Message-ID: <20180504154406.djxq6wqil3dwutvw@flea>
Date:   Fri, 4 May 2018 17:44:06 +0200
From:   Maxime Ripard <maxime.ripard@...tlin.com>
To:     Paul Kocialkowski <paul.kocialkowski@...tlin.com>
Cc:     linux-media@...r.kernel.org, devicetree@...r.kernel.org,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        linux-sunxi@...glegroups.com,
        Mauro Carvalho Chehab <mchehab@...nel.org>,
        Rob Herring <robh+dt@...nel.org>,
        Mark Rutland <mark.rutland@....com>,
        Chen-Yu Tsai <wens@...e.org>, Pawel Osciak <pawel@...iak.com>,
        Marek Szyprowski <m.szyprowski@...sung.com>,
        Kyungmin Park <kyungmin.park@...sung.com>,
        Hans Verkuil <hans.verkuil@...co.com>,
        Sakari Ailus <sakari.ailus@...ux.intel.com>,
        Philipp Zabel <p.zabel@...gutronix.de>,
        Arnd Bergmann <arnd@...db.de>,
        Alexandre Courbot <acourbot@...omium.org>,
        Tomasz Figa <tfiga@...omium.org>
Subject: Re: [PATCH v2 09/10] ARM: dts: sun7i-a20: Add Video Engine and
 reserved memory nodes

On Fri, May 04, 2018 at 03:57:48PM +0200, Paul Kocialkowski wrote:
> On Fri, 2018-05-04 at 15:40 +0200, Maxime Ripard wrote:
> > On Fri, May 04, 2018 at 02:04:38PM +0200, Paul Kocialkowski wrote:
> > > On Fri, 2018-05-04 at 11:15 +0200, Maxime Ripard wrote:
> > > > On Fri, May 04, 2018 at 10:47:44AM +0200, Paul Kocialkowski wrote:
> > > > > > > > > +			reg = <0x01c0e000 0x1000>;
> > > > > > > > > +			memory-region = <&ve_memory>;
> > > > > > > > 
> > > > > > > > Since you made the CMA region the default one, you don't
> > > > > > > > need
> > > > > > > > to
> > > > > > > > tie
> > > > > > > > it to that device in particular (and you can drop it being
> > > > > > > > mandatory
> > > > > > > > from your binding as well).
> > > > > > > 
> > > > > > > What if another driver (or the system) claims memory from
> > > > > > > that
> > > > > > > zone
> > > > > > > and
> > > > > > > that the reserved memory ends up not being available for the
> > > > > > > VPU
> > > > > > > anymore?
> > > > > > > 
> > > > > > > Acccording to the reserved-memory documentation, the
> > > > > > > reusable
> > > > > > > property
> > > > > > > (that we need for dmabuf) puts a limitation that the device
> > > > > > > driver
> > > > > > > owning the region must be able to reclaim it back.
> > > > > > > 
> > > > > > > How does that work out if the CMA region is not tied to a
> > > > > > > driver
> > > > > > > in
> > > > > > > particular?
> > > > > > 
> > > > > > I'm not sure to get what you're saying. You have the property
> > > > > > linux,cma-default in your reserved region, so the behaviour
> > > > > > you
> > > > > > described is what you explicitly asked for.
> > > > > 
> > > > > My point is that I don't see how the driver can claim back (part
> > > > > of)
> > > > > the
> > > > > reserved area if the area is not explicitly attached to it.
> > > > > 
> > > > > Or is that mechanism made in a way that all drivers wishing to
> > > > > use
> > > > > the
> > > > > reserved memory area can claim it back from the system, but
> > > > > there is
> > > > > no
> > > > > priority (other than first-come first-served) for which drivers
> > > > > claims
> > > > > it back in case two want to use the same reserved region (in a
> > > > > scenario
> > > > > where there isn't enough memory to allow both drivers)?
> > > > 
> > > > This is indeed what happens. Reusable is to let the system use the
> > > > reserved memory for things like caches that can easily be dropped
> > > > when
> > > > a driver wants to use the memory in that reserved area. Once that
> > > > memory has been allocated, there's no claiming back, unless that
> > > > memory segment was freed of course.
> > > 
> > > Thanks for the clarification. So in our case, perhaps the best fit
> > > would
> > > be to make that area the default CMA pool so that we can be ensured
> > > that
> > > the whole 96 MiB is available for the VPU and that no other consumer
> > > of
> > > CMA will use it?
> > 
> > The best fit for what use case ? We already discussed this, and I
> > don't see any point in having two separate CMA regions. If you have a
> > reasonably sized region that will accomodate for both the VPU and
> > display engine, why would we want to split them?
> 
> The use case I have in mind is boilerplate use of the VPU with the
> display engine, say with DMAbuf.
> 
> It wasn't exactly clear in my memory whether we had decided that the CMA
> pool we use for the VPU should also be used for other CMA consumers (I
> realize that this is what we've been doing all along by having
> linux,cma-default; though).
> 
> The fact that the memory region will accomodate for both the display
> engine and the VPU is not straightforward IMO and I think this has to be
> an explicit choice that we take. I was under the impression that we
> chose the 96 MiB because that's what the Allwinner reference code does.
> Does the reference code also use this pool for display?

Yes

> I liked the idea of having 96 MiB fully reserved to the VPU because it
> allows us to provide a limit on the use case, such as "this guarantees N
> buffers for resolution foo in format bar". If the display engine also
> uses it, then the limit also depends on how many GEM buffers are
> allocated (unless I'm missing something).

This also guarantees that you take away a fifth of the RAM on some
boards. If we had yet another CMA pool of 64MB as is the multi_v7
defconfig, that's a third of your RAM that's gone, possibly for no
particular reason.

If we make the math, let's say that we are running a system with 4
planes used in 1080p, with 4 Bpp, in double buffering (which is
already an unlikely setup). Let's add on top of that that we're
decoding a 1080p video with 8 buffers pre-allocated with 2Bpp (in
YUV422). Which really seems extreme now :)

And we're at 80MB. My guess is that your memory bus is going to be
dead before you need to use all that memory.

Maxime

-- 
Maxime Ripard, Bootlin (formerly Free Electrons)
Embedded Linux and Kernel engineering
https://bootlin.com

Download attachment "signature.asc" of type "application/pgp-signature" (834 bytes)

Powered by blists - more mailing lists