Re: Issue #5876 : assertion failure in rbd_img_obj_callback()

[Olivier Bonvalet <ceph.list@daevel.fr>]

Le jeudi 27 mars 2014 à 10:45 +0200, Ilya Dryomov a écrit :
> On Thu, Mar 27, 2014 at 9:48 AM, Olivier Bonvalet <ceph.list@daevel.fr> wrote:
> > Le mercredi 26 mars 2014 à 15:58 -0500, Alex Elder a écrit :
> >> Olivier reports that with the simple patch I provided
> >> (which changed a "<" to a "!=" and removed an assertion)
> >> he is running successfully.
> >>
> >> To me this is fantastic news, and you can see I posted
> >> a patch with the fix.
> >>
> >> There remains a race condition though, one which I described
> >> in a separate message earlier today.  I don't think it will
> >> prove to be a problem in practice, but I agreed to work on
> >> a fix to ensure the race condition is eliminated.  It will
> >> require some work with reference counting image and object
> >> requests.
> >>
> >> The fix won't be coming today.  But I aim to provide it
> >> in a matter of several days.
> >>
> >>                                       -Alex
> >>
> >
> > One question from one of my customers : why am I the only one to
> > complain about that problem ?
> > I know that Ceph users often use qemu/librbd instead of kernel client,
> > but what is the trigger of those «race condition» ? Having "multiple
> > requests" per RBD image ? It should be a normal use, no ?
> >
> > If someone can help me give an explanation, thanks :)
> 
> We've had a couple more, similar reports in the last few months.
> However you are the first reporter who was able to trigger this race
> often enough to track it down.  This race condition (read: bug) is
> kernel client specific, qemu/librbd is unaffected.  Having an rbd
> request that spans multiple RADOS objects and therefore results in
> multiple object requests is normal use, it's just that particular piece
> of code turned out to be prone to a subtle race.  You have to keep in
> mind that races are all about timing and relative order of events, so
> simply issuing a multi-object rbd request is not enough to trigger it,
> stars have to align too ;)
> 
> Thanks,
> 
>                 Ilya
> 

Great, thanks !

Olivier

--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html