+/* SPECIAL CASE for the $RECOVERY lock used by the recovery thread.
+ * We cannot wait for node recovery to complete to begin mastering this
+ * lockres because this lockres is used to kick off recovery! ;-)
+ * So, do a pre-check on all living nodes to see if any of those nodes
+ * think that $RECOVERY is currently mastered by a dead node. If so,
+ * we wait a short time to allow that node to get notified by its own
+ * heartbeat stack, then check again. All $RECOVERY lock resources
+ * mastered by dead nodes are purged when the hearbeat callback is
+ * fired, so we can know for sure that it is safe to continue once
+ * the node returns a live node or no node. */
+static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
+ struct dlm_lock_resource *res)
+{
+ struct dlm_node_iter iter;
+ int nodenum;
+ int ret = 0;
+ u8 master = DLM_LOCK_RES_OWNER_UNKNOWN;
+
+ spin_lock(&dlm->spinlock);
+ dlm_node_iter_init(dlm->domain_map, &iter);
+ spin_unlock(&dlm->spinlock);
+
+ while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
+ /* do not send to self */
+ if (nodenum == dlm->node_num)
+ continue;
+ ret = dlm_do_master_requery(dlm, res, nodenum, &master);
+ if (ret < 0) {
+ mlog_errno(ret);
+ if (!dlm_is_host_down(ret))
+ BUG();
+ /* host is down, so answer for that node would be
+ * DLM_LOCK_RES_OWNER_UNKNOWN. continue. */
+ ret = 0;
+ }
+
+ if (master != DLM_LOCK_RES_OWNER_UNKNOWN) {
+ /* check to see if this master is in the recovery map */
+ spin_lock(&dlm->spinlock);
+ if (test_bit(master, dlm->recovery_map)) {
+ mlog(ML_NOTICE, "%s: node %u has not seen "
+ "node %u go down yet, and thinks the "
+ "dead node is mastering the recovery "
+ "lock. must wait.\n", dlm->name,
+ nodenum, master);
+ ret = -EAGAIN;
+ }
+ spin_unlock(&dlm->spinlock);
+ mlog(0, "%s: reco lock master is %u\n", dlm->name,
+ master);
+ break;
+ }
+ }
+ return ret;
+}
+