Without a memory barrier, the following race can occur with a high-order
allocation:
wakeup_kcompactd(order == 1) kcompactd()
[L] waitqueue_active(kcompactd_wait)
[S]
prepare_to_wait_event(kcompactd_wait)
[L] (kcompactd_max_order == 0)
[S] kcompactd_max_order = order; schedule()
Where the waitqueue_active() check is speculatively re-ordered to before
setting the actual condition (max_order), not seeing the threads that's
going to block; making us miss a wakeup. There are a couple of options to
fix this, including calling wq_has_sleepers() which adds a full barrier,
or unconditionally doing the wake_up_interruptible() and serialize on the
q->lock. However, to make use of the control dependency, we just need to
add L->L guarantees.
While this bug is theoretical, there have been other offenders of the lockless
waitqueue_active() in the past -- this is also documented in the call itself.
Signed-off-by: Davidlohr Bueso <[email protected]>
---
mm/compaction.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/mm/compaction.c b/mm/compaction.c
index 949198d01260..fb0f87554eb9 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -1950,6 +1950,13 @@ void wakeup_kcompactd(pg_data_t *pgdat, int order, int
classzone_idx)
if (pgdat->kcompactd_max_order < order)
pgdat->kcompactd_max_order = order;
+ /*
+ * Pairs with implicit barrier in wait_event_freezable()
+ * such that wakeups are not missed in the lockless
+ * waitqueue_active() call.
+ */
+ smp_acquire__after_ctrl_dep();
+
if (pgdat->kcompactd_classzone_idx > classzone_idx)
pgdat->kcompactd_classzone_idx = classzone_idx;
--
2.6.6
