From: Peter Zijlstra <peterz@infradead.org>
To: Wen Yang <simon.wy@alibaba-inc.com>
Cc: Ingo Molnar <mingo@redhat.com>,
Arnaldo Carvalho de Melo <acme@kernel.org>,
Alexander Shishkin <alexander.shishkin@linux.intel.com>,
Thomas Gleixner <tglx@linutronix.de>,
Mark Rutland <mark.rutland@arm.com>, Jiri Olsa <jolsa@redhat.com>,
Namhyung Kim <namhyung@kernel.org>,
Borislav Petkov <bp@alien8.de>,
x86@kernel.org, "H. Peter Anvin" <hpa@zytor.com>,
linux-perf-users@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [RFC PATCH] perf/x86: improve the event scheduling to avoid unnecessary pmu_stop/start
Date: Fri, 11 Feb 2022 12:36:43 +0100 [thread overview]
Message-ID: <20220211113643.GT23216@worktop.programming.kicks-ass.net> (raw)
In-Reply-To: <20220210043930.34311-1-simon.wy@alibaba-inc.com>
On Thu, Feb 10, 2022 at 12:39:30PM +0800, Wen Yang wrote:
> This issue has been there for a long time, we could reproduce it as follows:
>
> 1, run a script that periodically collects perf data, eg:
> while true
> do
> perf stat -e cache-misses,cache-misses,cache-misses -C 1 sleep 2
> perf stat -e cache-misses -C 1 sleep 2
> sleep 1
> done
>
> 2, run another one to capture the IPC, eg:
> perf stat -e cycles:D,instructions:D -C 1 -I 1000
>
> Then we could observe that the counter used by cycles:D changes frequently:
> crash> struct cpu_hw_events.n_events,assign,event_list,events ffff88bf7f44f420
> n_events = 3
> assign = {33, 1, 32, 0, 1, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
> event_list = {0xffff88bf77b85000, 0xffff88b72db82000, 0xffff88b72db85800, 0xffff88ff6cfcb000, 0xffff88ff609f1800, 0xffff88ff609f1800, 0xffff88ff5f46a800, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
> events = {0x0, 0xffff88b72db82000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0xffff88b72db85800, 0xffff88bf77b85000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
>
> crash> struct cpu_hw_events.n_events,assign,event_list,events ffff88bf7f44f420
> n_events = 6
> assign = {33, 3, 32, 0, 1, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
> event_list = {0xffff88bf77b85000, 0xffff88b72db82000, 0xffff88b72db85800, 0xffff88bf46c34000, 0xffff88bf46c35000, 0xffff88bf46c30000, 0xffff88ff5f46a800, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
> events = {0xffff88bf46c34000, 0xffff88bf46c35000, 0xffff88bf46c30000, 0xffff88b72db82000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0xffff88b72db85800, 0xffff88bf77b85000, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
The above is unreadable; consider rewriting it in something other than a
raw data dump.
> The reason is that NMI watchdog permanently consumes one FP,
> so cycles can only use one GP, and its hweight is 5,
> but the hweight of other events (cache misses) is 4,
> so the counter used by cycles will be frequently taken away,
> resulting in unnecessary pmu_stop/start.
And the solution..... ? Or should I try and reverse engineer from the
proposed patch? If I do, how do I know it is what you intended?
next prev parent reply other threads:[~2022-02-11 11:37 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-02-10 4:39 [RFC PATCH] perf/x86: improve the event scheduling to avoid unnecessary pmu_stop/start Wen Yang
2022-02-11 11:36 ` Peter Zijlstra [this message]
-- strict thread matches above, loose matches on Subject: below --
2022-02-17 2:45 Wen Yang
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20220211113643.GT23216@worktop.programming.kicks-ass.net \
--to=peterz@infradead.org \
--cc=acme@kernel.org \
--cc=alexander.shishkin@linux.intel.com \
--cc=bp@alien8.de \
--cc=hpa@zytor.com \
--cc=jolsa@redhat.com \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-perf-users@vger.kernel.org \
--cc=mark.rutland@arm.com \
--cc=mingo@redhat.com \
--cc=namhyung@kernel.org \
--cc=simon.wy@alibaba-inc.com \
--cc=tglx@linutronix.de \
--cc=x86@kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.