[patch rfc] memcg: correctly order reading PCG_USED andpc->mem_cgroup

[Johannes Weiner]

The placement of the read-side barrier is confused: the writer first
sets pc->mem_cgroup, then PCG_USED.  The read-side barrier has to be
between testing PCG_USED and reading pc->mem_cgroup.

Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
---
 mm/memcontrol.c |   27 +++++++++------------------
 1 files changed, 9 insertions(+), 18 deletions(-)

I am a bit dumbfounded as to why this has never had any impact.  I see
two scenarios where charging can race with LRU operations:

One is shmem pages on swapoff.  They are on the LRU when charged as
page cache, which could race with isolation/putback.  This seems
sufficiently rare.

The other case is a swap cache page being charged while somebody else
had it isolated.  mem_cgroup_lru_del_before_commit_swapcache() would
see the page isolated and skip it.  The commit then has to race with
putback, which could see PCG_USED but not pc->mem_cgroup, and crash
with a NULL pointer dereference.  This does sound a bit more likely.

Any idea?  Am I missing something?

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 5b562b3..db76ef7 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -836,13 +836,12 @@ void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
 		return;
 
 	pc = lookup_page_cgroup(page);
-	/*
-	 * Used bit is set without atomic ops but after smp_wmb().
-	 * For making pc->mem_cgroup visible, insert smp_rmb() here.
-	 */
-	smp_rmb();
 	/* unused or root page is not rotated. */
-	if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
+	if (!PageCgroupUsed(pc))
+		return;
+	/* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
+	smp_rmb();
+	if (mem_cgroup_is_root(pc->mem_cgroup))
 		return;
 	mz = page_cgroup_zoneinfo(pc);
 	list_move(&pc->lru, &mz->lists[lru]);
@@ -857,14 +856,10 @@ void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
 		return;
 	pc = lookup_page_cgroup(page);
 	VM_BUG_ON(PageCgroupAcctLRU(pc));
-	/*
-	 * Used bit is set without atomic ops but after smp_wmb().
-	 * For making pc->mem_cgroup visible, insert smp_rmb() here.
-	 */
-	smp_rmb();
 	if (!PageCgroupUsed(pc))
 		return;
-
+	/* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
+	smp_rmb();
 	mz = page_cgroup_zoneinfo(pc);
 	/* huge page split is done under lru_lock. so, we have no races. */
 	MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
@@ -1031,14 +1026,10 @@ mem_cgroup_get_reclaim_stat_from_page(struct page *page)
 		return NULL;
 
 	pc = lookup_page_cgroup(page);
-	/*
-	 * Used bit is set without atomic ops but after smp_wmb().
-	 * For making pc->mem_cgroup visible, insert smp_rmb() here.
-	 */
-	smp_rmb();
 	if (!PageCgroupUsed(pc))
 		return NULL;
-
+	/* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
+	smp_rmb();
 	mz = page_cgroup_zoneinfo(pc);
 	if (!mz)
 		return NULL;
-- 
1.7.3.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/