return 0;
}
+/* returns true when the recovery master has contacted us */
+static int dlm_reco_master_ready(struct dlm_ctxt *dlm)
+{
+ int ready;
+ spin_lock(&dlm->spinlock);
+ ready = (dlm->reco.new_master != O2NM_INVALID_NODE_NUM);
+ spin_unlock(&dlm->spinlock);
+ return ready;
+}
+
+/* returns true if node is no longer in the domain
+ * could be dead or just not joined */
+int dlm_is_node_dead(struct dlm_ctxt *dlm, u8 node)
+{
+ int dead;
+ spin_lock(&dlm->spinlock);
+ dead = test_bit(node, dlm->domain_map);
+ spin_unlock(&dlm->spinlock);
+ return dead;
+}
+
/* callers of the top-level api calls (dlmlock/dlmunlock) should
* block on the dlm->reco.event when recovery is in progress.
* the dlm recovery thread will set this state when it begins
static int dlm_do_recovery(struct dlm_ctxt *dlm)
{
int status = 0;
+ int ret;
spin_lock(&dlm->spinlock);
goto master_here;
if (dlm->reco.new_master == O2NM_INVALID_NODE_NUM) {
- /* choose a new master */
- if (!dlm_pick_recovery_master(dlm)) {
+ /* choose a new master, returns 0 if this node
+ * is the master, -EEXIST if it's another node.
+ * this does not return until a new master is chosen
+ * or recovery completes entirely. */
+ ret = dlm_pick_recovery_master(dlm);
+ if (!ret) {
/* already notified everyone. go. */
- dlm->reco.new_master = dlm->node_num;
goto master_here;
}
mlog(0, "another node will master this recovery session.\n");
if (status < 0) {
mlog(ML_ERROR, "error %d remastering locks for node %u, "
"retrying.\n", status, dlm->reco.dead_node);
+ /* yield a bit to allow any final network messages
+ * to get handled on remaining nodes */
+ msleep(100);
} else {
/* success! see if any other nodes need recovery */
+ mlog(0, "DONE mastering recovery of %s:%u here(this=%u)!\n",
+ dlm->name, dlm->reco.dead_node, dlm->node_num);
dlm_reset_recovery(dlm);
}
dlm_end_recovery(dlm);
BUG();
break;
case DLM_RECO_NODE_DATA_DEAD:
- mlog(0, "node %u died after "
+ mlog(ML_NOTICE, "node %u died after "
"requesting recovery info for "
"node %u\n", ndata->node_num,
dead_node);
// start all over
destroy = 1;
status = -EAGAIN;
+ /* instead of spinning like crazy here,
+ * wait for the domain map to catch up
+ * with the network state. otherwise this
+ * can be hit hundreds of times before
+ * the node is really seen as dead. */
+ wait_event_timeout(dlm->dlm_reco_thread_wq,
+ dlm_is_node_dead(dlm,
+ ndata->node_num),
+ msecs_to_jiffies(1000));
+ mlog(0, "waited 1 sec for %u, "
+ "dead? %s\n", ndata->node_num,
+ dlm_is_node_dead(dlm, ndata->node_num) ?
