On Mon, Aug 13, 2012 at 04:40:03PM +0200, Jakub Hrozek wrote:
> On Mon, Aug 13, 2012 at 01:30:21PM +0200, Sumit Bose wrote:
> > On Fri, Aug 10, 2012 at 06:40:35PM +0200, Jakub Hrozek wrote:
> > >
https://fedorahosted.org/sssd/ticket/1460
> > >
> > > Please see the commit. I'm wondering if there is still a (small) race
> > > condition between the call to pthread_cleanup_pop() and unlocking the
> > > mutex. Would it be better to i.e. always call the cleanup handler with
> > > pthread_cleanup_pop(1) and disconnect from the fd based on some other
> > > condition?
> >
> > this patch works as expected for me on F17. I have some issues on F16,
> > but this might have other reason I still try to investigate. While
> > reading a bit about pthread I can across robust mutexes
> >
http://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_mutexat...
> > I wonder if we can better avoid races with those and maybe even make the
> > implementation cleaner? E.g. if we get EOWNERDEAD we can call
> > sss_cli_close_socket(); pthread_mutex_consistent(); run the request
> > which finally will call pthread_mutex_unlock().
> >
>
> As discussed briefly on IRC, I agree. I like any approach that removes
> the pthread_cleanup_push/pop hack.
>
> > There are also pthread_mutexattr_getrobust_np() and
> > pthread_mutex_consistent_np() in glibc. I do not have an idea how they
> > differ and which one should be preferred. But the *_np version only need
> > __USE_GNU while the others need __USE_XOPEN2K8.
> >
> > bye,
> > Sumit
>
> RHEL5 only contains the _np versions, but they still seem to work. If
> the approach is deemed OK, I'll detect the available versions during
> configure.
>
> Traditionally the _np suffix means "non-portable GNU extension".
>
> Please see the attached patch. Would that work for you? If so, I'll
> extend it to work on all supported RHEL releases.
Sumit nacked the patch on IRC. I was missing the disconnect logic when
lock returns EOWNERDEAD. I also moved the mutexattr to sss_mt_init, I
don't think there is any reason to keep it separate. It doesn't have to
live as long as the mutex itself.
This patch works as expected on F17, so ACK.
I still have some strange behaviour on F16 and strangely in the case
where the fast cache files in the /var/lib/sss/mc directory are
available. Here the reproducer get stuck in a futex() call with or
without the patch. I looks that glibc itself also uses a mutex here
which is neither robust nor unlocked if the tread holding it is killed.
I would be nice if someone with a F16 system can verify this.
bye,
Sumit