Re: [PATCH 4.5 007/238] KVM: i8254: change PIT discard tick policy

[Paolo Bonzini <pbonzini@redhat.com>]

On 12/04/2016 16:13, Greg Kroah-Hartman wrote:
> On Tue, Apr 12, 2016 at 03:30:23PM +0200, Radim Krčmář wrote:
>> 2016-04-11 18:21-0700, Greg Kroah-Hartman:
>>> On Tue, Apr 12, 2016 at 12:23:35AM +0100, Ben Hutchings wrote:
>>>> On Wed, 2016-03-02 at 22:56 +0100, Radim Krčmář wrote:
>>>>> Even though there is a chance of regressions, I think we can fix the
>>>>> LVT0 NMI bug without introducing a new tick policy.
>>>> [...]
>>>>
>>>> Given the 'chance of regressions', should we let this sit in mainline
>>>> longer before including it in stable updates?
>>>
>>> Hm, good point, Radim, what do you think, is this good to go to stable
>>> now?  This has been in since 4.6-rc1, so it's been a few weeks with
>>> people running it already...
>>
>> I think it is good to go.  No reasonable workload should regress and the
>> fixed use-case is common on old linux guest.
>>
>> This patch makes a difference if the guest doesn't EOI in PIT interrupts
>> before the next one arrives.  PIT would have been unreliable in that
>> situation, so all worloads that that could regress have likely been
>> avoided.  Changes to NMI injection would need even crazier guest to
>> regress.
> 
> Ok, thanks, leaving it in.

I agree.

FWIW the behavior after the patch is consistent with what actual
hardware does.  The old behavior was _documented_ to be consistent with
actual hardware, but instead it was crazy.

Paolo