On Tue, 2010-06-22 at 12:12 +0300, Avi Kivity wrote:
> On 06/22/2010 05:08 AM, Zhang, Yanmin wrote:
> >
> >> Something that is worrying is that we don't expose group information.
> >> perf will multiplex the events for us, but there will be a loss in accuracy.
> >>
> >>
> >>> #ifdef CONFIG_HAVE_HW_BREAKPOINT
> >>> #include<asm/hw_breakpoint.h>
> >>> #endif
> >>> @@ -753,6 +752,20 @@ struct perf_event {
> >>>
> >>> perf_overflow_handler_t overflow_handler;
> >>>
> >>> + /*
> >>> + * pointers used by kvm perf paravirt interface.
> >>> + *
> >>> + * 1) Used in host kernel and points to host_perf_shadow which
> >>> + * has information about guest perf_event
> >>> + */
> >>> + void *host_perf_shadow;
> >>>
> >>>
> >> Can we have real types instead of void pointers?
> >>
> > I just want perf generic codes have less dependency on KVM codes.
> >
>
> One way to do that and retain type safety is to have
>
> struct perf_client {
> struct perf_client_ops *ops;
> ...
> }
>
> The client (kvm) can do
>
> struct kvm_perf_client {
> struct perf_client pc;
> // kvm specific stuff
> };
>
> the callbacks receive struct perf_client and use container_of to reach
> the kvm_perf_client that contains it.
Let me double check it.
>
> >>> + /*
> >>> + * 2) Used in guest kernel and points to guest_perf_shadow which
> >>> + * is used as a communication area with host kernel. Host kernel
> >>> + * copies overflow data to it when an event overflows.
> >>> + */
> >>> + void *guest_perf_shadow;
> >>>
> >>>
> >> It's strange to see both guest and host parts in the same patch.
> >> Splitting to separate patches will really help review.
> >>
> > It's a little hard to split the patches if they change the same file. Perhaps
> > I could add more statements before the patch when I send it out.
> >
>
> With git, it's easy (once you're used to it):
>
> # go back one commit:
> git reset HEAD^
> # selectively add bits:
> git add -p
> # commit first patch
> git commit -s
> # selectively add bits:
> git add -p
> # commit second patch
> git commit -s
Thanks for your teaching.
>
>
> >>> @@ -1626,9 +1629,22 @@ void perf_event_task_tick(struct task_st
> >>> if (ctx&& ctx->nr_events&& ctx->nr_events != ctx->nr_active)
> >>> rotate = 1;
> >>>
> >>> - perf_ctx_adjust_freq(&cpuctx->ctx);
> >>> - if (ctx)
> >>> - perf_ctx_adjust_freq(ctx);
> >>> +#ifdef CONFIG_KVM_PERF
> >>> + if (kvm_para_available()) {
> >>> + /*
> >>> + * perf_ctx_adjust_freq causes lots of pmu->read which would
> >>> + * trigger too many vmexit to host kernel. We disable it
> >>> + * under para virt situation
> >>> + */
> >>> + adjust_freq = 0;
> >>> + }
> >>> +#endif
> >>>
> >>>
> >> Perhaps we can have a batch read interface which will read many counters
> >> at once.
> >>
> > It's a good idea. But that will touch many perf generic codes which causes it's hard
> > to maintain or follow future changes.
> >
>
> I'm talking about the guest/host interface. So you have one vmexit and
> many host perf calls.
I understood what you were speaking. I mean, perf generic codes operate perf_event
one by one. At low layer, we just know one perf_event before calling hypercall to
vmexit to host kernel.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
