On 08.02.2022 14:27, Jane Malalane wrote:
> On 31/01/2022 12:05, Jan Beulich wrote:
>> On 27.01.2022 17:01, Jane Malalane wrote:
>>> Introduce a new per-domain creation x86 specific flag to
>>> select whether hardware assisted virtualization should be used for
>>> x{2}APIC.
>>>
>>> A per-domain option is added to xl in order to select the usage of
>>> x{2}APIC hardware assisted vitualization, as well as a global
>>> configuration option.
>>>
>>> Having all APIC interaction exit to Xen for emulation is slow and can
>>> induce much overhead. Hardware can speed up x{2}APIC by running APIC
>>> read/write accesses without taking a VM exit.
>>
>> This is odd to read for a patch which makes it possible to _turn off_
>> acceleration. Instead it would be interesting to know what problems
>> you have encountered making it desirable to have a way to turn this off.
>
> After speaking to Andrew he told me of a performance regression that was
> reported some time ago when enabling apicv related to the pass-through
> LAPIC timer of a HVM guest causing Xen to intercept the LAPIC timer MSR,
> making anything that uses the LAPIC timer end up slower than it was
> before. So, adressing your comment here, other than mentioning how being
> able to disable acceleration for apicv can be useful when testing and
> debugging, do you think it's worth mentioning (in more detail) that this
> perf problem exists, in the commit message.
Yes, I think it would be worth mentioning, as then the purpose of this
change is also to be a workaround there, not just testing/debugging. In
fact this workaround may then be viewed as the primary purpose.
Jan
