Re: can device drivers return non-ram via vm_ops->nopage?

[William Lee Irwin III <wli@holomorphy.com>]

On Sat, Mar 20, 2004 at 11:54:45PM +0000, Russell King wrote:
> The issues are:
> 1. ALSA wants to mmap the buffer used to transfer data to/from the
>    card into user space.  This buffer may be direct-mapped RAM,
>    memory allocated via dma_alloc_coherent(), an on-device buffer,
>    or anything else.
>    The user space mapping must likewise be DMA-coherent.
>    Currently, ALSA just does virt_to_page() on whatever address it
>    feels like in its nopage() function, which is obviously not
>    acceptable for two out of the three specific cases above.
> 2. ALSA wants to _coherently_ share data between the kernel-side
>    drivers, and user space ALSA library, mainly the DMA buffer
>    head/tail pointers so both kernel space and user space knows
>    when the buffer is full/empty.

Okay, so we've got these pinned down. So I've got two small ideas
(I mentioned them earlier, but maybe vger dropped the message):

(a) I think prefaulting should work for that in general, though the API
doesn't fit the extra things needed for e.g. DMA. Is there some way we
could extend remap_area_pages() (or provide an alternative interface to
similar functionality with the missing pieces included) to do the extra
things needed to make the coherency and/or DMA (or whatever else is
missing) work?

(b) Alternatively, would dma_coherent_to_pfn() instead of
dma_coherent_to_page() and making ->nopage() return pfns help salvage
the method using non-cachable and/or dma-coherent page protections in
vma->vm_page_prot?


-- wli