IPC::Run - Online Manual Page Of Unix/Linux

  Command: man perldoc info search(apropos)

WebSearch:
Our Recommended Sites: Full-Featured Editor
 

IPC::Run(3)                    User Contributed Perl Documentation                    IPC::Run(3)



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>', \$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,  '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', \*OUT,
                   '2>pipe', \*ERR
                or die "cat returned $?" ;
             print IN "some input\n" ;
             close IN ;
             print ,  ;
             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, \<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 sub-
       command'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 pro-
       cesses' result codes, for instance.

       Spawning synchronization, child exception propagation

       start() pauses the parent until the child executes the command or CODE reference and prop-
       agates any exceptions thrown (including exec() failure) back to the parent.  This has sev-
       eral 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 com-
       mands 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 redi-
       rected-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 syn-
       tax 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 high-
       est 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 redi-
       rected-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 com-
       pletion 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 opti-
           mization 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>').

       ?   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 =~ /^\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>').

       ?   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 con-
           nected 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 sub-
       process 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>', \$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

          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.

       '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]pump() to feed incremental bits of input to a copro-
           cess.  See "Coprocesses" below for more information.

           The IPC::Run does nothing further with it: you are respon-
           sible 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 redi-
           rection 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 call-
           back:

           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  ;
              print  ;
              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 descrip-
           tor number m

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

              run \@cmd, \undef ;
              run \@cmd, '<&-' ;
              run \@cmd, ', >&, &>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 file-
       handle 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 pars-
       ing 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 "", 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 granu-
           larity, 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 build-
           ing 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 informa-
           tion.

           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 pro-
           ceed 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 proces-
           sor 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 sig-
           nal 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 sys-
           tem 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 sys-
           tem()).

           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 chil-
           dren.  Call "<$h-"kill_kill>> in this case if you care.  This differs from the behav-
           ior 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 sig-
           nal 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 car-
           riage 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 pro-
           grams 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 wor-
           ried 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, sugges-
       tions 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 sim-
           ple 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>
           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 inte-
           ger 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 send-
           ing 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 anony-
           mous 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 con-
           tact 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 file-
           handles 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 clos-
           ing 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 sec-
           onds 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 debug-
           ging, 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 con-
       tributed by Borislav Deianov <> 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',
                      ', print "\n" x '.$_[0].'; while(){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), '', \$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 danger-
       ous 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: "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 , on 2000/02/04.

AUTHOR
       Barrie Slaymaker <>, with numerous suggestions by p5p.



perl v5.8.8                                 2006-05-10                                IPC::Run(3)