At Fri, 25 Mar 2011 21:23:26 +1100,
Benjamin Herrenschmidt wrote:

On Fri, 2011-03-25 at 11:12 +0100, Takashi Iwai wrote:

quoted

At Fri, 25 Mar 2011 20:15:33 +1100,
Benjamin Herrenschmidt wrote:

quoted

quoted

quoted

We
must also make sure we don't go down that path for vmalloc memory
though.

Yes.

I haven't actually checked, but I assume that the test

substream->dma_buffer.dev.type == SNDRV_DMA_TYPE_DEV

In snd_pcm_default_mmap() takes care of that, please correct me if
I'm wrong in which case we'll need something else there.

Well, in the case of usb-audio, it's not handled via
dma_mmap_coherent(), as the page isn't allocated via
dma_alloc_coherent() but vmalloc().

Right, I just wanted to make sure I was right to assume that a page
allocated by vmalloc() was going to fail the above test in 
snd_pcm_default_mmap() and thus -not- get into dma_mmap_coherent()...
just double checking as I'm not totally familiar with the intricacies of
the pcm code :-)

quoted

The bad commit was to overcome some problems on SH platform, IIRC,
when it's used with dmix -- i.e. concurrent accesses on the mmapped
buffer from multiple processes.  But, this looks obviously like a
wrong approach.

Is this a vivt architecture ? Maybe enforcing some restrictions on the
virtual addresses so they hit the same cache congruence classes ?

Possibly, yes.  But we need to re-test the problem first with the
recent kernel.

quoted

Actually, the buffer allocated there in usb-audio is an intermediate
buffer, that isn't directly transferred to hardware.  We may need a
bit more consideration what is the best way to solve that issue (if
it's still really present).

Right. I wouldn't expect vmalloc stuff to hit HW in most cases anyways,
though I do wonder why you don't pass the buffer directly to the HCD and
avoid that intermediate step but that's a completely different
question :-)

The direct buffer was used partly in the older versions.  But there
was a problem in the accuracy of the buffer playback position, so this
was switched to full double-buffering.


thanks,

Takashi