README

NAME
    IPC::Run - system() and background procs w/ piping, redirs, ptys (Unix,
    Win32)

SYNOPSIS
       ## First,a command to run:
          my @cat = qw( cat );

       ## Using run() instead of system():
          use IPC::Run qw( run timeout );

          run \@cmd, \$in, \$out, \$err, timeout( 10 ) or die "cat: $?"

          # Can do I/O to sub refs and filenames, too:
          run \@cmd, '<', "in.txt", \&out, \&err or die "cat: $?"
          run \@cat, '<', "in.txt", '>>', "out.txt", '2>>', "err.txt";


          # Redirecting using psuedo-terminals instad of pipes.
          run \@cat, '<pty<', \$in,  '>pty>', \$out_and_err;

       ## Scripting subprocesses (like Expect):

          use IPC::Run qw( start pump finish timeout );

          # Incrementally read from / write to scalars. 
          # $in is drained as it is fed to cat's stdin,
          # $out accumulates cat's stdout
          # $err accumulates cat's stderr
          # $h is for "harness".
          my $h = start \@cat, \$in, \$out, \$err, timeout( 10 );

          $in .= "some input\n";
          pump $h until $out =~ /input\n/g;

          $in .= "some more input\n";
          pump $h until $out =~ /\G.*more input\n/;

          $in .= "some final input\n";
          finish $h or die "cat returned $?";

          warn $err if $err; 
          print $out;         ## All of cat's output

       # Piping between children
          run \@cat, '|', \@gzip;

       # Multiple children simultaneously (run() blocks until all
       # children exit, use start() for background execution):
          run \@foo1, '&', \@foo2;

       # Calling \&set_up_child in the child before it executes the
       # command (only works on systems with true fork() & exec())
       # exceptions thrown in set_up_child() will be propagated back
       # to the parent and thrown from run().
          run \@cat, \$in, \$out,
             init => \&set_up_child;

       # Read from / write to file handles you open and close
          open IN,  '<in.txt'  or die $!;
          open OUT, '>out.txt' or die $!;
          print OUT "preamble\n";
          run \@cat, \*IN, \*OUT or die "cat returned $?";
          print OUT "postamble\n";
          close IN;
          close OUT;

       # Create pipes for you to read / write (like IPC::Open2 & 3).
          $h = start
             \@cat,
                '<pipe', \*IN,
                '>pipe', \*OUT,
                '2>pipe', \*ERR 
             or die "cat returned $?";
          print IN "some input\n";
          close IN;
          print <OUT>, <ERR>;
          finish $h;

       # Mixing input and output modes
          run \@cat, 'in.txt', \&catch_some_out, \*ERR_LOG );

       # Other redirection constructs
          run \@cat, '>&', \$out_and_err;
          run \@cat, '2>&1';
          run \@cat, '0<&3';
          run \@cat, '<&-';
          run \@cat, '3<', \$in3;
          run \@cat, '4>', \$out4;
          # etc.

       # Passing options:
          run \@cat, 'in.txt', debug => 1;

       # Call this system's shell, returns TRUE on 0 exit code
       # THIS IS THE OPPOSITE SENSE OF system()'s RETURN VALUE
          run "cat a b c" or die "cat returned $?";

       # Launch a sub process directly, no shell.  Can't do redirection
       # with this form, it's here to behave like system() with an
       # inverted result.
          $r = run "cat a b c";

       # Read from a file in to a scalar
          run io( "filename", 'r', \$recv );
          run io( \*HANDLE,   'r', \$recv );

DESCRIPTION
    IPC::Run allows you run and interact with child processes using files,
    pipes, and pseudo-ttys. Both system()-style and scripted usages are
    supported and may be mixed. Likewise, functional and OO API styles are
    both supported and may be mixed.

    Various redirection operators reminiscent of those seen on common Unix
    and DOS command lines are provided.

    Before digging in to the details a few LIMITATIONS are important enough
    to be mentioned right up front:

    Win32 Support
        Win32 support is working but EXPERIMENTAL, but does pass all
        relevant tests on NT 4.0. See "Win32 LIMITATIONS".

    pty Support
        If you need pty support, IPC::Run should work well enough most of
        the time, but IO::Pty is being improved, and IPC::Run will be
        improved to use IO::Pty's new features when it is release.

        The basic problem is that the pty needs to initialize itself before
        the parent writes to the master pty, or the data written gets lost.
        So IPC::Run does a sleep(1) in the parent after forking to
        (hopefully) give the child a chance to run. This is a kludge that
        works well on non heavily loaded systems :(.

        ptys are not supported yet under Win32, but will be emulated...

    Debugging Tip
        You may use the environment variable "IPCRUNDEBUG" to see what's
        going on under the hood:

           $ IPCRUNDEBUG=basic   myscript     # prints minimal debugging
           $ IPCRUNDEBUG=data    myscript     # prints all data reads/writes
           $ IPCRUNDEBUG=details myscript     # prints lots of low-level details
           $ IPCRUNDEBUG=gory    myscript     # (Win32 only) prints data moving through
                                              # the helper processes.

    We now return you to your regularly scheduled documentation.

  Harnesses
    Child processes and I/O handles are gathered in to a harness, then
    started and run until the processing is finished or aborted.

  run() vs. start(); pump(); finish();
    There are two modes you can run harnesses in: run() functions as an
    enhanced system(), and start()/pump()/finish() allow for background
    processes and scripted interactions with them.

    When using run(), all data to be sent to the harness is set up in
    advance (though one can feed subprocesses input from subroutine refs to
    get around this limitation). The harness is run and all output is
    collected from it, then any child processes are waited for:

       run \@cmd, \<<IN, \$out;
       blah
       IN

       ## To precompile harnesses and run them later:
       my $h = harness \@cmd, \<<IN, \$out;
       blah
       IN

       run $h;

    The background and scripting API is provided by start(), pump(), and
    finish(): start() creates a harness if need be (by calling harness())
    and launches any subprocesses, pump() allows you to poll them for
    activity, and finish() then monitors the harnessed activities until they
    complete.

       ## Build the harness, open all pipes, and launch the subprocesses
       my $h = start \@cat, \$in, \$out;
       $in = "first input\n";

       ## Now do I/O.  start() does no I/O.
       pump $h while length $in;  ## Wait for all input to go

       ## Now do some more I/O.
       $in = "second input\n";
       pump $h until $out =~ /second input/;

       ## Clean up
       finish $h or die "cat returned $?";

    You can optionally compile the harness with harness() prior to
    start()ing or run()ing, and you may omit start() between harness() and
    pump(). You might want to do these things if you compile your harnesses
    ahead of time.

  Using regexps to match output
    As shown in most of the scripting examples, the read-to-scalar facility
    for gathering subcommand's output is often used with regular expressions
    to detect stopping points. This is because subcommand output often
    arrives in dribbles and drabs, often only a character or line at a time.
    This output is input for the main program and piles up in variables like
    the $out and $err in our examples.

