[MODERATED] Re: [PATCH v2 0/8] PERFv2

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 07/02/2019 17:00, speck for Peter Zijlstra wrote:
> On Thu, Feb 07, 2019 at 04:35:12PM +0000, speck for Andrew Cooper wrote:
>> On 07/02/2019 15:37, speck for Peter Zijlstra wrote:
>>> On Thu, Feb 07, 2019 at 06:37:59AM -0800, speck for Andi Kleen wrote:
>>>
>>>> Okay I see you rely on it being guarded by the global, but for virtualization
>>>> support will need the copy.
>>>>
>>>> Would need a copy at least.
>>> I've not thought abou virt yet; do we _have_ to expose the thing to
>>> guests?
>> Virt has been discussed at length.
>>
>> This feature and MSR are not architectural, and not expected to exist on
>> future hardware with a silicon fix for the issue (CascadeLake B1
>> stepping and later, IIRC).
>>
>> By exposing the features to guests, you automatically limit their
>> migrateability to a subset of the Skylake uarch processors.
>>
>> For Xen, I've just gone with a boot time global flag.  PMU deliberately
>> isn't available to guests in general, but developers wanting to do
>> profiling/other can trivially flip the default.
> Is that a per-guest or per-hv setting? My current patch does per-hv, but
> I suppose it would not be too hard to make that per-guest. The only
> question is where/how does the guest communicate this desire.

This is server-wide boot time setting.  I did consider per-guest, but
this TSX issue isn't the only thing on my plate right now, and anyone
wanting to use PMU also has to flip a different server-wide boot time
setting anyway.

In my otherwise copious free time, I am working on improvements to our
APIs for creation, so will make it a per-guest as soon as we can
sensibly express it.

~Andrew