Subject: Re: [ast-developers] ksh file io race condition
--------
> Hi,
>
> we've found that ksh can fail to read file it should create with previous
> comman
> d. This occurs even when the file
> creation is inside parentheses and the access occurs outside the parentheses.
>
>
> Reproducer:
>
> #!/bin/ksh
>
> n=0
> TEErc=Trcfile
> while [ $n -lt 1000 ]
> do
> echo $n
> >$TEErc
> (ls /tmp > /dev/null 2>&1; echo $? >$TEErc) | tee -a junk
> rc=$(<$TEErc)
> if [ -z "$rc" ]
> then
> echo hit the bug, rc is empty
> ls -l $TEErc
> sync
> sleep 1
> echo after sleep and sync
> ls -l $TEErc
> exit
> fi
> rm $TEErc
> n=$(expr $n + 1)
> done
> rm junk
First of all, let me explain what causes this.
When you run a pipeline, the shell only waits for the last process
to complete unless the pipefail option is on.
Thus, as soon as tee completes, the pipeline is complete and the command
rc=$(<$TEErc)
can run.
Now, the
echo $? >$TEErc
causes the writer of the pipeline to close since it is no longer needed.
This in turn cause the redirection causes tee to complete and there is now
a race between the parent shell running
rc=$(<$TEErc)
and the child contining after to redirection to run
echo $?
If you put a sleep before
rc=$(<$TEErc)
you will see the exit status in rc.
Now, the standard does allow (but does not require) the shell to wait
for all processes (the way pipefail) does with ksh93.
For example
sleep 2 | :
can either complete immediately or in two seconds as far as the standard
is concerned.
What I have done for the next release of ksh93u is to modify the shell so
that it will not close the write end of the pipe when file descriptor 1 is
redirected on the left hand side. This removes the race conditon in this case.
However, using set -o pipefail is a better solution but that is not
in the standard.
I will consider changing the default behavior of pipefail to be enabled in
ksh93v. I would like to know if any existing script break.
David Korn
[email protected]
_______________________________________________
ast-developers mailing list
[email protected]
https://mailman.research.att.com/mailman/listinfo/ast-developers
