Re: [RFC 1/4] perf kvm: Enable 'record' on powerpc
From: Arnaldo Carvalho de Melo
Date: Mon Mar 28 2016 - 08:28:55 EST
Em Mon, Mar 28, 2016 at 04:28:45PM +0530, Ravi Bangoria escreveu:
> Thanks Arnaldo for putting the effort.
> I've tested this patch on powerpc and it looks fine to me. Please find my
> below comments.
> On Friday 25 March 2016 02:45 AM, Arnaldo Carvalho de Melo wrote:
> >Em Tue, Mar 22, 2016 at 11:19:21PM -0300, Arnaldo Carvalho de Melo escreveu:
> >>Em Tue, Mar 22, 2016 at 04:12:11PM -0300, Arnaldo Carvalho de Melo escreveu:
> >>>Em Wed, Feb 24, 2016 at 02:37:42PM +0530, Ravi Bangoria escreveu:
> >>>>'perf kvm record' is not available on powerpc because 'perf' relies on
> >>>>the 'cycles' event (a PMU event) to profile the guest. However, for
> >>>>powerpc, this can't be used from the host because the PMUs are controlled
> >>>>by the guest rather than the host.
> >>>>There exists a tracepoint 'kvm_hv:kvm_guest_exit' in powerpc which is
> >>>>hit whenever any of the threads exit the guest context. The guest
> >>>>instruction pointer dumped along with this tracepoint data in the field
> >>>>'pc', can be used as guest instruction pointer.
> >>>>This patch changes default event as kvm_hv:kvm_guest_exit for recording
> >>>>guest data in host on powerpc. As we are using host event to record guest
> >>>>data, this approach will enable only --guest option of 'perf kvm'. Still
> >>>>--host --guest together won't work.
> >>>It should, i.e. --host --guest should translate to:
> >>> -e cycles:H,kvm_hv:kvm_guest_exit
> >>>I.e. both collect cycles only in the host, and also the tracepoint that
> >>>will allow us to get the guest approximation for the unavailable cycles
> >>>event, no?
> >>>I'm putting the infrastructure work needed for this the perf/cpumode
> >>>branch. More work will be put there soon.
> >>So I took a different path and made perf_evsel__parse_sample set a new
> >>perf_sample.cpumode field, this way we'll end up having just to set a
> >>per-evsel ->post_parse_sample() callback for the event that replaces
> >>"cycles" for PPC guests where we'll just set data->ip and data->cpumode,
> >>the rest of the code remains unchanged.
> >>The changes I made looks useful in itself, as, IIRC more code was
> >>removed than added.
> >>I'll continue tomorrow and will test with the kvm:kvm_exit on x86_64 for
> >>testing, that has:
> >Ok, so the infrastructure got merged already and from there the next
> >steps are in running with:
> > perf kvm --guest record -a -e cycles:H,kvm:kvm_exit
> >And then, with the patch below applied, try:
> >perf kvm --guestkallsyms kallsyms.guest --guestmodules modules.guest report -i perf.data.guest --munge-ppc-guest-sample kvm:kvm_exit
> The initial proposal was to change the default event as "kvm_guest_exit" for
> kvm recording/reporting
> on ppc. If I understand it correctly, your patch creates a handler for
> reporting kvm events
> based on "munge_ppc_guest_event" and the required tracepoint i.e., we need
> to mention the
> required tracepoint event name for recording and reporting.
What is present in my patch is the creation of infrastructure that is as
much generic as possible, to enable doing this event conversion for PPC.
Doing it transparently in PPC systems and on other systems when handling
a perf.data file generated by 'perf kvm record' on PPC will be done on
top of this, thanks for testing, I'll send a complete series soon.
> There might be a little bit of an issue here. For scripts which depend on
> generic perf kvm record/report,
> we need to change those appropriately to prevent those from failing on
> powerpc. Otherwise, (just a
> thought) can we create some kind of an alias to map the ppc specific perf
> kvm commands with the
> generic perf kvm.
> For e.g :
> perf kvm record -e "kvm_hv:kvm_guest_exit" mapped to perf kvm record
> perf kvm report --munge-ppc-guest-sample kvm_hv:kvm_guest_exit mapped to
> perf kvm report.