Re: [Xen-devel] vcpu_block() and do_poll() question

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 30.04.19 at 12:31, <julien.grall@xxxxxxx> wrote:

> 
> On 30/04/2019 11:27, Jan Beulich wrote:
>>>>> On 30.04.19 at 11:13, <jgross@xxxxxxxx> wrote:
>>> In xen/common/schedule.c there is a weird "#ifndef CONFIG_X86" in
>>> do_poll().
>>>
>>> It was introduced way before anyone would think about ARM by commit
>>> ef4c6b079cc55e (I couldn't find any xen-devel mail related to that
>>> commit), so I guess it is related to IA64?
>> 
>> Quite certainly, yes (or PPC as an alternative, but I know nothing
>> about their memory ordering model).
>> 
>>> Question is: can we just drop it, or does ARM depend on it? And if ARM
>>> really needs it, is it the memory barrier only? And why wouldn't a
>>> similar barrier be needed in vcpu_block() then?
>> 
>> It's not the memory barrier, but the subsequent checks. The barrier
>> is there only to make sure the checks don't happen before the
>> earlier writes. And as the comment says - it's the combination of
>> the 3 writes that may get mis-ordered and hence mis-interpreted by
>> other (reading) parties. vcpu_block(), otoh, sets the VPF_blocked
>> bit only, so there's no potential ordering issue there.
> 
> Are you suggesting the comment in vcpu_block is not correct?

No. But I'm not talking about the local_events_need_delivery()
checks (which exist in both functions), but the ones inside the
#ifdef in do_poll() (which have no equivalent in vcpu_block()).
That's not to say that there may not need to be a barrier ahead
of local_events_need_delivery() as well - I've simply not thought
through that further case (yet).

Jan



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel