exec_pty.c   [plain text]

/*
 * Copyright (c) 2009-2010 Todd C. Miller <Todd.Miller@courtesan.com>
 *
 * Permission to use, copy, modify, and distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 *
 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 */

#include <config.h>

#include <sys/types.h>
#include <sys/param.h>
#include <sys/socket.h>
#include <sys/time.h>
#include <sys/wait.h>
#ifdef HAVE_TERMIOS_H
# include <termios.h>
#else
# include <termio.h>
#endif /* HAVE_TERMIOS_H */
#include <sys/ioctl.h>
#ifdef HAVE_SYS_SELECT_H
# include <sys/select.h>
#endif /* HAVE_SYS_SELECT_H */
#include <stdio.h>
#ifdef STDC_HEADERS
# include <stdlib.h>
# include <stddef.h>
#else
# ifdef HAVE_STDLIB_H
#  include <stdlib.h>
# endif
#endif /* STDC_HEADERS */
#ifdef HAVE_STRING_H
# if defined(HAVE_MEMORY_H) && !defined(STDC_HEADERS)
#  include <memory.h>
# endif
# include <string.h>
#endif /* HAVE_STRING_H */
#ifdef HAVE_STRINGS_H
# include <strings.h>
#endif /* HAVE_STRINGS_H */
#ifdef HAVE_UNISTD_H
# include <unistd.h>
#endif /* HAVE_UNISTD_H */
#if TIME_WITH_SYS_TIME
# include <time.h>
#endif
#include <errno.h>
#include <fcntl.h>
#include <signal.h>

#include "sudo.h"
#include "sudo_exec.h"

#define SFD_STDIN	0
#define SFD_STDOUT	1
#define SFD_STDERR	2
#define SFD_MASTER	3
#define SFD_SLAVE	4
#define SFD_USERTTY	5

#define TERM_COOKED	0
#define TERM_RAW	1

/* Compatibility with older tty systems. */
#if !defined(TIOCGSIZE) && defined(TIOCGWINSZ)
# define TIOCGSIZE	TIOCGWINSZ
# define TIOCSSIZE	TIOCSWINSZ
# define ttysize	winsize
# define ts_cols	ws_col
#endif

struct io_buffer {
    struct io_buffer *next;
    int len; /* buffer length (how much produced) */
    int off; /* write position (how much already consumed) */
    int rfd;  /* reader (producer) */
    int wfd; /* writer (consumer) */
    int (*action) __P((const char *buf, unsigned int len));
    char buf[16 * 1024];
};

static char slavename[PATH_MAX];
static int foreground;
static int io_fds[6] = { -1, -1, -1, -1, -1, -1};
static int pipeline = FALSE;
static int tty_initialized;
static int ttymode = TERM_COOKED;
static pid_t ppgrp, child;
static struct io_buffer *iobufs;

static void flush_output __P((void));
static int exec_monitor __P((const char *path, char *argv[],
    char *envp[], int, int));
static void exec_pty __P((const char *path, char *argv[],
    char *envp[], int));
static void sigwinch __P((int s));
static void sync_ttysize __P((int src, int dst));
static void deliver_signal __P((pid_t pid, int signo));
static int safe_close __P((int fd));

/*
 * Allocate a pty if /dev/tty is a tty.
 * Fills in io_fds[SFD_USERTTY], io_fds[SFD_MASTER], io_fds[SFD_SLAVE]
 * and slavename globals.
 */
void
pty_setup(uid)
    uid_t uid;
{
    io_fds[SFD_USERTTY] = open(_PATH_TTY, O_RDWR|O_NOCTTY, 0);
    if (io_fds[SFD_USERTTY] != -1) {
	if (!get_pty(&io_fds[SFD_MASTER], &io_fds[SFD_SLAVE],
	    slavename, sizeof(slavename), uid))
	    error(1, "Can't get pty");
    }
}

/*
 * Check whether we are running in the foregroup.
 * Updates the foreground global and does lazy init of the
 * the pty slave as needed.
 */
static void
check_foreground()
{
    if (io_fds[SFD_USERTTY] != -1) {
	foreground = tcgetpgrp(io_fds[SFD_USERTTY]) == ppgrp;
	if (foreground && !tty_initialized) {
	    if (term_copy(io_fds[SFD_USERTTY], io_fds[SFD_SLAVE])) {
		tty_initialized = 1;
		sync_ttysize(io_fds[SFD_USERTTY], io_fds[SFD_SLAVE]);
	    }
	}
    }
}

