os/exec: use pidfd for waiting and signaling of processes

Using pidfd allows us to have a handle on the process and
poll the handle to non-blocking wait for the process to
exit.

Fixes golang#34396
Fixes golang#60321
Fixes golang#60320

src/os/exec.go

@@ -85,7 +85,7 @@ func Getppid() int { return syscall.Getppid() }
 // The Process it returns can be used to obtain information
 // about the underlying operating system process.
 //
-// On Unix systems, FindProcess always succeeds and returns a Process
+// On non-Linux Unix systems, FindProcess always succeeds and returns a Process
 // for the given pid, regardless of whether the process exists.
 func FindProcess(pid int) (*Process, error) {
 	return findProcess(pid)

src/os/exec_posix.go

@@ -69,9 +69,9 @@ func (p *Process) kill() error {
 
 // ProcessState stores information about a process, as reported by Wait.
 type ProcessState struct {
-	pid    int                // The process's id.
-	status syscall.WaitStatus // System-dependent status info.
-	rusage *syscall.Rusage
+	pid     int     // The process's id.
+	siginfo Siginfo // System-dependent status info.
+	rusage  *syscall.Rusage
 }
 
 // Pid returns the process id of the exited process.
@@ -80,15 +80,15 @@ func (p *ProcessState) Pid() int {
 }
 
 func (p *ProcessState) exited() bool {
-	return p.status.Exited()
+	return p.siginfo.Exited()
 }
 
 func (p *ProcessState) success() bool {
-	return p.status.ExitStatus() == 0
+	return p.siginfo.ExitStatus() == 0
 }
 
 func (p *ProcessState) sys() any {
-	return p.status
+	return p.siginfo
 }
 
 func (p *ProcessState) sysUsage() any {
@@ -99,27 +99,27 @@ func (p *ProcessState) String() string {
 	if p == nil {
 		return "<nil>"
 	}
-	status := p.Sys().(syscall.WaitStatus)
+	siginfo := p.Sys().(Siginfo)
 	res := ""
 	switch {
-	case status.Exited():
-		code := status.ExitStatus()
+	case siginfo.Exited():
+		code := siginfo.ExitStatus()
 		if runtime.GOOS == "windows" && uint(code) >= 1<<16 { // windows uses large hex numbers
 			res = "exit status " + uitox(uint(code))
 		} else { // unix systems use small decimal integers
 			res = "exit status " + itoa.Itoa(code) // unix
 		}
-	case status.Signaled():
-		res = "signal: " + status.Signal().String()
-	case status.Stopped():
-		res = "stop signal: " + status.StopSignal().String()
-		if status.StopSignal() == syscall.SIGTRAP && status.TrapCause() != 0 {
-			res += " (trap " + itoa.Itoa(status.TrapCause()) + ")"
+	case siginfo.Signaled():
+		res = "signal: " + siginfo.Signal().String()
+	case siginfo.Stopped():
+		res = "stop signal: " + siginfo.StopSignal().String()
+		if siginfo.Trapped() && siginfo.TrapCause() != 0 {
+			res += " (trap " + itoa.Itoa(siginfo.TrapCause()) + ")"
 		}
-	case status.Continued():
+	case siginfo.Continued():
 		res = "continued"
 	}
-	if status.CoreDump() {
+	if siginfo.CoreDump() {
 		res += " (core dumped)"
 	}
 	return res
@@ -132,5 +132,5 @@ func (p *ProcessState) ExitCode() int {
 	if p == nil {
 		return -1
 	}
-	return p.status.ExitStatus()
+	return p.siginfo.ExitStatus()
 }

src/os/exec_unix.go

@@ -11,8 +11,64 @@ import (
 	"runtime"
 	"syscall"
 	"time"
+	"unsafe"
 )
 
