Re: [Xen-devel] [PATCH v10 5/6] passthrough/io: don't migrate pirq when it is delivered through VT-d PI

[Chao Gao <chao.gao@xxxxxxxxx>]

On Mon, Mar 20, 2017 at 06:50:37AM -0600, Jan Beulich wrote:
>>>> On 20.03.17 at 06:22, <chao.gao@xxxxxxxxx> wrote:
>> On Mon, Mar 20, 2017 at 04:26:10AM -0600, Jan Beulich wrote:
>>>>>> On 20.03.17 at 03:38, <chao.gao@xxxxxxxxx> wrote:
>>>> On Mon, Mar 20, 2017 at 03:18:18AM -0600, Jan Beulich wrote:
>>>>>>>>          spin_unlock(&d->event_lock);
>>>>>>>>          if ( dest_vcpu_id >= 0 )
>>>>>>>>              hvm_migrate_pirqs(d->vcpu[dest_vcpu_id]);
>>>
>>>... right after the lock release.
>> 
>> @via_pi is also consumed during vCPU migration.
>
>But the event lock isn't being held around the checking of the
>field, so putting the setting of the field under lock is of no use.
>
>> I just think the event_lock protects R/W operations on struct hvm_pirq_dpci.
>> To prohibit potential race(we can't use VT-d PI in 1st binding, but we can 
>> use
>> in the 2nd binding. But somehow the first update to via_pi overrides the 
>> second one),
>> and don't complicate the fields event_lock protects,
>> I'm inclined to put it in event_lock-ed region as long as it doesn't 
>> introduce other issues.
>
>I certainly don't object to properly synchronizing things here,
>but then both producing and consuming side need to hold
>respective locks. Otherwise the best you can hope for is to
>reduce timing windows; you won't eliminate them though.

You are right. I am convinced.
Thanks
Chao

>
>Jan
>

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