    Regular expressions can be used to wait for appropriate output in
    several ways. The "cat" example in the previous section demonstrates how
    to pump() until some string appears in the output. Here's an example
    that uses "smb" to fetch files from a remote server:

       $h = harness \@smbclient, \$in, \$out;

       $in = "cd /src\n";
       $h->pump until $out =~ /^smb.*> \Z/m;
       die "error cding to /src:\n$out" if $out =~ "ERR";
       $out = '';

       $in = "mget *\n";
       $h->pump until $out =~ /^smb.*> \Z/m;
       die "error retrieving files:\n$out" if $out =~ "ERR";

       $in = "quit\n";
       $h->finish;

    Notice that we carefully clear $out after the first command/response
    cycle? That's because IPC::Run does not delete $out when we continue,
    and we don't want to trip over the old output in the second
    command/response cycle.

    Say you want to accumulate all the output in $out and analyze it
    afterwards. Perl offers incremental regular expression matching using
    the "m//gc" and pattern matching idiom and the "\G" assertion. IPC::Run
    is careful not to disturb the current "pos()" value for scalars it
    appends data to, so we could modify the above so as not to destroy $out
    by adding a couple of "/gc" modifiers. The "/g" keeps us from tripping
    over the previous prompt and the "/c" keeps us from resetting the prior
    match position if the expected prompt doesn't materialize immediately:

       $h = harness \@smbclient, \$in, \$out;

       $in = "cd /src\n";
       $h->pump until $out =~ /^smb.*> \Z/mgc;
       die "error cding to /src:\n$out" if $out =~ "ERR";

       $in = "mget *\n";
       $h->pump until $out =~ /^smb.*> \Z/mgc;
       die "error retrieving files:\n$out" if $out =~ "ERR";

       $in = "quit\n";
       $h->finish;

       analyze( $out );

    When using this technique, you may want to preallocate $out to have
    plenty of memory or you may find that the act of growing $out each time
    new input arrives causes an "O(length($out)^2)" slowdown as $out grows.
    Say we expect no more than 10,000 characters of input at the most. To
    preallocate memory to $out, do something like:

       my $out = "x" x 10_000;
       $out = "";

    "perl" will allocate at least 10,000 characters' worth of space, then
    mark the $out as having 0 length without freeing all that yummy RAM.

  Timeouts and Timers
    More than likely, you don't want your subprocesses to run forever, and
    sometimes it's nice to know that they're going a little slowly. Timeouts
    throw exceptions after a some time has elapsed, timers merely cause
    pump() to return after some time has elapsed. Neither is reset/restarted
    automatically.

    Timeout objects are created by calling timeout( $interval ) and passing
    the result to run(), start() or harness(). The timeout period starts
    ticking just after all the child processes have been fork()ed or
    spawn()ed, and are polled for expiration in run(), pump() and finish().
    If/when they expire, an exception is thrown. This is typically useful to
    keep a subprocess from taking too long.

    If a timeout occurs in run(), all child processes will be terminated and
    all file/pipe/ptty descriptors opened by run() will be closed. File
    descriptors opened by the parent process and passed in to run() are not
    closed in this event.

    If a timeout occurs in pump(), pump_nb(), or finish(), it's up to you to
    decide whether to kill_kill() all the children or to implement some more
    graceful fallback. No I/O will be closed in pump(), pump_nb() or
    finish() by such an exception (though I/O is often closed down in those
    routines during the natural course of events).

    Often an exception is too harsh. timer( $interval ) creates timer
    objects that merely prevent pump() from blocking forever. This can be
    useful for detecting stalled I/O or printing a soothing message or "."
    to pacify an anxious user.

    Timeouts and timers can both be restarted at any time using the timer's
    start() method (this is not the start() that launches subprocesses). To
    restart a timer, you need to keep a reference to the timer:

       ## Start with a nice long timeout to let smbclient connect.  If
       ## pump or finish take too long, an exception will be thrown.

     my $h;
     eval {
       $h = harness \@smbclient, \$in, \$out, \$err, ( my $t = timeout 30 );
       sleep 11;  # No effect: timer not running yet

       start $h;
       $in = "cd /src\n";
       pump $h until ! length $in;

       $in = "ls\n";
       ## Now use a short timeout, since this should be faster
       $t->start( 5 );
       pump $h until ! length $in;

       $t->start( 10 );  ## Give smbclient a little while to shut down.
       $h->finish;
     };
     if ( $@ ) {
       my $x = $@;    ## Preserve $@ in case another exception occurs
       $h->kill_kill; ## kill it gently, then brutally if need be, or just
                       ## brutally on Win32.
       die $x;
     }

    Timeouts and timers are *not* checked once the subprocesses are shut
    down; they will not expire in the interval between the last valid
    process and when IPC::Run scoops up the processes' result codes, for
    instance.

  Spawning synchronization, child exception propagation
    start() pauses the parent until the child executes the command or CODE
    reference and propagates any exceptions thrown (including exec()
    failure) back to the parent. This has several pleasant effects: any
    exceptions thrown in the child, including exec() failure, come flying
    out of start() or run() as though they had ocurred in the parent.

    This includes exceptions your code thrown from init subs. In this
    example:

       eval {
          run \@cmd, init => sub { die "blast it! foiled again!" };
       };
       print $@;

    the exception "blast it! foiled again" will be thrown from the child
    process (preventing the exec()) and printed by the parent.

    In situations like

       run \@cmd1, "|", \@cmd2, "|", \@cmd3;

    @cmd1 will be initted and exec()ed before @cmd2, and @cmd2 before @cmd3.
    This can save time and prevent oddball errors emitted by later commands
    when earlier commands fail to execute. Note that IPC::Run doesn't start
    any commands unless it can find the executables referenced by all
    commands. These executables must pass both the "-f" and "-x" tests
    described in perlfunc.

    Another nice effect is that init() subs can take their time doing things
    and there will be no problems caused by a parent continuing to execute
    before a child's init() routine is complete. Say the init() routine
    needs to open a socket or a temp file that the parent wants to connect
    to; without this synchronization, the parent will need to implement a
    retry loop to wait for the child to run, since often, the parent gets a
    lot of things done before the child's first timeslice is allocated.

    This is also quite necessary for pseudo-tty initialization, which needs
    to take place before the parent writes to the child via pty. Writes that
    occur before the pty is set up can get lost.

    A final, minor, nicety is that debugging output from the child will be
    emitted before the parent continues on, making for much clearer
    debugging output in complex situations.

    The only drawback I can conceive of is that the parent can't continue to
    operate while the child is being initted. If this ever becomes a problem
    in the field, we can implement an option to avoid this behavior, but I
    don't expect it to.

    Win32: executing CODE references isn't supported on Win32, see "Win32
    LIMITATIONS" for details.

  Syntax
    run(), start(), and harness() can all take a harness specification as
    input. A harness specification is either a single string to be passed to
    the systems' shell:

       run "echo 'hi there'";

    or a list of commands, io operations, and/or timers/timeouts to execute.
    Consecutive commands must be separated by a pipe operator '|' or an '&'.
    External commands are passed in as array references, and, on systems
    supporting fork(), Perl code may be passed in as subs:

       run \@cmd;
       run \@cmd1, '|', \@cmd2;
       run \@cmd1, '&', \@cmd2;
       run \&sub1;
       run \&sub1, '|', \&sub2;
       run \&sub1, '&', \&sub2;

    '|' pipes the stdout of \@cmd1 the stdin of \@cmd2, just like a shell
    pipe. '&' does not. Child processes to the right of a '&' will have
    their stdin closed unless it's redirected-to.

    IPC::Run::IO objects may be passed in as well, whether or not child
    processes are also specified:

       run io( "infile", ">", \$in ), io( "outfile", "<", \$in );

    as can IPC::Run::Timer objects:

       run \@cmd, io( "outfile", "<", \$in ), timeout( 10 );

    Commands may be followed by scalar, sub, or i/o handle references for
    redirecting child process input & output:

       run \@cmd,  \undef,            \$out;
       run \@cmd,  \$in,              \$out;
       run \@cmd1, \&in, '|', \@cmd2, \*OUT;
       run \@cmd1, \*IN, '|', \@cmd2, \&out;

    This is known as succinct redirection syntax, since run(), start() and
    harness(), figure out which file descriptor to redirect and how. File
    descriptor 0 is presumed to be an input for the child process, all
    others are outputs. The assumed file descriptor always starts at 0,
    unless the command is being piped to, in which case it starts at 1.

    To be explicit about your redirects, or if you need to do more complex
    things, there's also a redirection operator syntax:

       run \@cmd, '<', \undef, '>',  \$out;
       run \@cmd, '<', \undef, '>&', \$out_and_err;
       run(
          \@cmd1,
             '<', \$in,
          '|', \@cmd2,
             \$out
       );

    Operator syntax is required if you need to do something other than
    simple redirection to/from scalars or subs, like duping or closing file
    descriptors or redirecting to/from a named file. The operators are
    covered in detail below.

    After each \@cmd (or \&foo), parsing begins in succinct mode and toggles
    to operator syntax mode when an operator (ie plain scalar, not a ref) is
    seen. Once in operator syntax mode, parsing only reverts to succinct
    mode when a '|' or '&' is seen.

    In succinct mode, each parameter after the \@cmd specifies what to do
    with the next highest file descriptor. These File descriptor start with
    0 (stdin) unless stdin is being piped to ("'|', \@cmd"), in which case
    they start with 1 (stdout). Currently, being on the left of a pipe
    ("\@cmd, \$out, \$err, '|'") does *not* cause stdout to be skipped,
    though this may change since it's not as DWIMerly as it could be. Only
    stdin is assumed to be an input in succinct mode, all others are assumed
    to be outputs.

    If no piping or redirection is specified for a child, it will inherit
    the parent's open file handles as dictated by your system's
    close-on-exec behavior and the $^F flag, except that processes after a
    '&' will not inherit the parent's stdin. Also note that $^F does not
    affect file desciptors obtained via POSIX, since it only applies to
    full-fledged Perl file handles. Such processes will have their stdin
    closed unless it has been redirected-to.

    If you want to close a child processes stdin, you may do any of:

       run \@cmd, \undef;
       run \@cmd, \"";
       run \@cmd, '<&-';
       run \@cmd, '0<&-';

    Redirection is done by placing redirection specifications immediately
    after a command or child subroutine:

       run \@cmd1,      \$in, '|', \@cmd2,      \$out;
       run \@cmd1, '<', \$in, '|', \@cmd2, '>', \$out;

    If you omit the redirection operators, descriptors are counted starting
    at 0. Descriptor 0 is assumed to be input, all others are outputs. A
    leading '|' consumes descriptor 0, so this works as expected.

       run \@cmd1, \$in, '|', \@cmd2, \$out;

    The parameter following a redirection operator can be a scalar ref, a
    subroutine ref, a file name, an open filehandle, or a closed filehandle.

    If it's a scalar ref, the child reads input from or sends output to that
    variable:

       $in = "Hello World.\n";
       run \@cat, \$in, \$out;
       print $out;

    Scalars used in incremental (start()/pump()/finish()) applications are
    treated as queues: input is removed from input scalers, resulting in
    them dwindling to '', and output is appended to output scalars. This is
    not true of harnesses run() in batch mode.

    It's usually wise to append new input to be sent to the child to the
    input queue, and you'll often want to zap output queues to '' before
    pumping.

       $h = start \@cat, \$in;
       $in = "line 1\n";
       pump $h;
       $in .= "line 2\n";
       pump $h;
       $in .= "line 3\n";
       finish $h;

    The final call to finish() must be there: it allows the child
    process(es) to run to completion and waits for their exit values.

OBSTINATE CHILDREN
    Interactive applications are usually optimized for human use. This can
    help or hinder trying to interact with them through modules like
    IPC::Run. Frequently, programs alter their behavior when they detect
    that stdin, stdout, or stderr are not connected to a tty, assuming that
    they are being run in batch mode. Whether this helps or hurts depends on
    which optimizations change. And there's often no way of telling what a
    program does in these areas other than trial and error and,
    occasionally, reading the source. This includes different versions and
    implementations of the same program.

    All hope is not lost, however. Most programs behave in reasonably
    tractable manners, once you figure out what it's trying to do.

    Here are some of the issues you might need to be aware of.

    *   fflush()ing stdout and stderr

        This lets the user see stdout and stderr immediately. Many programs
        undo this optimization if stdout is not a tty, making them harder to
        manage by things like IPC::Run.

        Many programs decline to fflush stdout or stderr if they do not
        detect a tty there. Some ftp commands do this, for instance.

        If this happens to you, look for a way to force interactive
        behavior, like a command line switch or command. If you can't, you
        will need to use a pseudo terminal ('<pty<' and '>pty>').

    *   false prompts

        Interactive programs generally do not guarantee that output from
        user commands won't contain a prompt string. For example, your shell
        prompt might be a '$', and a file named '$' might be the only file
        in a directory listing.

        This can make it hard to guarantee that your output parser won't be
        fooled into early termination of results.

        To help work around this, you can see if the program can alter it's
        prompt, and use something you feel is never going to occur in actual
        practice.

        You should also look for your prompt to be the only thing on a line:

           pump $h until $out =~ /^<SILLYPROMPT>\s?\z/m;

        (use "(?!\n)\Z" in place of "\z" on older perls).

        You can also take the approach that IPC::ChildSafe takes and emit a
        command with known output after each 'real' command you issue, then
        look for this known output. See new_appender() and new_chunker() for
        filters that can help with this task.

        If it's not convenient or possibly to alter a prompt or use a known
        command/response pair, you might need to autodetect the prompt in
        case the local version of the child program is different then the
        one you tested with, or if the user has control over the look & feel
        of the prompt.

    *   Refusing to accept input unless stdin is a tty.

        Some programs, for security reasons, will only accept certain types
        of input from a tty. su, notable, will not prompt for a password
        unless it's connected to a tty.

        If this is your situation, use a pseudo terminal ('<pty<' and
        '>pty>').

    *   Not prompting unless connected to a tty.

        Some programs don't prompt unless stdin or stdout is a tty. See if
        you can turn prompting back on. If not, see if you can come up with
        a command that you can issue after every real command and look for
        it's output, as IPC::ChildSafe does. There are two filters included
        with IPC::Run that can help with doing this: appender and chunker
        (see new_appender() and new_chunker()).

    *   Different output format when not connected to a tty.

        Some commands alter their formats to ease machine parsability when
        they aren't connected to a pipe. This is actually good, but can be
        surprising.

PSEUDO TERMINALS
    On systems providing pseudo terminals under /dev, IPC::Run can use
    IO::Pty (available on CPAN) to provide a terminal environment to
    subprocesses. This is necessary when the subprocess really wants to
    think it's connected to a real terminal.

  CAVEATS
    Psuedo-terminals are not pipes, though they are similar. Here are some
    differences to watch out for.

    Echoing
        Sending to stdin will cause an echo on stdout, which occurs before
        each line is passed to the child program. There is currently no way
        to disable this, although the child process can and should disable
        it for things like passwords.

    Shutdown
        IPC::Run cannot close a pty until all output has been collected.
        This means that it is not possible to send an EOF to stdin by
        half-closing the pty, as we can when using a pipe to stdin.

        This means that you need to send the child process an exit command
        or signal, or run() / finish() will time out. Be careful not to
        expect a prompt after sending the exit command.

    Command line editing
        Some subprocesses, notable shells that depend on the user's prompt
        settings, will reissue the prompt plus the command line input so far
        once for each character.

    '>pty>' means '&>pty>', not '1>pty>'
        The pseudo terminal redirects both stdout and stderr unless you
        specify a file descriptor. If you want to grab stderr separately, do
        this:

           start \@cmd, '<pty<', \$in, '>pty>', \$out, '2>', \$err;

    stdin, stdout, and stderr not inherited
        Child processes harnessed to a pseudo terminal have their stdin,
        stdout, and stderr completely closed before any redirection
        operators take effect. This casts of the bonds of the controlling
        terminal. This is not done when using pipes.

        Right now, this affects all children in a harness that has a pty in
        use, even if that pty would not affect a particular child. That's a
        bug and will be fixed. Until it is, it's best not to mix-and-match
        children.

  Redirection Operators
       Operator       SHNP   Description
       ========       ====   ===========
       <, N<          SHN    Redirects input to a child's fd N (0 assumed)

       >, N>          SHN    Redirects output from a child's fd N (1 assumed)
       >>, N>>        SHN    Like '>', but appends to scalars or named files
       >&, &>         SHN    Redirects stdout & stderr from a child process

       <pty, N<pty    S      Like '<', but uses a pseudo-tty instead of a pipe
       >pty, N>pty    S      Like '>', but uses a pseudo-tty instead of a pipe

       N<&M                  Dups input fd N to input fd M
       M>&N                  Dups output fd N to input fd M
       N<&-                  Closes fd N

       <pipe, N<pipe     P   Pipe opens H for caller to read, write, close.
       >pipe, N>pipe     P   Pipe opens H for caller to read, write, close.

    'N' and 'M' are placeholders for integer file descriptor numbers. The
    terms 'input' and 'output' are from the child process's perspective.

    The SHNP field indicates what parameters an operator can take:

       S: \$scalar or \&function references.  Filters may be used with
          these operators (and only these).
       H: \*HANDLE or IO::Handle for caller to open, and close
       N: "file name".
       P: \*HANDLE opened by IPC::Run as the parent end of a pipe, but read
          and written to and closed by the caller (like IPC::Open3).

    Redirecting input: [n]<, [n]<pipe
        You can input the child reads on file descriptor number n to come
        from a scalar variable, subroutine, file handle, or a named file. If
        stdin is not redirected, the parent's stdin is inherited.

           run \@cat, \undef          ## Closes child's stdin immediately
              or die "cat returned $?"; 

           run \@cat, \$in;

           run \@cat, \<<TOHERE;
           blah
           TOHERE

           run \@cat, \&input;       ## Calls &input, feeding data returned
                                      ## to child's.  Closes child's stdin
                                      ## when undef is returned.

        Redirecting from named files requires you to use the input
        redirection operator:

           run \@cat, '<.profile';
           run \@cat, '<', '.profile';

           open IN, "<foo";
           run \@cat, \*IN;
           run \@cat, *IN{IO};

        The form used second example here is the safest, since filenames
        like "0" and "&more\n" won't confuse &run:

        You can't do either of

           run \@a, *IN;      ## INVALID
           run \@a, '<', *IN; ## BUGGY: Reads file named like "*main::A"

        because perl passes a scalar containing a string that looks like
        "*main::A" to &run, and &run can't tell the difference between that
        and a redirection operator or a file name. &run guarantees that any
        scalar you pass after a redirection operator is a file name.

        If your child process will take input from file descriptors other
        than 0 (stdin), you can use a redirection operator with any of the
        valid input forms (scalar ref, sub ref, etc.):

           run \@cat, '3<', \$in3;

        When redirecting input from a scalar ref, the scalar ref is used as
        a queue. This allows you to use &harness and pump() to feed
        incremental bits of input to a coprocess. See "Coprocesses" below
        for more information.

        The <pipe operator opens the write half of a pipe on the filehandle
        glob reference it takes as an argument:

           $h = start \@cat, '<pipe', \*IN;
           print IN "hello world\n";
           pump $h;
           close IN;
           finish $h;

        Unlike the other '<' operators, IPC::Run does nothing further with
        it: you are responsible for it. The previous example is functionally
        equivalent to:

           pipe( \*R, \*IN ) or die $!;
           $h = start \@cat, '<', \*IN;
           print IN "hello world\n";
           pump $h;
           close IN;
           finish $h;

        This is like the behavior of IPC::Open2 and IPC::Open3.

        Win32: The handle returned is actually a socket handle, so you can
        use select() on it.

    Redirecting output: [n]>, [n]>>, [n]>&[m], [n]>pipe
        You can redirect any output the child emits to a scalar variable,
        subroutine, file handle, or file name. You can have &run truncate or
        append to named files or scalars. If you are redirecting stdin as
        well, or if the command is on the receiving end of a pipeline ('|'),
        you can omit the redirection operator:

           @ls = ( 'ls' );
           run \@ls, \undef, \$out
              or die "ls returned $?"; 

           run \@ls, \undef, \&out;  ## Calls &out each time some output
                                      ## is received from the child's 
                                      ## when undef is returned.

           run \@ls, \undef, '2>ls.err';
           run \@ls, '2>', 'ls.err';

        The two parameter form guarantees that the filename will not be
        interpreted as a redirection operator:

           run \@ls, '>', "&more";
           run \@ls, '2>', ">foo\n";

        You can pass file handles you've opened for writing:

           open( *OUT, ">out.txt" );
           open( *ERR, ">err.txt" );
           run \@cat, \*OUT, \*ERR;

        Passing a scalar reference and a code reference requires a little
        more work, but allows you to capture all of the output in a scalar
        or each piece of output by a callback:

        These two do the same things:

           run( [ 'ls' ], '2>', sub { $err_out .= $_[0] } );

        does the same basic thing as:

           run( [ 'ls' ], '2>', \$err_out );

        The subroutine will be called each time some data is read from the
        child.

        The >pipe operator is different in concept than the other '>'
        operators, although it's syntax is similar:

           $h = start \@cat, $in, '>pipe', \*OUT, '2>pipe', \*ERR;
           $in = "hello world\n";
           finish $h;
           print <OUT>;
           print <ERR>;
           close OUT;
           close ERR;

        causes two pipe to be created, with one end attached to cat's stdout
        and stderr, respectively, and the other left open on OUT and ERR, so
        that the script can manually read(), select(), etc. on them. This is
        like the behavior of IPC::Open2 and IPC::Open3.

        Win32: The handle returned is actually a socket handle, so you can
        use select() on it.

