Re: [Xen-devel] [PATCH 1/3] VMX: Properly adjuest the status of pi descriptor

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 23.05.16 at 11:21, <feng.wu@xxxxxxxxx> wrote:
>> From: Jan Beulich [mailto:JBeulich@xxxxxxxx]
>> Sent: Monday, May 23, 2016 5:04 PM
>> >>> On 23.05.16 at 09:16, <feng.wu@xxxxxxxxx> wrote:
>> >> From: Tian, Kevin
>> >> Sent: Monday, May 23, 2016 2:52 PM
>> >> Somehow I'm thinking whether we really need such dynamic
>> >> callback changes in a SMP environment. Is it safer to introduce
>> >> some per-domain flag to indicate above scenario so each callback
>> >> can return early with negligible cost so no need to reset them once
>> >> registered?
>> >
>> > Yes, I agree with you. As you can see, the current issue is kind of because
>> > of the dynamically assigning/deassigning the pi hooks, and there are
>> > lots of concern cases by using it. I think it worth try what you mentioned
>> > above. However, this method was sugguested by Jan before, so I'd like
>> > to know what is Jan's option about this.
>> 
>> Avoiding the indirect calls would certainly be nice. Apart from
>> reading the function pointer into a local variable, use of which is
>> separated from the read by a barrier() (to address the NULL
>> deref concern), did you consider zapping the pointers in an RCU
>> callback instead of right when being requested?
> 
> Is there any example for the RCU callback in current Xen code,
> so I can first have some investigation on it.

Did you try to find any before asking, e.g. by grep-ing for "rcu",
or (more precisely) "call_rcu"?

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel