[PATCH] overcommit: use totalreserve_pages for nommu

This patch is an enhancement of OVERCOMMIT_GUESS algorithm in
__vm_enough_memory() in mm/nommu.c.

When the OVERCOMMIT_GUESS algorithm calculates the number of free pages,
the algorithm subtracts the number of reserved pages from the result
nr_free_pages().

Signed-off-by: Hideo Aoki <haoki@redhat.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

mm/nommu.c

@@ -1147,14 +1147,26 @@ int __vm_enough_memory(long pages, int cap_sys_admin)
 		 * only call if we're about to fail.
 		 */
 		n = nr_free_pages();
+
+		/*
+		 * Leave reserved pages. The pages are not for anonymous pages.
+		 */
+		if (n <= totalreserve_pages)
+			goto error;
+		else
+			n -= totalreserve_pages;
+
+		/*
+		 * Leave the last 3% for root
+		 */
 		if (!cap_sys_admin)
 			n -= n / 32;
 		free += n;
 
 		if (free > pages)
 			return 0;
-		vm_unacct_memory(pages);
-		return -ENOMEM;
+
+		goto error;
 	}
 
 	allowed = totalram_pages * sysctl_overcommit_ratio / 100;
@@ -1175,7 +1187,7 @@ int __vm_enough_memory(long pages, int cap_sys_admin)
 	 */
 	if (atomic_read(&vm_committed_space) < (long)allowed)
 		return 0;
-
+error:
 	vm_unacct_memory(pages);
 
 	return -ENOMEM;