+ "yes" : "no");
goto leave;
case DLM_RECO_NODE_DATA_RECEIVING:
case DLM_RECO_NODE_DATA_REQUESTED:
dlm = item->dlm;
dead_node = item->u.ral.dead_node;
reco_master = item->u.ral.reco_master;
+ mres = (struct dlm_migratable_lockres *)data;
+
+ if (dead_node != dlm->reco.dead_node ||
+ reco_master != dlm->reco.new_master) {
+ /* show extra debug info if the recovery state is messed */
+ mlog(ML_ERROR, "%s: bad reco state: reco(dead=%u, master=%u), "
+ "request(dead=%u, master=%u)\n",
+ dlm->name, dlm->reco.dead_node, dlm->reco.new_master,
+ dead_node, reco_master);
+ mlog(ML_ERROR, "%s: name=%.*s master=%u locks=%u/%u flags=%u "
+ "entry[0]={c=%"MLFu64",l=%u,f=%u,t=%d,ct=%d,hb=%d,n=%u}\n",
+ dlm->name, mres->lockname_len, mres->lockname, mres->master,
+ mres->num_locks, mres->total_locks, mres->flags,
+ mres->ml[0].cookie, mres->ml[0].list, mres->ml[0].flags,
+ mres->ml[0].type, mres->ml[0].convert_type,
+ mres->ml[0].highest_blocked, mres->ml[0].node);
+ BUG();
+ }
BUG_ON(dead_node != dlm->reco.dead_node);
BUG_ON(reco_master != dlm->reco.new_master);
- mres = (struct dlm_migratable_lockres *)data;
-
/* lock resources should have already been moved to the
* dlm->reco.resources list. now move items from that list
* to a temp list if the dead owner matches. note that the
continue;
switch (ndata->state) {
+ /* should have moved beyond INIT but not to FINALIZE yet */
case DLM_RECO_NODE_DATA_INIT:
case DLM_RECO_NODE_DATA_DEAD:
- case DLM_RECO_NODE_DATA_DONE:
case DLM_RECO_NODE_DATA_FINALIZE_SENT:
mlog(ML_ERROR, "bad ndata state for node %u:"
" state=%d\n", ndata->node_num,
ndata->state);
BUG();
break;
+ /* these states are possible at this point, anywhere along
+ * the line of recovery */
+ case DLM_RECO_NODE_DATA_DONE:
case DLM_RECO_NODE_DATA_RECEIVING:
case DLM_RECO_NODE_DATA_REQUESTED:
case DLM_RECO_NODE_DATA_REQUESTING:
{
struct dlm_lock_resource *res;
struct list_head *iter, *iter2;
+ struct dlm_lock *lock;
spin_lock(&dlm->spinlock);
list_for_each_safe(iter, iter2, &dlm->reco.resources) {
res = list_entry (iter, struct dlm_lock_resource, recovering);
+ /* always prune any $RECOVERY entries for dead nodes,
+ * otherwise hangs can occur during later recovery */
if (dlm_is_recovery_lock(res->lockname.name,
- res->lockname.len))
+ res->lockname.len)) {
+ spin_lock(&res->spinlock);
+ list_for_each_entry(lock, &res->granted, list) {
+ if (lock->ml.node == dead_node) {
+ mlog(0, "AHA! there was "
+ "a $RECOVERY lock for dead "
+ "node %u (%s)!\n",
+ dead_node, dlm->name);
+ list_del_init(&lock->list);
+ dlm_lock_put(lock);
+ break;
+ }
+ }
+ spin_unlock(&res->spinlock);
continue;
+ }
+
if (res->owner == dead_node) {
mlog(0, "found lockres owned by dead node while "
"doing recovery for node %u. sending it.\n",
again:
ret = dlm_lockres_master_requery(dlm, res, &real_master);
if (ret < 0) {
- mlog(0, "dlm_lockres_master_requery failure: %d\n",
+ mlog(0, "dlm_lockres_master_requery ret=%d\n",
ret);
goto again;
}
struct dlm_lock_resource *res;
int i;
struct list_head *bucket;
+ struct dlm_lock *lock;
/* purge any stale mles */
bucket = &(dlm->resources[i]);
list_for_each(iter, bucket) {
res = list_entry (iter, struct dlm_lock_resource, list);
+ /* always prune any $RECOVERY entries for dead nodes,
+ * otherwise hangs can occur during later recovery */
if (dlm_is_recovery_lock(res->lockname.name,
- res->lockname.len))
+ res->lockname.len)) {
+ spin_lock(&res->spinlock);
+ list_for_each_entry(lock, &res->granted, list) {
+ if (lock->ml.node == dead_node) {
+ mlog(0, "AHA! there was "
+ "a $RECOVERY lock for dead "
+ "node %u (%s)!\n",
+ dead_node, dlm->name);
+ list_del_init(&lock->list);
+ dlm_lock_put(lock);
+ break;
+ }
+ }
+ spin_unlock(&res->spinlock);
continue;
-
+ }
spin_lock(&res->spinlock);
/* zero the lvb if necessary */
dlm_revalidate_lvb(dlm, res, dead_node);
return;
spin_lock(&dlm->spinlock);
-
set_bit(idx, dlm->live_nodes_map);
-
- /* notify any mles attached to the heartbeat events */
- dlm_hb_event_notify_attached(dlm, idx, 1);
-
+ /* do NOT notify mle attached to the heartbeat events.
