mmap & no cache regions

mmap & no cache regions

[Steve Rossi]

Hi All -

I'm writing a driver for custom hardware on an 8xx based system. I'm
using MontaVista's Journeyman 2.0 (2.4.2 kernel). The driver uses
consistent_alloc to allocate a region of memory which is used for DMA. I
mmap that region to user space (using remap_page_range) such that the
user-level application can directly access the data that has been DMA'd
from the device. I learned that consistent_alloc is not sufficent to
guarantee cache coherency on the mmaped region - in fact I had to mark
these pages PAGE_NO_CACHE in the user space application's memory map as
well. Having figured that out, it all works great now - but my question
is: do I need to pay attention to how I deallocate and unmap that
space? I use consistent_free in the driver so that takes care of marking
the pages cacheable again in the kernel level mapping - but is it safe
to assume that the application's memory map will be destroyed so I don't
have to worry about marking the pages in that map as cacheable again?

Thanks,
Steve

--
-------------------------------------------------------
Steven K. Rossi                     srossi@labs.mot.com
Staff Engineer
Multimedia Communications Research Laboratory
Motorola Labs
-------------------------------------------------------


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/

Re: mmap & no cache regions

[Dan Malek]

Steve Rossi wrote:

> .....I learned that consistent_alloc is not sufficent to
> guarantee cache coherency on the mmaped region

Well, it has no clue about what you are doing.  It's purpose is to
ensure the kernel's mapping of that page is set properly.

> .... - in fact I had to mark
> these pages PAGE_NO_CACHE in the user space application's memory map as
> well.

Yep, that's because the kernel and applications have their own page
tables and page attributes.

> .....- but is it safe
> to assume that the application's memory map will be destroyed so I don't
> have to worry about marking the pages in that map as cacheable again?

I was about to write a couple of pages when it is best to answer this
with a simple "yes."  When the application releases this space, the page
tables are simply destroyed.  The kernel does it's weird attribute change
because it doesn't really "release" the page as an application does.  That's
all I'll say for now :-).

Thanks.


	-- Dan


** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/