+const CLD_EXITED = 1
+const CLD_KILLED = 2
+const CLD_DUMPED = 3
+const CLD_TRAPPED = 4
+const CLD_STOPPED = 5
+const CLD_CONTINUED = 6
+
+type Siginfo struct {
+	Signo int32
+	Errno int32
+	Code  int32
+	_     int32
+	_     [112]byte
+}
+
+func (s Siginfo) Exited() bool { return s.Code == CLD_EXITED }
+
+func (s Siginfo) Signaled() bool { return s.Code == CLD_KILLED }
+
+func (s Siginfo) Stopped() bool { return s.Code == CLD_STOPPED }
+
+func (s Siginfo) Trapped() bool { return s.Code == CLD_TRAPPED }
+
+func (s Siginfo) Continued() bool { return s.Code == CLD_CONTINUED }
+
+func (s Siginfo) CoreDump() bool { return s.Code == CLD_DUMPED }
+
+func (s Siginfo) ExitStatus() int {
+	if !s.Exited() {
+		return -1
+	}
+	return int(s.Errno)
+}
+
+func (s Siginfo) Signal() syscall.Signal {
+	if !s.Signaled() {
+		return -1
+	}
+	return syscall.Signal(s.Errno)
+}
+
+func (s Siginfo) StopSignal() syscall.Signal {
+	if !s.Signaled() {
+		return -1
+	}
+	return syscall.Signal(s.Errno)
+}
+
+func (s Siginfo) TrapCause() int {
+	if !s.Trapped() {
+		return -1
+	}
+	return int(s.Errno)
+}
+
 func (p *Process) wait() (ps *ProcessState, err error) {
 	if p.Pid == -1 {
 		return nil, syscall.EINVAL
@@ -34,27 +90,37 @@ func (p *Process) wait() (ps *ProcessState, err error) {
 	}
 
 	var (
-		status syscall.WaitStatus
-		rusage syscall.Rusage
-		pid1   int
-		e      error
+		siginfo Siginfo
+		rusage  syscall.Rusage
+		e       syscall.Errno
 	)
 	for {
-		pid1, e = syscall.Wait4(p.Pid, &status, 0, &rusage)
-		if e != syscall.EINTR {
+		_, _, e = syscall.Syscall6(syscall.SYS_WAITID, _P_PIDFD, p.handle, uintptr(unsafe.Pointer(&siginfo)), syscall.WEXITED, uintptr(unsafe.Pointer(&rusage)), 0)
+		if e == syscall.EINTR {
+			continue
+		} else if e == syscall.ENOSYS {
+			// waitid has been available since Linux 2.6.9, but
+			// reportedly is not available in Ubuntu on Windows.
+			// See issue 16610.
+			panic("TODO: Implement fallback")
+		} else if e != 0 {
+			break
+		}
+		// During ptrace the wait might return also for non-exit reasons. In that case we retry.
+		// See: https://lwn.net/Articles/688624/
+		if siginfo.Exited() || siginfo.Signaled() || siginfo.CoreDump() {
 			break
 		}
 	}
-	if e != nil {
-		return nil, NewSyscallError("wait", e)
-	}
-	if pid1 != 0 {
-		p.setDone()
+	runtime.KeepAlive(p)
+	if e != 0 {
+		return nil, NewSyscallError("waitid", e)
 	}
+	p.setDone()
 	ps = &ProcessState{
-		pid:    pid1,
-		status: status,
-		rusage: &rusage,
+		pid:     p.Pid,
+		siginfo: siginfo,
+		rusage:  &rusage,
 	}
 	return ps, nil
 }
@@ -75,26 +141,35 @@ func (p *Process) signal(sig Signal) error {
 	if !ok {
 		return errors.New("os: unsupported signal type")
 	}
-	if e := syscall.Kill(p.Pid, s); e != nil {
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_PIDFD_SEND_SIGNAL, p.handle, uintptr(s), 0, 0, 0, 0); e != 0 {
 		if e == syscall.ESRCH {
 			return ErrProcessDone
 		}
-		return e
+		return NewSyscallError("pidfd_send_signal", e)
 	}
+	runtime.KeepAlive(p)
 	return nil
 }
 
