Re: [RFC/PATCH]Allow agp memory allocations to use node information

[Christoph Hellwig <hch@infradead.org>]

not a lot of description here, he? :-)


On Wed, May 12, 2004 at 11:47:39AM -0700, Mike Werner wrote:
> diff -u linux-2.6.6.orig/drivers/char/agp/agp.h
> linux-2.6.6/drivers/char/agp/agp.h
> --- linux-2.6.6.orig/drivers/char/agp/agp.h     Sun May  9 19:32:28 2004
> 
> +++ linux-2.6.6/drivers/char/agp/agp.h  Wed May 12 11:43:19 2004
> @@ -111,7 +111,7 @@
>         int (*remove_memory)(struct agp_memory *, off_t, int);
>         struct agp_memory *(*alloc_by_type) (size_t, int);
>         void (*free_by_type)(struct agp_memory *);
> -       void *(*agp_alloc_page)(void);
> +       void *(*agp_alloc_page)(int);
>         void (*agp_destroy_page)(void *);

Wrong interface.  We need to pass a struct agp_bridge_data to
each of the methods anyway for proper support for multiple AGP
bridges in a single system.  And once you have that you can simply
use the node id your patches stores in agp_bridge_data.

> +++ linux-2.6.6/drivers/char/agp/intel-mch-agp.c        Wed May 12
> 11:16:02 2004
> @@ -43,7 +43,7 @@
>         if (pg_count != 1)
>                 return NULL;
> 
> -       addr = agp_bridge->driver->agp_alloc_page();
> +       addr = agp_bridge->driver->agp_alloc_page(AGPGART_DEFAULT_NODE);

wrong default again.  for the agp bridges that have an associated pci
device (aka about everything except hpzx and alpha IIRC) you should
use pcibus_to_cpumask (or what the thing was called).  For the others
you should try to contact the maintainers for a proper choice.