Re: [RFC PATCH 0/2] KVM: VMX: Fix VM entry failure on PT_MODE_HOST_GUEST while host is using PT

[Liang, Kan]

On 2022-09-14 10:46 p.m., Wang, Wei W wrote:
> On Thursday, September 15, 2022 4:26 AM, Liang, Kan wrote:
>> The perf_event_disable() eventually invokes the intel_pt_stop().
>> We already expose the intel_pt_stop()/cpu_emergency_stop_pt() to other
>> modules. I don't think we have to use the perf_event_disable(). Also, the
>> perf_event_disable() requires extra codes.
>>
>> I went through the discussions. I agree with Sean's suggestion.
>> We should only put the logic in the KVM but all the MSR access details into the PT
>> driver. 
> 
> Even the driver itself doesn’t drive the save/restore of the MSRs, it is drived by perf.

It through perf_event, not driven by perf_event. The perf_event generic
code never knows when should invokes each driver to save/restore
information. It should be driven by the other subsystem e.g., scheduler.

For this case, KVM should drive the save/restore, and the PT driver
eventually does all the MSR access details.

> 1. If we make KVM a user of perf, we should do this via perf_event_disable/enable_*.
> 2. If we make KVM an alternative to perf (i.e. have direct control over PMU HW),
> we can do this via driver interfaces like perf.
> Per my experience, we should go for 1. Probably need Peter's opinions on this.
>

For 1, the perf_event_disable/enable_* are not enough. They don't
save/restore MSRs. If we go to this way, we have to introduce a new
generic interface to ask each driver to save/restore their MSRs when the
guest is entering/exiting. We'd better combine the new interface with
the existing perf_guest_get_msrs() of the core driver.
I think that's an ideal solution, but requires big changes in the code.

2 is the current KVM implementation. See pt_save_msr()/pt_load_msr(). I
don't think it's a right way. We'd better fix it.

The suggestion should be 3. The KVM notify the PT driver via the
interface provided by PT. The PT driver save/restore all the registers.
I think it's an acceptable solution with small code changes.

So I prefer 3.

Thanks,
Kan

>> But I prefer a more generic and straightforward function name, e.g.,
>> intel_pt_stop_save()/intel_pt_start_load(), in case other modules may want to
>> save/restore the PT information in their context switch later.
>>
>> Thanks,
>> Kan
>>
>>>
>>>> It seems perf_event_disable() is not used widely by other kernel
>>>> component. 
> 
> Because there are not lots of kernel users.
> You can check another user, watchdog_hld.c, perf_event_enable/disable are used there.