Messages in this thread | | | Date | Mon, 6 May 2013 21:00:24 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH] perf, x86: Add workaround for MEM_*_RETIRED errata BV98 |
| |
On Mon, May 06, 2013 at 07:41:54PM +0200, Stephane Eranian wrote:
> Problem is, those events are really important. We need to be able to > measure them.
At what cost? If the code we need to make it work is horrendous we need to maintain that too -- possibly quite a lot longer than Intel thinks IVB is relevant.
Anyway; since these events are immediately dangerous (they corrupt state) blocking them is the right thing; we can always merge a patch that undoes that once we've figured out a sane way.
Currently I'm still waiting for someone to explain (in complete detail) how the corruption manifests. Until that is clear I can't even begin formulating what might be needed to make it work again.
One mode that makes sense to me is where CNTx on SMTi corrupts CNTx on SMTj for i != j. However there's also been said that the only case that doesn't corrupt is for SMTj to have all counters disabled -- which contradicts the previous case.
The former case we could maybe make work; the latter I can't see how we could possible make work at all.
| |