Re: [RFC v2 3/5] i386/kvm: Support event with select & umask format in KVM PMU filter

[Markus Armbruster <armbru@redhat.com>]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Thu, Feb 06, 2025 at 05:54:32PM +0800, Zhao Liu wrote:
>> On Wed, Feb 05, 2025 at 11:07:10AM +0100, Markus Armbruster wrote:
>> > Date: Wed, 05 Feb 2025 11:07:10 +0100
>> > From: Markus Armbruster <armbru@redhat.com>
>> > Subject: Re: [RFC v2 3/5] i386/kvm: Support event with select & umask
>> >  format in KVM PMU filter
>> > 
>> > Zhao Liu <zhao1.liu@intel.com> writes:
>> > 
>> > > The select&umask is the common way for x86 to identify the PMU event,
>> > > so support this way as the "x86-default" format in kvm-pmu-filter
>> > > object.
>> > 
>> > So, format 'raw' lets you specify the PMU event code as a number, wheras
>> > 'x86-default' lets you specify it as select and umask, correct?
>> 
>> Yes!
>> 
>> > Why do we want both?
>> 
>> This 2 formats are both wildly used in x86(for example, in perf tool).
>> 
>> x86 documents usually specify the umask and select fields.
>> 
>> But raw format could also be applied for ARM since ARM just uses a number
>> to encode event.

Is it really too much to ask of the client to compute a raw value from
umask and select values?

>> > > Signed-off-by: Zhao Liu <zhao1.liu@intel.com>
>> > 
>> > [...]
>> > 
>> > > diff --git a/qapi/kvm.json b/qapi/kvm.json
>> > > index d51aeeba7cd8..93b869e3f90c 100644
>> > > --- a/qapi/kvm.json
>> > > +++ b/qapi/kvm.json
>> > > @@ -27,11 +27,13 @@
>> > >  #
>> > >  # @raw: the encoded event code that KVM can directly consume.
>> > >  #
>> > > +# @x86-default: standard x86 encoding format with select and umask.
>> > 
>> > Why is this named -default?
>> 
>> Intel and AMD both use umask+select to encode events, but this format
>> doesn't have a name... so I call it `default`, or what about
>> "x86-umask-select"?

Works for me.

>> > > +#
>> > >  # Since 10.0
>> > >  ##
>> > >  { 'enum': 'KVMPMUEventEncodeFmt',
>> > >    'prefix': 'KVM_PMU_EVENT_FMT',
>> > > -  'data': ['raw'] }
>> > > +  'data': ['raw', 'x86-default'] }
>> > >  
>> > >  ##
>> > >  # @KVMPMURawEvent:
>> > > @@ -46,6 +48,25 @@
>> > >  { 'struct': 'KVMPMURawEvent',
>> > >    'data': { 'code': 'uint64' } }
>> > >  
>> > > +##
>> > > +# @KVMPMUX86DefalutEvent:
>> > 
>> > Default, I suppose.
>> 
>> Thanks!
>> 
>> > > +#
>> > > +# x86 PMU event encoding with select and umask.
>> > > +# raw_event = ((select & 0xf00UL) << 24) | \
>> > > +#              (select) & 0xff) | \
>> > > +#              ((umask) & 0xff) << 8)
>> > 
>> > Sphinx rejects this with "Unexpected indentation."
>> > 
>> > Is the formula needed here?
>> 
>> I tried to explain the relationship between raw format and umask+select.
>> 
>> Emm, where do you think is the right place to put the document like
>> this?

Do users need to know how to compute the raw event value from @select
and @umask?

If yes, is C code the best way?

Here's another way:

    bits  0..7 : bits 0..7 of @select
    bits  8..15: @umask
    bits 24..27: bits 8..11 of @select
    all other bits: zero

>> ...
>> 
>> > > +##
>> > > +# @KVMPMUX86DefalutEventVariant:
>
> Typo   s/Defalut/Default/ - repeated many times in this patch.
>
>> > > +#
>> > > +# The variant of KVMPMUX86DefalutEvent with the string, rather than
>> > > +# the numeric value.
>> > > +#
>> > > +# @select: x86 PMU event select field.  This field is a 12-bit
>> > > +#     unsigned number string.
>> > > +#
>> > > +# @umask: x86 PMU event umask field. This field is a uint8 string.
>> > 
>> > Why are these strings?  How are they parsed into numbers?
>> 
>> In practice, the values associated with PMU events (code for arm, select&
>> umask for x86) are often expressed in hexadecimal. Further, from linux
>> perf related information (tools/perf/pmu-events/arch/*/*/*.json), x86/
>> arm64/riscv/nds32/powerpc all prefer the hexadecimal numbers and only
>> s390 uses decimal value.
>> 
>> Therefore, it is necessary to support hexadecimal in order to honor PMU
>> conventions.
>
> IMHO having a data format that matches an arbitrary external tool is not
> a goal for QMP. It should be neutral and exclusively use the normal JSON
> encoding, ie base-10 decimal. Yes, this means a user/client may have to
> convert from hex to dec before sending data over QMP. This is true of
> many areas of QMP/QEMU config though and thus normal/expected behaviour.

Concur.