Re: [PATCH: x86] Add dma_mmap_coherent()

[Russell King <rmk@arm.linux.org.uk>]

On Sun, Mar 28, 2004 at 01:36:33PM +0100, Russell King wrote:
> On Sun, Mar 28, 2004 at 03:40:47AM -0800, William Lee Irwin III wrote:
> > I think we're just missing an API for it e.g. break_compound_page();
> > I don't see why we shouldn't arrange for it. Alternatively, clearing
> > PG_compound by hand is possible. Though I guess it raises the question
> > of what we're trying to do; e.g. why not allocate pieces separately?
> 
> dma_alloc_coherent() is supposed to allocate a contiguous memory area.
> 
> However, x86 can get around this problem if it uses the ->nopage
> method of mapping the pages into memory.
> 
> This all brings up one fundamental question: what really are the
> semantics of dma_alloc_coherent() - if it returns a kernel direct
> mapped address, are the pages supposed to be marked reserved or not.
> 
> And this leads on to this question: if the pages are not marked
> reserved, how do we mmap() the pages given that remap_page_range()
> won't touch them?

Furthermore, should these pages contribute to the processes RSS?
If pages are faulted in via ->nopage, and the pages are not marked
reserved, they will contribute to the RSS.  I suspect that reviewing
existing implementations will reveal some which do and some which
don't - certainly ALSA marks pages reserved, and thus does not
contribute to RSS.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 PCMCIA      - http://pcmcia.arm.linux.org.uk/
                 2.6 Serial core