/*
 * Suspend sudo if the underlying command is suspended.
 * Returns SIGUSR1 if the child should be resume in foreground else SIGUSR2.
 */
int
suspend_parent(signo)
    int signo;
{
    sigaction_t sa, osa;
    int n, oldmode = ttymode, rval = 0;

    switch (signo) {
    case SIGTTOU:
    case SIGTTIN:
	/*
	 * If we are the foreground process, just resume the child.
	 * Otherwise, re-send the signal with the handler disabled.
	 */
	if (!foreground)
	    check_foreground();
	if (foreground) {
	    if (ttymode != TERM_RAW) {
		do {
		    n = term_raw(io_fds[SFD_USERTTY], 0);
		} while (!n && errno == EINTR);
		ttymode = TERM_RAW;
	    }
	    rval = SIGUSR1; /* resume child in foreground */
	    break;
	}
	ttymode = TERM_RAW;
	/* FALLTHROUGH */
    case SIGSTOP:
    case SIGTSTP:
	/* Flush any remaining output before suspending. */
	flush_output();

	/* Restore original tty mode before suspending. */
	if (oldmode != TERM_COOKED) {
	    do {
		n = term_restore(io_fds[SFD_USERTTY], 0);
	    } while (!n && errno == EINTR);
	}

	/* Suspend self and continue child when we resume. */
	sa.sa_handler = SIG_DFL;
	sigaction(signo, &sa, &osa);
	if (killpg(ppgrp, signo) != 0)
	    warning("killpg(%d, %d)", ppgrp, signo);

	/* Check foreground/background status on resume. */
	check_foreground();

	/*
	 * Only modify term if we are foreground process and either
	 * the old tty mode was not cooked or child got SIGTT{IN,OU}
	 */
	if (ttymode != TERM_COOKED) {
	    if (foreground) {
		/* Set raw mode. */
		do {
		    n = term_raw(io_fds[SFD_USERTTY], 0);
		} while (!n && errno == EINTR);
	    } else {
		/* Background process, no access to tty. */
		ttymode = TERM_COOKED;
	    }
	}

	sigaction(signo, &osa, NULL);
	rval = ttymode == TERM_RAW ? SIGUSR1 : SIGUSR2;
	break;
    }

    return(rval);
}

/*
 * Kill child with increasing urgency.
 */
static void
terminate_child(pid, use_pgrp)
    pid_t pid;
    int use_pgrp;
{
    /*
     * Note that SIGCHLD will interrupt the sleep()
     */
    if (use_pgrp) {
	killpg(pid, SIGHUP);
	killpg(pid, SIGTERM);
	sleep(2);
	killpg(pid, SIGKILL);
    } else {
	kill(pid, SIGHUP);
	kill(pid, SIGTERM);
	sleep(2);
	kill(pid, SIGKILL);
    }
}

/*
 * Allocate a new io_buffer struct and insert it at the head of the list.
 * Returns the new head element.
 */
static struct io_buffer *
io_buf_new(rfd, wfd, action, head)
    int rfd;
    int wfd;
    int (*action) __P((const char *, unsigned int));
    struct io_buffer *head;
{
    struct io_buffer *iob;

    iob = emalloc(sizeof(*iob));
    zero_bytes(iob, sizeof(*iob));
    iob->rfd = rfd;
    iob->wfd = wfd;
    iob->action = action;
    iob->next = head;
    return iob;
}

/*
 * Read/write iobufs depending on fdsr and fdsw.
 * Fills in cstat on error.
 * Returns the number of errors.
 */
int
perform_io(fdsr, fdsw, cstat)
    fd_set *fdsr;
    fd_set *fdsw;
    struct command_status *cstat;
{
    struct io_buffer *iob;
    int n, errors = 0;

