CONSISTENT_DMA_SIZE limitations

CONSISTENT_DMA_SIZE limitations

[Magnus Damm]

Hi everyone,

Are there any special CONSISTENT_DMA_SIZE limitiations in the latest
upstrem ARM kernel?

On SH-Mobile ARM a single 256 MB I/O window is mapped virt:phys 1:1 at
0xe6000000 -> 0xf5ffffff at ->map_io() time.

The virtual area on top of the I/O window at 0xf6000000 -> 0xffdfffff
seems suitable for consistent dma mappings. Extending
CONSISTENT_DMA_SIZE from the default 2MB to 158 MB makes efficient use
of the entire range. Empty 2MB space is left at the top as usual.

Extending CONSISTENT_DMA_SIZE seems like a good idea to me, but I'm
not sure if doing so will break something else. Perhaps I need to
rework some code in arch/arm/mm/dma-mapping.c?

I'm asking because most platforms only export up to 14 MB, and
Documentation/arm/memory.txt mentions 0xff000000 -> 0xffbfffff as
reserved.

Thanks!

/ magnus

CONSISTENT_DMA_SIZE limitations

[Eric Miao]

On Wed, Apr 28, 2010 at 2:25 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
> Hi everyone,
>
> Are there any special CONSISTENT_DMA_SIZE limitiations in the latest
> upstrem ARM kernel?
>
> On SH-Mobile ARM a single 256 MB I/O window is mapped virt:phys 1:1 at
> 0xe6000000 -> 0xf5ffffff at ->map_io() time.
>
> The virtual area on top of the I/O window at 0xf6000000 -> 0xffdfffff
> seems suitable for consistent dma mappings. Extending
> CONSISTENT_DMA_SIZE from the default 2MB to 158 MB makes efficient use
> of the entire range. Empty 2MB space is left at the top as usual.
>
> Extending CONSISTENT_DMA_SIZE seems like a good idea to me, but I'm
> not sure if doing so will break something else. Perhaps I need to
> rework some code in arch/arm/mm/dma-mapping.c?
>

Hi Magnus,

For a hackish way of enabling something, I don't think that's a bad idea
to extend CONSISTENT_DMA_SIZE. And it's often seen by framebuffer
and video/graphics related drivers, which is normally hungry for large
amount of consistent memory. However, I'm a little bit concerned about
such usage, (esp. I'm not sure how this is done in ARMv7), my preference
would be:

1) either use the dma API
2) or reserve a large chunk of memory during boot and get it remapped later

> I'm asking because most platforms only export up to 14 MB, and
> Documentation/arm/memory.txt mentions 0xff000000 -> 0xffbfffff as
> reserved.
>
> Thanks!
>
> / magnus
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel at lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
>

CONSISTENT_DMA_SIZE limitations

[Shilimkar, Santosh]

> -----Original Message-----
> From: linux-arm-kernel-bounces at lists.infradead.org [mailto:linux-arm-kernel-
> bounces at lists.infradead.org] On Behalf Of Eric Miao
> Sent: Wednesday, April 28, 2010 12:17 PM
> To: Magnus Damm
> Cc: linux-arm-kernel at lists.infradead.org
> Subject: Re: CONSISTENT_DMA_SIZE limitations
> 
> On Wed, Apr 28, 2010 at 2:25 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
> > Hi everyone,
> >
> > Are there any special CONSISTENT_DMA_SIZE limitiations in the latest
> > upstrem ARM kernel?
> >
> > On SH-Mobile ARM a single 256 MB I/O window is mapped virt:phys 1:1 at
> > 0xe6000000 -> 0xf5ffffff at ->map_io() time.
> >
> > The virtual area on top of the I/O window at 0xf6000000 -> 0xffdfffff
> > seems suitable for consistent dma mappings. Extending
> > CONSISTENT_DMA_SIZE from the default 2MB to 158 MB makes efficient use
> > of the entire range. Empty 2MB space is left at the top as usual.
> >
> > Extending CONSISTENT_DMA_SIZE seems like a good idea to me, but I'm
> > not sure if doing so will break something else. Perhaps I need to
> > rework some code in arch/arm/mm/dma-mapping.c?
> >
> 
> Hi Magnus,
> 
> For a hackish way of enabling something, I don't think that's a bad idea
> to extend CONSISTENT_DMA_SIZE. And it's often seen by framebuffer
> and video/graphics related drivers, which is normally hungry for large
> amount of consistent memory. However, I'm a little bit concerned about
> such usage, (esp. I'm not sure how this is done in ARMv7), my preference
> would be:
> 
> 1) either use the dma API
> 2) or reserve a large chunk of memory during boot and get it remapped later
> 
> > I'm asking because most platforms only export up to 14 MB, and
> > Documentation/arm/memory.txt mentions 0xff000000 -> 0xffbfffff as
> > reserved.

