Re: [libvirt] Strange periodic failures of the commandtest case
On 02/14/2011 06:45 AM, Daniel P. Berrange wrote:
On Mon, Feb 14, 2011 at 12:32:18PM +0000, Daniel P. Berrange wrote:
> I'm getting periodic failures of the 'commandtest' case where the diff
> is
>
>> DAEMON:no
> --
> < DAEMON:yes
>
> For test cases 'test3' through to 'test15', except 'test'4.
Tests 1,
> 2, 4, 16, 17, 18 are all unaffected. I can never reproduce it when
> I run just that one test case manually, but the automated builds
> I'm doing hit it > 50% of the time so it is clearly some kind of race
> condition. The line of code that's getting confused is this
>
> fprintf(log, "DAEMON:%s\n", getpgrp() == getsid(0) ? "yes" :
"no");
>
> but I'm not clear how this can be going wrong. Ideas ?
Hmm, this seems to occur when you run the build under cron. Under an
interactive shell getpgrp() != getsid() by default. Under non-interactive
shell, then getpgrp() == getsid() even if we didn't ask to daemonize.
So this approach to detecting whether we're daemonized isn't reliable.
It's more reliable than the previous test, which was getppid() == 1,
since that was even more racy depending on whether the intermediate
process had exited yet to result in the grandchild being reparented to
init. But yes, I think you correctly analyzed the problem of running
from cron rather than an interactive shell.
Perhaps the 'commandtest' program itself needs to call setsid()+setpgid()
before running the child processes ?
That sounds like a reasonable solution. Do you want to tackle it, or
shall I?
I'm also seeing that test fail on cygwin; probably due to .exe suffixes.
But I haven't yet had time to investigate how easy that would be to
patch, yet.
--
Eric Blake eblake(a)redhat.com +1-801-349-2682
Libvirt virtualization library http://libvirt.org
Attachments:
- signature.asc (application/pgp-signature — 619 bytes)