Re: ALSA vs. non coherent DMA

[Takashi Iwai <tiwai@suse.de>]

At Thu, 08 May 2008 07:53:11 +1000,
Benjamin Herrenschmidt wrote:
> 
> On Wed, 2008-05-07 at 09:22 -0500, Timur Tabi wrote:
> > Takashi Iwai wrote:
> > 
> > > This is a mmap of the data record to be shared in realtime with apps.
> > > The app updates its data pointer (appl_ptr) on the mmapped buffer
> > > while the driver updates the data (e.g. DMA position, called hwptr) on
> > > the fly on the mmapped record.  Due to its real-time nature, it has to
> > > be coherent -- at least, it was a problem on ARM.
> > 
> > This doesn't sound like a coherency problem to me, and least not one you'd find
> > on PowerPC.  Both the driver and the application run on the host CPU, so there
> > shouldn't be any coherency problem.  My understanding is that a "non coherent"
> > platform is one where the host CPU isn't aware when a *hardware device* writes
> > directly to memory, e.g. via DMA.
> 
> Yes, precisely. I was about to make a reply here. There is some
> confusion at least in terminology, in Alsa. This is not DMA coherency,
> though it is a problem with virtually tagged data caches that some archs
> such as ARM have.

Right.  The words should be corrected.
Since the only way to get a certain non-cached map was the (ab-)use of
dma_mmap_coherent(), such a confusing wording was chosen.

> So this is ok for all PowerPC since they all have a physically tagged
> data cache.

OK, so that part should work as is for PPC.

> The real problem -is- still the DMA coherency issue and as I see it, is
> two fold:
> 
>  - mmap'ing of the result of dma_alloc_coherent() doesn't work. There
> are two issues at play here, one is the pgprot that -must- be set to
> uncached for such a mapping on non coherent architectures (and non
> coherent architectures only), and the other is our virt_to_page() that
> will puke on virtual addresses coming from dma_alloc_coherent().

And dma_mmap_coherent() would be a solution for it, I suppose.

>  - mmap'ing of SG lists for non coherent DMA. There the problem is a
> mixture of how Alsa allocate the SG buffers mixes with the previous
> problem.

Yes.

> I think it's never valid to create an SG list with the output of
> dma_alloc_coherent though. We would need a dma_alloc_sg() for that...
> 
> sglists are made of pages, thus allocated with GFP, and later DMA mapped
> with dma_map_*, however this brings a whole other set of issues/constra
> ints such as bouce bufferring on some MMU less platforms if the memory
> happens to come out of the wrong place. Also, such mapped buffers are
> -not- coherent as they must not be modified via their virtual address
> while mapped, -unless- they are also mapped in kernel and/or user space
> (vmap & mmap) using some kind of "coherent" attributes such as
> pgprot_noncached. (and provided that is possible at all in kernel place
> for archs like MIPS).
> 
> I don't have an easy answer there, it seems the bogosity roots deep in
> alsa, at least for the SG bits. For the non-SG bits, we can probably
> work around with an accessor to get the right pgprot and maybe some
> variant of virt_to_page() (dma_virt_to_page() ?) that would walk the
> kernel page tables to obtain the pfn.

The vmap() in sound/core/sgbuf.c can be omitted by adding proper PCM
callbacks (copy and silent) to handle SG-buffers.  These are only guys
that access the linear buffer runtime->area.

Then we'll just need a proepr mmap PCM callback just calling
dma_mmap_coherent() for each SG page.  Also, the default PCM mmap
should be fixed to use dma_mmap_coherent() appropriately.  That's
all.

So, what we really need is dma_mmap_coherent() implementations...


thanks,

Takashi