drbd: fix potential deadlock during "restart" of conflicting writes

w_restart_write(), run from worker context, calls __drbd_make_request()
and further drbd_al_begin_io(, delegate=true), which then
potentially deadlocks.  The previous patch moved a BUG_ON to expose
such call paths, which would now be triggered.

Also, if we call __drbd_make_request() from resource worker context,
like w_restart_write() did, and that should block for whatever reason
(!drbd_state_is_stable(), resource suspended, ...),
we potentially deadlock the whole resource, as the worker
is needed for state changes and other things.

Create a dedicated retry workqueue for this instead.

Also make sure that inc_ap_bio()/dec_ap_bio() are properly paired,
even if do_retry() needs to retry itself,
in case __drbd_make_request() returns != 0.
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

drivers/block/drbd/drbd_main.c

@@ -2383,6 +2383,73 @@ void drbd_minor_destroy(struct kref *kref)
 	kref_put(&tconn->kref, &conn_destroy);
 }
 
+/* One global retry thread, if we need to push back some bio and have it
+ * reinserted through our make request function.
+ */
+static struct retry_worker {
+	struct workqueue_struct *wq;
+	struct work_struct worker;
+
+	spinlock_t lock;
+	struct list_head writes;
+} retry;
+
+static void do_retry(struct work_struct *ws)
+{
+	struct retry_worker *retry = container_of(ws, struct retry_worker, worker);
+	LIST_HEAD(writes);
+	struct drbd_request *req, *tmp;
+
+	spin_lock_irq(&retry->lock);
+	list_splice_init(&retry->writes, &writes);
+	spin_unlock_irq(&retry->lock);
+
+	list_for_each_entry_safe(req, tmp, &writes, tl_requests) {
+		struct drbd_conf *mdev = req->w.mdev;
+		struct bio *bio = req->master_bio;
+		unsigned long start_time = req->start_time;
+
+		/* We have exclusive access to this request object.
+		 * If it had not been RQ_POSTPONED, the code path which queued
+		 * it here would have completed and freed it already.
+		 */
+		mempool_free(req, drbd_request_mempool);
+
+		/* A single suspended or otherwise blocking device may stall
+		 * all others as well.  Fortunately, this code path is to
+		 * recover from a situation that "should not happen":
+		 * concurrent writes in multi-primary setup.
+		 * In a "normal" lifecycle, this workqueue is supposed to be
+		 * destroyed without ever doing anything.
+		 * If it turns out to be an issue anyways, we can do per
+		 * resource (replication group) or per device (minor) retry
+		 * workqueues instead.
+		 */
+
+		/* We are not just doing generic_make_request(),
+		 * as we want to keep the start_time information. */
+		do {
+			inc_ap_bio(mdev);
+		} while(__drbd_make_request(mdev, bio, start_time));
+	}
+}
+
+void drbd_restart_write(struct drbd_request *req)
+{
+	unsigned long flags;
+	spin_lock_irqsave(&retry.lock, flags);
+	list_move_tail(&req->tl_requests, &retry.writes);
+	spin_unlock_irqrestore(&retry.lock, flags);
+
+	/* Drop the extra reference that would otherwise
+	 * have been dropped by complete_master_bio.
+	 * do_retry() needs to grab a new one. */
+	dec_ap_bio(req->w.mdev);
+
+	queue_work(retry.wq, &retry.worker);
+}
+
+
 static void drbd_cleanup(void)
 {
 	unsigned int i;
@@ -2402,6 +2469,9 @@ static void drbd_cleanup(void)
 	if (drbd_proc)
 		remove_proc_entry("drbd", NULL);
 
+	if (retry.wq)
+		destroy_workqueue(retry.wq);
+
 	drbd_genl_unregister();
 
 	idr_for_each_entry(&minors, mdev, i) {
@@ -2851,6 +2921,15 @@ int __init drbd_init(void)
 	rwlock_init(&global_state_lock);
 	INIT_LIST_HEAD(&drbd_tconns);
 
+	retry.wq = create_singlethread_workqueue("drbd-reissue");
+	if (!retry.wq) {
+		printk(KERN_ERR "drbd: unable to create retry workqueue\n");
+		goto fail;
+	}
+	INIT_WORK(&retry.worker, do_retry);
+	spin_lock_init(&retry.lock);
+	INIT_LIST_HEAD(&retry.writes);
+
 	printk(KERN_INFO "drbd: initialized. "
 	       "Version: " REL_VERSION " (api:%d/proto:%d-%d)\n",
 	       API_VERSION, PRO_VERSION_MIN, PRO_VERSION_MAX);

drivers/block/drbd/drbd_receiver.c

@@ -1748,30 +1748,6 @@ static int receive_RSDataReply(struct drbd_tconn *tconn, struct packet_info *pi)
 	return err;
 }
 