    for (iob = iobufs; iob; iob = iob->next) {
	if (iob->rfd != -1 && FD_ISSET(iob->rfd, fdsr)) {
	    do {
		n = read(iob->rfd, iob->buf + iob->len,
		    sizeof(iob->buf) - iob->len);
	    } while (n == -1 && errno == EINTR);
	    switch (n) {
		case -1:
		    if (errno == EAGAIN)
			break;
		    if (errno != ENXIO && errno != EBADF) {
			errors++;
			break;
		    }
		    /* FALLTHROUGH */
		case 0:
		    /* got EOF or pty has gone away */
		    safe_close(iob->rfd);
		    iob->rfd = -1;
		    break;
		default:
		    if (!iob->action(iob->buf + iob->len, n))
			terminate_child(child, TRUE);
		    iob->len += n;
		    break;
	    }
	}
	if (iob->wfd != -1 && FD_ISSET(iob->wfd, fdsw)) {
	    do {
		n = write(iob->wfd, iob->buf + iob->off,
		    iob->len - iob->off);
	    } while (n == -1 && errno == EINTR);
	    if (n == -1) {
		if (errno == EPIPE || errno == ENXIO || errno == EBADF) {
		    /* other end of pipe closed or pty revoked */
		    if (iob->rfd != -1) {
			safe_close(iob->rfd);
			iob->rfd = -1;
		    }
		    safe_close(iob->wfd);
		    iob->wfd = -1;
		    continue;
		}
		if (errno != EAGAIN)
		    errors++;
	    } else {
		iob->off += n;
	    }
	}
    }
    if (errors && cstat != NULL) {
	cstat->type = CMD_ERRNO;
	cstat->val = errno;
    }
    return errors;
}

/*
 * Fork a monitor process which runs the actual command as its own child
 * process with std{in,out,err} hooked up to the pty or pipes as appropriate.
 * Returns the child pid.
 */
int
fork_pty(path, argv, envp, sv, rbac_enabled, maxfd)
    const char *path;
    char *argv[];
    char *envp[];
    int sv[2];
    int rbac_enabled;
    int *maxfd;
{
    struct command_status cstat;
    struct io_buffer *iob;
    int io_pipe[3][2], n;
    sigaction_t sa;

    ppgrp = getpgrp(); /* parent's pgrp, so child can signal us */

    zero_bytes(&sa, sizeof(sa));
    sigemptyset(&sa.sa_mask);

    if (io_fds[SFD_USERTTY] != -1) {
	sa.sa_flags = SA_RESTART;
	sa.sa_handler = sigwinch;
	sigaction(SIGWINCH, &sa, NULL);
    }

    /*
     * Setup stdin/stdout/stderr for child, to be duped after forking.
     */
    io_fds[SFD_STDIN] = io_fds[SFD_SLAVE];
    io_fds[SFD_STDOUT] = io_fds[SFD_SLAVE];
    io_fds[SFD_STDERR] = io_fds[SFD_SLAVE];

    /* Copy /dev/tty -> pty master */
    if (io_fds[SFD_USERTTY] != -1) {
	iobufs = io_buf_new(io_fds[SFD_USERTTY], io_fds[SFD_MASTER],
	    log_ttyin, iobufs);

	/* Copy pty master -> /dev/tty */
	iobufs = io_buf_new(io_fds[SFD_MASTER], io_fds[SFD_USERTTY],
	    log_ttyout, iobufs);

	/* Are we the foreground process? */
	foreground = tcgetpgrp(io_fds[SFD_USERTTY]) == ppgrp;
    }

    /*
     * If either stdin, stdout or stderr is not a tty we use a pipe
     * to interpose ourselves instead of duping the pty fd.
     */
    memset(io_pipe, 0, sizeof(io_pipe));
    if (io_fds[SFD_STDIN] == -1 || !isatty(STDIN_FILENO)) {
	pipeline = TRUE;
	if (pipe(io_pipe[STDIN_FILENO]) != 0)
	    error(1, "unable to create pipe");
	iobufs = io_buf_new(STDIN_FILENO, io_pipe[STDIN_FILENO][1],
	    log_stdin, iobufs);
	io_fds[SFD_STDIN] = io_pipe[STDIN_FILENO][0];
    }
    if (io_fds[SFD_STDOUT] == -1 || !isatty(STDOUT_FILENO)) {
	pipeline = TRUE;
	if (pipe(io_pipe[STDOUT_FILENO]) != 0)
	    error(1, "unable to create pipe");
	iobufs = io_buf_new(io_pipe[STDOUT_FILENO][0], STDOUT_FILENO,
	    log_stdout, iobufs);
	io_fds[SFD_STDOUT] = io_pipe[STDOUT_FILENO][1];
    }
    if (io_fds[SFD_STDERR] == -1 || !isatty(STDERR_FILENO)) {
	if (pipe(io_pipe[STDERR_FILENO]) != 0)
	    error(1, "unable to create pipe");
	iobufs = io_buf_new(io_pipe[STDERR_FILENO][0], STDERR_FILENO,
	    log_stderr, iobufs);
	io_fds[SFD_STDERR] = io_pipe[STDERR_FILENO][1];
    }

