Andrew,
Mel's latest V9 regressed slightly for ia64 FLATMEM+VIRTUAL_MEM_MAP. When
the largest hole is greater than LARGE_GAP, vmem_map is allocated before
free_area_init_nodes; resultant crash follows. Rather than complicate
alloc_node_mem_map just for this ia64 case, add an adjustment to node_mem_map
which is later negated by alloc_node_mem_map.
Previous to V9, the mem_map adjustment was done in the scope where allocation
is achieved in alloc_node_mem_map. The current code is more appropriate but
unfortunately caused an issue for ia64.
Please add this to the next -mm.
thanks,
bob
Acked-by: Mel Gorman <[email protected]>
Signed-off-by: Bob Picco <[email protected]>
arch/ia64/mm/contig.c | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
Index: linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c
===================================================================
--- linux-2.6.18-rc4-mm3.orig/arch/ia64/mm/contig.c 2006-08-28 13:10:00.000000000 -0400
+++ linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c 2006-08-28 18:18:54.000000000 -0400
@@ -252,7 +252,12 @@ paging_init (void)
vmem_map = (struct page *) vmalloc_end;
efi_memmap_walk(create_mem_map_page_table, NULL);
- NODE_DATA(0)->node_mem_map = vmem_map;
+ /*
+ * alloc_node_mem_map makes an adjustment for mem_map
+ * which isn't compatible with vmem_map.
+ */
+ NODE_DATA(0)->node_mem_map = vmem_map +
+ find_min_pfn_with_active_regions();
free_area_init_nodes(max_zone_pfns);
printk("Virtual mem_map starts at 0x%p\n", mem_map);
Bob Picco wrote:
> Andrew,
>
> Mel's latest V9 regressed slightly for ia64 FLATMEM+VIRTUAL_MEM_MAP. When
> the largest hole is greater than LARGE_GAP, vmem_map is allocated before
> free_area_init_nodes; resultant crash follows. Rather than complicate
> alloc_node_mem_map just for this ia64 case, add an adjustment to node_mem_map
> which is later negated by alloc_node_mem_map.
>
> Previous to V9, the mem_map adjustment was done in the scope where allocation
> is achieved in alloc_node_mem_map. The current code is more appropriate but
> unfortunately caused an issue for ia64.
>
> Please add this to the next -mm.
>
> thanks,
>
> bob
>
> Acked-by: Mel Gorman <[email protected]>
> Signed-off-by: Bob Picco <[email protected]>
>
> arch/ia64/mm/contig.c | 7 ++++++-
> 1 file changed, 6 insertions(+), 1 deletion(-)
>
> Index: linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c
> ===================================================================
> --- linux-2.6.18-rc4-mm3.orig/arch/ia64/mm/contig.c 2006-08-28 13:10:00.000000000 -0400
> +++ linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c 2006-08-28 18:18:54.000000000 -0400
> @@ -252,7 +252,12 @@ paging_init (void)
> vmem_map = (struct page *) vmalloc_end;
> efi_memmap_walk(create_mem_map_page_table, NULL);
>
> - NODE_DATA(0)->node_mem_map = vmem_map;
> + /*
> + * alloc_node_mem_map makes an adjustment for mem_map
> + * which isn't compatible with vmem_map.
> + */
Bob, which adjustment is this that is incompatible? Is it the one in
the final stanza, the FLATMEM mem_map instantiation? This one?
#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
mem_map -= pgdat->node_start_pfn;
#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
-apw
> + NODE_DATA(0)->node_mem_map = vmem_map +
> + find_min_pfn_with_active_regions();
> free_area_init_nodes(max_zone_pfns);
>
> printk("Virtual mem_map starts at 0x%p\n", mem_map);
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
Andy Wihitcroft wrote: [Sat Sep 09 2006, 09:03:09PM EDT]
> Bob Picco wrote:
> > Andrew,
> >
> > Mel's latest V9 regressed slightly for ia64 FLATMEM+VIRTUAL_MEM_MAP. When
> > the largest hole is greater than LARGE_GAP, vmem_map is allocated before
> > free_area_init_nodes; resultant crash follows. Rather than complicate
> > alloc_node_mem_map just for this ia64 case, add an adjustment to node_mem_map
> > which is later negated by alloc_node_mem_map.
> >
> > Previous to V9, the mem_map adjustment was done in the scope where allocation
> > is achieved in alloc_node_mem_map. The current code is more appropriate but
> > unfortunately caused an issue for ia64.
> >
> > Please add this to the next -mm.
> >
> > thanks,
> >
> > bob
> >
> > Acked-by: Mel Gorman <[email protected]>
> > Signed-off-by: Bob Picco <[email protected]>
> >
> > arch/ia64/mm/contig.c | 7 ++++++-
> > 1 file changed, 6 insertions(+), 1 deletion(-)
> >
> > Index: linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c
> > ===================================================================
> > --- linux-2.6.18-rc4-mm3.orig/arch/ia64/mm/contig.c 2006-08-28 13:10:00.000000000 -0400
> > +++ linux-2.6.18-rc4-mm3/arch/ia64/mm/contig.c 2006-08-28 18:18:54.000000000 -0400
> > @@ -252,7 +252,12 @@ paging_init (void)
> > vmem_map = (struct page *) vmalloc_end;
> > efi_memmap_walk(create_mem_map_page_table, NULL);
> >
> > - NODE_DATA(0)->node_mem_map = vmem_map;
> > + /*
> > + * alloc_node_mem_map makes an adjustment for mem_map
> > + * which isn't compatible with vmem_map.
> > + */
>
> Bob, which adjustment is this that is incompatible? Is it the one in
> the final stanza, the FLATMEM mem_map instantiation? This one?
Andy, yes this is the one.
>
> #ifdef CONFIG_ARCH_POPULATES_NODE_MAP
> if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
> mem_map -= pgdat->node_start_pfn;
> #endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
>
> -apw
bob
>
> > + NODE_DATA(0)->node_mem_map = vmem_map +
> > + find_min_pfn_with_active_regions();
> > free_area_init_nodes(max_zone_pfns);
> >
> > printk("Virtual mem_map starts at 0x%p\n", mem_map);
> > -
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to [email protected]
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at http://www.tux.org/lkml/
>