    Duplicating output descriptors: >&m, n>&m
        This duplicates output descriptor number n (default is 1 if n is
        omitted) from descriptor number m.

    Duplicating input descriptors: <&m, n<&m
        This duplicates input descriptor number n (default is 0 if n is
        omitted) from descriptor number m

    Closing descriptors: <&-, 3<&-
        This closes descriptor number n (default is 0 if n is omitted). The
        following commands are equivalent:

           run \@cmd, \undef;
           run \@cmd, '<&-';
           run \@cmd, '<in.txt', '<&-';

        Doing

           run \@cmd, \$in, '<&-';    ## SIGPIPE recipe.

        is dangerous: the parent will get a SIGPIPE if $in is not empty.

    Redirecting both stdout and stderr: &>, >&, &>pipe, >pipe&
        The following pairs of commands are equivalent:

           run \@cmd, '>&', \$out;       run \@cmd, '>', \$out,     '2>&1';
           run \@cmd, '>&', 'out.txt';   run \@cmd, '>', 'out.txt', '2>&1';

        etc.

        File descriptor numbers are not permitted to the left or the right
        of these operators, and the '&' may occur on either end of the
        operator.

        The '&>pipe' and '>pipe&' variants behave like the '>pipe' operator,
        except that both stdout and stderr write to the created pipe.

    Redirection Filters
        Both input redirections and output redirections that use scalars or
        subs as endpoints may have an arbitrary number of filter subs placed
        between them and the child process. This is useful if you want to
        receive output in chunks, or if you want to massage each chunk of
        data sent to the child. To use this feature, you must use operator
        syntax:

           run(
              \@cmd
                 '<', \&in_filter_2, \&in_filter_1, $in,
                 '>', \&out_filter_1, \&in_filter_2, $out,
           );

        This capability is not provided for IO handles or named files.

        Two filters are provided by IPC::Run: appender and chunker. Because
        these may take an argument, you need to use the constructor
        functions new_appender() and new_chunker() rather than using \&
        syntax:

           run(
              \@cmd
                 '<', new_appender( "\n" ), $in,
                 '>', new_chunker, $out,
           );

  Just doing I/O
    If you just want to do I/O to a handle or file you open yourself, you
    may specify a filehandle or filename instead of a command in the harness
    specification:

       run io( "filename", '>', \$recv );

       $h = start io( $io, '>', \$recv );

       $h = harness \@cmd, '&', io( "file", '<', \$send );

  Options
    Options are passed in as name/value pairs:

       run \@cat, \$in, debug => 1;

    If you pass the debug option, you may want to pass it in first, so you
    can see what parsing is going on:

       run debug => 1, \@cat, \$in;

    debug
        Enables debugging output in parent and child. Debugging info is
        emitted to the STDERR that was present when IPC::Run was first
        "use()"ed (it's "dup()"ed out of the way so that it can be
        redirected in children without having debugging output emitted on
        it).

RETURN VALUES
    harness() and start() return a reference to an IPC::Run harness. This is
    blessed in to the IPC::Run package, so you may make later calls to
    functions as members if you like:

       $h = harness( ... );
       $h->start;
       $h->pump;
       $h->finish;

       $h = start( .... );
       $h->pump;
       ...

    Of course, using method call syntax lets you deal with any IPC::Run
    subclasses that might crop up, but don't hold your breath waiting for
    any.

    run() and finish() return TRUE when all subcommands exit with a 0 result
    code. This is the opposite of perl's system() command.

    All routines raise exceptions (via die()) when error conditions are
    recognized. A non-zero command result is not treated as an error
    condition, since some commands are tests whose results are reported in
    their exit codes.

ROUTINES
    run Run takes a harness or harness specification and runs it, pumping
        all input to the child(ren), closing the input pipes when no more
        input is available, collecting all output that arrives, until the
        pipes delivering output are closed, then waiting for the children to
        exit and reaping their result codes.

        You may think of "run( ... )" as being like

           start( ... )->finish();

        , though there is one subtle difference: run() does not set
        \$input_scalars to '' like finish() does. If an exception is thrown
        from run(), all children will be killed off "gently", and then
        "annihilated" if they do not go gently (in to that dark night.
        sorry).

        If any exceptions are thrown, this does a "kill_kill" before
        propogating them.

    signal
           ## To send it a specific signal by name ("USR1"):
           signal $h, "USR1";
           $h->signal ( "USR1" );

        If $signal is provided and defined, sends a signal to all child
        processes. Try not to send numeric signals, use "KILL" instead of 9,
        for instance. Numeric signals aren't portable.

        Throws an exception if $signal is undef.

        This will *not* clean up the harness, "finish" it if you kill it.

        Normally TERM kills a process gracefully (this is what the command
        line utility "kill" does by default), INT is sent by one of the keys
        "^C", "Backspace" or "<Del>", and "QUIT" is used to kill a process
        and make it coredump.

        The "HUP" signal is often used to get a process to "restart",
        rereading config files, and "USR1" and "USR2" for really
        application-specific things.

        Often, running "kill -l" (that's a lower case "L") on the command
        line will list the signals present on your operating system.

        WARNING: The signal subsystem is not at all portable. We *may* offer
        to simulate "TERM" and "KILL" on some operating systems, submit code
        to me if you want this.

        WARNING 2: Up to and including perl v5.6.1, doing almost anything in
        a signal handler could be dangerous. The most safe code avoids all
        mallocs and system calls, usually by preallocating a flag before
        entering the signal handler, altering the flag's value in the
        handler, and responding to the changed value in the main system:

           my $got_usr1 = 0;
           sub usr1_handler { ++$got_signal }

           $SIG{USR1} = \&usr1_handler;
           while () { sleep 1; print "GOT IT" while $got_usr1--; }

        Even this approach is perilous if ++ and -- aren't atomic on your
        system (I've never heard of this on any modern CPU large enough to
        run perl).

    kill_kill
           ## To kill off a process:
           $h->kill_kill;
           kill_kill $h;

           ## To specify the grace period other than 30 seconds:
           kill_kill $h, grace => 5;

           ## To send QUIT instead of KILL if a process refuses to die:
           kill_kill $h, coup_d_grace => "QUIT";

        Sends a "TERM", waits for all children to exit for up to 30 seconds,
        then sends a "KILL" to any that survived the "TERM".

        Will wait for up to 30 more seconds for the OS to sucessfully "KILL"
        the processes.

        The 30 seconds may be overriden by setting the "grace" option, this
        overrides both timers.

        The harness is then cleaned up.

        The doubled name indicates that this function may kill again and
        avoids colliding with the core Perl "kill" function.

        Returns a 1 if the "TERM" was sufficient, or a 0 if "KILL" was
        required. Throws an exception if "KILL" did not permit the children
        to be reaped.

        NOTE: The grace period is actually up to 1 second longer than that
        given. This is because the granularity of "time" is 1 second. Let me
        know if you need finer granularity, we can leverage Time::HiRes
        here.

        Win32: Win32 does not know how to send real signals, so "TERM" is a
        full-force kill on Win32. Thus all talk of grace periods, etc. do
        not apply to Win32.

    harness
        Takes a harness specification and returns a harness. This harness is
        blessed in to IPC::Run, allowing you to use method call syntax for
        run(), start(), et al if you like.

        harness() is provided so that you can pre-build harnesses if you
        would like to, but it's not required..

