Index: linux-2.6.16-rc1-mm3/mm/page_alloc.c
===================================================================
--- linux-2.6.16-rc1-mm3.orig/mm/page_alloc.c	2006-01-25 18:29:02.000000000 -0800
+++ linux-2.6.16-rc1-mm3/mm/page_alloc.c	2006-01-25 18:29:18.000000000 -0800
@@ -12,6 +12,7 @@
  *  Zone balancing, Kanoj Sarcar, SGI, Jan 2000
  *  Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
  *          (lots of bits borrowed from Ingo Molnar & Andrew Morton)
+ *  Page zeroing by Christoph Lameter, Silicon Graphics, Inc, January 2006
  */
 
 #include <linux/config.h>
@@ -794,19 +795,28 @@ static struct page *buffered_rmqueue(str
 {
 	unsigned long flags;
 	struct page *page;
-	int cold = !!(gfp_flags & __GFP_COLD);
+	int area, list;
 	int cpu;
 
+	/* Determine the per cpu cache and the area to use */
+	list = PER_CPU_HOT;
+	area = FREE_AREA_DEFAULT;
+	if (gfp_flags & __GFP_COLD)
+		list = PER_CPU_COLD;
+	if (gfp_flags & __GFP_ZERO) {
+		list = PER_CPU_ZERO;
+		area = FREE_AREA_ZEROED;
+	}
 again:
 	cpu  = get_cpu();
 	if (likely(order == 0)) {
 		struct per_cpu_pages *pcp;
 
-		pcp = &zone_pcp(zone, cpu)->pcp[cold];
+		pcp = &zone_pcp(zone, cpu)->pcp[list];
 		local_irq_save(flags);
 		if (!pcp->count) {
 			pcp->count += rmqueue_bulk(zone, 0,
-						pcp->batch, &pcp->list, FREE_AREA_DEFAULT);
+						pcp->batch, &pcp->list, area);
 			if (unlikely(!pcp->count))
 				goto failed;
 		}
@@ -815,7 +825,21 @@ again:
 		pcp->count--;
 	} else {
 		spin_lock_irqsave(&zone->lock, flags);
-		page = __rmqueue(zone, order, FREE_AREA_DEFAULT);
+		page = __rmqueue(zone, order, area);
+
+		/*
+		 * If we failed to obtain a zero and/or unzeroed page
+		 * then we may still be able to obtain the page from
+		 * another area.
+		 */
+		if (!page) {
+			if (area == FREE_AREA_ZEROED) {
+				page = __rmqueue(zone, order, FREE_AREA_DEFAULT);
+				list = PER_CPU_COLD;
+			} else
+				page = __rmqueue(zone, order, FREE_AREA_ZEROED);
+		}
+
 		spin_unlock(&zone->lock);
 		if (!page)
 			goto failed;
@@ -830,7 +854,11 @@ again:
 	if (prep_new_page(page, order))
 		goto again;
 
-	if (gfp_flags & __GFP_ZERO)
+	/*
+	 * If we were able to get the page from a prezeroed list then
+	 * we can avoid zeroing here.
+	 */
+	if ((gfp_flags & __GFP_ZERO) && list != PER_CPU_ZERO)
 		prep_zero_page(page, order, gfp_flags);
 
 	if (order && (gfp_flags & __GFP_COMP))
@@ -1458,7 +1486,7 @@ void si_meminfo_node(struct sysinfo *val
 
 #define K(x) ((x) << (PAGE_SHIFT-10))
 
-const char *temperature_descr[] = { "cold", "hot" };
+const char *temperature_descr[] = { "cold", "hot", "zero" };
 
 /*
  * Show free area list (used inside shift_scroll-lock stuff)
@@ -1952,6 +1980,12 @@ inline void setup_pageset(struct per_cpu
 	pcp->high = 2 * batch;
 	pcp->batch = max(1UL, batch/2);
 	INIT_LIST_HEAD(&pcp->list);
+
+	pcp = &p->pcp[PER_CPU_ZERO];
+	pcp->count = 0;
+	pcp->high = 2 * batch;
+	pcp->batch = max(1UL, batch/2);
+	INIT_LIST_HEAD(&pcp->list);
 }
 
 /*
Index: linux-2.6.16-rc1-mm3/include/linux/mmzone.h
===================================================================
--- linux-2.6.16-rc1-mm3.orig/include/linux/mmzone.h	2006-01-25 17:15:18.000000000 -0800
+++ linux-2.6.16-rc1-mm3/include/linux/mmzone.h	2006-01-25 18:29:05.000000000 -0800
@@ -22,16 +22,18 @@
 #define MAX_ORDER CONFIG_FORCE_MAX_ZONEORDER
 #endif
 
-#define NR_PER_CPU_PAGES 2
+#define NR_PER_CPU_PAGES 3
 
 /* Types of per cpu pages */
 #define PER_CPU_HOT 0
 #define PER_CPU_COLD 1
+#define PER_CPU_ZERO 2
 
-#define NR_FREE_AREAS 1
+#define NR_FREE_AREAS 2
 
 /* Types of free areas */
 #define FREE_AREA_DEFAULT 0
+#define FREE_AREA_ZEROED 1
 
 struct free_area {
 	struct list_head	free_list;