omap3isp buffer alignment

omap3isp buffer alignment

[Michael Jones]

Hi Laurent,

If I understood your discussion with Russell [1] correctly, user pointer
buffers are required to be page-aligned because of the IOMMU API, and
it's desirable to keep the IOMMU driver that way for other subsystems
which may use it. So we're stuck with user buffers needing to be
page-aligned.

There's a check in ispvideo.c:isp_video_buffer_prepare() that the buffer
address is 32-byte aligned. Isn't this superfluous considering the
page-aligned restriction?

-Michael

[1]
http://www.mail-archive.com/linux-omap%40vger.kernel.org/msg50611.html

MATRIX VISION GmbH, Talstrasse 16, DE-71570 Oppenweiler
Registergericht: Amtsgericht Stuttgart, HRB 271090
Geschaeftsfuehrer: Gerhard Thullner, Werner Armingeon, Uwe Furtner, Erhard Meier

Re: omap3isp buffer alignment

[Sakari Ailus]

On Thu, Aug 11, 2011 at 11:57:04AM +0200, Michael Jones wrote:
> Hi Laurent,
> 
> If I understood your discussion with Russell [1] correctly, user pointer
> buffers are required to be page-aligned because of the IOMMU API, and
> it's desirable to keep the IOMMU driver that way for other subsystems
> which may use it. So we're stuck with user buffers needing to be
> page-aligned.

My understanding is that this is actually a hardware requirement. You only
can map pages of 4 kiB (at least).

> There's a check in ispvideo.c:isp_video_buffer_prepare() that the buffer
> address is 32-byte aligned. Isn't this superfluous considering the
> page-aligned restriction?

I guess the ISP driver isn't assuming that ispmmu_vmap always give page
aligned mappings --- or that the page size couls theoretically be smaller.
The assumptions might not hold in another implementation of the IOMMU API,
which however will be replaced (hopefully at some point) by the improved DMA
mapping API.

Cheers,

-- 
Sakari Ailus
sakari.ailus@iki.fi

Re: omap3isp buffer alignment

[Laurent Pinchart]

On Thursday 11 August 2011 23:21:45 Sakari Ailus wrote:
> On Thu, Aug 11, 2011 at 11:57:04AM +0200, Michael Jones wrote:
> > Hi Laurent,
> > 
> > If I understood your discussion with Russell [1] correctly, user pointer
> > buffers are required to be page-aligned because of the IOMMU API, and
> > it's desirable to keep the IOMMU driver that way for other subsystems
> > which may use it. So we're stuck with user buffers needing to be
> > page-aligned.
> 
> My understanding is that this is actually a hardware requirement. You only
> can map pages of 4 kiB (at least).

The IOMMU works on a 4kiB page granularity, but that doesn't require buffers 
to be page-aligned. If the buffer start address isn't aligned on a page 
boundary the IOMMU can map the whole page, and the driver can then just add an 
offset to the virtual address.

> > There's a check in ispvideo.c:isp_video_buffer_prepare() that the buffer
> > address is 32-byte aligned. Isn't this superfluous considering the
> > page-aligned restriction?
> 
> I guess the ISP driver isn't assuming that ispmmu_vmap always give page
> aligned mappings --- or that the page size couls theoretically be smaller.
> The assumptions might not hold in another implementation of the IOMMU API,
> which however will be replaced (hopefully at some point) by the improved
> DMA mapping API.

-- 
Regards,

Laurent Pinchart