drbd: fix resend/resubmit of frozen IO

DRBD can freeze IO, due to fencing policy (fencing resource-and-stonith),
or because we lost access to data (on-no-data-accessible suspend-io).

Resuming from there (re-connect, or re-attach, or explicit admin
intervention) should "just work".

Unfortunately, if the re-attach/re-connect did not happen within
the timeout, since the commit

  drbd: Implemented real timeout checking for request processing time

if so configured, the request_timer_fn() would timeout and
detach/disconnect virtually immediately.

This change tracks the most recent attach and connect, and does not
timeout within <configured timeout interval> after attach/connect.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>

drivers/block/drbd/drbd_int.h

@@ -859,6 +859,7 @@ struct drbd_tconn {			/* is a resource from the config file */
 	unsigned int epochs;
 	enum write_ordering_e write_ordering;
 
+	unsigned long last_reconnect_jif;
 	struct drbd_thread receiver;
 	struct drbd_thread worker;
 	struct drbd_thread asender;
@@ -881,6 +882,7 @@ struct drbd_conf {
 	struct block_device *this_bdev;
 	struct gendisk	    *vdisk;
 
+	unsigned long last_reattach_jif;
 	struct drbd_work  resync_work,
 			  unplug_work,
 			  go_diskless,

drivers/block/drbd/drbd_req.c

@@ -1171,12 +1171,14 @@ void request_timer_fn(unsigned long data)
 	struct list_head *le;
 	struct net_conf *nc;
 	unsigned long ent = 0, dt = 0, et, nt; /* effective timeout = ko_count * timeout */
+	unsigned long now;
 
 	rcu_read_lock();
 	nc = rcu_dereference(tconn->net_conf);
-	ent = nc ? nc->timeout * HZ/10 * nc->ko_count : 0;
+	if (nc && mdev->state.conn >= C_WF_REPORT_PARAMS)
+		ent = nc->timeout * HZ/10 * nc->ko_count;
 
-	if (get_ldev(mdev)) {
+	if (get_ldev(mdev)) { /* implicit state.disk >= D_INCONSISTENT */
 		dt = rcu_dereference(mdev->ldev->disk_conf)->disk_timeout * HZ / 10;
 		put_ldev(mdev);
 	}
@@ -1184,32 +1186,51 @@ void request_timer_fn(unsigned long data)
 
 	et = min_not_zero(dt, ent);
 
-	if (!et || (mdev->state.conn < C_WF_REPORT_PARAMS && mdev->state.disk <= D_FAILED))
+	if (!et)
 		return; /* Recurring timer stopped */
 
+	now = jiffies;
+
 	spin_lock_irq(&tconn->req_lock);
 	le = &tconn->oldest_tle->requests;
 	if (list_empty(le)) {
 		spin_unlock_irq(&tconn->req_lock);
-		mod_timer(&mdev->request_timer, jiffies + et);
+		mod_timer(&mdev->request_timer, now + et);
 		return;
 	}
 
 	le = le->prev;
 	req = list_entry(le, struct drbd_request, tl_requests);
-	if (ent && req->rq_state & RQ_NET_PENDING) {
+
-		if (time_is_before_eq_jiffies(req->start_time + ent)) {
+	/* The request is considered timed out, if
+	 * - we have some effective timeout from the configuration,
+	 *   with above state restrictions applied,
+	 * - the oldest request is waiting for a response from the network
+	 *   resp. the local disk,
+	 * - the oldest request is in fact older than the effective timeout,
+	 * - the connection was established (resp. disk was attached)
+	 *   for longer than the timeout already.
+	 * Note that for 32bit jiffies and very stable connections/disks,
+	 * we may have a wrap around, which is catched by
+	 *   !time_in_range(now, last_..._jif, last_..._jif + timeout).
+	 *
+	 * Side effect: once per 32bit wrap-around interval, which means every
+	 * ~198 days with 250 HZ, we have a window where the timeout would need
+	 * to expire twice (worst case) to become effective. Good enough.
+	 */
+	if (ent && req->rq_state & RQ_NET_PENDING &&
+		 time_after(now, req->start_time + ent) &&
+		!time_in_range(now, tconn->last_reconnect_jif, tconn->last_reconnect_jif + ent)) {
 		dev_warn(DEV, "Remote failed to finish a request within ko-count * timeout\n");
 		_drbd_set_state(_NS(mdev, conn, C_TIMEOUT), CS_VERBOSE | CS_HARD, NULL);
 	}
-	}
+	if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev &&
-	if (dt && req->rq_state & RQ_LOCAL_PENDING && req->w.mdev == mdev) {
+		 time_after(now, req->start_time + dt) &&
-		if (time_is_before_eq_jiffies(req->start_time + dt)) {
+		!time_in_range(now, mdev->last_reattach_jif, mdev->last_reattach_jif + dt)) {
 		dev_warn(DEV, "Local backing device failed to meet the disk-timeout\n");
 		__drbd_chk_io_error(mdev, 1);
 	}
-	}
+	nt = (time_after(now, req->start_time + et) ? now : req->start_time) + et;
-	nt = (time_is_before_eq_jiffies(req->start_time + et) ? jiffies : req->start_time) + et;
 	spin_unlock_irq(&tconn->req_lock);
 	mod_timer(&mdev->request_timer, nt);
 }

drivers/block/drbd/drbd_state.c

@@ -1075,6 +1075,13 @@ __drbd_set_state(struct drbd_conf *mdev, union drbd_state ns,
 	if (os.conn < C_CONNECTED && ns.conn >= C_CONNECTED)
 		drbd_resume_al(mdev);
 
+	/* remember last attach time so request_timer_fn() won't
+	 * kill newly established sessions while we are still trying to thaw
+	 * previously frozen IO */
+	if ((os.disk == D_ATTACHING || os.disk == D_NEGOTIATING) &&
+	    ns.disk > D_NEGOTIATING)
+		mdev->last_reattach_jif = jiffies;
+
 	ascw = kmalloc(sizeof(*ascw), GFP_ATOMIC);
 	if (ascw) {
 		ascw->os = os;
@@ -1609,8 +1616,15 @@ conn_set_state(struct drbd_tconn *tconn, union drbd_state mask, union drbd_state
 	enum drbd_state_rv rv;
 	int vnr, number_of_volumes = 0;
 
-	if (mask.conn == C_MASK)
+	if (mask.conn == C_MASK) {
+		/* remember last connect time so request_timer_fn() won't
+		 * kill newly established sessions while we are still trying to thaw
+		 * previously frozen IO */
+		if (tconn->cstate != C_WF_REPORT_PARAMS && val.conn == C_WF_REPORT_PARAMS)
+			tconn->last_reconnect_jif = jiffies;
+
 		tconn->cstate = val.conn;
+	}
 
 	rcu_read_lock();
 	idr_for_each_entry(&tconn->volumes, mdev, vnr) {