[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>

diff --git a/mm/nommu.c b/mm/nommu.c
index db45efa..029fada 100644 (file)
--- a/mm/nommu.c
+++ b/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;