 func (p *Process) release() error {
-	// NOOP for unix.
+	e := syscall.Close(int(p.handle))
+	if e != nil {
+		return NewSyscallError("close", e)
+
+	}
 	p.Pid = -1
 	// no need for a finalizer anymore
 	runtime.SetFinalizer(p, nil)
 	return nil
 }
 
 func findProcess(pid int) (p *Process, err error) {
-	// NOOP for unix.
-	return newProcess(pid, 0), nil
+	fd, _, e := syscall.Syscall(syscall.SYS_PIDFD_OPEN, uintptr(pid), 0, 0)
+	runtime.KeepAlive(p)
+	if e != 0 {
+		return nil, NewSyscallError("pidfd_open", e)
+	}
+	return newProcess(pid, fd), nil
 }
 
 func (p *ProcessState) userTime() time.Duration {

src/os/wait_waitid.go

@@ -10,39 +10,32 @@
 package os
 
 import (
-	"runtime"
+	"internal/poll"
 	"syscall"
-	"unsafe"
 )
 
-const _P_PID = 1
+const _P_PIDFD = 3
 
 // blockUntilWaitable attempts to block until a call to p.Wait will
 // succeed immediately, and reports whether it has done so.
 // It does not actually call p.Wait.
 func (p *Process) blockUntilWaitable() (bool, error) {
-	// The waitid system call expects a pointer to a siginfo_t,
-	// which is 128 bytes on all Linux systems.
-	// On darwin/amd64, it requires 104 bytes.
-	// We don't care about the values it returns.
-	var siginfo [16]uint64
-	psig := &siginfo[0]
-	var e syscall.Errno
-	for {
-		_, _, e = syscall.Syscall6(syscall.SYS_WAITID, _P_PID, uintptr(p.Pid), uintptr(unsafe.Pointer(psig)), syscall.WEXITED|syscall.WNOWAIT, 0, 0)
-		if e != syscall.EINTR {
-			break
-		}
+	fd := poll.FD{
+		Sysfd: int(p.handle),
 	}
-	runtime.KeepAlive(p)
-	if e != 0 {
-		// waitid has been available since Linux 2.6.9, but
-		// reportedly is not available in Ubuntu on Windows.
-		// See issue 16610.
-		if e == syscall.ENOSYS {
-			return false, nil
-		}
-		return false, NewSyscallError("waitid", e)
+	err := fd.Init("pidfd", false)
+	if err != nil {
+		return false, err
+	}
+	// We just want to make sure fd is ready for reading, but that is not yet available.
+	// See: https://github.com/golang/go/issues/15735
+	buf := make([]byte, 1)
+	_, err = fd.Read(buf)
+	if err == syscall.EINVAL {
+		// fd is ready for reading, but reading failed, as expected on pidfd.
+		return true, nil
+	} else if err != nil {
+		return false, err
 	}
 	return true, nil
 }

src/syscall/exec_unix.go

@@ -16,6 +16,9 @@ import (
 	"unsafe"
 )
 
+const SYS_PIDFD_OPEN = 434
+const SYS_PIDFD_SEND_SIGNAL = 424
+
 // ForkLock is used to synchronize creation of new file descriptors
 // with fork.
 //
@@ -332,7 +335,15 @@ func ForkExec(argv0 string, argv []string, attr *ProcAttr) (pid int, err error)
 // StartProcess wraps ForkExec for package os.
 func StartProcess(argv0 string, argv []string, attr *ProcAttr) (pid int, handle uintptr, err error) {
 	pid, err = forkExec(argv0, argv, attr)
-	return pid, 0, err
+	if err != nil {
+		return
+	}
+	fd, _, e := Syscall(SYS_PIDFD_OPEN, uintptr(pid), 0, 0)
+	if e != 0 {
+		err = errnoErr(e)
+		return
+	}
+	return pid, fd, err
 }
 
 // Implemented in runtime package.