mm/mmap.c: fix arithmetic overflow in __vm_enough_memory()
authorRoman Gushchin <>
Wed, 11 Feb 2015 23:28:39 +0000 (15:28 -0800)
committerBen Hutchings <>
Sat, 9 May 2015 22:16:16 +0000 (23:16 +0100)
commit 5703b087dc8eaf47bfb399d6cf512d471beff405 upstream.

I noticed, that "allowed" can easily overflow by falling below 0,
because (total_vm / 32) can be larger than "allowed".  The problem
occurs in OVERCOMMIT_NONE mode.

In this case, a huge allocation can success and overcommit the system
(despite OVERCOMMIT_NONE mode).  All subsequent allocations will fall
(system-wide), so system become unusable.

The problem was masked out by commit c9b1d0981fcc
("mm: limit growth of 3% hardcoded other user reserve"),
but it's easy to reproduce it on older kernels:
1) set overcommit_memory sysctl to 2
2) mmap() large file multiple times (with VM_SHARED flag)
3) try to malloc() large amount of memory

It also can be reproduced on newer kernels, but miss-configured
sysctl_user_reserve_kbytes is required.

Fix this issue by switching to signed arithmetic here.

[ use min_t]
Signed-off-by: Roman Gushchin <>
Cc: Andrew Shewmaker <>
Cc: Rik van Riel <>
Cc: Konstantin Khlebnikov <>
Reviewed-by: Michal Hocko <>
Signed-off-by: Andrew Morton <>
Signed-off-by: Linus Torvalds <>
[bwh: Backported to 3.2: there is no 'reserved' variable]
Signed-off-by: Ben Hutchings <>

index 13b5685..48f6efd 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -111,7 +111,7 @@ struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
-       unsigned long free, allowed;
+       long free, allowed;