DMARC-Filter: OpenDMARC Filter v1.3.2 mx1.redhat.com 387FCC04B320
Date: Mon, 12 Jun 2017 12:36:31 +0200
From: Jiri Olsa <jolsa@redhat.com>
To: kan.liang@intel.com
Cc: peterz@infradead.org, mingo@redhat.com, eranian@google.com,
        linux-kernel@vger.kernel.org, alexander.shishkin@linux.intel.com,
        acme@redhat.com, torvalds@linux-foundation.org, tglx@linutronix.de,
        vincent.weaver@maine.edu, ak@linux.intel.com
Subject: Re: [PATCH V2 1/2] perf/x86/intel: enable CPU ref_cycles for GP
 counter
Message-ID: <20170612103631.GA6099@krava>
References: <1497029283-3332-1-git-send-email-kan.liang@intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1497029283-3332-1-git-send-email-kan.liang@intel.com>
User-Agent: Mutt/1.8.0 (2017-02-23)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1969
Lines: 45

On Fri, Jun 09, 2017 at 10:28:02AM -0700, kan.liang@intel.com wrote:

SNIP

> +	/*
> +	 * Correct the count number if applying ref_cycles replacement.
> +	 * There is a constant ratio between ref_cycles (event A) and
> +	 * ref_cycles replacement (event B). The delta result is from event B.
> +	 * To get accurate event A count, the real delta result must be
> +	 * multiplied with the constant ratio.
> +	 *
> +	 * It is handled differently for sampling and counting.
> +	 * - Fixed period Sampling: The period is already adjusted in
> +	 *   scheduling for event B. The period_left is the remaining period
> +	 *   for event B. Don't need to modify period_left.
> +	 *   It's enough to only adjust event->count here.
> +	 *
> +	 * - Fixed freq Sampling: The adaptive frequency algorithm needs
> +	 *   the last_period and event counts for event A to estimate the next
> +	 *   period for event A. So the period_left is the remaining period
> +	 *   for event A. It needs to multiply the result with the ratio.
> +	 *   However, the period_left is also used to reload the event counter
> +	 *   for event B in x86_perf_event_set_period. It has to be adjusted to
> +	 *   event B's remaining period there.
> +	 *
> +	 * - Counting: It's enough to only adjust event->count for perf stat.
> +	 *
> +	 * - RDPMC: User can also read the counter directly by rdpmc/mmap.
> +	 *   Users have to handle the adjustment themselves.
> +	 *   For kernel, it only needs to guarantee that the offset is correct.
> +	 *   In x86's arch_perf_update_userpage, the offset will be corrected
> +	 *   if event B is used.
> +	 */
> +	adjust_delta = delta;
> +	if (hwc->flags & PERF_X86_EVENT_REF_CYCLES_REP) {
> +		adjust_delta = delta * x86_pmu.ref_cycles_factor;
> +
> +		if (is_sampling_event(event) && event->attr.freq)
> +			local64_sub(adjust_delta, &hwc->period_left);
> +		else
> +			local64_sub(delta, &hwc->period_left);

shouldn't you use adjust_delta in here also ^^^ ?

jirka