Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220411221536.283312-1-dmitry.osipenko@collabora.com>
 <064d8958-a288-64e1-b2a4-c2302a456d5b@amd.com> <a04733af-5ff1-a1b4-527d-68b28a037231@collabora.com>
 <f6de2b9a-005f-d1f5-9818-cfbee2bdddc5@amd.com>
In-Reply-To: <f6de2b9a-005f-d1f5-9818-cfbee2bdddc5@amd.com>
From:   Erico Nunes <nunes.erico@gmail.com>
Date:   Wed, 13 Apr 2022 00:59:17 +0200
Message-ID: <CAK4VdL3VOtVGi36SY0TEL4P2jW33dM4TOmFXYmewE7cGNhY4Zw@mail.gmail.com>
Subject: Re: [PATCH v1] drm/scheduler: Don't kill jobs in interrupt context
To:     Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Cc:     Dmitry Osipenko <dmitry.osipenko@collabora.com>,
        David Airlie <airlied@linux.ie>,
        Daniel Vetter <daniel@ffwll.ch>,
        Tomeu Vizoso <tomeu.vizoso@collabora.com>,
        Steven Price <steven.price@arm.com>,
        Rob Herring <robh@kernel.org>,
        Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>,
        Rob Clark <robdclark@gmail.com>,
        Dmitry Osipenko <digetx@gmail.com>,
        linux-kernel <linux-kernel@vger.kernel.org>,
        dri-devel <dri-devel@lists.freedesktop.org>,
        Qiang Yu <yuq825@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Apr 12, 2022 at 9:41 PM Andrey Grodzovsky
<andrey.grodzovsky@amd.com> wrote:
>
>
> On 2022-04-12 14:20, Dmitry Osipenko wrote:
> > On 4/12/22 19:51, Andrey Grodzovsky wrote:
> >> On 2022-04-11 18:15, Dmitry Osipenko wrote:
> >>> Interrupt context can't sleep. Drivers like Panfrost and MSM are taking
> >>> mutex when job is released, and thus, that code can sleep. This results
> >>> into "BUG: scheduling while atomic" if locks are contented while job is
> >>> freed. There is no good reason for releasing scheduler's jobs in IRQ
> >>> context, hence use normal context to fix the trouble.
> >>
> >> I am not sure this is the beast Idea to leave job's sw fence signalling
> >> to be
> >> executed in system_wq context which is prone to delays of executing
> >> various work items from around the system. Seems better to me to leave the
> >> fence signaling within the IRQ context and offload only the job freeing or,
> >> maybe handle rescheduling to thread context within drivers implemention
> >> of .free_job cb. Not really sure which is the better.
> > We're talking here about killing jobs when driver destroys context,
> > which doesn't feel like it needs to be a fast path. I could move the
> > signalling into drm_sched_entity_kill_jobs_cb() and use unbound wq, but
> > do we really need this for a slow path?
>
>
> You can't move the signaling back to drm_sched_entity_kill_jobs_cb
> since this will bring back the lockdep splat that 'drm/sched: Avoid
> lockdep spalt on killing a processes'
> was fixing.
>
> I see your point and i guess we can go this way too. Another way would
> be to add to
> panfrost and msm job a  work_item and reschedule to thread context from
> within their
> .free_job callbacks but that probably to cumbersome to be justified here.

FWIW since this mentioned individual drivers, commit 'drm/sched: Avoid
lockdep spalt on killing a processes' also introduced problems for
lima.
There were some occurrences in our CI
https://gitlab.freedesktop.org/mesa/mesa/-/jobs/20980982/raw .
Later I found it also reproducible on normal usage when just closing
applications, so it may be affecting users too.

I tested this patch and looks like it fixes things for lima.

Thanks

Erico