaining", 2*settleTime)
   755  				}
   756  				subTimeout = timeout - (timeout / 10) // Leave 10% headroom for cleaning up subprocess.
   757  			}
   758  
   759  			args := []string{
   760  				"-test.v",
   761  				"-test.run=^TestNotifyContextNotifications$",
   762  				"-check_notify_ctx",
   763  				fmt.Sprintf("-ctx_notify_times=%d", tc.n),
   764  			}
   765  			if subTimeout != 0 {
   766  				args = append(args, fmt.Sprintf("-test.timeout=%v", subTimeout))
   767  			}
   768  			out, err := testenv.Command(t, os.Args[0], args...).CombinedOutput()
   769  			if err != nil {
   770  				t.Errorf("ran test with -check_notify_ctx_notification and it failed with %v.\nOutput:\n%s", err, out)
   771  			}
   772  			if want := []byte("received SIGINT\n"); !bytes.Contains(out, want) {
   773  				t.Errorf("got %q, wanted %q", out, want)
   774  			}
   775  		})
   776  	}
   777  }
   778  
   779  func TestNotifyContextStop(t *testing.T) {
   780  	Ignore(syscall.SIGHUP)
   781  	if !Ignored(syscall.SIGHUP) {
   782  		t.Errorf("expected SIGHUP to be ignored when explicitly ignoring it.")
   783  	}
   784  
   785  	parent, cancelParent := context.WithCancel(context.Background())
   786  	defer cancelParent()
   787  	c, stop := NotifyContext(parent, syscall.SIGHUP)
   788  	defer stop()
   789  
   790  	// If we're being notified, then the signal should not be ignored.
   791  	if Ignored(syscall.SIGHUP) {
   792  		t.Errorf("expected SIGHUP to not be ignored.")
   793  	}
   794  
   795  	if want, got := "signal.NotifyContext(context.Background.WithCancel, [hangup])", fmt.Sprint(c); want != got {
   796  		t.Errorf("c.String() = %q, wanted %q", got, want)
   797  	}
   798  
   799  	stop()
   800  	<-c.Done()
   801  	if got := c.Err(); got != context.Canceled {
   802  		t.Errorf("c.Err() = %q, want %q", got, context.Canceled)
   803  	}
   804  }
   805  
   806  func TestNotifyContextCancelParent(t *testing.T) {
   807  	parent, cancelParent := context.WithCancel(context.Background())
   808  	defer cancelParent()
   809  	c, stop := NotifyContext(parent, syscall.SIGINT)
   810  	defer stop()
   811  
   812  	if want, got := "signal.NotifyContext(context.Background.WithCancel, [interrupt])", fmt.Sprint(c); want != got {
   813  		t.Errorf("c.String() = %q, want %q", got, want)
   814  	}
   815  
   816  	cancelParent()
   817  	<-c.Done()
   818  	if got := c.Err(); got != context.Canceled {
   819  		t.Errorf("c.Err() = %q, want %q", got, context.Canceled)
   820  	}
   821  }
   822  
   823  func TestNotifyContextPrematureCancelParent(t *testing.T) {
   824  	parent, cancelParent := context.WithCancel(context.Background())
   825  	defer cancelParent()
   826  
   827  	cancelParent() // Prematurely cancel context before calling NotifyContext.
   828  	c, stop := NotifyContext(parent, syscall.SIGINT)
   829  	defer stop()
   830  
   831  	if want, got := "signal.NotifyContext(context.Background.WithCancel, [interrupt])", fmt.Sprint(c); want != got {
   832  		t.Errorf("c.String() = %q, want %q", got, want)
   833  	}
   834  
   835  	<-c.Done()
   836  	if got := c.Err(); got != context.Canceled {
   837  		t.Errorf("c.Err() = %q, want %q", got, context.Canceled)
   838  	}
   839  }
   840  
   841  func TestNotifyContextSimultaneousStop(t *testing.T) {
   842  	c, stop := NotifyContext(context.Background(), syscall.SIGINT)
   843  	defer stop()
   844  
   845  	if want, got := "signal.NotifyContext(context.Background, [interrupt])", fmt.Sprint(c); want != got {
   846  		t.Errorf("c.String() = %q, want %q", got, want)
   847  	}
   848  
   849  	var wg sync.WaitGroup
   850  	n := 10
   851  	wg.Add(n)
   852  	for i := 0; i < n; i++ {
   853  		go func() {
   854  			stop()
   855  			wg.Done()
   856  		}()
   857  	}
   858  	wg.Wait()
   859  	<-c.Done()
   860  	if got := c.Err(); got != context.Canceled {
   861  		t.Errorf("c.Err() = %q, want %q", got, context.Canceled)
   862  	}
   863  }
   864  
   865  func TestNotifyContextStringer(t *testing.T) {
   866  	parent, cancelParent := context.WithCancel(context.Background())
   867  	defer cancelParent()
   868  	c, stop := NotifyContext(parent, syscall.SIGHUP, syscall.SIGINT, syscall.SIGTERM)
   869  	defer stop()
   870  
   871  	want := `signal.NotifyContext(context.Background.WithCancel, [hangup interrupt terminated])`
   872  	if got := fmt.Sprint(c); got != want {
   873  		t.Errorf("c.String() = %q, want %q", got, want)
   874  	}
   875  }
   876  
   877  // #44193 test signal handling while stopping and starting the world.
   878  func TestSignalTrace(t *testing.T) {
   879  	done := make(chan struct{})
   880  	quit := make(chan struct{})
   881  	c := make(chan os.Signal, 1)
   882  	Notify(c, syscall.SIGHUP)
   883  
   884  	// Source and sink for signals busy loop unsynchronized with
   885  	// trace starts and stops. We are ultimately validating that
   886  	// signals and runtime.(stop|start)TheWorldGC are compatible.
   887  	go func() {
   888  		defer close(done)
   889  		defer Stop(c)
   890  		pid := syscall.Getpid()
   891  		for {
   892  			select {
   893  			case <-quit:
   894  				return
   895  			default:
   896  				syscall.Kill(pid, syscall.SIGHUP)
   897  			}
   898  			waitSig(t, c, syscall.SIGHUP)
   899  		}
   900  	}()
   901  
   902  	for i := 0; i < 100; i++ {
   903  		buf := new(bytes.Buffer)
   904  		if err := trace.Start(buf); err != nil {
   905  			t.Fatalf("[%d] failed to start tracing: %v", i, err)
   906  		}
   907  		trace.Stop()
   908  		size := buf.Len()
   909  		if size == 0 {
   910  			t.Fatalf("[%d] trace is empty", i)
   911  		}
   912  	}
   913  	close(quit)
   914  	<-done
   915  }
   916  

View as plain text