[linuxsh-dev] Problems with remap_page_range()

[David M. <Dav...@st...>]

Folks,
	I've been trying to get PCI sound working on our platform, and I've
come across some interesting problems. I'm using a standard soundblaster AC97.
(driver is in es1371.c).  The problem is with applications that
mmap() the sound device. The sound driver call pci_alloc_consistent(), which
returns a P2 pointer. This is then passed into virt_to_page(), which gets it
wrong, as it get the physical page by the __pa() macro which works subtracting
from PAGE_OFFSET.

To fix this I've changed __pa() and __va() to do PHYSADDR()
and P1SEGADDR(). This fixes the problem nicely. However, what I'm not sure
about is the cachability, when we call remap_page_range(), the pages will be
marked as cacheable. I don't think this is what you want to do! I think we need
something like pgprot_noncached(), or perhaps pgprot_pci(), if you are
remapping memory that comes from the pci_alloc() routines.

The other problem of course is that if you do intend to make them cacheable,
the dreaded D-cache alias raises it's ugly head again. I haven't started
looking
at this yet, but unfortunately I've got a driver to port that expects this to
work. I suspect the only solution is to adjust the VMA to match the alignment
of the memory being remapped. Anybody got any other ideas?


-- 
Dave McKay 
Software Engineer STMicroelectronics
Email: dav...@st...