Re: [PATCH V2 5/6] x86/intel_rdt: Use perf infrastructure for measurements

From: Reinette Chatre
Date: Thu Sep 06 2018 - 16:05:08 EST


Hi Peter,

On 9/6/2018 12:44 PM, Peter Zijlstra wrote:
> On Thu, Sep 06, 2018 at 12:21:59PM -0700, Reinette Chatre wrote:
>> If you do have suggestions on how I can improve the implementation while
>> maintaining (or improving) the accuracy of the measurements I would
>> greatly appreciate it.
>
> You can reduce the code duplication by using __always_inline functions
> with const function arguments.
>

Could you please elaborate? I am unable to see how that would help in,
for example:

if (need_l2) {
rdpmcl(l2_hit_pmcnum, l2_hits_after);
rdpmcl(l2_miss_pmcnum, l2_miss_after);
}
if (need_l3) {
rdpmcl(l3_hit_pmcnum, l3_hits_after);
rdpmcl(l3_miss_pmcnum, l3_miss_after);
}

Two issues with the above are:
- the measurements captured in l2_hits_after and l2_miss_after would
include the hits/misses associated with the "if (need_l2)" test
- the measurements captured in l3_hits_after and l3_miss_after would
include the hits/misses associated with both the "if (need_l2)" and "if
(need_l3)" tests.

When I separate the above into the two functions it just becomes either:
rdpmcl(l2_hit_pmcnum, l2_hits_after);
rdpmcl(l2_miss_pmcnum, l2_miss_after);
or:
rdpmcl(l3_hit_pmcnum, l3_hits_after);
rdpmcl(l3_miss_pmcnum, l3_miss_after);

Reinette