Messages in this thread | | | Date | Wed, 1 May 2013 21:21:55 +0200 | Subject | Re: [PATCH] perf, x86: Add workaround for MEM_*_RETIRED errata BV98 | From | Stephane Eranian <> |
| |
On Wed, May 1, 2013 at 9:13 PM, Andi Kleen <andi@firstfloor.org> wrote: >> Unfortunately, Andi's patch fails to correct the problem though >> it is the right approach. As he said, the problem is unfortunately > > perf stat -a / record -a / top will work. > But it does not for perf stat -a. Look at my example, run it on your machine.
> The case it doesn't fix is per thread when the other sibling > is not running a counter. > > Essentially with the patch you may still get under-counting > on these specific events if they run per thread (or are really > unlucky in multiplexing), but never over-counting on anything. > > In the past Ingo's guidance was to focus on the most common > use cases, which I think this patch does. > > It would be possible to fix the per thread mode too, but it would > need quite a bit of new infrastructure and I have doubts that's > worth it. > > -Andi
| |