Messages in this thread | | | Date | Tue, 10 May 2016 14:15:38 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH 2/3] perf/x86/mbm: Fix mbm counting for RMID reuse |
| |
On Fri, May 06, 2016 at 04:44:14PM -0700, Vikas Shivappa wrote: > This patch tries to fix the issue where multiple perf instances try to > monitor the same PID.
> MBM cannot count directly in the usual perf way of continuously adding > the diff of current h/w counter and the prev count to the event->count > because of some h/w dependencies:
And yet the patch appears to do exactly that; *confused*.
> (1) the mbm h/w counters overflow.
As do most other counters.. so your point is? You also have the software timer < overflow period..
> (2) There are limited h/w RMIDs and hence we recycle the RMIDs due to > which an event may count from different RMIDs.
This fails to explain why this is a problem.
> (3) Also we may not want to count at every sched_in and sched_out > because the MSR reads involve quite a bit of overhead.
Every single other PMU driver just does this; why are you special?
You list 3 issues of why you think you cannot do the regular thing, but completely fail to explain how these issues are a problem.
> However we try to do something similar to usual perf way in this patch > and mainly handle (1) and (3).
> update_sample takes care of the overflow in the hardware counters and > provides abstraction by returning total bytes counted as if there was no > overflow. We use this abstraction to count as below: > > init: > event->prev_count = update_sample(rmid) //returns current total_bytes > > count: // MBM right now uses count instead of read > cur_count = update_sample(rmid) > event->count += cur_count - event->prev_count > event->prev_count = cur_count
So where does cqm_prev_count come from and why do you need it? What's wrong with event->hw.prev_count ?
In fact, I cannot seem to find any event->hw.prev_count usage in this or the next patch, so can we simply use that and not add pointless new members?
| |