Fix off by one in pfn_valid

From Jim Paradis <jparadis@redhat.com>

When I gave proposed the fix to pfn_valid() for RHEL4, Stephen Tweedie's
sharp eyes caught this:

--- quote ---
> -#define pfn_valid(pfn)          ((pfn) < num_physpages)
> +#define pfn_valid(pfn) ((pfn) >= num_physpages ? 0 : \
> +                     ({ u8 nid__ = pfn_to_nid(pfn); \
> +                        nid__ != 0xff && (pfn) >= node_start_pfn(nid__) && (pfn) <= node_end_pfn(nid__); }))

That's an off-by-one error, isn't it?  Looking at some of the setup code
for i386 srat.c:

        start_pfn = paddr >> PAGE_SHIFT;
        end_pfn = (paddr + size) >> PAGE_SHIFT;

which gets pushed unmodified into the node data; so end_pfn is the pfn
*after* the last valid one, isn't it?  If so,

        (pfn) <= node_end_pfn(nid__)

needs to use "<", not "<=".  Other code in x86_64 numa also seems to be
making the assumption that size = end_pfn - start_pfn.

--- end quote ---

(even though x86_64 node_end_pfn() uses node_spanned_pages rather than
end_pfn, the issue is exactly the same; node_end_pfn(nid) points to
the page *after* the end of the node).


Signed-off-by: Andi Kleen <ak@suse.de>

Index: linux/include/asm-x86_64/mmzone.h
===================================================================
--- linux.orig/include/asm-x86_64/mmzone.h
+++ linux/include/asm-x86_64/mmzone.h
@@ -54,7 +54,7 @@ static inline __attribute__((pure)) int 
 
 #define pfn_valid(pfn) ((pfn) >= num_physpages ? 0 : \
 			({ u8 nid__ = pfn_to_nid(pfn); \
-			   nid__ != 0xff && (pfn) >= node_start_pfn(nid__) && (pfn) <= node_end_pfn(nid__); }))
+			   nid__ != 0xff && (pfn) >= node_start_pfn(nid__) && (pfn) < node_end_pfn(nid__); }))
 #endif
 
 #define local_mapnr(kvaddr) \