[PATCH] KVM: arm64: pmu: Resync EL0 state on counter rotation

[Shijie Huang]

Hi Leo,

在 2023/8/14 15:16, Leo Yan 写道:
> On Fri, Aug 11, 2023 at 07:05:20PM +0100, Marc Zyngier wrote:
>> Huang Shijie reports that, when profiling a guest from the host
>> with a number of events that exceeds the number of available
>> counters, the reported counts are wildly inaccurate. Without
>> the counter oversubscription, the reported counts are correct.
>>
>> Their investigation indicates that upon counter rotation (which
>> takes place on the back of a timer interrupt), we fail to
>> re-apply the guest EL0 enabling, leading to the counting of host
>> events instead of guest events.
> Seems to me, it's not clear for why the counter rotation will cause
> the issue.
>
> In the example shared by Shijie in [1], the cycle counter is enabled for
> both host and guest, and cycle counter is a dedicated event which does
> not share counter with other events.  Even there have counter rotation,
> it should not impact the cycle counter.

Just take a simple case:

    perf stat -e cycles:G,cycles:H, e2,e3,e4,e5,e6,e7 ....


Assume we have 8 events, but PMU only privides 7 counters(cycle + 6 normal)

1.) The initial:

          event 0 (cycles:G) ---> used cycle counter

          event 1 (cycles:H)  ---> used counter 1

           event 2 ---> used counter 2

           event 3 ---> used counter 3

            event 4 ---> used counter 4

            event 5 ---> used counter 5

            event 6---> used counter 6

  2.) After the event rotation , the event0 will put to the tail of the 
list, please see rotate_ctx()

the first round, it will like this:

          event 1(cycles:H) ---> used cycle counter

          event 2 ---> used counter 1

         event 3 ---> used counter 2

         event 4 ---> used counter 3

         event 5 ---> used counter 4

          event 6 ---> used counter 5

          event 7 ---> used counter 6


  3.) Rotation it again, event 1 will in the tail.

      In the second round, it will like this:

         evnet 0(cycles:G) ---> used cycle counter

        event 2 ---> used counter 1

         event 3 ---> used counter 2

         event 4 ---> used counter 3

         event 5 ---> used counter 4

          event 6 ---> used counter 5

          event 7 ---> used counter 6


  4.) Rotation it again, in the third round, it will like this:

          evnet 0(cycles:G) ---> used cycle counter

        event 3 ---> used counter 1

         event 4 ---> used counter 2

         event 5 ---> used counter 3

         event 6 ---> used counter 4

          event 7 ---> used counter 5

          event 2(cycles:H) ---> used counter 6


....

So it will impact the cycle counter..:)


Thanks

Huang Shijie