[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>

1 files changed, 15 insertions, 3 deletions

diff --git a/mm/nommu.c b/mm/nommu.c
index db45efac17cc..029fadac0fb5 100644
--- 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;