Re: [RFC PATCH v2 2/3] perf/x86/intel/pt: Introduce and export pt_get_curr_event()

["Liang, Kan" <kan.liang@linux.intel.com>]

On 2022-09-27 10:27 a.m., Wang, Wei W wrote:
> On Tuesday, September 27, 2022 2:09 AM, Liang, Kan wrote:
>> From my understanding of the host-guest mode, the host PT event never
>> traces the guest no matter whether the guest enables PT. So when VM-entry,
>> there is only a guest PT event or no event. If so, I think the perf tool should
>> warn the user if they try to create a host event with !exclude_guest, since the
>> host event never traces a guest.
> 
> Probably not from the perf side. It's actually an issue of the KVM's pt_mode:
> How can KVM prevent host from profiling the guest when in host-guest mode?

I don't think it's KVM's job to prevent host from profiling the guest.
However, the current KVM implementation implicitly disables the guest
profiling from the host. If I understand correct, this patch series just
change it to an explicit way (Just for kernel. Perf tool still doesn't
know about it.). There is no substantial change.

I think it should be PT driver who decide which event should be
scheduled. So the first step is to let the PT driver understand the
host-guest mode. Then if a perf user in the host tries to profile with
!exclude_guest with the host-guest mode, the pt driver can deny the
request. The perf tool than throw a warning, e.g., "The VMM is running
with host-guest mode. Perf cannot profile guest. Please remove the guest
profiling setting or switch to the host-only mode.".

> Just a warning from the perf side wouldn't be enough. I think that's a wrong
> assumption from KVM side. I had a plan to fix this one. Exposing the host event
> can tell KVM if the host is profiling the guest or not (i.e. host-guest is allowed).

I don't think KVM should know such information.

Thanks,
Kan