[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <d64f85e8-98ca-a81d-bf2d-13dd72fc411e@intel.com>
Date: Thu, 2 Aug 2018 13:06:19 -0700
From: Dave Hansen <dave.hansen@...el.com>
To: Peter Zijlstra <peterz@...radead.org>
Cc: Reinette Chatre <reinette.chatre@...el.com>, tglx@...utronix.de,
mingo@...hat.com, fenghua.yu@...el.com, tony.luck@...el.com,
vikas.shivappa@...ux.intel.com, gavin.hindman@...el.com,
jithu.joseph@...el.com, hpa@...or.com, x86@...nel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH 0/2] x86/intel_rdt and perf/x86: Fix lack of coordination
with perf
On 08/02/2018 12:54 PM, Peter Zijlstra wrote:
>> I totally understand not wanting to fill the tree with code hijacking
>> the raw PMU. Is your reaction to this really around not wanting to
>> start down the slippery slope that ends up with lots of raw PMU "owners"?
> That and the fact that multiple owner directly contradicts what perf set
> out to do, provide resource arbitration for the PMU.
>
> Not being able to use both perf and this resctl thing at the same time
> is utter crap. You will not get special dispensation.
Is there something we could do in the middle, like have perf itself be
in charge of doing all the programming, but the psuedo-locking could
still _read_ the counters?
I guess we'd still end up needing to have perf expose which counter got
assigned to which event, and make sure things like multiplexing are not
in play.
But, either way, I think we can live with this measurement being fragile
(like it already is). If it only works when all the planets are
aligned, I think it's still usable enough.
Powered by blists - more mailing lists