Re: mmap() implementation for pci_alloc_consistent() memory?

[Leon Woestenberg]

Hello Clemens, Takashi,

On Fri, May 20, 2011 at 10:17 AM, Takashi Iwai <tiwai@xxxxxxx> wrote:
> At Fri, 20 May 2011 08:51:03 +0200,
> Clemens Ladisch wrote:
>>
>> Leon Woestenberg wrote:
>> > On Thu, May 19, 2011 at 5:58 PM, Clemens Ladisch <clemens@xxxxxxxxxx> wrote:
>> >>> On Thu, May 19, 2011 at 12:14:40AM +0200, Leon Woestenberg wrote:
>> >>> >     vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
>> >>
>> >> So is this an architecture without coherent caches?
>> >
>> > My aim is to have an architecture independent driver.
>>
>> Please note that most MMU architectures forbid mapping the same memory
>> with different attributes, so you must use pgprot_noncached if and only
>> if dma_alloc_coherent actually uses it.  Something like the code below.
>>
>> And I'm not sure if you have to do some additional cache flushes when
>> mapping on some architectures.
>>
>> >> Or would you want to use pgprot_dmacoherent, if available?
>> >
>> > Hmm, let me check that.
>>
>> It's available only on ARM and Unicore32.
>>
>> There's also dma_mmap_coherent(), which does exactly what you want if
>> your buffer is physically contiguous, but it's ARM only.
>> Takashi tried to implement it for other architectures; I don't know
>> what came of it.
>
> PPC got this recently (thanks to Ben), but still missing in other
> areas.
>
> There was little uncertain issue on MIPS, and it looks difficult to
> achieve it on PA-RISC at all.  The development was stuck due to lack
> of time since then.
>
Thanks for all the insights, I wasn't aware there were arch-specific
calls that already solved the topic issue.

Having dma_mmap_coherent() there is good for one or two archs, but how
can we built portable drivers if the others arch's are still missing?

I assume this call is thus not officially DMA-API (yet)?

Clemens showed some pretty amazing preprocessor #if(def)s to cater for
the all the different arch's and their mapping/cache-coherency
behaviour, but that's not something I would like to put in a driver.

How would dma_mmap_coherent() look like on x86?


Regards,
-- 
Leon
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/