-static int w_restart_write(struct drbd_work *w, int cancel)
-{
-	struct drbd_request *req = container_of(w, struct drbd_request, w);
-	struct drbd_conf *mdev = w->mdev;
-	struct bio *bio;
-	unsigned long start_time;
-	unsigned long flags;
-
-	spin_lock_irqsave(&mdev->tconn->req_lock, flags);
-	if (!expect(req->rq_state & RQ_POSTPONED)) {
-		spin_unlock_irqrestore(&mdev->tconn->req_lock, flags);
-		return -EIO;
-	}
-	bio = req->master_bio;
-	start_time = req->start_time;
-	/* Postponed requests will not have their master_bio completed!  */
-	__req_mod(req, DISCARD_WRITE, NULL);
-	spin_unlock_irqrestore(&mdev->tconn->req_lock, flags);
-
-	while (__drbd_make_request(mdev, bio, start_time))
-		/* retry */ ;
-	return 0;
-}
-
 static void restart_conflicting_writes(struct drbd_conf *mdev,
 				       sector_t sector, int size)
 {
@@ -1785,11 +1761,9 @@ static void restart_conflicting_writes(struct drbd_conf *mdev,
 		if (req->rq_state & RQ_LOCAL_PENDING ||
 		    !(req->rq_state & RQ_POSTPONED))
 			continue;
-		if (expect(list_empty(&req->w.list))) {
-			req->w.mdev = mdev;
-			req->w.cb = w_restart_write;
-			drbd_queue_work(&mdev->tconn->data.work, &req->w);
-		}
+		/* as it is RQ_POSTPONED, this will cause it to
+		 * be queued on the retry workqueue. */
+		__req_mod(req, DISCARD_WRITE, NULL);
 	}
 }
 

drivers/block/drbd/drbd_req.c

@@ -104,7 +104,7 @@ static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const
 	 * and never sent), it should still be "empty" as
 	 * initialized in drbd_req_new(), so we can list_del() it
 	 * here unconditionally */
-	list_del(&req->tl_requests);
+	list_del_init(&req->tl_requests);
 
 	/* if it was a write, we may have to set the corresponding
 	 * bit(s) out-of-sync first. If it had a local part, we need to
@@ -143,7 +143,10 @@ static void _req_is_done(struct drbd_conf *mdev, struct drbd_request *req, const
 		}
 	}
 
-	drbd_req_free(req);
+	if (s & RQ_POSTPONED)
+		drbd_restart_write(req);
+	else
+		drbd_req_free(req);
 }
 
 static void queue_barrier(struct drbd_conf *mdev)
@@ -289,8 +292,16 @@ void _req_may_be_done(struct drbd_request *req, struct bio_and_error *m)
 		if (!(s & RQ_POSTPONED)) {
 			m->error = ok ? 0 : (error ?: -EIO);
 			m->bio = req->master_bio;
+			req->master_bio = NULL;
+		} else {
+			/* Assert that this will be _req_is_done()
+			 * with this very invokation. */
+			/* FIXME:
+			 * what about (RQ_LOCAL_PENDING | RQ_LOCAL_ABORTED)?
+			 */
+			D_ASSERT(!(s & RQ_LOCAL_PENDING));
+			D_ASSERT(s & RQ_NET_DONE);
 		}
-		req->master_bio = NULL;
 	}
 
 	if (s & RQ_LOCAL_PENDING)

drivers/block/drbd/drbd_req.h

@@ -268,6 +268,9 @@ extern void request_timer_fn(unsigned long data);
 extern void tl_restart(struct drbd_tconn *tconn, enum drbd_req_event what);
 extern void _tl_restart(struct drbd_tconn *tconn, enum drbd_req_event what);
 
+/* this is in drbd_main.c */
+extern void drbd_restart_write(struct drbd_request *req);
+
 /* use this if you don't want to deal with calling complete_master_bio()
  * outside the spinlock, e.g. when walking some list on cleanup. */
 static inline int _req_mod(struct drbd_request *req, enum drbd_req_event what)