[PATCH] Re: Testsuite races

[Jeff Law <law at redhat dot com>]

On 08/15/2012 12:55 PM, Roland McGrath wrote:
> > It looks like it's just tst-codeset, tst-gettext3 and tst-gettext5 which
> > conflict on ${objpfx}domaindir/de_DE/LC_MESSAGES/codeset.mo
> >
> > Changing the .sh file to look like:
> >
> > (
> >     flock 200
> >     [ generate data & run test ]
> > ) 200>${objpfx}lockfile
> >
> > In each of the three tests seems to do the trick.  Similarly creating a
> > dependency in the Makefile where tst-gettext3.out depends on
> > tst-codeset.out and tst-gettext5.out depends on tst-gettext3.out works
> > just as well.
>
> The right answer is to have makefile rules to generate .../codeset.mo
> and have each of those .out targets depend on the file directly.
>
> With that, it looks like these tests won't need shell script wrappers at
> all, just the vanilla pattern rule should dtrt.
OK.  Here's a patch which implements that approach.  This is my first 
twiddle to glibc's build machinery, so if I've goof'd something it's 
purely from inexperience.

I've confirmed that after these changes we're using the vanilla pattern 
rule with a proper environment passed to the test program.  I've also 
run the intl/ tests repeatedly in a make -j loop for considerable time 
periods without seeing spurious failures.

Attachment: patch
Description: Text document