    /* Job control signals to relay from parent to child. */
    sa.sa_flags = SA_INTERRUPT; /* do not restart syscalls */
    sa.sa_handler = handler;
    sigaction(SIGTSTP, &sa, NULL);

    if (foreground) {
	/* Copy terminal attrs from user tty -> pty slave. */
	if (term_copy(io_fds[SFD_USERTTY], io_fds[SFD_SLAVE])) {
	    tty_initialized = 1;
	    sync_ttysize(io_fds[SFD_USERTTY], io_fds[SFD_SLAVE]);
	}

	/* Start out in raw mode if we are not part of a pipeline. */
	if (!pipeline) {
	    ttymode = TERM_RAW;
	    do {
		n = term_raw(io_fds[SFD_USERTTY], 0);
	    } while (!n && errno == EINTR);
	    if (!n)
		error(1, "Can't set terminal to raw mode");
	}
    }

    child = fork();
    switch (child) {
    case -1:
	error(1, "fork");
	break;
    case 0:
	/* child */
	close(sv[0]);
	fcntl(sv[1], F_SETFD, FD_CLOEXEC);
	if (exec_setup(rbac_enabled, slavename, io_fds[SFD_SLAVE]) == TRUE) {
	    /* Close the other end of the stdin/stdout/stderr pipes and exec. */
	    if (io_pipe[STDIN_FILENO][1])
		close(io_pipe[STDIN_FILENO][1]);
	    if (io_pipe[STDOUT_FILENO][0])
		close(io_pipe[STDOUT_FILENO][0]);
	    if (io_pipe[STDERR_FILENO][0])
		close(io_pipe[STDERR_FILENO][0]);
	    exec_monitor(path, argv, envp, sv[1], rbac_enabled);
	}
	cstat.type = CMD_ERRNO;
	cstat.val = errno;
	send(sv[1], &cstat, sizeof(cstat), 0);
	_exit(1);
    }

    /* Close the other end of the stdin/stdout/stderr pipes. */
    if (io_pipe[STDIN_FILENO][0])
	close(io_pipe[STDIN_FILENO][0]);
    if (io_pipe[STDOUT_FILENO][1])
	close(io_pipe[STDOUT_FILENO][1]);
    if (io_pipe[STDERR_FILENO][1])
	close(io_pipe[STDERR_FILENO][1]);

    for (iob = iobufs; iob; iob = iob->next) {
	/* Adjust maxfd. */
	if (iob->rfd > *maxfd)
	    *maxfd = iob->rfd;
	if (iob->wfd > *maxfd)
	    *maxfd = iob->wfd;

	/* Set non-blocking mode. */
	n = fcntl(iob->rfd, F_GETFL, 0);
	if (n != -1 && !ISSET(n, O_NONBLOCK))
	    (void) fcntl(iob->rfd, F_SETFL, n | O_NONBLOCK);
	n = fcntl(iob->wfd, F_GETFL, 0);
	if (n != -1 && !ISSET(n, O_NONBLOCK))
	    (void) fcntl(iob->wfd, F_SETFL, n | O_NONBLOCK);
    }

    return child;
}

/*
 * Flush any remaining output and restore /dev/tty to the way we found it.
 * If the command died due to a signal, writes the reason to stdout.
 */
void
pty_close(cstat)
    struct command_status *cstat;
{
    int n;

    /* Flush any remaining output (the plugin already got it) */
    if (io_fds[SFD_USERTTY] != -1) {
	n = fcntl(io_fds[SFD_USERTTY], F_GETFL, 0);
	if (n != -1 && ISSET(n, O_NONBLOCK)) {
	    CLR(n, O_NONBLOCK);
	    (void) fcntl(io_fds[SFD_USERTTY], F_SETFL, n);
	}
    }
    flush_output();

    if (io_fds[SFD_USERTTY] != -1) {
	do {
	    n = term_restore(io_fds[SFD_USERTTY], 0);
	} while (!n && errno == EINTR);
    }

