On 11/09/15 14:35, Shannon Zhao wrote:

On 2015/9/11 19:04, Marc Zyngier wrote:

[...]

quoted

Having had a chat with Will, it appears that a much better solution
would be to ask perf to use raw events instead of trying to map things
to perf events (which the guest has already done).

See drivers/oprofile/oprofile_perf.c::op_perf_setup().

Thoughts?

Yeah, directly using PERF_TYPE_RAW event looks simpler. But should we 
check whether the value of event number written to PMXEVTYPER_EL0 is 
valid? Or That is guaranteed by guest?

I don't think we need to check for anything. Userspace is allowed to
request any event and the PMU will count it if this event exists, or
won't if it doesn't.

We shouldn't be doing more validation than that, if only because most
events are microarchitectural, and they are mostly undocumented.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...