12.4 Two-Way Communications with Another Process

It is often useful to be able to send data to a separate program for processing and then read the result. This can always be done with temporary files:

# Write the data for processing
tempfile = ("mydata." PROCINFO["pid"])
while (not done with data)
    print data | ("subprogram > " tempfile)
close("subprogram > " tempfile)

# Read the results, remove tempfile when done
while ((getline newdata < tempfile) > 0)
    process newdata appropriately
close(tempfile)
system("rm " tempfile)

This works, but not elegantly. Among other things, it requires that the program be run in a directory that cannot be shared among users; for example, /tmp will not do, as another user might happen to be using a temporary file with the same name.87

However, with gawk, it is possible to open a two-way pipe to another process. The second process is termed a coprocess, as it runs in parallel with gawk. The two-way connection is created using the ‘|&’ operator (borrowed from the Korn shell, ksh):88

do {
    print data |& "subprogram"
    "subprogram" |& getline results
} while (data left to process)
close("subprogram")

The first time an I/O operation is executed using the ‘|&’ operator, gawk creates a two-way pipeline to a child process that runs the other program. Output created with print or printf is written to the program’s standard input, and output from the program’s standard output can be read by the gawk program using getline. As is the case with processes started by ‘|’, the subprogram can be any program, or pipeline of programs, that can be started by the shell.

There are some cautionary items to be aware of:

It is possible to close just one end of the two-way pipe to a coprocess, by supplying a second argument to the close() function of either "to" or "from" (see Closing Input and Output Redirections). These strings tell gawk to close the end of the pipe that sends data to the coprocess or the end that reads from it, respectively.

This is particularly necessary in order to use the system sort utility as part of a coprocess; sort must read all of its input data before it can produce any output. The sort program does not receive an end-of-file indication until gawk closes the write end of the pipe.

When you have finished writing data to the sort utility, you can close the "to" end of the pipe, and then start reading sorted data via getline. For example:

BEGIN {
    command = "LC_ALL=C sort"
    n = split("abcdefghijklmnopqrstuvwxyz", a, "")

    for (i = n; i > 0; i--)
        print a[i] |& command
    close(command, "to")

    while ((command |& getline line) > 0)
        print "got", line
    close(command)
}

This program writes the letters of the alphabet in reverse order, one per line, down the two-way pipe to sort. It then closes the write end of the pipe, so that sort receives an end-of-file indication. This causes sort to sort the data and write the sorted data back to the gawk program. Once all of the data has been read, gawk terminates the coprocess and exits.

As a side note, the assignment ‘LC_ALL=C’ in the sort command ensures traditional Unix (ASCII) sorting from sort. This is not strictly necessary here, but it’s good to know how to do this.

Be careful when closing the "from" end of a two-way pipe; in this case gawk waits for the child process to exit, which may cause your program to hang. (Thus, this particular feature is of much less use in practice than being able to close the "to" end.)

CAUTION: Normally, it is a fatal error to write to the "to" end of a two-way pipe which has been closed, and it is also a fatal error to read from the "from" end of a two-way pipe that has been closed.

You may set PROCINFO["command", "NONFATAL"] to make such operations become nonfatal. If you do so, you then need to check ERRNO after each print, printf, or getline. See Enabling Nonfatal Output, for more information.

You may also use pseudo-ttys (ptys) for two-way communication instead of pipes, if your system supports them. This is done on a per-command basis, by setting a special element in the PROCINFO array (see Built-in Variables That Convey Information), like so:

command = "sort -nr"           # command, save in convenience variable
PROCINFO[command, "pty"] = 1   # update PROCINFO
print ... |& command           # start two-way pipe
...

If your system does not have ptys, or if all the system’s ptys are in use, gawk automatically falls back to using regular pipes.

Using ptys usually avoids the buffer deadlock issues described earlier, at some loss in performance. This is because the tty driver buffers and sends data line-by-line. On systems with the stdbuf (part of the GNU Coreutils package), you can use that program instead of ptys.

Note also that ptys are not fully transparent. Certain binary control codes, such Ctrl-d for end-of-file, are interpreted by the tty driver and not passed through.

CAUTION: Finally, coprocesses open up the possibility of deadlock between gawk and the program running in the coprocess. This can occur if you send “too much” data to the coprocess before reading any back; each process is blocked writing data with no one available to read what they’ve already written. There is no workaround for deadlock; careful programming and knowledge of the behavior of the coprocess are required.

The following example, due to Andrew Schorr, demonstrates how using ptys can help deal with buffering deadlocks.

Suppose gawk were unable to add numbers. You could use a coprocess to do it. Here’s an exceedingly simple program written for that purpose:

$ cat add.c
#include <stdio.h>

int
main(void)
{
    int x, y;
    while (scanf("%d %d", & x, & y) == 2)
        printf("%d\n", x + y);
    return 0;
}
$ cc -O add.c -o add      Compile the program

You could then write an exceedingly simple gawk program to add numbers by passing them to the coprocess:

$ echo 1 2 |
> gawk -v cmd=./add '{ print |& cmd; cmd |& getline x; print x }'

And it would deadlock, because add.c fails to call ‘setlinebuf(stdout)’. The add program freezes.

Now try instead:

$ echo 1 2 |
> gawk -v cmd=add 'BEGIN { PROCINFO[cmd, "pty"] = 1 }
>                  { print |& cmd; cmd |& getline x; print x }'
-| 3

By using a pty, gawk fools the standard I/O library into thinking it has an interactive session, so it defaults to line buffering. And now, magically, it works!


Footnotes

(87)

Michael Brennan suggests the use of rand() to generate unique file names. This is a valid point; nevertheless, temporary files remain more difficult to use than two-way pipes.

(88)

This is very different from the same operator in the C shell and in Bash.