On Tue, May 26, 2015 at 10:53:53 +0200, Martin Kletzander wrote:
On Tue, May 26, 2015 at 10:08:58AM +0200, Peter Krempa wrote:
>On Mon, May 25, 2015 at 17:43:05 +0200, Martin Kletzander wrote:
>> On Mon, May 25, 2015 at 05:26:01PM +0200, Peter Krempa wrote:
>> >On Mon, May 25, 2015 at 16:59:08 +0200, Andrea Bolognani wrote:
...
>> >And just conditionally call this when the
>> >QEMU_CAPS_RTC_RESET_REINJECTION is present and not in an architecture
>> >specific way.
>> >
>> >By this you get rid of the arch specific hackery.
>> >
>>
>> But on x86 we don't even want to call the SetTime command when we
>> cannot reset the rtc reinjection. On ppc there is no reinjection
>> being done, hence nothing to reset.
>
>Well I don't think that the missing reinjection request should be a hard
>requirement on x86 since the guest can modify time independently and in
>that case the reinjection command won't be even issued. The guest agent
>command is technically a guest issued time change too, so there's only
>the added bonus that we'd reset the interrupt backlog. One slight issue
>is that libvirt does not allow to reset the interrupt backlog while not
>setting the time though.
>
The implementer of the check might remember more reasons for arguing
on this, let him shed the light into this discussion. Michal?
Anyway, this is broken all the way because the command also makes
sense only with some particular timers and their settins, not with all
of them.
In that case it would make more sense to have the interrupt reinjection
reset as a separate option or it to be enabled with a flag.
>Even if we decide that the check for the existance of the command should
>be arch-specific and should report error, the actual place where the
>command is called should be made dependant on the presence of the
>command rather than the ARCH check, so that it does not need to be
>modified once it's implemented somewhere else.
>
I don't quite catch your drift here. If you mean that the second
condition added should be based on virQEMUCapsGet(), then I agree.
Yes that was my intent. The actual call should be based on the
capability. Sanity checks that are arch specific should be present
before.
Peter