On Wed, 2018-04-11 at 13:25 +0100, George Dunlap wrote:
> Some compile-tested-only sketches of what I'm talking about. Let me
> know what you think.
>
So, patches 1 and 2 of this series solves what I think was one of the
nastiest races I've ever had to chase in the scheduler. :-)
Having figured out what the exact root cause of the race itself is,
this is the _proper_ fix, as it puts setting of VPF_migrate and
SCHED_op(sleep) inside the same critical section, which is what closes
the race window.
I'd like to argue for this series to be considered a bugfix, and
included in 4.11 (and backported as far as possible, which has been
already proved to be feasible, e.g., until 4.7).
The alternative would be to come up with something else which kind of
works around the race, within sched_credit.c... But I don't really see
a reason for doing that. Code-wise, it may probably be a bit more self-
contained, but it's not like this series is that spread/intrusive in
the first place.
And the net effect would be basically the same. I.e., in both cases, we
need to change what happens when vcpu_migrate() is called, and I don't
see much difference between doing that by changing vcpu_migrate()
itself, or by changing how Credit react to vcpu_migrate() being called
(especially considering that Credit is the default scheduler).
And therefore, between a proper fix and a workaround, which have
similar impact and effects, I think we should go for the former. :-)
Regards,
Dario
--
<> (Raistlin Majere)
-
Dario Faggioli, Ph.D, http://about.me/dario.faggioli
Software Engineer @ SUSE https://www.suse.com/
signature.asc
Description: This is a digitally signed message part
___
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel
