Re: [PATCH v1 01/11] perf/x86/core: Support KVM to assign a dedicated counter for guest PEBS

[Andi Kleen]

> Suppose your KVM thing claims counter 0/2 (ICL/SKL) for some random PEBS
> event, and then the host wants to use PREC_DIST.. Then one of them will
> be screwed for no reason what so ever.

It's no different from some user using an event that requires some
specific counter.

> 
> How is that not destroying scheduling freedom? Any other situation we'd
> have moved the !PREC_DIST PEBS event to another counter.

Anyways what are you suggesting to do instead? Do you have a better proposal?

The only alternative I know to doing this would be to go through the PEBS
buffer in the guest and patch the applicable counter field up on each PMI.

I tried that at some point (still have code somewhere), but it was
quite complicated and tricky and somewhat slow, so I gave up eventually.

It's also inherently racy because if the guest starts looking at
the PEBS buffer before an PMI it could see the unpatched values
Given I don't know any real software which would break from this,
but such "polled PEBS" usages are certainly concievable.

The artificial constraint is a lot simpler and straight forward,
and also doesn't have any holes like this.

-Andi