[KERNEL] Fix possible stall when switching server

If we switch to a different server when we only have something in
the send list but nothing in the recv list, the send worker would
not have gotten invoked. Now we unconditionally trigger the send
worker when asked to re-queue any pending requests.

1 files changed, 9 insertions, 1 deletions

diff --git a/src/kernel/blk.c b/src/kernel/blk.c
index 5795c03..bc18c44 100644
--- a/src/kernel/blk.c
+++ b/src/kernel/blk.c
@@ -536,8 +536,16 @@ void dnbd3_blk_requeue_all_requests(dnbd3_device_t *dev)
 	while (!list_empty(&local_copy)) {
 		blk_request = list_entry(local_copy.next, struct request, queuelist);
 		list_del_init(&blk_request->queuelist);
-		dnbd3_add_queue(dev, blk_request);
+		spin_lock_irqsave(&dev->send_queue_lock, flags);
+		list_add_tail(&blk_request->queuelist, &dev->send_queue);
+		spin_unlock_irqrestore(&dev->send_queue_lock, flags);
 	}
+	// Do this even if we didn't move anything from the recv list to the send
+	// list. It might have already contained something, which needs to be
+	// re-requested anyways if this was called because of a server switch.
+	spin_lock_irqsave(&dev->blk_lock, flags);
+	queue_work(dev->send_wq, &dev->send_work);
+	spin_unlock_irqrestore(&dev->blk_lock, flags);
 }
 
 void dnbd3_blk_fail_all_requests(dnbd3_device_t *dev)