Not the best way but works is :
You can override the default setting of 2MB to whatever memory you need and
then use the dma_alloc_coherent(), to allocate from this space.

Just ensure that you don't ioremap this memory with different cache-attribute
otherwise you will end up in the alias mapping of same memory which is not
allowed on newer ARMv7.

Regards,
Santosh

CONSISTENT_DMA_SIZE limitations

[Magnus Damm]

On Wed, Apr 28, 2010 at 3:47 PM, Eric Miao <eric.y.miao@gmail.com> wrote:
> On Wed, Apr 28, 2010 at 2:25 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
>> Extending CONSISTENT_DMA_SIZE seems like a good idea to me, but I'm
>> not sure if doing so will break something else. Perhaps I need to
>> rework some code in arch/arm/mm/dma-mapping.c?
>>
>
> Hi Magnus,
>
> For a hackish way of enabling something, I don't think that's a bad idea
> to extend CONSISTENT_DMA_SIZE. And it's often seen by framebuffer
> and video/graphics related drivers, which is normally hungry for large
> amount of consistent memory. However, I'm a little bit concerned about
> such usage, (esp. I'm not sure how this is done in ARMv7), my preference
> would be:
>
> 1) either use the dma API
> 2) or reserve a large chunk of memory during boot and get it remapped later

Hi Eric,

So we already have a bunch of multimedia-related drivers upstream for
SH, and they often make use of dma_alloc_coherent() and friends.
Chunks of physically contiguous memory are allocated during boot and
passed on to the platform drivers as a struct resource. This is true
for the LCDC fbdev and the CEU v4l2 code, as well as all the uio
devices. To avoid fragmentation issues the v4l2 driver also creates a
memory pool from the struct resource using
dma_declare_coherent_memory().

Eventually I'd like to tie in the IOMMU, but first i'd like to enable
the existing drivers by extending CONSISTENT_DMA_SIZE.

Thanks for your comments!

/ magnus

CONSISTENT_DMA_SIZE limitations

[Magnus Damm]

Hi Santosh,

On Wed, Apr 28, 2010 at 4:07 PM, Shilimkar, Santosh
<santosh.shilimkar@ti.com> wrote:
>> On Wed, Apr 28, 2010 at 2:25 PM, Magnus Damm <magnus.damm@gmail.com> wrote:
>> > Extending CONSISTENT_DMA_SIZE seems like a good idea to me, but I'm
>> > not sure if doing so will break something else. Perhaps I need to
>> > rework some code in arch/arm/mm/dma-mapping.c?

> Not the best way but works is :
> You can override the default setting of 2MB to whatever memory you need and
> then use the dma_alloc_coherent(), to allocate from this space.
>
> Just ensure that you don't ioremap this memory with different cache-attribute
> otherwise you will end up in the alias mapping of same memory which is not
> allowed on newer ARMv7.

Good idea. =) Thanks for your suggestions.

/ magnus