[PATCH] workqueue: fix possible race condition when rescuer VS pwq-release

[Lai Jiangshan]

There is a race condition between rescuer_thread() and
pwq_unbound_release_workfn().

The works of the @pwq may be processed by some other workers,
and @pwq is scheduled to release(due to its wq's attr is changed)
before the rescuer starts to process. In this case
pwq_unbound_release_workfn() will corrupt wq->maydays list,
and rescuer_thead() will access to corrupted data.

Using get_unbound_pwq() when send_mayday() will keep @pwq's lifetime
and avoid the race condition.

Changed from V1:
	1) Introduce get_unbound_pwq() for beter readibility. Since
	   get_pwq() is considerred no-op for percpu workqueue,
	   so the patch are the same behavior in functionality.
	2) More precise comments.

Signed-off-by: Lai Jiangshan <laijs@xxxxxxxxxxxxxx>
---
 kernel/workqueue.c |   30 ++++++++++++++++++++++++++++++
 1 files changed, 30 insertions(+), 0 deletions(-)

diff --git a/kernel/workqueue.c b/kernel/workqueue.c
index 0c74979..d845bdd 100644
--- a/kernel/workqueue.c
+++ b/kernel/workqueue.c
@@ -1050,6 +1050,12 @@ static void get_pwq(struct pool_workqueue *pwq)
 	pwq->refcnt++;
 }
 
+static inline void get_unbound_pwq(struct pool_workqueue *pwq)
+{
+	if (pwq->wq->flags & WQ_UNBOUND)
+		get_pwq(pwq);
+}
+
 /**
  * put_pwq - put a pool_workqueue reference
  * @pwq: pool_workqueue to put
@@ -1075,6 +1081,12 @@ static void put_pwq(struct pool_workqueue *pwq)
 	schedule_work(&pwq->unbound_release_work);
 }
 
+static inline void put_unbound_pwq(struct pool_workqueue *pwq)
+{
+	if (pwq->wq->flags & WQ_UNBOUND)
+		put_pwq(pwq);
+}
+
 /**
  * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
  * @pwq: pool_workqueue to put (can be %NULL)
@@ -1908,6 +1920,19 @@ static void send_mayday(struct work_struct *work)
 
 	/* mayday mayday mayday */
 	if (list_empty(&pwq->mayday_node)) {
+		/*
+		 * A pwq of an unbound wq may be released before wq's
+		 * destruction when the wq's attr is changed. In this case,
+		 * pwq_unbound_release_workfn() may execute earlier before
+		 * rescuer_thread() and corrupt wq->maydays list.
+		 *
+		 * get_unbound_pwq() keeps the unbound pwq until the rescuer
+		 * processes it and protects the pwq from being scheduled to
+		 * release when someone else processes all the works before
+		 * the rescuer starts to process.
+		 */
+		get_unbound_pwq(pwq);
+
 		list_add_tail(&pwq->mayday_node, &wq->maydays);
 		wake_up_process(wq->rescuer->task);
 	}
@@ -2424,6 +2449,7 @@ repeat:
 		/* migrate to the target cpu if possible */
 		worker_maybe_bind_and_lock(pool);
 		rescuer->pool = pool;
+		put_unbound_pwq(pwq);
 
 		/*
 		 * Slurp in all works issued via this workqueue and
@@ -4318,6 +4344,10 @@ void destroy_workqueue(struct workqueue_struct *wq)
 		/*
 		 * The base ref is never dropped on per-cpu pwqs.  Directly
 		 * free the pwqs and wq.
+		 *
+		 * The wq->maydays list maybe still have some pwqs linked,
+		 * but it is safe to free them all together since the rescuer
+		 * is stopped.
 		 */
 		free_percpu(wq->cpu_pwqs);
 		kfree(wq);
-- 
1.7.4.4

--
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/