    /* If child was signalled, write the reason to stdout like the shell. */
    if (cstat->type == CMD_WSTATUS && WIFSIGNALED(cstat->val)) {
	int signo = WTERMSIG(cstat->val);
	if (signo && signo != SIGINT && signo != SIGPIPE) {
	    const char *reason = strsignal(signo);
	    n = io_fds[SFD_USERTTY] != -1 ?
		io_fds[SFD_USERTTY] : STDOUT_FILENO;
	    write(n, reason, strlen(reason));
	    if (WCOREDUMP(cstat->val))
		write(n, " (core dumped)", 14);
	    write(n, "\n", 1);
	}
    }
}


/*
 * Fill in fdsr and fdsw based on the io buffers list.
 * Called prior to select().
 */
void
fd_set_iobs(fdsr, fdsw)
    fd_set *fdsr;
    fd_set *fdsw;
{
    struct io_buffer *iob;

    for (iob = iobufs; iob; iob = iob->next) {
	if (iob->rfd == -1 && iob->wfd == -1)
	    continue;
	if (iob->off == iob->len) {
	    iob->off = iob->len = 0;
	    /* Forward the EOF from reader to writer. */
	    if (iob->rfd == -1) {
		safe_close(iob->wfd);
		iob->wfd = -1;
	    }
	}
	/* Don't read/write /dev/tty if we are not in the foreground. */
	if (iob->rfd != -1 &&
	    (ttymode == TERM_RAW || iob->rfd != io_fds[SFD_USERTTY])) {
	    if (iob->len != sizeof(iob->buf))
		FD_SET(iob->rfd, fdsr);
	}
	if (iob->wfd != -1 &&
	    (foreground || iob->wfd != io_fds[SFD_USERTTY])) {
	    if (iob->len > iob->off)
		FD_SET(iob->wfd, fdsw);
	}
    }
}

/*
 * Deliver a relayed signal to the command.
 */
static void
deliver_signal(pid, signo)
    pid_t pid;
    int signo;
{
    int status;

    /* Handle signal from parent. */
    switch (signo) {
    case SIGKILL:
	_exit(1); /* XXX */
	/* NOTREACHED */
    case SIGPIPE:
    case SIGHUP:
    case SIGTERM:
    case SIGINT:
    case SIGQUIT:
    case SIGTSTP:
	/* relay signal to child */
	killpg(pid, signo);
	break;
    case SIGALRM:
	terminate_child(pid, TRUE);
	break;
    case SIGUSR1:
	/* foreground process, grant it controlling tty. */
	do {
	    status = tcsetpgrp(io_fds[SFD_SLAVE], pid);
	} while (status == -1 && errno == EINTR);
	killpg(pid, SIGCONT);
	break;
    case SIGUSR2:
	/* background process, I take controlling tty. */
	do {
	    status = tcsetpgrp(io_fds[SFD_SLAVE], getpid());
	} while (status == -1 && errno == EINTR);
	killpg(pid, SIGCONT);
	break;
    default:
	warningx("unexpected signal from child: %d", signo);
	break;
    }
}

/*
 * Send status to parent over socketpair.
 * Return value is the same as send(2).
 */
static int
send_status(fd, cstat)
    int fd;
    struct command_status *cstat;
{
    int n = -1;

    if (cstat->type != CMD_INVALID) {
	do {
	    n = send(fd, cstat, sizeof(*cstat), 0);
	} while (n == -1 && errno == EINTR);
	cstat->type = CMD_INVALID; /* prevent re-sending */
    }
    return n;
}

/*
 * Wait for child status after receiving SIGCHLD.
 * If the child was stopped, the status is send back to the parent.
 * Otherwise, cstat is filled in but not sent.
 * Returns TRUE if child is still alive, else FALSE.
 */
static int
handle_sigchld(backchannel, cstat)
    int backchannel;
    struct command_status *cstat;
{
    int status, alive = TRUE;
    pid_t pid;

    /* read child status */
    do {
#ifdef sudo_waitpid
	pid = sudo_waitpid(child, &status, WUNTRACED|WNOHANG);
#else
	pid = wait(&status);
#endif
    } while (pid == -1 && errno == EINTR);
    if (pid == child) {
	if (cstat->type != CMD_ERRNO) {
	    cstat->type = CMD_WSTATUS;
	    cstat->val = status;
	    if (WIFSTOPPED(status)) {
		if (send_status(backchannel, cstat) == -1)
		    return alive; /* XXX */
	    }
	}
	if (!WIFSTOPPED(status))
	    alive = FALSE;
    }
    return alive;
}

