Re: AGP and PAT (induced?) problem (on AMD family 6)

[Venki Pallipadi <venkatesh.pallipadi@intel.com>]

On Wed, Aug 20, 2008 at 02:46:49PM -0700, Dave Airlie wrote:
> On Thu, Aug 21, 2008 at 7:40 AM, Rene Herman <rene.herman@keyaccess.nl> wrote:
> > On 20-08-08 21:41, Venki Pallipadi wrote:
> >
> >> OK. I have reproduced this list size issue locally and this order 1
> >> allocation and set_memory_uc on that allocation is actually coming
> >> from agp_allocate_memory() -> agp_generic_alloc_page() ->
> >> map_page_into_agp() agp_allocate_memory breaks higher order page
> >> requests into order 1 allocs.
> >>
> >> On my system I see multiple agp_allocate_memory requests for nrpages 8841,
> >> 1020, 16, 2160, 2160, 8192. Together they end up resulting in more than 22K
> >> entries in PAT pages.
> >
> > Okay, thanks for the confirmation.
> >
> > Now, how to fix...
> >
> > Firstly, it seems we can conclude that any expectancy of a short PAT list is
> > simply destroyed by AGP. I believe the best thing migh be to look into
> > "fixing" AGP rather than PAT for now?
> >
> > In a sense the entire purpose of the AGP GART is collecting non contiguous
> > pages but given that in practice it's generally still just one or at most a
> > few regions, going to multi-page allocs sounds most appetising to me.
> >
> > All in tree AGP drivers except sgi-agp use agp_generic_alloc_page(), ali via
> > m1541_alloc_page and i460 via i460_alloc_page.
> 
> In the future we will be getting more smaller AGP allocs, so the other
> problem needs a fix as well.
> 
> http://git.kernel.org/?p=linux/kernel/git/airlied/agp-2.6.git;a=shortlog;h=agp-pageattr2
> 
> contains some code I started on before that moves the interfaces
> around,  Shaohua has been looking at
> it as it needs the changes to the set_pages interface as well, which
> is where I ran out of time/steam last time.
> 
> However with alloc/free pages we could change to a higher order
> allocation function as long as it fell back to lower
> orders internally.
> 

Yes. Atleast during the bootup, we should be able to get higher order pages.
And if we cannot get that, agp can internally fall back to zero order pages.
That will reduce the number of times set_memory_* gets called, even without
pat.

We are also looking at changing the reserve_memtype in PAT, not to use linked
list for RAM backed pages and track them in page struct. That way we will be
using the list only for reserved region which should still be less in
number and agp set_memory_* call will not have the list overhead.

BTW, Rene: Did the patch from yday, caching the last list add, help in
eliminating the slowdown in X startup?

Thanks,
Venki