IFS and read

I find myself continually underestimating the Korn shell. I always thought that Bash was neat because we could read into an array with read –a, but of course we can do the same in ksh with read –A. The effect is to split up the input into fields around $IFS. But did you know we can set IFS just for a read statement? Look carefully:

IFS=','
while IFS=':' read -A line
do
end=$(( ${#line[*]} - 1 ))
if [[ ${line[$end]} != /sbin/nologin ]]
then
echo "${line[*]}"
fi
done < /etc/passwd

(display all the lines in /etc/passwd whose last field is not /sbin/nologin, changing the field delimiter from colon to comma).

The first IFS=',' sets IFS for the echo expansion and is not overridden by the IFS on the read line. That IFS only applies for the read, nothing else.

By the way, did you know that the expansion (including $*) uses the first character from IFS, so order matters? So IFS=',.:' gives different results to IFS=':,.'.

All this applies to the standard POSIX shell, as well as Korn shell and Bash, and I was alerted to these features by Classic Shell Scripting (see a previous post).

Finally, a significant bug in pdksh has been fixed in the version with Fedora core 4 (1993-12-28 q). In early versions a pipe generated a child process on each side, so piping into a while loop was not useful, since the while loop ran in a different process to the rest of the script, and so variables could not be saved. Now it all works, consider:

#!/bin/ksh

ps -ef | while read uid pid ppid c stime tty time cmd
do
if [[ $cmd == *xinetd* ]]
then
break
fi
done

echo "Pid of xinetd is: $pid"

Unfortunately this will still not work in Bash, for the same reason as it did not work in the old pdksh. In Bash we can use process substitution instead, and while we are about it we may as well use an ERE (because we can):

while read uid pid ppid c stime tty time cmd
do
if [[ $cmd =~ '^xinetd +' ]]
then
break
fi
done < <(ps -ef)

echo "Pid of xinetd is: $pid"

The man pages for the new pdksh say process substitution works in Korn shell as well, and I am told this works on Solaris, but it does not appear to work on Linux.