/*
 * Monitor process that creates a new session with the controlling tty,
 * resets signal handlers and forks a child to call exec_pty().
 * Waits for status changes from the command and relays them to the
 * parent and relays signals from the parent to the command.
 * Returns an error if fork(2) fails, else calls _exit(2).
 */
static int
exec_monitor(path, argv, envp, backchannel, rbac)
    const char *path;
    char *argv[];
    char *envp[];
    int backchannel;
    int rbac;
{
    struct command_status cstat;
    struct timeval tv;
    fd_set *fdsr;
    sigaction_t sa;
    int errpipe[2], maxfd, n, status;
    int alive = TRUE;

    /* Close unused fds. */
    if (io_fds[SFD_MASTER] != -1)
	close(io_fds[SFD_MASTER]);
    if (io_fds[SFD_USERTTY] != -1)
	close(io_fds[SFD_USERTTY]);

    /* Reset SIGWINCH and SIGALRM. */
    zero_bytes(&sa, sizeof(sa));
    sigemptyset(&sa.sa_mask);
    sa.sa_flags = SA_RESTART;
    sa.sa_handler = SIG_DFL;
    sigaction(SIGWINCH, &sa, NULL);
    sigaction(SIGALRM, &sa, NULL);

    /* Ignore any SIGTTIN or SIGTTOU we get. */
    sa.sa_handler = SIG_IGN;
    sigaction(SIGTTIN, &sa, NULL);
    sigaction(SIGTTOU, &sa, NULL);

    /* Note: HP-UX select() will not be interrupted if SA_RESTART set */
    sa.sa_flags = SA_INTERRUPT;
    sa.sa_handler = handler;
    sigaction(SIGCHLD, &sa, NULL);

    /*
     * Start a new session with the parent as the session leader
     * and the slave pty as the controlling terminal.
     * This allows us to be notified when the child has been suspended.
     */
    if (setsid() == -1) {
	warning("setsid");
	goto bad;
    }
    if (io_fds[SFD_SLAVE] != -1) {
#ifdef TIOCSCTTY
	if (ioctl(io_fds[SFD_SLAVE], TIOCSCTTY, NULL) != 0)
	    error(1, "unable to set controlling tty");
#else
	/* Set controlling tty by reopening slave. */
	if ((n = open(slavename, O_RDWR)) >= 0)
	    close(n);
#endif
    }

    /*
     * If stdin/stdout is not a tty, start command in the background
     * since it might be part of a pipeline that reads from /dev/tty.
     * In this case, we rely on the command receiving SIGTTOU or SIGTTIN
     * when it needs access to the controlling tty.
     */
    if (pipeline)
	foreground = 0;

    /* Start command and wait for it to stop or exit */
    if (pipe(errpipe) == -1)
	error(1, "unable to create pipe");
    child = fork();
    if (child == -1) {
	warning("Can't fork");
	goto bad;
    }
    if (child == 0) {
	/* We pass errno back to our parent via pipe on exec failure. */
	close(backchannel);
	close(errpipe[0]);
	fcntl(errpipe[1], F_SETFD, FD_CLOEXEC);

	/* setup tty and exec command */
	exec_pty(path, argv, envp, rbac);
	cstat.type = CMD_ERRNO;
	cstat.val = errno;
	write(errpipe[1], &cstat, sizeof(cstat));
	_exit(1);
    }
    close(errpipe[1]);

    /* If any of stdin/stdout/stderr are pipes, close them in parent. */
    if (io_fds[SFD_STDIN] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDIN]);
    if (io_fds[SFD_STDOUT] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDOUT]);
    if (io_fds[SFD_STDERR] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDERR]);

    /*
     * Put child in its own process group.  If we are starting the command
     * in the foreground, assign its pgrp to the tty.
     */
    setpgid(child, child);
    if (foreground) {
	do {
	    status = tcsetpgrp(io_fds[SFD_SLAVE], child);
	} while (status == -1 && errno == EINTR);
    }

