drbd: fix for spurious fullsync (uuids rotated too fast)
authorLars Ellenberg <lars.ellenberg@linbit.com>
Thu, 7 Oct 2010 14:07:55 +0000 (16:07 +0200)
committerPhilipp Reisner <philipp.reisner@linbit.com>
Thu, 14 Oct 2010 16:38:48 +0000 (18:38 +0200)
If it was an "empty" resync, the SyncSource may have already "finished"
the resync and rotated the UUIDs, before noticing the connection loss
(and generating a new uuid, if Primary, rotating again), while the
SyncTarget did not change its uuids at all, or only got to the previous
sync-uuid.
This would then again lead to a full sync on next handshake
(see also Bug #251).

Fix:
Use explicit resync finished notification even for empty resyncs,
do not finish an empty resync implicitly on the SyncSource.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
drivers/block/drbd/drbd_main.c
drivers/block/drbd/drbd_worker.c

index accb37d..63f45d7 100644 (file)
@@ -1426,6 +1426,11 @@ static void after_state_ch(struct drbd_conf *mdev, union drbd_state os,
            (os.user_isp && !ns.user_isp))
                resume_next_sg(mdev);
 
+       /* sync target done with resync.  Explicitly notify peer, even though
+        * it should (at least for non-empty resyncs) already know itself. */
+       if (os.disk < D_UP_TO_DATE && os.conn >= C_SYNC_SOURCE && ns.conn == C_CONNECTED)
+               drbd_send_state(mdev);
+
        /* free tl_hash if we Got thawed and are C_STANDALONE */
        if (ns.conn == C_STANDALONE && !is_susp(ns) && mdev->tl_hash)
                drbd_free_tl_hash(mdev);
index 166b51e..88be45a 100644 (file)
@@ -522,6 +522,12 @@ int w_make_resync_request(struct drbd_conf *mdev,
                dev_err(DEV, "%s in w_make_resync_request\n",
                        drbd_conn_str(mdev->state.conn));
 
+       if (mdev->rs_total == 0) {
+               /* empty resync? */
+               drbd_resync_finished(mdev);
+               return 1;
+       }
+
        if (!get_ldev(mdev)) {
                /* Since we only need to access mdev->rsync a
                   get_ldev_if_state(mdev,D_FAILED) would be sufficient, but
@@ -768,6 +774,14 @@ static int w_resync_finished(struct drbd_conf *mdev, struct drbd_work *w, int ca
        return 1;
 }
 
+static void ping_peer(struct drbd_conf *mdev)
+{
+       clear_bit(GOT_PING_ACK, &mdev->flags);
+       request_ping(mdev);
+       wait_event(mdev->misc_wait,
+                  test_bit(GOT_PING_ACK, &mdev->flags) || mdev->state.conn < C_CONNECTED);
+}
+
 int drbd_resync_finished(struct drbd_conf *mdev)
 {
        unsigned long db, dt, dbdt;
@@ -807,6 +821,8 @@ int drbd_resync_finished(struct drbd_conf *mdev)
        if (!get_ldev(mdev))
                goto out;
 
+       ping_peer(mdev);
+
        spin_lock_irq(&mdev->req_lock);
        os = mdev->state;
 
@@ -1420,14 +1436,6 @@ int drbd_alter_sa(struct drbd_conf *mdev, int na)
        return retcode;
 }
 
-static void ping_peer(struct drbd_conf *mdev)
-{
-       clear_bit(GOT_PING_ACK, &mdev->flags);
-       request_ping(mdev);
-       wait_event(mdev->misc_wait,
-                  test_bit(GOT_PING_ACK, &mdev->flags) || mdev->state.conn < C_CONNECTED);
-}
-
 /**
  * drbd_start_resync() - Start the resync process
  * @mdev:      DRBD device.
@@ -1527,9 +1535,21 @@ void drbd_start_resync(struct drbd_conf *mdev, enum drbd_conns side)
                     (unsigned long) mdev->rs_total << (BM_BLOCK_SHIFT-10),
                     (unsigned long) mdev->rs_total);
 
-               if (mdev->rs_total == 0) {
-                       /* Peer still reachable? Beware of failing before-resync-target handlers! */
-                       ping_peer(mdev);
+               if (mdev->agreed_pro_version < 95 && mdev->rs_total == 0) {
+                       /* This still has a race (about when exactly the peers
+                        * detect connection loss) that can lead to a full sync
+                        * on next handshake. In 8.3.9 we fixed this with explicit
+                        * resync-finished notifications, but the fix
+                        * introduces a protocol change.  Sleeping for some
+                        * time longer than the ping interval + timeout on the
+                        * SyncSource, to give the SyncTarget the chance to
+                        * detect connection loss, then waiting for a ping
+                        * response (implicit in drbd_resync_finished) reduces
+                        * the race considerably, but does not solve it. */
+                       if (side == C_SYNC_SOURCE)
+                               schedule_timeout_interruptible(
+                                       mdev->net_conf->ping_int * HZ +
+                                       mdev->net_conf->ping_timeo*HZ/9);
                        drbd_resync_finished(mdev);
                }