        You may proceed to run(), start() or pump() after calling harness()
        (pump() calls start() if need be). Alternatively, you may pass your
        harness specification to run() or start() and let them harness() for
        you. You can't pass harness specifications to pump(), though.

    close_terminal
        This is used as (or in) an init sub to cast off the bonds of a
        controlling terminal. It must precede all other redirection ops that
        affect STDIN, STDOUT, or STDERR to be guaranteed effective.

    start
           $h = start(
              \@cmd, \$in, \$out, ...,
              timeout( 30, name => "process timeout" ),
              $stall_timeout = timeout( 10, name => "stall timeout"   ),
           );

           $h = start \@cmd, '<', \$in, '|', \@cmd2, ...;

        start() accepts a harness or harness specification and returns a
        harness after building all of the pipes and launching (via
        fork()/exec(), or, maybe someday, spawn()) all the child processes.
        It does not send or receive any data on the pipes, see pump() and
        finish() for that.

        You may call harness() and then pass it's result to start() if you
        like, but you only need to if it helps you structure or tune your
        application. If you do call harness(), you may skip start() and
        proceed directly to pump.

        start() also starts all timers in the harness. See IPC::Run::Timer
        for more information.

        start() flushes STDOUT and STDERR to help you avoid duplicate
        output. It has no way of asking Perl to flush all your open
        filehandles, so you are going to need to flush any others you have
        open. Sorry.

        Here's how if you don't want to alter the state of $| for your
        filehandle:

           $ofh = select HANDLE; $of = $|; $| = 1; $| = $of; select $ofh;

        If you don't mind leaving output unbuffered on HANDLE, you can do
        the slightly shorter

           $ofh = select HANDLE; $| = 1; select $ofh;

        Or, you can use IO::Handle's flush() method:

           use IO::Handle;
           flush HANDLE;

        Perl needs the equivalent of C's fflush( (FILE *)NULL ).

    pump
           pump $h;
           $h->pump;

        Pump accepts a single parameter harness. It blocks until it delivers
        some input or recieves some output. It returns TRUE if there is
        still input or output to be done, FALSE otherwise.

        pump() will automatically call start() if need be, so you may call
        harness() then proceed to pump() if that helps you structure your
        application.

        If pump() is called after all harnessed activities have completed, a
        "process ended prematurely" exception to be thrown. This allows for
        simple scripting of external applications without having to add lots
        of error handling code at each step of the script:

           $h = harness \@smbclient, \$in, \$out, $err;

           $in = "cd /foo\n";
           $h->pump until $out =~ /^smb.*> \Z/m;
           die "error cding to /foo:\n$out" if $out =~ "ERR";
           $out = '';

           $in = "mget *\n";
           $h->pump until $out =~ /^smb.*> \Z/m;
           die "error retrieving files:\n$out" if $out =~ "ERR";

           $h->finish;

           warn $err if $err;

    pump_nb
           pump_nb $h;
           $h->pump_nb;

        "pump() non-blocking", pumps if anything's ready to be pumped,
        returns immediately otherwise. This is useful if you're doing some
        long-running task in the foreground, but don't want to starve any
        child processes.

    pumpable
        Returns TRUE if calling pump() won't throw an immediate "process
        ended prematurely" exception. This means that there are open I/O
        channels or active processes. May yield the parent processes' time
        slice for 0.01 second if all pipes are to the child and all are
        paused. In this case we can't tell if the child is dead, so we yield
        the processor and then attempt to reap the child in a nonblocking
        way.

    reap_nb
        Attempts to reap child processes, but does not block.

        Does not currently take any parameters, one day it will allow
        specific children to be reaped.

        Only call this from a signal handler if your "perl" is recent enough
        to have safe signal handling (5.6.1 did not, IIRC, but it was beign
        discussed on perl5-porters). Calling this (or doing any significant
        work) in a signal handler on older "perl"s is asking for seg faults.

    finish
        This must be called after the last start() or pump() call for a
        harness, or your system will accumulate defunct processes and you
        may "leak" file descriptors.

        finish() returns TRUE if all children returned 0 (and were not
        signaled and did not coredump, ie ! $?), and FALSE otherwise (this
        is like run(), and the opposite of system()).

        Once a harness has been finished, it may be run() or start()ed
        again, including by pump()s auto-start.

        If this throws an exception rather than a normal exit, the harness
        may be left in an unstable state, it's best to kill the harness to
        get rid of all the child processes, etc.

        Specifically, if a timeout expires in finish(), finish() will not
        kill all the children. Call "<$h-"kill_kill>> in this case if you
        care. This differs from the behavior of "run".

    result
           $h->result;

        Returns the first non-zero result code (ie $? >> 8). See
        "full_result" to get the $? value for a child process.

        To get the result of a particular child, do:

           $h->result( 0 );  # first child's $? >> 8
           $h->result( 1 );  # second child

        or

           ($h->results)[0]
           ($h->results)[1]

        Returns undef if no child processes were spawned and no child number
        was specified. Throws an exception if an out-of-range child number
        is passed.

    results
        Returns a list of child exit values. See "full_results" if you want
        to know if a signal killed the child.

        Throws an exception if the harness is not in a finished state.

    full_result
           $h->full_result;

        Returns the first non-zero $?. See "result" to get the first $? >> 8
        value for a child process.

        To get the result of a particular child, do:

           $h->full_result( 0 );  # first child's $? >> 8
           $h->full_result( 1 );  # second child

        or

           ($h->full_results)[0]
           ($h->full_results)[1]

        Returns undef if no child processes were spawned and no child number
        was specified. Throws an exception if an out-of-range child number
        is passed.

    full_results
        Returns a list of child exit values as returned by "wait". See
        "results" if you don't care about coredumps or signals.

        Throws an exception if the harness is not in a finished state.

FILTERS
    These filters are used to modify input our output between a child
    process and a scalar or subroutine endpoint.

    binary
           run \@cmd, ">", binary, \$out;
           run \@cmd, ">", binary, \$out;  ## Any TRUE value to enable
           run \@cmd, ">", binary 0, \$out;  ## Any FALSE value to disable

        This is a constructor for a "binmode" "filter" that tells IPC::Run
        to keep the carriage returns that would ordinarily be edited out for
        you (binmode is usually off). This is not a real filter, but an
        option masquerading as a filter.

        It's not named "binmode" because you're likely to want to call
        Perl's binmode in programs that are piping binary data around.

    new_chunker
        This breaks a stream of data in to chunks, based on an optional
        scalar or regular expression parameter. The default is the Perl
        input record separator in $/, which is a newline be default.

           run \@cmd, '>', new_chunker, \&lines_handler;
           run \@cmd, '>', new_chunker( "\r\n" ), \&lines_handler;

        Because this uses $/ by default, you should always pass in a
        parameter if you are worried about other code (modules, etc)
        modifying $/.

        If this filter is last in a filter chain that dumps in to a scalar,
        the scalar must be set to '' before a new chunk will be written to
        it.

