Re: [Dri-devel] Re: agp_ioremap() Implementation

[Michel <mic...@ii...>]

Chris Ahna wrote:
> 
> On Fri, Aug 17, 2001 at 09:29:20PM +0200, Benjamin Herrenschmidt wrote:
> > Hrm... who exactly touch the AGP memory ? I beleive both the in-kernel
> > DRM driver and the userland part. They must take care of it. Also, it
> > looks like the r128 driver also reads from the AGP aperture for the
> > ring pointer. That can be handled with a simple flush, the cache line
> > must be properly invalidated, and any possible contention between
> > card access and CPU access solved, which might even be impossible.
> 
> I think you're right, bother user and kernel level code touches AGP
> memory.  However, it seems that a <driver>_flush_write_combine() is
> issued every time the driver wants the card to pick up new data.  On
> most hardware, <driver>_flush_write_combine() will push all
> writecombined transactions out to memory.  It would probably be tough to
> make this efficiently flush exactly the memory you need though.

At least r128_flush_write_combine() is defined as mb(), does that serve the
purpose on PPC?


-- 
Earthling Michel Dänzer (MrCooper)    \   Debian GNU/Linux (powerpc) developer
CS student, Free Software enthusiast   \        XFree86 and DRI project member