The coproc keyword

 coproc [NAME] command [redirections]

Bash 4.0 introduced coprocesses, a feature certainly familiar to ksh users. The coproc keyword starts a command as a background job, setting up pipes connected to both its stdin and stdout so that you can interact with it bidirectionally. Optionally, the co-process can have a name NAME. If NAME is given, the command that follows must be a compound command. If no NAME is given, then the command can be either simple or compound.

The process ID of the shell spawned to execute the coprocess is available through the value of the variable named by NAME followed by a _PID suffix. For example, the variable name used to store the PID of a coproc started with no NAME given would be COPROC_PID (because COPROC is the default NAME). The wait builtin command may be used to wait for the coprocess to terminate. Additionally, coprocesses may be manipulated through their jobspec.

The return status of a coprocess is the exit status of its command.

The optional redirections are applied after the pipes have been set up. Some examples:

# redirecting stderr in the pipe
$ coproc { ls thisfiledoesntexist; read; } 2>&1
[2] 23084
$ IFS= read -ru ${COPROC[0]} x; printf '%s\n' "$x"
ls: cannot access thisfiledoesntexist: No such file or directory

#let the output of the coprocess go to stdout
$ { coproc mycoproc { awk '{print "foo" $0;fflush()}'; } >&3; } 3>&1
[2] 23092
$ echo bar >&${mycoproc[1]}
$ foobar

Here we need to save the previous file descriptor of stdout, because by the time we redirect the fds of the coprocess, stdout has already been redirected to the pipe.

The traditional Ksh workaround to avoid the subshell when doing command | while read is to use a coprocess. Unfortunately, Bash's behavior differs.

In Ksh you would do:

# ksh93 or mksh/pdksh derivatives
ls |& # start a coprocess
while IFS= read -rp file; do print -r -- "$file"; done # read its output

In bash:

#DOESN'T WORK
$ coproc ls
[1] 23232
$ while IFS= read -ru ${COPROC[0]} line; do printf '%s\n' "$line"; done
bash: read: line: invalid file descriptor specification
[1]+  Done                    coproc COPROC ls

By the time we start reading from the output of the coprocess, the file descriptor has been closed.

See this FAQ entry on Greg's wiki for other pipeline subshell workarounds.

In the first example, we GNU awk's fflush() command. As always, when you use pipes the I/O operations are buffered. Let's see what happens with sed:

$ coproc sed s/^/foo/
[1] 22981
$ echo bar >&${COPROC[1]}
$ read -t 3 -ru ${COPROC[0]} _; (( $? > 127 )) && echo "nothing read"
nothing read

Even though this example is the same as the first awk example, the read doesn't return because the output is waiting in a buffer.

See this faq entry on Greg's wiki for some workarounds and more information on buffering issues.

A coprocess' file descriptors are accessible only to the process from which the coproc was started. They are not inherited by subshells.

Here is a not-so-meaningful illustration. Suppose we want to continuously read the output of a coprocess and echo the result:

#NOT WORKING
$ coproc awk '{print "foo" $0;fflush()}'
[2] 23100
$ while IFS= read -ru ${COPROC[0]} x; do printf '%s\n' "$x"; done &
[3] 23104
bash: line 243: read: 61: invalid file descriptor: Bad file descriptor

This fails because the file descriptors created by the parent are not available to the subshell created by &.

A possible workaround:

#WARNING: for illustration purpose ONLY
# this is not the way to make the coprocess print its output
# to stdout, see the redirections above.
$ coproc awk '{print "foo" $0;fflush()}'
[2] 23109
$ exec 3<&${COPROC[0]}
$ while IFS= read -ru 3 x; do printf '%s\n' "$x"; done &
[3] 23110
$ echo bar >&${COPROC[1]}
$ foobar

Here, fd 3 is inherited.

Unlike ksh, Bash doesn't have true anonymous coprocesses. Instead, Bash assigns FDs to a default array named COPROC if no NAME is supplied. Here's an example:

$ coproc awk '{print "foo" $0;fflush()}'
[1] 22978

This command starts in the background, and coproc returns immediately. Two new file descriptors are now available via the COPROC array. We can send data to our command:

$ echo bar >&${COPROC[1]}

And then read its output:

$ IFS= read -ru ${COPROC[0]} x; printf '%s\n' "$x"
foobar

When we don't need our command anymore, we can kill it via its pid:

$ kill $COPROC_PID
$
[1]+  Terminated              coproc COPROC awk '{print "foo" $0;fflush()}'

Using a named coprocess is simple. We just need a compound command (like when defining a function), and the resulting FDs will be assigned to the indexed array NAME we supply instead.

$ coproc mycoproc { awk '{print "foo" $0;fflush()}' ;}
[1] 23058
$ echo bar >&${mycoproc[1]}
$ IFS= read -ru ${mycoproc[0]} x; printf '%s\n' "$x"
foobar
$ kill $mycoproc_PID
$
[1]+  Terminated              coproc mycoproc { awk '{print "foo" $0;fflush()}'; }

#!/bin/bash
# we start tee in the background
# redirecting its output to the stdout of the script
{ coproc tee { tee logfile ;} >&3 ;} 3>&1 
# we redirect stding and stdout of the script to our coprocess
exec >&${tee[1]} 2>&1

• The coproc keyword is not specified by POSIX(R)
• The coproc keyword appeared in Bash version 4.0-alpha
• The -p option to Bash's print loadable is a NOOP and not connected to Bash coprocesses in any way. It is only recognized as an option for ksh compatibility, and has no effect.
• The -p option to Bash's read builtin conflicts with that of all kshes and zsh. The equivalent in those shells is to add a \?prompt suffix to the first variable name argument to read. i.e., if the first variable name given contains a ? character, the remainder of the argument is used as the prompt string. Since this feature is pointless and redundant, I suggest not using it in either shell. Simply precede the read command with a printf %s prompt >&2.

ksh93, mksh, zsh, and Bash all support something called "coprocesses" which all do approximately the same thing. ksh93 and mksh have virtually identical syntax and semantics for coprocs. A list operator: |& is added to the language which runs the preceding pipeline as a coprocess (This is another reason not to use the special |& pipe operator in Bash – its syntax is conflicting). The -p option to the read and print builtins can then be used to read and write to the pipe of the coprocess (whose FD isn't yet known). Special redirects are added to move the last spawned coprocess to a different FD: <&p and >&p, at which point it can be accessed at the new FD using ordinary redirection, and another coprocess may then be started, again using |&.

zsh coprocesses are very similar to ksh except in the way they are started. zsh adds the shell reserved word coproc to the pipeline syntax (similar to the way Bash's time keyword works), so that the pipeline that follows is started as a coproc. The coproc's input and output FDs can then be accessed and moved using the same read/print -p and redirects used by the ksh shells.

It is unfortunate that Bash chose to go against existing practice in their coproc implementation, especially considering it was the last of the major shells to incorporate this feature. However, Bash's method accomplishes the same without requiring nearly as much additional syntax. The coproc keyword is easy enough to wrap in a function such that it takes Bash code as an ordinary argument and/or stdin like eval. Coprocess functionality in other shells can be similarly wrapped to create a COPROC array automatically.

The title says it all, complain to the bug-bash mailing list if you want more. See http://lists.gnu.org/archive/html/bug-bash/2011-04/msg00056.html for more details

The ability to use multiple coprocesses in Bash is considered "experimental". Bash will throw an error if you attempt to start more than one. This may be overridden at compile-time with the MULTIPLE_COPROCS option. However, at this time there are still issues – see the above mailing list discussion.

• Anthony Thyssen's Coprocess Hints - excellent summary of everything around the topic