        As an example of how a filter like this can be written, here's a
        chunker that splits on newlines:

           sub line_splitter {
              my ( $in_ref, $out_ref ) = @_;

              return 0 if length $$out_ref;

              return input_avail && do {
                 while (1) {
                    if ( $$in_ref =~ s/\A(.*?\n)// ) {
                       $$out_ref .= $1;
                       return 1;
                    }
                    my $hmm = get_more_input;
                    unless ( defined $hmm ) {
                       $$out_ref = $$in_ref;
                       $$in_ref = '';
                       return length $$out_ref ? 1 : 0;
                    }
                    return 0 if $hmm eq 0;
                 }
              }
           };

    new_appender
        This appends a fixed string to each chunk of data read from the
        source scalar or sub. This might be useful if you're writing
        commands to a child process that always must end in a fixed string,
        like "\n":

           run( \@cmd,
              '<', new_appender( "\n" ), \&commands,
           );

        Here's a typical filter sub that might be created by new_appender():

           sub newline_appender {
              my ( $in_ref, $out_ref ) = @_;

              return input_avail && do {
                 $$out_ref = join( '', $$out_ref, $$in_ref, "\n" );
                 $$in_ref = '';
                 1;
              }
           };

    io  Takes a filename or filehandle, a redirection operator, optional
        filters, and a source or destination (depends on the redirection
        operator). Returns an IPC::Run::IO object suitable for harness()ing
        (including via start() or run()).

        This is shorthand for

           require IPC::Run::IO;

              ... IPC::Run::IO->new(...) ...

    timer
           $h = start( \@cmd, \$in, \$out, $t = timer( 5 ) );

           pump $h until $out =~ /expected stuff/ || $t->is_expired;

        Instantiates a non-fatal timer. pump() returns once each time a
        timer expires. Has no direct effect on run(), but you can pass a
        subroutine to fire when the timer expires.

        See "timeout" for building timers that throw exceptions on
        expiration.

        See "timer" in IPC::Run::Timer for details.

    timeout
           $h = start( \@cmd, \$in, \$out, $t = timeout( 5 ) );

           pump $h until $out =~ /expected stuff/;

        Instantiates a timer that throws an exception when it expires. If
        you don't provide an exception, a default exception that matches
        /^IPC::Run: .*timed out/ is thrown by default. You can pass in your
        own exception scalar or reference:

           $h = start(
              \@cmd, \$in, \$out,
              $t = timeout( 5, exception => 'slowpoke' ),
           );

        or set the name used in debugging message and in the default
        exception string:

           $h = start(
              \@cmd, \$in, \$out,
              timeout( 50, name => 'process timer' ),
              $stall_timer = timeout( 5, name => 'stall timer' ),
           );

           pump $h until $out =~ /started/;

           $in = 'command 1';
           $stall_timer->start;
           pump $h until $out =~ /command 1 finished/;

           $in = 'command 2';
           $stall_timer->start;
           pump $h until $out =~ /command 2 finished/;

           $in = 'very slow command 3';
           $stall_timer->start( 10 );
           pump $h until $out =~ /command 3 finished/;

           $stall_timer->start( 5 );
           $in = 'command 4';
           pump $h until $out =~ /command 4 finished/;

           $stall_timer->reset; # Prevent restarting or expirng
           finish $h;

        See "timer" for building non-fatal timers.

        See "timer" in IPC::Run::Timer for details.

FILTER IMPLEMENTATION FUNCTIONS
    These functions are for use from within filters.

    input_avail
        Returns TRUE if input is available. If none is available, then
        &get_more_input is called and its result is returned.

        This is usually used in preference to &get_more_input so that the
        calling filter removes all data from the $in_ref before more data
        gets read in to $in_ref.

        "input_avail" is usually used as part of a return expression:

           return input_avail && do {
              ## process the input just gotten
              1;
           };

        This technique allows input_avail to return the undef or 0 that a
        filter normally returns when there's no input to process. If a
        filter stores intermediate values, however, it will need to react to
        an undef:

           my $got = input_avail;
           if ( ! defined $got ) {
              ## No more input ever, flush internal buffers to $out_ref
           }
           return $got unless $got;
           ## Got some input, move as much as need be
           return 1 if $added_to_out_ref;

    get_more_input
        This is used to fetch more input in to the input variable. It
        returns undef if there will never be any more input, 0 if there is
        none now, but there might be in the future, and TRUE if more input
        was gotten.

        "get_more_input" is usually used as part of a return expression, see
        "input_avail" for more information.

TODO
    These will be addressed as needed and as time allows.

    Stall timeout.

    Expose a list of child process objects. When I do this, each child
    process is likely to be blessed into IPC::Run::Proc.

    $kid->abort(), $kid->kill(), $kid->signal( $num_or_name ).

    Write tests for /(full_)?results?/ subs.

    Currently, pump() and run() only work on systems where select() works on
    the filehandles returned by pipe(). This does *not* include ActiveState
    on Win32, although it does work on cygwin under Win32 (thought the tests
    whine a bit). I'd like to rectify that, suggestions and patches welcome.

    Likewise start() only fully works on fork()/exec() machines (well, just
    fork() if you only ever pass perl subs as subprocesses). There's some
    scaffolding for calling Open3::spawn_with_handles(), but that's
    untested, and not that useful with limited select().

    Support for "\@sub_cmd" as an argument to a command which gets replaced
    with /dev/fd or the name of a temporary file containing foo's output.
    This is like <(sub_cmd ...) found in bash and csh (IIRC).

    Allow multiple harnesses to be combined as independant sets of processes
    in to one 'meta-harness'.

    Allow a harness to be passed in place of an \@cmd. This would allow
    multiple harnesses to be aggregated.

    Ability to add external file descriptors w/ filter chains and endpoints.

    Ability to add timeouts and timing generators (i.e. repeating timeouts).

    High resolution timeouts.

Win32 LIMITATIONS
    Fails on Win9X
        If you want Win9X support, you'll have to debug it or fund me
        because I don't use that system any more. The Win32 subsysem has
        been extended to use temporary files in simple run() invocations and
        these may actually work on Win9X too, but I don't have time to work
        on it.

    May deadlock on Win2K (but not WinNT4 or WinXPPro)
        Spawning more than one subprocess on Win2K causes a deadlock I
        haven't figured out yet, but simple uses of run() often work. Passes
        all tests on WinXPPro and WinNT.

    no support yet for <pty< and >pty>
        These are likely to be implemented as "<" and ">" with binmode on,
        not sure.

    no support for file descriptors higher than 2 (stderr)
        Win32 only allows passing explicit fds 0, 1, and 2. If you really,
        really need to pass file handles, us Win32API:: GetOsFHandle() or
        ::FdGetOsFHandle() to get the integer handle and pass it to the
        child process using the command line, environment, stdin,
        intermediary file, or other IPC mechnism. Then use that handle in
        the child (Win32API.pm provides ways to reconstitute Perl file
        handles from Win32 file handles).

    no support for subroutine subprocesses (CODE refs)
        Can't fork(), so the subroutines would have no context, and closures
        certainly have no meaning

        Perhaps with Win32 fork() emulation, this can be supported in a
        limited fashion, but there are other very serious problems with
        that: all parent fds get dup()ed in to the thread emulating the
        forked process, and that keeps the parent from being able to close
        all of the appropriate fds.

    no support for init => sub {} routines.
        Win32 processes are created from scratch, there is no way to do an
        init routine that will affect the running child. Some limited
        support might be implemented one day, do chdir() and %ENV changes
        can be made.

    signals
        Win32 does not fully support signals. signal() is likely to cause
        errors unless sending a signal that Perl emulates, and "kill_kill()"
        is immediately fatal (there is no grace period).

    helper processes
        IPC::Run uses helper processes, one per redirected file, to adapt
        between the anonymous pipe connected to the child and the TCP socket
        connected to the parent. This is a waste of resources and will
        change in the future to either use threads (instead of helper
        processes) or a WaitForMultipleObjects call (instead of select).
        Please contact me if you can help with the WaitForMultipleObjects()
        approach; I haven't figured out how to get at it without C code.

    shutdown pause
        There seems to be a pause of up to 1 second between when a child
        program exits and the corresponding sockets indicate that they are
        closed in the parent. Not sure why.

    binmode
        binmode is not supported yet. The underpinnings are implemented,
        just ask if you need it.

