From: Larry Woodman <lwoodman@redhat.com>
Date: Thu, 13 Mar 2008 14:10:11 -0500
Subject: [misc] allow hugepage allocation to use most of memory
Message-id: 47D97C13.80104@redhat.com
O-Subject: [RHEL5 patch] Allow hugepage allocation to use most of memory.
Bugzilla: 438889
RH-Acked-by: Pete Zaitcev <zaitcev@redhat.com>
RH-Acked-by: Rik van Riel <riel@redhat.com>

In RHEL5-U2 we included 2 patches that conflict and potentially prevent
hugepages from using all or most of memory.   This can result in database
restart failures when the shared cache/SGA is large enough to consume most
of the RAM and hugepages are requested.

1.) linux-2.6-ppc64-unequal-allocation_of_huge_pages.patch added the
alloc_pages_thisnode() routine which builds a private zonelist that includes
only the zones on the node passed in as the "nid" argument.

2.) linux-2.6-mm-make-zonelist-order-selectable-in-numa.patch adds a the
boot cmdline argument numa_zonelist_order to allow you to select the default
zonelist ordered by zones or nodes.

If all of the zones for a given node are not contiguous in the zonelist
alloc_pages_thisnode() will terminate pre-maturely and build a private
zonelist
that does not include all of the nodes for the specified zone.  This can
result in
the system failing to allocate most of the memory for hugepages even
though it
is free.

The attached patch fixes this problem:

 include/linux/gfp.h |   12 ++++++------
 1 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index ab4fb53..f35b414 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h
@@ -114,7 +114,7 @@ static inline struct page *alloc_pages_thisnode(int nid, gfp_t gfp_mask,
 {
 	struct zonelist *zl;
 	struct zonelist thisnode_zl;
-	int i;
+	int i, j;
 
 	if (unlikely(order >= MAX_ORDER))
 		return NULL;
@@ -131,12 +131,12 @@ static inline struct page *alloc_pages_thisnode(int nid, gfp_t gfp_mask,
 	if (zl->zones[0]->zone_pgdat->node_id != nid)
 		return NULL;
 
-	for (i = 0; zl->zones[i] != NULL; i++) {
-		if (zl->zones[i]->zone_pgdat->node_id != nid)
-			break;
-		thisnode_zl.zones[i] = zl->zones[i];
+	/* make zonelist with every zone on this node and null terminate */
+	for (i = 0, j = 0; zl->zones[i] != NULL; i++) {
+		if (zl->zones[i]->zone_pgdat->node_id == nid)
+			thisnode_zl.zones[j++] = zl->zones[i];
 	}
-	thisnode_zl.zones[i] = NULL;
+	thisnode_zl.zones[j] = NULL;
 
 	return __alloc_pages(gfp_mask, order, &thisnode_zl);
 }