Use O_CLOEXEC when dup'ing FDs

The sentry doesn't allow execve, but it's a good defense
in-depth measure.

PiperOrigin-RevId: 305958737

pkg/sentry/fs/gofer/inode.go

@@ -273,7 +273,7 @@ func (i *inodeFileState) recreateReadHandles(ctx context.Context, writer *handle
 	// operations on the old will see the new data. Then, make the new handle take
 	// ownereship of the old FD and mark the old readHandle to not close the FD
 	// when done.
-	if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), 0); err != nil {
+	if err := syscall.Dup3(h.Host.FD(), i.readHandles.Host.FD(), syscall.O_CLOEXEC); err != nil {
 		return err
 	}
 

pkg/sentry/fsimpl/gofer/gofer.go

@@ -1089,7 +1089,7 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool
 				// description, but this doesn't matter since they refer to the
 				// same file (unless d.fs.opts.overlayfsStaleRead is true,
 				// which we handle separately).
-				if err := syscall.Dup3(int(h.fd), int(d.handle.fd), 0); err != nil {
+				if err := syscall.Dup3(int(h.fd), int(d.handle.fd), syscall.O_CLOEXEC); err != nil {
 					d.handleMu.Unlock()
 					ctx.Warningf("gofer.dentry.ensureSharedHandle: failed to dup fd %d to fd %d: %v", h.fd, d.handle.fd, err)
 					h.close(ctx)

runsc/boot/filter/config.go

@@ -44,7 +44,7 @@ var allowedSyscalls = seccomp.SyscallRules{
 		{
 			seccomp.AllowAny{},
 			seccomp.AllowAny{},
-			seccomp.AllowValue(0),
+			seccomp.AllowValue(syscall.O_CLOEXEC),
 		},
 	},
 	syscall.SYS_EPOLL_CREATE1: {},

runsc/main.go

@@ -291,7 +291,7 @@ func main() {
 		// want with them. Since Docker and Containerd both eat boot's stderr, we
 		// dup our stderr to the provided log FD so that panics will appear in the
 		// logs, rather than just disappear.
-		if err := syscall.Dup3(fd, int(os.Stderr.Fd()), 0); err != nil {
+		if err := syscall.Dup3(fd, int(os.Stderr.Fd()), syscall.O_CLOEXEC); err != nil {
 			cmd.Fatalf("error dup'ing fd %d to stderr: %v", fd, err)
 		}
 	}