    IPC::Run::IO
        IPC::Run::IO objects can be used on Unix to read or write arbitrary
        files. On Win32, they will need to use the same helper processes to
        adapt from non-select()able filehandles to select()able ones (or
        perhaps WaitForMultipleObjects() will work with them, not sure).

    startup race conditions
        There seems to be an occasional race condition between child process
        startup and pipe closings. It seems like if the child is not fully
        created by the time CreateProcess returns and we close the TCP
        socket being handed to it, the parent socket can also get closed.
        This is seen with the Win32 pumper applications, not the "real"
        child process being spawned.

        I assume this is because the kernel hasn't gotten around to
        incrementing the reference count on the child's end (since the child
        was slow in starting), so the parent's closing of the child end
        causes the socket to be closed, thus closing the parent socket.

        Being a race condition, it's hard to reproduce, but I encountered it
        while testing this code on a drive share to a samba box. In this
        case, it takes t/run.t a long time to spawn it's chile processes
        (the parent hangs in the first select for several seconds until the
        child emits any debugging output).

        I have not seen it on local drives, and can't reproduce it at will,
        unfortunately. The symptom is a "bad file descriptor in select()"
        error, and, by turning on debugging, it's possible to see that
        select() is being called on a no longer open file descriptor that
        was returned from the _socket() routine in Win32Helper. There's a
        new confess() that checks for this ("PARENT_HANDLE no longer open"),
        but I haven't been able to reproduce it (typically).

LIMITATIONS
    On Unix, requires a system that supports "waitpid( $pid, WNOHANG )" so
    it can tell if a child process is still running.

    PTYs don't seem to be non-blocking on some versions of Solaris. Here's a
    test script contributed by Borislav Deianov <borislav@ensim.com> to see
    if you have the problem. If it dies, you have the problem.

       #!/usr/bin/perl

       use IPC::Run qw(run);
       use Fcntl;
       use IO::Pty;

       sub makecmd {
           return ['perl', '-e', 
                   '<STDIN>, print "\n" x '.$_[0].'; while(<STDIN>){last if /end/}'];
       }

       #pipe R, W;
       #fcntl(W, F_SETFL, O_NONBLOCK);
       #while (syswrite(W, "\n", 1)) { $pipebuf++ };
       #print "pipe buffer size is $pipebuf\n";
       my $pipebuf=4096;
       my $in = "\n" x ($pipebuf * 2) . "end\n";
       my $out;

       $SIG{ALRM} = sub { die "Never completed!\n" };

       print "reading from scalar via pipe...";
       alarm( 2 );
       run(makecmd($pipebuf * 2), '<', \$in, '>', \$out);
       alarm( 0 );
       print "done\n";

       print "reading from code via pipe... ";
       alarm( 2 );
       run(makecmd($pipebuf * 3), '<', sub { $t = $in; undef $in; $t}, '>', \$out);
       alarm( 0 );
       print "done\n";

       $pty = IO::Pty->new();
       $pty->blocking(0);
       $slave = $pty->slave();
       while ($pty->syswrite("\n", 1)) { $ptybuf++ };
       print "pty buffer size is $ptybuf\n";
       $in = "\n" x ($ptybuf * 3) . "end\n";

       print "reading via pty... ";
       alarm( 2 );
       run(makecmd($ptybuf * 3), '<pty<', \$in, '>', \$out);
       alarm(0);
       print "done\n";

    No support for ';', '&&', '||', '{ ... }', etc: use perl's, since run()
    returns TRUE when the command exits with a 0 result code.

    Does not provide shell-like string interpolation.

    No support for "cd", "setenv", or "export": do these in an init() sub

       run(
          \cmd,
             ...
             init => sub {
                chdir $dir or die $!;
                $ENV{FOO}='BAR'
             }
       );

    Timeout calculation does not allow absolute times, or specification of
    days, months, etc.

    WARNING: Function coprocesses ("run \&foo, ...") suffer from two
    limitations. The first is that it is difficult to close all filehandles
    the child inherits from the parent, since there is no way to scan all
    open FILEHANDLEs in Perl and it both painful and a bit dangerous to
    close all open file descriptors with "POSIX::close()". Painful because
    we can't tell which fds are open at the POSIX level, either, so we'd
    have to scan all possible fds and close any that we don't want open
    (normally "exec()" closes any non-inheritable but we don't "exec()" for
    &sub processes.

    The second problem is that Perl's DESTROY subs and other on-exit cleanup
    gets run in the child process. If objects are instantiated in the parent
    before the child is forked, the the DESTROY will get run once in the
    parent and once in the child. When coprocess subs exit, POSIX::exit is
    called to work around this, but it means that objects that are still
    referred to at that time are not cleaned up. So setting package vars or
    closure vars to point to objects that rely on DESTROY to affect things
    outside the process (files, etc), will lead to bugs.

    I goofed on the syntax: "<pipe" vs. "<pty<" and ">filename" are both
    oddities.

TODO
    Allow one harness to "adopt" another:
           $new_h = harness \@cmd2;
           $h->adopt( $new_h );

    Close all filehandles not explicitly marked to stay open.
        The problem with this one is that there's no good way to scan all
        open FILEHANDLEs in Perl, yet you don't want child processes
        inheriting handles willy-nilly.

INSPIRATION
    Well, select() and waitpid() badly needed wrapping, and open3() isn't
    open-minded enough for me.

    The shell-like API inspired by a message Russ Allbery sent to
    perl5-porters, which included:

       I've thought for some time that it would be
       nice to have a module that could handle full Bourne shell pipe syntax
       internally, with fork and exec, without ever invoking a shell.  Something
       that you could give things like:

       pipeopen (PIPE, [ qw/cat file/ ], '|', [ 'analyze', @args ], '>&3');

    Message ylln51p2b6.fsf@windlord.stanford.edu, on 2000/02/04.

SUPPORT
    Bugs should always be submitted via the CPAN bug tracker

    <http://rt.cpan.org/NoAuth/ReportBug.html?Queue=IPC-Run>

    For other issues, contact the maintainer (the first listed author)

AUTHORS
    Adam Kennedy <adamk@cpan.org>

    Barrie Slaymaker <barries@slaysys.com>

COPYRIGHT
    Some parts copyright 2008 - 2009 Adam Kennedy.

    Copyright 1999 Barrie Slaymaker.

    You may distribute under the terms of either the GNU General Public
    License or the Artistic License, as specified in the README file.