[alsa-devel] Audio buffer mmap-ing on sh arch vs. cache aliasing problem

[Takashi Iwai]

At Fri, 26 Sep 2008 12:25:17 +0100,
Pawel MOLL wrote:
> 
> > > IIRC, a similar buffer handling (via vmalloc) is used in video
> > > drivers.  I suppose they don't work as well, right?
> 
> Well, our FB driver is aware of the problem. I can't say much about
> other ones...

FB driver has some generic API for that, defined in asm/fb.h.

> > > DRM driver uses __vmalloc() with PAGE_KERNEL_NOCACHE, but it's only
> > > for PPC32 non-coherent.
> 
> Hm. PPC with non-coherent cache?
> 
> /*
>  * Only on coherent architectures, we can mmap the status and the control records
>  * for effcient data transfer.  On others, we have to use HWSYNC ioctl...
>  */
> #if defined(CONFIG_X86) || defined(CONFIG_PPC) || defined(CONFIG_ALPHA)
>                                    ^^^^^^^^^^
> /*
>  * mmap status record
>  */
> static struct page * snd_pcm_mmap_status_nopage(struct vm_area_struct *area,
>                                                 unsigned long address, int *type)

Yeah, it's a known problem, and I already have patches.
But, it's still under discussion how to solve it better.
I'm going to post the theme to linux-arch, which James Bottomley and
others suggested in the previous discussions.

> > x86 does it also, via its PAGE_AGP definition.
> > 
> > There are not that many platforms that define PAGE_KERNEL_NOCACHE though,
> > so this gets a bit messy..
> 
> Oh, bloody hell, I haven't realised that!
> 
> > Yeah, that's not sexy, but maybe the only working case right now
> > > (better with arch-specific ifdefs).
> 
> The question is what ifdefs to use ;-) Because the cached/uncached
> decision must be consistent between snd_pcm_default_mmap() and buffer
> allocation...

Right.  Both must have matching ifdefs.  The vmalloc buffer handling
can be in sound/core common part, so that it won't be scattered in
different sub directories.

Or, let's define a common generic API (oh let me dream on)

> What a mess! ;-)

Indeed ;)


Takashi