Messages in this thread Patch in this message | | | From | Johannes Weiner <> | Subject | [patch 1/2] mm: memcg: use proper memcg in limit bypass | Date | Thu, 24 Oct 2013 05:58:14 -0400 |
| |
84235de ("fs: buffer: move allocation failure loop into the allocator") allowed __GFP_NOFAIL allocations to bypass the limit if they fail to reclaim enough memory for the charge. Because the main test case was on a 3.2-based system, this patch missed the fact that on newer kernels the charge function needs to return root_mem_cgroup when bypassing the limit, and not NULL. This will corrupt whatever memory is at NULL + percpu pointer offset. Fix this quickly before problems are reported.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> --- mm/memcontrol.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c index 34d3ca9..13a9c80 100644 --- a/mm/memcontrol.c +++ b/mm/memcontrol.c @@ -2765,10 +2765,10 @@ done: *ptr = memcg; return 0; nomem: - *ptr = NULL; - if (gfp_mask & __GFP_NOFAIL) - return 0; - return -ENOMEM; + if (!(gfp_mask & __GFP_NOFAIL)) { + *ptr = NULL; + return -ENOMEM; + } bypass: *ptr = root_mem_cgroup; return -EINTR; -- 1.8.4.1
| |