[Linaro-mm-sig] [RFC] ARM DMA mapping TODO, v1

[Jerome Glisse]

On Fri, Apr 29, 2011 at 9:42 AM, Joerg Roedel <joro at 8bytes.org> wrote:
> On Thu, Apr 28, 2011 at 03:37:00PM -0400, Jerome Glisse wrote:
>> As Jesse pointed out already, for performance reasons it's lot better
>> if you let the driver decide even if you have an iommu capable of
>> handling coherency for you. My understanding is that each time
>> coherency is asked for it trigger bus activities of some kind (i think
>> snoop is the term used for pci) this traffic can slow down both the
>> cpu and the device. For graphic driver we have a lot of write once and
>> use (once or more) buffer and it makes a lot of sense to have those
>> buffer allocated using uncached memory so we can tell the device (in
>> case of drm driver) that there is no need to trigger snoop activities
>> for coherency. So i believe the decision should ultimately be in the
>> driver side.
>
> Stupid question: Couldn't these write-once-read-often buffers just stay
> in the memory of the GPU instead of refetching them every time from main
> memory? Or is that necessary because of the limited space on some GPUs?
>
> Regards,
>
>        Joerg
>

We might be talking about several G of data, so using system is not
uncommon. Also when uploading data to GPU vram is better to let the
GPU do dma from system memory rather than having the CPU do memcpy.

Cheers,
Jerome