RDMA/cma: Using the standard locking pattern when delivering the removal event

Whenever an event is delivered to the handler it should be done under the
handler_mutex and upon any non-zero return from the handler it should
trigger destruction of the cm_id.

cma_process_remove() skips some steps here, it is not necessarily wrong
since the state change should prevent any races, but it is confusing and
unnecessary.

Follow the standard pattern here, with the slight twist that the
transition to RDMA_CM_DEVICE_REMOVAL includes a cma_cancel_operation().

Link: https://lore.kernel.org/r/20200723070707.1771101-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

1 files changed, 36 insertions, 26 deletions

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index 77743ce5216b..b51a0acd672b 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -1921,6 +1921,8 @@ static int cma_cm_event_handler(struct rdma_id_private *id_priv,
 {
 	int ret;
 
+	lockdep_assert_held(&id_priv->handler_mutex);
+
 	trace_cm_event_handler(id_priv, event);
 	ret = id_priv->id.event_handler(&id_priv->id, event);
 	trace_cm_event_done(id_priv, event, ret);
@@ -4775,50 +4777,58 @@ free_cma_dev:
 	return ret;
 }
 
-static int cma_remove_id_dev(struct rdma_id_private *id_priv)
+static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
 {
-	struct rdma_cm_event event = {};
+	struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
 	enum rdma_cm_state state;
-	int ret = 0;
-
-	/* Record that we want to remove the device */
-	state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
-	if (state == RDMA_CM_DESTROYING)
-		return 0;
+	unsigned long flags;
 
-	cma_cancel_operation(id_priv, state);
 	mutex_lock(&id_priv->handler_mutex);
+	/* Record that we want to remove the device */
+	spin_lock_irqsave(&id_priv->lock, flags);
+	state = id_priv->state;
+	if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
+		spin_unlock_irqrestore(&id_priv->lock, flags);
+		mutex_unlock(&id_priv->handler_mutex);
+		cma_id_put(id_priv);
+		return;
+	}
+	id_priv->state = RDMA_CM_DEVICE_REMOVAL;
+	spin_unlock_irqrestore(&id_priv->lock, flags);
 
-	/* Check for destruction from another callback. */
-	if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
-		goto out;
-
-	event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
-	ret = cma_cm_event_handler(id_priv, &event);
-out:
+	if (cma_cm_event_handler(id_priv, &event)) {
+		/*
+		 * At this point the ULP promises it won't call
+		 * rdma_destroy_id() concurrently
+		 */
+		cma_id_put(id_priv);
+		mutex_unlock(&id_priv->handler_mutex);
+		rdma_destroy_id(&id_priv->id);
+		return;
+	}
 	mutex_unlock(&id_priv->handler_mutex);
-	return ret;
+
+	/*
+	 * If this races with destroy then the thread that first assigns state
+	 * to a destroying does the cancel.
+	 */
+	cma_cancel_operation(id_priv, state);
+	cma_id_put(id_priv);
 }
 
 static void cma_process_remove(struct cma_device *cma_dev)
 {
-	struct rdma_id_private *id_priv;
-	int ret;
-
 	mutex_lock(&lock);
 	while (!list_empty(&cma_dev->id_list)) {
-		id_priv = list_entry(cma_dev->id_list.next,
-				     struct rdma_id_private, list);
+		struct rdma_id_private *id_priv = list_first_entry(
+			&cma_dev->id_list, struct rdma_id_private, list);
 
 		list_del(&id_priv->listen_list);
 		list_del_init(&id_priv->list);
 		cma_id_get(id_priv);
 		mutex_unlock(&lock);
 
-		ret = cma_remove_id_dev(id_priv);
-		cma_id_put(id_priv);
-		if (ret)
-			rdma_destroy_id(&id_priv->id);
+		cma_send_device_removal_put(id_priv);
 
 		mutex_lock(&lock);
 	}