Graphics memory management in ARM/Embedded Linux

[Alex Deucher]

2011/4/20 Tom Cooksey <tomcooksey at gmail.com>:
>
>
> 2011/4/20 microcai <microcai at fedoraproject.org>
>>
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> 于 2011年04月20日 20:39, Tom Cooksey 写道:
>> > Hi All,
>> >
>> > Please excuse the spam... just trying to spread the word: Linaro is
>> > currently developing a graphics memory manager for Linux which allows
>> > buffers to be shared between different devices and different processes.
>> > On
>> > ARM SoCs, many different hw devices need to access the same buffer
>> > (which
>> > E.g. GEM doesn't support). As a simple example, a GPU needs to write to
>> > a
>> > buffer which a display controller reads from. At least in ARM-Linux
>> > land,
>> > this is a fairly hot-topic and a dedicated mailing list has been created
>> > to
>> > discuss different requirements, figure out if TTM/GEM can be adapted,
>> > etc.
>> >
>> >
>> > The mailing list is here:
>> >
>> > http://lists.linaro.org/mailman/listinfo/linaro-mm-sig
>> >
>> >
>> > There's also a wiki page gathering requirements here:
>> >
>> >
>> > https://wiki.linaro.org/WorkingGroups/Middleware/Graphics/Projects/UnifiedMemoryManagement
>> >
>> >
>> > Hopefully there's people on the Wayland list which this will be
>> > interesting
>> > to. Would be cool if a Wayland compositor used the mechanism Linaro
>> > comes up
>> > with to share window buffers between different processes!
>> >
>> >
>> >
>> > Cheers,
>> >
>> > Tom
>> >
>> >
>> >
>>
>> Not again, due!
>> Does DRI has some mentally wrong that cann't be fixed other than
>> re-create?
>
> Yes, possibly. The existing memory managers in Linux (GEM & TTM) don't:
>
> a) Allow buffers to be shared between different device drivers

This being looked at by a number of people.

> b) Don't allow physically contiguous memory allocations (something simple
> hardware without MMUs need)

ttm supports discrete vram which is contiguous.

>
> So while I agree re-inventing something just for the sake of it is bad,
> that's not what's happening. In fact, I think there's a good chance TTM will
> be ripped out, stuck into its own driver (with its own device node /dev/ttm)
> and extended to meet everyone's requirements. At least there has been some
> feasibility work done by Dave Airlie to partially address issue (a):
> http://airlied.livejournal.com/71734.html.

There are a number of people interested in allowing buffers to be
shared between devices and writing a memory manager is a LOT of work.
Would it not make more sense to fix the short comings in ttm to do
what you need rather than writing a new one?

Alex

>
>
> Cheers,
>
> Tom
>
>
> _______________________________________________
> wayland-devel mailing list
> wayland-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/wayland-devel
>
>