Re: [Linux-fbdev-devel] ___ioremap() / memremap() question

[Paul M. <le...@Ch...>]

On Sat, Aug 11, 2001 at 01:58:17AM -0400, David T Eger wrote:
> I'm developing a framebuffer driver for embedded PowerPC where the
> framebuffer resides in system RAM.  I'd like to follow what the arm branch
> does in their sa1100fb.c driver and do the following:
> 
> (1) call alloc_pages(GFP_KERNEL|GFP_DMA, order) to get a chunk of ram
> (2) call free_page() repeatedly on the extra that I won't use
> (3) call mem_map_reserve() on all of the pages I am using
> (4) call memremap() to remap this to a high virtual address that is marked
> non-cacheable
> (5) poke around at this remapped area in my graphics code.
> 
> I tried this, but am getting data TLB misses.  Am I missing
> something?  I'm running on a 2.4.2 derived kernel that's been patched by
> Montavista.
> 
Which embedded PowerPC is this?

For MPC823 (RPXLite and LinuxPlanet) the base address is also retrieved from
a chunk of memory, but it's done a little differently then how SA-1100 FB and
what you're describing do it.

The size is the line length * width, rounded to the nearest full page, and
then handed off to alloc_bootmem_pages(). The result from that becomes the
base address of the frame buffer. (Calling virt_to_phys() and such where
necessary).

The only catch is is that allocation should happen relatively early. In the
case of the MPC823, page allocation was done in m8xx_setup_arch().

Regards,

-- 
Paul Mundt <le...@ch...>