[PATCH 1/2] memcg, oom: no oom-kill for __GFP_RETRY_MAYFAIL

[PATCH 1/2] memcg, oom: no oom-kill for __GFP_RETRY_MAYFAIL

[Shakeel Butt]

The documentation of __GFP_RETRY_MAYFAIL clearly mentioned that the
OOM killer will not be triggered and indeed the page alloc does not
invoke OOM killer for such allocations. However we do trigger memcg
OOM killer for __GFP_RETRY_MAYFAIL. Fix that. This flag will used later
to not trigger oom-killer in the charging path for fanotify and inotify
event allocations.

Signed-off-by: Shakeel Butt <shakeelb@google.com>
Acked-by: Michal Hocko <mhocko@suse.com>
---
Changelog since v1:
- commit message updated.

 mm/memcontrol.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 2713b45ec3f0..99eca724ed3b 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -2294,7 +2294,6 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
 	unsigned long nr_reclaimed;
 	bool may_swap = true;
 	bool drained = false;
-	bool oomed = false;
 	enum oom_status oom_status;
 
 	if (mem_cgroup_is_root(memcg))
@@ -2381,7 +2380,7 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
 	if (nr_retries--)
 		goto retry;
 
-	if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
+	if (gfp_mask & __GFP_RETRY_MAYFAIL)
 		goto nomem;
 
 	if (gfp_mask & __GFP_NOFAIL)
@@ -2400,7 +2399,6 @@ static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
 	switch (oom_status) {
 	case OOM_SUCCESS:
 		nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
-		oomed = true;
 		goto retry;
 	case OOM_FAILED:
 		goto force;
-- 
2.21.0.593.g511ec345e18-goog

[PATCH 2/2] memcg, fsnotify: no oom-kill for remote memcg charging

[Shakeel Butt]

The commit d46eb14b735b ("fs: fsnotify: account fsnotify metadata to
kmemcg") added remote memcg charging for fanotify and inotify event
objects. The aim was to charge the memory to the listener who is
interested in the events but without triggering the OOM killer.
Otherwise there would be security concerns for the listener. At the
time, oom-kill trigger was not in the charging path. A parallel work
added the oom-kill back to charging path i.e. commit 29ef680ae7c2
("memcg, oom: move out_of_memory back to the charge path"). So to not
trigger oom-killer in the remote memcg, explicitly add
__GFP_RETRY_MAYFAIL to the fanotify and inotify event allocations.

Signed-off-by: Shakeel Butt <shakeelb@google.com>
---
 fs/notify/fanotify/fanotify.c        | 4 +++-
 fs/notify/inotify/inotify_fsnotify.c | 7 +++++--
 2 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/fs/notify/fanotify/fanotify.c b/fs/notify/fanotify/fanotify.c
index 6b9c27548997..9aa5d325e6d8 100644
--- a/fs/notify/fanotify/fanotify.c
+++ b/fs/notify/fanotify/fanotify.c
@@ -282,13 +282,15 @@ struct fanotify_event *fanotify_alloc_event(struct fsnotify_group *group,
 					    __kernel_fsid_t *fsid)
 {
 	struct fanotify_event *event = NULL;
-	gfp_t gfp = GFP_KERNEL_ACCOUNT;
+	gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL;
 	struct inode *id = fanotify_fid_inode(inode, mask, data, data_type);
 
 	/*
 	 * For queues with unlimited length lost events are not expected and
 	 * can possibly have security implications. Avoid losing events when
 	 * memory is short.
+	 *
+	 * Note: __GFP_NOFAIL takes precedence over __GFP_RETRY_MAYFAIL.
 	 */
 	if (group->max_events == UINT_MAX)
 		gfp |= __GFP_NOFAIL;
diff --git a/fs/notify/inotify/inotify_fsnotify.c b/fs/notify/inotify/inotify_fsnotify.c
index ff30abd6a49b..17c08daa1ba7 100644
--- a/fs/notify/inotify/inotify_fsnotify.c
+++ b/fs/notify/inotify/inotify_fsnotify.c
@@ -99,9 +99,12 @@ int inotify_handle_event(struct fsnotify_group *group,
 	i_mark = container_of(inode_mark, struct inotify_inode_mark,
 			      fsn_mark);
 
-	/* Whoever is interested in the event, pays for the allocation. */
+	/*
+	 * Whoever is interested in the event, pays for the allocation. However
+	 * do not trigger the OOM killer in the target memcg.
+	 */
 	memalloc_use_memcg(group->memcg);
-	event = kmalloc(alloc_len, GFP_KERNEL_ACCOUNT);
+	event = kmalloc(alloc_len, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
 	memalloc_unuse_memcg();
 
 	if (unlikely(!event)) {
-- 
2.21.0.593.g511ec345e18-goog

Re: [PATCH 2/2] memcg, fsnotify: no oom-kill for remote memcg charging

[Michal Hocko]

On Mon 29-04-19 10:13:32, Shakeel Butt wrote:
[...]
>  	/*
>  	 * For queues with unlimited length lost events are not expected and
>  	 * can possibly have security implications. Avoid losing events when
>  	 * memory is short.
> +	 *
> +	 * Note: __GFP_NOFAIL takes precedence over __GFP_RETRY_MAYFAIL.
>  	 */

No, I there is no rule like that. Combining the two is undefined
currently and I do not think we want to legitimize it. What does it even
mean?
-- 
Michal Hocko
SUSE Labs

Re: [PATCH 2/2] memcg, fsnotify: no oom-kill for remote memcg charging

[Shakeel Butt]

On Mon, Apr 29, 2019 at 5:41 PM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Mon 29-04-19 10:13:32, Shakeel Butt wrote:
> [...]
> >       /*
> >        * For queues with unlimited length lost events are not expected and
> >        * can possibly have security implications. Avoid losing events when
> >        * memory is short.
> > +      *
> > +      * Note: __GFP_NOFAIL takes precedence over __GFP_RETRY_MAYFAIL.
> >        */
>
> No, I there is no rule like that. Combining the two is undefined
> currently and I do not think we want to legitimize it. What does it even
> mean?
>

Actually the code is doing that but I agree this is not documented and
weird. I will fix this.

Shakeel