Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Sat, 31 Aug 2019 11:18:02 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     "Liang, Kan" <kan.liang@linux.intel.com>
Cc:     acme@kernel.org, mingo@redhat.com, linux-kernel@vger.kernel.org,
        tglx@linutronix.de, jolsa@kernel.org, eranian@google.com,
        alexander.shishkin@linux.intel.com, ak@linux.intel.com
Subject: Re: [RESEND PATCH V3 3/8] perf/x86/intel: Support hardware TopDown
 metrics
Message-ID: <20190831091802.GI2369@hirez.programming.kicks-ass.net>
References: <20190826144740.10163-1-kan.liang@linux.intel.com>
 <20190826144740.10163-4-kan.liang@linux.intel.com>
 <20190828151921.GD17205@worktop.programming.kicks-ass.net>
 <fd95a255-4499-2907-8af9-d340f157da68@linux.intel.com>
 <20190829135256.GW2369@hirez.programming.kicks-ass.net>
 <cc2ee16d-b10f-a31e-7411-320e90413ceb@linux.intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <cc2ee16d-b10f-a31e-7411-320e90413ceb@linux.intel.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Aug 29, 2019 at 12:56:02PM -0400, Liang, Kan wrote:
> On 8/29/2019 9:52 AM, Peter Zijlstra wrote:

> > What what? The PERF_METRICS contents depends on the FIXCTR3 value ?!
> 
> Yes.
> 
> For current implementation, PERF_METRIC MSR is composed by four fields,
> backend bound, frontend bound, bad speculation and retiring.
> Each of the fields are populated using the below formula for eg:
> PERF_METRIC[RETIRING] = (0xFF *
> PERF_METRICS_RETIRING_INTERNAL_48bit_COUNTER)
> / FIXCTR3

So it really depends on the actual exposed FIXCTR3 _value_ to compute
the PERF_METRIC field? *mind boggles*, that's really unspeakable crap.
And this isn't documented anywhere afaict.

I was thinking they've have an internal counter for the SLOTS value too,
so the PERF_METRIC fields are indenpendent; which would be like 'sane'.

Exposing the internal counters would've been _soooo_ much better, just
add 4 more fixed counters and call it a day.

> The METRICS_OVF indicates the overflow of any internal counters.

OK, but I'm thinking that by that time the fraction in PERF_METRIC will
be too coarse and we're loosing precision. Reconstruction will be
inaccurate.

> The internal counters only start counting from 0, which cannot be programmed
> by SW. But resetting the PERF_METRIC would implicitly resetting the internal
> counters.

The only possible option given the choices.