lkml.org 
[lkml]   [2012]   [Dec]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/1] Change IBS PMU to use perf_hw_context
On 16.12.12 10:04:10, Ingo Molnar wrote:
>
> * suravee.suthikulpanit@amd.com <suravee.suthikulpanit@amd.com> wrote:
>
> > From: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
> >
> > Currently, the AMD IBS PMU initialize pmu.task_ctx_nr to
> > perf_invalid_context which only allows IBS to be running only
> > in system-wide mode (e.g. perf record -a). IBS hardware is
> > available in each core and should be per-context. This patch
> > modifies the task_ctx_nr to use the perf_hw_context (default)
> > instead.
>
> I'm wondering how extensively was it tested/verified that it's
> safe to enable IBS in per context mode as well, and that the
> profiling results are precise and accurate?

From the implementation's point of view this is very similar to hw
perf counters. I wouldn't expect any issues here. Since IBS can be
immediatly started/stopped and there is no caching, there won't be any
incomming sample that is not related to that context.

The only potential problem I see could be a security risk in a way
that an IBS sample might expose data related to other contexts such as
cache information. This is similar to uncore/northbridge events so I
don't think this is an issue, but we might want to evaluate this.

> We never used the IBS hardware in this fashion before, so some
> extra care is prudent - and traces of that extra care should be
> visible in the changelog as well.

Yeah, a comparison of numbers for IBS and hw counter (-e r076:p,r076
and -e r0C1:p,r0C1) in per-context mode would be useful here.

-Robert


\
 
 \ /
  Last update: 2012-12-17 11:21    [W:0.084 / U:2.108 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site