    /* Wait for errno on pipe, signal on backchannel or for SIGCHLD */
    maxfd = MAX(errpipe[0], backchannel);
    fdsr = (fd_set *)emalloc2(howmany(maxfd + 1, NFDBITS), sizeof(fd_mask));
    zero_bytes(fdsr, howmany(maxfd + 1, NFDBITS) * sizeof(fd_mask));
    zero_bytes(&cstat, sizeof(cstat));
    tv.tv_sec = 0;
    tv.tv_usec = 0;
    for (;;) {
	/* Read child status. */
	if (recvsig[SIGCHLD]) {
	    recvsig[SIGCHLD] = FALSE;
	    alive = handle_sigchld(backchannel, &cstat);
	}

	/* Check for signal on backchannel or errno on errpipe. */
	FD_SET(backchannel, fdsr);
	if (errpipe[0] != -1)
	    FD_SET(errpipe[0], fdsr);
	maxfd = MAX(errpipe[0], backchannel);

	if (recvsig[SIGCHLD])
	    continue;
	/* If command exited we just poll, there may be data on errpipe. */
	n = select(maxfd + 1, fdsr, NULL, NULL, alive ? NULL : &tv);
	if (n <= 0) {
	    if (n == 0)
		goto done;
	    if (errno == EINTR)
		continue;
	    error(1, "select failed");
	}

	if (errpipe[0] != -1 && FD_ISSET(errpipe[0], fdsr)) {
	    /* read errno or EOF from command pipe */
	    n = read(errpipe[0], &cstat, sizeof(cstat));
	    if (n == -1) {
		if (errno == EINTR)
		    continue;
		warning("error reading from pipe");
		goto done;
	    }
	    /* Got errno or EOF, either way we are done with errpipe. */
	    FD_CLR(errpipe[0], fdsr);
	    close(errpipe[0]);
	    errpipe[0] = -1;
	}
	if (FD_ISSET(backchannel, fdsr)) {
	    struct command_status cstmp;

	    /* read command from backchannel, should be a signal */
	    n = recv(backchannel, &cstmp, sizeof(cstmp), 0);
	    if (n == -1) {
		if (errno == EINTR)
		    continue;
		warning("error reading from socketpair");
		goto done;
	    }
	    if (cstmp.type != CMD_SIGNO) {
		warningx("unexpected reply type on backchannel: %d", cstmp.type);
		continue;
	    }
	    deliver_signal(child, cstmp.val);
	}
    }

done:
    if (alive) {
	/* XXX An error occurred, should send an error back. */
	kill(child, SIGKILL);
    } else {
	/* Send parent status. */
	send_status(backchannel, &cstat);
    }
    _exit(1);

bad:
    return errno;
}

/*
 * Flush any output buffered in iobufs or readable from the fds.
 * Does not read from /dev/tty.
 */
static void
flush_output()
{
    struct io_buffer *iob;
    struct timeval tv;
    fd_set *fdsr, *fdsw;
    int nready, nwriters, maxfd = -1;

    /* Determine maxfd */
    for (iob = iobufs; iob; iob = iob->next) {
	if (iob->rfd > maxfd)
	    maxfd = iob->rfd;
	if (iob->wfd > maxfd)
	    maxfd = iob->wfd;
    }
    if (maxfd == -1)
	return;

    fdsr = (fd_set *)emalloc2(howmany(maxfd + 1, NFDBITS), sizeof(fd_mask));
    fdsw = (fd_set *)emalloc2(howmany(maxfd + 1, NFDBITS), sizeof(fd_mask));
    for (;;) {
	zero_bytes(fdsw, howmany(maxfd + 1, NFDBITS) * sizeof(fd_mask));
	zero_bytes(fdsr, howmany(maxfd + 1, NFDBITS) * sizeof(fd_mask));

	nwriters = 0;
	for (iob = iobufs; iob; iob = iob->next) {
	    /* Don't read from /dev/tty while flushing. */
	    if (io_fds[SFD_USERTTY] != -1 && iob->rfd == io_fds[SFD_USERTTY])
		continue;
	    if (iob->rfd == -1 && iob->wfd == -1)
	    	continue;
	    if (iob->off == iob->len) {
		iob->off = iob->len = 0;
		/* Forward the EOF from reader to writer. */
		if (iob->rfd == -1) {
		    safe_close(iob->wfd);
		    iob->wfd = -1;
		}
	    }
	    if (iob->rfd != -1) {
		if (iob->len != sizeof(iob->buf))
		    FD_SET(iob->rfd, fdsr);
	    }
	    if (iob->wfd != -1) {
		if (iob->len > iob->off) {
		    nwriters++;
		    FD_SET(iob->wfd, fdsw);
		}
	    }
	}

