Received-SPF: pass (google.com: domain of linux-kernel+bounces-34580-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:40f1:3f00::1 as permitted sender) client-ip=2604:1380:40f1:3f00::1;
Precedence: bulk
MIME-Version: 1.0
References: <20240117031212.1104034-1-nunes.erico@gmail.com>
 <20240117031212.1104034-5-nunes.erico@gmail.com> <20240121095604.2368-1-hdanton@sina.com>
 <CAKGbVbtww65w+sr7OiccawRkGt-p3MRWaWNU=nkQTaqdyMxqVA@mail.gmail.com> <CAK4VdL2oRsvCMAWsMyvT+MpPz5e=7=mVtMWAbJKopLEOM0_A+w@mail.gmail.com>
In-Reply-To: <CAK4VdL2oRsvCMAWsMyvT+MpPz5e=7=mVtMWAbJKopLEOM0_A+w@mail.gmail.com>
From: Qiang Yu <yuq825@gmail.com>
Date: Tue, 23 Jan 2024 09:18:02 +0800
Message-ID: <CAKGbVbu18v_f4N0nGQAKmTZa5Dijy7u1GMqOrggthrXCsYzyRg@mail.gmail.com>
Subject: Re: [PATCH v1 4/6] drm/lima: handle spurious timeouts due to high irq latency
To: Erico Nunes <nunes.erico@gmail.com>
Cc: Hillf Danton <hdanton@sina.com>, dri-devel@lists.freedesktop.org, 
	lima@lists.freedesktop.org, David Airlie <airlied@gmail.com>, 
	Daniel Vetter <daniel@ffwll.ch>, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Sun, Jan 21, 2024 at 11:11=E2=80=AFPM Erico Nunes <nunes.erico@gmail.com=
> wrote:
>
> On Sun, Jan 21, 2024 at 12:20=E2=80=AFPM Qiang Yu <yuq825@gmail.com> wrot=
e:
> >
> > On Sun, Jan 21, 2024 at 5:56=E2=80=AFPM Hillf Danton <hdanton@sina.com>=
 wrote:
> > >
> > > On Wed, 17 Jan 2024 04:12:10 +0100 Erico Nunes <nunes.erico@gmail.com=
>
> > > >
> > > > @@ -401,9 +399,33 @@ static enum drm_gpu_sched_stat lima_sched_time=
dout_job(struct drm_sched_job *job
> > > >       struct lima_sched_pipe *pipe =3D to_lima_pipe(job->sched);
> > > >       struct lima_sched_task *task =3D to_lima_task(job);
> > > >       struct lima_device *ldev =3D pipe->ldev;
> > > > +     struct lima_ip *ip =3D pipe->processor[0];
> > > > +
> > > > +     /*
> > > > +      * If the GPU managed to complete this jobs fence, the timeou=
t is
> > > > +      * spurious. Bail out.
> > > > +      */
> > > > +     if (dma_fence_is_signaled(task->done_fence)) {
> > > > +             DRM_WARN("%s spurious timeout\n", lima_ip_name(ip));
> > > > +             return DRM_GPU_SCHED_STAT_NOMINAL;
> > > > +     }
> > >
> > > Given 500ms in lima_sched_pipe_init(), no timeout is spurious by defi=
ne,
> > > and stop selling bandaid like this because you have options like loca=
ting
> > > the reasons behind timeout.
> >
> > This chang do look like to aim for 2FPS apps. Maybe 500ms is too short
> > for week mali4x0 gpus (2FPS apps appear more likely). AMD/NV GPU uses
> > 10s timeout. So increasing the timeout seems to be an equivalent and be=
tter
> > way?
>
> Indeed 500ms might be too optimistic for the sort of applications that
> users expect to run on this hardware currently. For a more similar
> reference though, other embedded drivers like v3d and panfrost do
> still set it to 500ms. Note that this patch is just exactly the same
> as exists in Panfrost today and was already discussed with some common
> arguments in the patches of this series:
> https://patchwork.freedesktop.org/series/120820/
>
> But I would agree to bump the timeout to a higher value for lima. Some
> distributions are already doing this with module parameters anyway to
> even be able to run some more demanding application stacks on a Mali
> 400.
>
> Another thing we might consider (probably in a followup patchset to
> not delay these fixes forever for the people hitting this issue) is to
> configure the Mali hardware watchdog to the value that we would like
> as a timeout. That way we would get timeout jobs going through the
> same error irq path as other hardware error jobs and might be able to
> delete(?)/simplify this software timeout code.
>
This way should be much simpler and stable.

> In the meantime for v2 of this series I'll make the change to account
> for the multiple pp irqs. So do you agree it is ok to leave
> drm_sched_stop() as it is at least for this series?
>
I'm OK with this.

> Thanks all for the reviews
>
> Erico