[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20180124121114.GA17605@krava>
Date: Wed, 24 Jan 2018 13:11:14 +0100
From: Jiri Olsa <jolsa@...hat.com>
To: Jiri Olsa <jolsa@...nel.org>
Cc: Peter Zijlstra <a.p.zijlstra@...llo.nl>,
Ingo Molnar <mingo@...nel.org>,
lkml <linux-kernel@...r.kernel.org>,
Namhyung Kim <namhyung@...nel.org>,
David Ahern <dsahern@...il.com>,
Andi Kleen <ak@...ux.intel.com>,
Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
Andy Lutomirski <luto@...capital.net>,
Arnaldo Carvalho de Melo <acme@...nel.org>
Subject: Re: [RFC 00/21] perf tools: Add user data delayed processing
changing wrong subject :-\
On Wed, Jan 24, 2018 at 12:51:22PM +0100, Jiri Olsa wrote:
> hi,
> this RFC contains change to delay sample's user space
> data retrieval into task work, originally described and
> discussed by Peter and Ingo in here [1].
>
> This patchset tries to follow the original patch with
> some kernel changes (described below) and perf tool
> support included.
>
> Basically we allow the NMI event code to skip user data
> retrieval and schedule task work to do it, before the
> task resumes.
>
> Using the task work limits the window where we can do
> this. We can trigger the delayed task work only if the
> taskwork gets executed before the process executes again
> after NMI, because we need its stack as it was in NMI.
>
> That leaves us with window during the slow syscall path
> (check task_struct::perf_user_data_allowed in patches).
>
> The slow syscall processing is forced for task when
> the user data event is enabled, which makes the task
> slower.
>
> On the other hand I noticed roughly 100us drop in NMI
> processing times, which I plotted in here [2].
>
> Not sure it's worth to introduce this processing, which adds
> more processing time and does not show much improvement. On
> the other hand IIRC Peter mentioned it'd be nice to get user
> space data retrieval out of NMI.
>
> Also you guys could think of some other better/faster way ;-)
>
> NOTE I also implemented putting the user stack data into
> delayed processing, which showed nicer numbers. But it's
> little more tricky and brings more changes into this already
> big patchset. The logic stays, so I did not include it to
> keep the patchset simple.
>
> Also available in:
> https://git.kernel.org/pub/scm/linux/kernel/git/jolsa/perf.git
> perf/user_data
>
> thanks for comments,
> jirka
>
> [1] https://marc.info/?l=linux-kernel&m=150098372819938&w=2
> [2] http://people.redhat.com/~jolsa/ud-bench.png
>
> ---
> Jiri Olsa (21):
> perf tools: Add perf_evsel__is_sample_bit function
> perf tools: Add perf_sample__process function
> perf tools: Add callchain__printf for pure callchain dump
> perf tools: Add perf_sample__copy|free functions
> perf: Add TIF_PERF_USER_DATA bit
> perf: Add PERF_RECORD_USER_DATA event processing
> perf: Add PERF_SAMPLE_USER_DATA_ID sample type
> perf: Add PERF_SAMPLE_CALLCHAIN to user data event
> perf: Export running sample length values through debugfs
> perf tools: Sync perf_event.h uapi header
> perf tools: Add perf_sample__parse function
> perf tools: Add struct parse_args arg to perf_sample__parse
> perf tools: Add support to parse user data event
> perf tools: Add support to dump user data event info
> perf report: Add delayed user data event processing
> perf record: Enable delayed user data events
> perf script: Add support to display user data events
> perf script: Add support to display user data ID
> perf script: Display USER_DATA misc char for sample
> perf report: Add user data processing stats
> perf report: Add --stats=ud option to display user data debug info
>
> arch/x86/entry/common.c | 6 +++
> arch/x86/events/core.c | 18 ++++++++
> arch/x86/events/intel/ds.c | 4 +-
> arch/x86/include/asm/thread_info.h | 4 +-
> include/linux/init_task.h | 4 +-
> include/linux/perf_event.h | 3 ++
> include/linux/sched.h | 20 ++++++++
> include/uapi/linux/perf_event.h | 34 +++++++++++++-
> kernel/events/core.c | 283 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++--
> tools/include/uapi/linux/perf_event.h | 34 +++++++++++++-
> tools/perf/Documentation/perf-script.txt | 3 +-
> tools/perf/builtin-record.c | 2 +
> tools/perf/builtin-report.c | 301 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------
> tools/perf/builtin-script.c | 98 +++++++++++++++++++++++++++++++++++++++
> tools/perf/perf.h | 1 +
> tools/perf/util/event.c | 1 +
> tools/perf/util/event.h | 9 ++++
> tools/perf/util/evsel.c | 118 +++++++++++++++++++++++++++++++++++++----------
> tools/perf/util/evsel.h | 5 ++
> tools/perf/util/session.c | 60 +++++++++++++++++++-----
> tools/perf/util/thread.c | 1 +
> tools/perf/util/thread.h | 16 +++++++
> tools/perf/util/tool.h | 1 +
> 23 files changed, 954 insertions(+), 72 deletions(-)
Powered by blists - more mailing lists