	/* Don't sleep in select if there are no buffers that need writing. */
	tv.tv_sec = 0;
	tv.tv_usec = 0;
	nready = select(maxfd + 1, fdsr, fdsw, NULL, nwriters ? NULL : &tv);
	if (nready <= 0) {
	    if (nready == 0)
		break; /* all I/O flushed */
	    if (errno == EINTR)
		continue;
	    error(1, "select failed");
	}
	if (perform_io(fdsr, fdsw, NULL) != 0)
	    break;
    }
    efree(fdsr);
    efree(fdsw);
}

/*
 * Sets up std{in,out,err} and executes the actual command.
 * Returns only if execve() fails.
 */
static void
exec_pty(path, argv, envp, rbac_enabled)
    const char *path;
    char *argv[];
    char *envp[];
    int rbac_enabled;
{
    sigaction_t sa;
    pid_t self = getpid();

    /* Reset signal handlers. */
    zero_bytes(&sa, sizeof(sa));
    sigemptyset(&sa.sa_mask);
    sa.sa_flags = SA_RESTART;
    sa.sa_handler = SIG_DFL;
    sigaction(SIGHUP, &sa, NULL);
    sigaction(SIGTERM, &sa, NULL);
    sigaction(SIGINT, &sa, NULL);
    sigaction(SIGQUIT, &sa, NULL);
    sigaction(SIGTSTP, &sa, NULL);
    sigaction(SIGTTIN, &sa, NULL);
    sigaction(SIGTTOU, &sa, NULL);
    sigaction(SIGUSR1, &sa, NULL);
    sigaction(SIGUSR2, &sa, NULL);
    sigaction(SIGCHLD, &sa, NULL);

    /* Set child process group here too to avoid a race. */
    setpgid(0, self);

    /* Wire up standard fds, note that stdout/stderr may be pipes. */
    if (dup2(io_fds[SFD_STDIN], STDIN_FILENO) == -1 ||
	dup2(io_fds[SFD_STDOUT], STDOUT_FILENO) == -1 ||
	dup2(io_fds[SFD_STDERR], STDERR_FILENO) == -1)
	error(1, "dup2");

    /* Wait for parent to grant us the tty if we are foreground. */
    if (foreground) {
	while (tcgetpgrp(io_fds[SFD_SLAVE]) != self)
	    ; /* spin */
    }

    /* We have guaranteed that the slave fd is > 2 */
    if (io_fds[SFD_SLAVE] != -1)
	close(io_fds[SFD_SLAVE]);
    if (io_fds[SFD_STDIN] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDIN]);
    if (io_fds[SFD_STDOUT] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDOUT]);
    if (io_fds[SFD_STDERR] != io_fds[SFD_SLAVE])
	close(io_fds[SFD_STDERR]);

    closefrom(def_closefrom);
#ifdef HAVE_SELINUX
    if (rbac_enabled)
	selinux_execve(path, argv, envp);
    else
#endif
	my_execve(path, argv, envp);
}

/*
 * Propagates tty size change signals to pty being used by the command.
 */
static void
sync_ttysize(src, dst)
    int src;
    int dst;
{
#ifdef TIOCGSIZE
    struct ttysize tsize;
    pid_t pgrp;

    if (ioctl(src, TIOCGSIZE, &tsize) == 0) {
	    ioctl(dst, TIOCSSIZE, &tsize);
	    if ((pgrp = tcgetpgrp(dst)) != -1)
		killpg(pgrp, SIGWINCH);
    }
#endif
}

/*
 * Handler for SIGWINCH in parent.
 */
static void
sigwinch(s)
    int s;
{
    int serrno = errno;

    sync_ttysize(io_fds[SFD_USERTTY], io_fds[SFD_SLAVE]);
    errno = serrno;
}

/*
 * Only close the fd if it is not /dev/tty or std{in,out,err}.
 * Return value is the same as send(2).
 */
static int
safe_close(fd)
    int fd;
{
    /* Avoid closing /dev/tty or std{in,out,err}. */
    if (fd < 3 || fd == io_fds[SFD_USERTTY]) {
	errno = EINVAL;
	return -1;
    }
    return close(fd);
}