Re: [PATCH] MIPS: Fixups of ALSA memory maps

[Takashi Iwai <tiwai@suse.de>]

At Wed, 18 Nov 2009 15:20:56 +0100,
Ralf Baechle wrote:
> 
> On Tue, Nov 17, 2009 at 10:29:10AM +0100, Takashi Iwai wrote:
> 
> > > > Actually, this has been a looong-standing problem.
> > > > I have a series of patches to fix these issues, but it's more
> > > > intensively involved with dma_*() functions.
> > > > 
> > > > The patches can be found in test/dma-fix branch of sound GIT tree.
> > > >   git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound-2.6.git test/dma-fix
> > > > 
> > > > This basically adds dma_mmap_coherent() function to feasible
> > > > architectures, which is already implemented for ARM, so far.
> > > 
> > > Cool - but needs a little further tweaking to work right.  That's a
> > > solution which will use uncached accesses on all MIPS systems.
> > > 
> > > IP27/IP35-family machines will explode when you try that.  Eventually the
> > > cache coherency logic will notice that cache, directory caches and memory
> > > have become inconsistent and bombard the CPU with a bunch of nasty
> > > exceptions.
> > 
> > OK, that's really bad.
> 
> Hardware designers do such things to you.  Often even for a reason.

Heh, not only software engineers are so crazy ;)

> > > For cache-coherent machines otoh it's a big waste of performance.
> > > 
> > > int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
> > >                       void *cpu_addr, dma_addr_t handle, size_t size)
> > > {
> > >         struct page *pg;
> > > 
> > > 	if (!plat_device_is_coherent(dev))
> > > 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
> > >         cpu_addr = (void *)dma_addr_to_virt(handle);
> > >         pg = virt_to_page(cpu_addr);
> > > 
> > >         return remap_pfn_range(vma, vma->vm_start,
> > >                                page_to_pfn(pg) + vma->vm_pgoff,
> > >                                size, vma->vm_page_prot);
> > > }
> > > EXPORT_SYMBOL(dma_mmap_coherent);
> > > 
> > > Thomas - you're the IP28 specialist.  Would the plat_device_is_coherent()
> > > above have to become a cpu_is_noncoherent_r10000() call?  Any further
> > > nasties?
> > 
> > Thanks for checking!
> 
> You're welcome!
> 
> So basically I'd not mind putting this into the Linux/MIPS tree; we still
> can iron out the kinks from there on and probably much better than by
> having arch stuff in the ALSA tree.

Yes, that's the very purpose of my patchset.

> I recall this new API having been posted for discussion to linux-arch.
> What was the outcome?  I'd only like to add a new API if the other arch
> maintainers see it fit their needs also.

Well, we haven't reached the consensus.  The discussion faded away
somehow mainly because I had too little time to update and ping people
again.

In Tokyo, I talked with some guys regarding this.  Ben agreed to take
this approach for ppc, and David said that he doesn't mind for sparc
part.   Fujita-san mentioned it's no big problem to add one op from
the generic dma_ops.

So, maybe somehow need to convince James in the end (and ask Paul to
check SH part, too), then it'll be all up... theoretically :)

Anyway, I'm going to raise the discussion again on linux-arch.
I'm afraid it's a bit too late game for 2.6.33, but starting now is
better than too late again.


thanks,

Takashi