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;
From:   Alexander Shishkin <alexander.shishkin@linux.intel.com>
To:     Mathieu Poirier <mathieu.poirier@linaro.org>
Cc:     Peter Zijlstra <peterz@infradead.org>,
        Arnaldo Carvalho de Melo <acme@kernel.org>,
        Ingo Molnar <mingo@redhat.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Alexander Shishkin <alexander.shishkin@linux.intel.com>,
        schwidefsky@de.ibm.com, heiko.carstens@de.ibm.com,
        Will Deacon <will.deacon@arm.com>,
        Mark Rutland <mark.rutland@arm.com>,
        Jiri Olsa <jolsa@redhat.com>,
        Namhyung Kim <namhyung@kernel.org>,
        Adrian Hunter <adrian.hunter@intel.com>, ast@kernel.org,
        Greg KH <gregkh@linuxfoundation.org>,
        "H. Peter Anvin" <hpa@zytor.com>, linux-s390@vger.kernel.org,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        linux-arm-kernel <linux-arm-kernel@lists.infradead.org>
Subject: Re: [PATCH 5/6] perf/core: Use ioctl to communicate driver configuration to kernel
In-Reply-To: <CANLsYkw4w-zRczcb2uddO4UsxXrCCoXCSkPko2it-OJX2H0ecA@mail.gmail.com>
References: <1530570810-28929-1-git-send-email-mathieu.poirier@linaro.org> <1530570810-28929-6-git-send-email-mathieu.poirier@linaro.org> <20180703100348.fy43f4fosw3fdc6i@um.fi.intel.com> <20180703105652.psjq5cq4t7o2j6zk@um.fi.intel.com> <CANLsYkw4w-zRczcb2uddO4UsxXrCCoXCSkPko2it-OJX2H0ecA@mail.gmail.com>
User-Agent: Notmuch/0.27 (http://notmuchmail.org) Emacs/25.2.2 (x86_64-pc-linux-gnu)
Date:   Wed, 04 Jul 2018 13:34:41 +0300
Message-ID: <87fu0z70hq.fsf@ashishki-desk.ger.corp.intel.com>
MIME-Version: 1.0
Content-Type: text/plain
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Mathieu Poirier <mathieu.poirier@linaro.org> writes:

> On Tue, 3 Jul 2018 at 04:57, Alexander Shishkin
> <alexander.shishkin@linux.intel.com> wrote:
>>
>> On Tue, Jul 03, 2018 at 01:03:48PM +0300, Alexander Shishkin wrote:
>> > On Mon, Jul 02, 2018 at 04:33:29PM -0600, Mathieu Poirier wrote:
>> > > +/*
>> > > + * PMU driver configuration works the same way as filter management above,
>> > > + * but without the need to deal with memory mapping.  Driver configuration
>> > > + * arrives through the SET_DRV_CONFIG ioctl() where it is validated and applied
>> > > + * to the event.  When the PMU is ready it calls perf_event_drv_config_sync() to
>> > > + * bring the configuration information within reach of the PMU.
>> >
>> > Wait a second. The reason why we dance around with the generations of filters
>> > is the locking order of ctx::mutex vs mmap_sem. In an mmap path, where we're
>> > notified about mapping changes, we're called under the latter, and we'd need
>> > to grab the former to update the event configuration. In your case, the
>> > update comes in via perf_ioctl(), where we're already holding the ctx::mutex,
>> > so you can just kick the PMU right there, via an event_function_call() or
>> > perf_event_stop(restart=1). In the latter case, your pmu::start() would just
>> > grab the new configuration. Should also be about 90% less code. :)
>>
>> Also, since it affects the AUX buffer configuration, it is probably a one
>> time ioctl command that you issue before you mmap the buffer. If that's
>> the case, you don't even have to worry about stopping the event, as it
>> shouldn't be running, because without the buffer perf_aux_output_begin()
>> should fail and so should the pmu::add() iirc.
>>
>
> The idea behind the current approach was to make the SET_DRV_CONFIG
> ioctl() usable by other drivers where multiple ioctl() calls could be
> performed while a session in ongoing.  I also opted to introduce a
> _sync() function to let the PMU refresh its configuration at the time
> of its own choosing rather than having to interrupt the session.

Yes, but the times of PMU's own choosing would still be more or less
limited to ->start()/->stop().

You can also do an event_function_call(), which would call
->config_sync(), which would be free to decide what to do with the new
information, up to and including doing a ->stop()/->start() sequence. My
guess is that you'd want to do either of the following:
 * decide to apply the new configuration immediately, and do the
 start-stop thing,
 * decide to defer the new configuration until the next ->start().

Both should work via cross call directly from the ioctl() call.

> But all I need for coresight is to have available the sink information
> and PMU configuration (in an upcoming patchset) by the time
> setup_aux() is called.  You are correct, this is a one time
> configuration and since the event isn't running there is no need for
> locking - I should be able to access the PMU when the ioctl is called.
>
> If you are fine with this bare-bone scenario and don't care much about
> usability in different situation, I'll do a respin with minimal
> functionality that cover my needs.

It doesn't have to be bare-bones, what I'm saying is that you shouldn't
need the event->drv_config, as you can directly call
pmu::config(new_config) (or config_sync(), but I'm guessing the _sync
part is redundant if you don't keep the configuration in 2 parts) from
the ioctl() and it should cover all your bases.

Regards,
--
Alex