block: remove retry loop in ioc_release_fn()

The reverse-order double lock dance in ioc_release_fn() is using a
retry loop. This is a problem on PREEMPT_RT because it could preempt
the task that would release q->queue_lock and thus live lock in the
retry loop.

RCU is already managing the freeing of the request queue and icq. If
the trylock fails, use RCU to guarantee that the request queue and
icq are not freed and re-acquire the locks in the correct order,
allowing forward progress.

Signed-off-by: John Ogness <john.ogness@linutronix.de>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/block/blk-ioc.c b/block/blk-ioc.c
index 5dbcfa1..57299f8 100644 (file)
--- a/block/blk-ioc.c
+++ b/block/blk-ioc.c
@@ -107,9 +107,23 @@ static void ioc_release_fn(struct work_struct *work)
                        ioc_destroy_icq(icq);
                        spin_unlock(&q->queue_lock);
                } else {
-                       spin_unlock_irq(&ioc->lock);
-                       cpu_relax();
-                       spin_lock_irq(&ioc->lock);
+                       /* Make sure q and icq cannot be freed. */
+                       rcu_read_lock();
+
+                       /* Re-acquire the locks in the correct order. */
+                       spin_unlock(&ioc->lock);
+                       spin_lock(&q->queue_lock);
+                       spin_lock(&ioc->lock);
+
+                       /*
+                        * The icq may have been destroyed when the ioc lock
+                        * was released.
+                        */
+                       if (!(icq->flags & ICQ_DESTROYED))
+                               ioc_destroy_icq(icq);
+
+                       spin_unlock(&q->queue_lock);
+                       rcu_read_unlock();
                }
        }