+ * new nodes are not interesting in mastery until joined. */
spin_unlock(&dlm->spinlock);
dlm_put(dlm);
mlog(0, "unlockast for recovery lock fired!\n");
}
-
+/*
+ * dlm_pick_recovery_master will continually attempt to use
+ * dlmlock() on the special "$RECOVERY" lockres with the
+ * LKM_NOQUEUE flag to get an EX. every thread that enters
+ * this function on each node racing to become the recovery
+ * master will not stop attempting this until either:
+ * a) this node gets the EX (and becomes the recovery master),
+ * or b) dlm->reco.new_master gets set to some nodenum
+ * != O2NM_INVALID_NODE_NUM (another node will do the reco).
+ * so each time a recovery master is needed, the entire cluster
+ * will sync at this point. if the new master dies, that will
+ * be detected in dlm_do_recovery */
static int dlm_pick_recovery_master(struct dlm_ctxt *dlm)
{
enum dlm_status ret;
mlog(0, "starting recovery of %s at %lu, dead=%u, this=%u\n",
dlm->name, jiffies, dlm->reco.dead_node, dlm->node_num);
-retry:
+again:
memset(&lksb, 0, sizeof(lksb));
ret = dlmlock(dlm, LKM_EXMODE, &lksb, LKM_NOQUEUE|LKM_RECOVERY,
DLM_RECOVERY_LOCK_NAME, dlm_reco_ast, dlm, dlm_reco_bast);
+ mlog(0, "%s: dlmlock($RECOVERY) returned %d, lksb=%d\n",
+ dlm->name, ret, lksb.status);
+
if (ret == DLM_NORMAL) {
mlog(0, "dlm=%s dlmlock says I got it (this=%u)\n",
dlm->name, dlm->node_num);
- /* I am master, send message to all nodes saying
- * that I am beginning a recovery session */
- status = dlm_send_begin_reco_message(dlm,
- dlm->reco.dead_node);
+
+ /* got the EX lock. check to see if another node
+ * just became the reco master */
+ if (dlm_reco_master_ready(dlm)) {
+ mlog(0, "%s: got reco EX lock, but %u will "
+ "do the recovery\n", dlm->name,
+ dlm->reco.new_master);
+ status = -EEXIST;
+ } else {
+ status = dlm_send_begin_reco_message(dlm,
+ dlm->reco.dead_node);
+ /* this always succeeds */
+ BUG_ON(status);
+
+ /* set the new_master to this node */
+ spin_lock(&dlm->spinlock);
+ dlm->reco.new_master = dlm->node_num;
+ spin_unlock(&dlm->spinlock);
+ }
/* recovery lock is a special case. ast will not get fired,
* so just go ahead and unlock it. */
ret = dlmunlock(dlm, &lksb, 0, dlm_reco_unlock_ast, dlm);
+ if (ret == DLM_DENIED) {
+ mlog(0, "got DLM_DENIED, trying LKM_CANCEL\n");
+ ret = dlmunlock(dlm, &lksb, LKM_CANCEL, dlm_reco_unlock_ast, dlm);
+ }
if (ret != DLM_NORMAL) {
/* this would really suck. this could only happen
* if there was a network error during the unlock
* is actually "done" and the lock structure is
* even freed. we can continue, but only
* because this specific lock name is special. */
- mlog(0, "dlmunlock returned %d\n", ret);
- }
-
- if (status < 0) {
- mlog(0, "failed to send recovery message. "
- "must retry with new node map.\n");
- goto retry;
+ mlog(ML_ERROR, "dlmunlock returned %d\n", ret);
}
} else if (ret == DLM_NOTQUEUED) {
mlog(0, "dlm=%s dlmlock says another node got it (this=%u)\n",
dlm->name, dlm->node_num);
/* another node is master. wait on
- * reco.new_master != O2NM_INVALID_NODE_NUM */
+ * reco.new_master != O2NM_INVALID_NODE_NUM
+ * for at most one second */
+ wait_event_timeout(dlm->dlm_reco_thread_wq,
+ dlm_reco_master_ready(dlm),
+ msecs_to_jiffies(1000));
+ if (!dlm_reco_master_ready(dlm)) {
+ mlog(0, "%s: reco master taking awhile\n",
+ dlm->name);
+ goto again;
+ }
+ /* another node has informed this one that it is reco master */
+ mlog(0, "%s: reco master %u is ready to recover %u\n",
+ dlm->name, dlm->reco.new_master, dlm->reco.dead_node);
status = -EEXIST;
+ } else {
+ struct dlm_lock_resource *res;
+
+ /* dlmlock returned something other than NOTQUEUED or NORMAL */
+ mlog(ML_ERROR, "%s: got %s from dlmlock($RECOVERY), "
+ "lksb.status=%s\n", dlm->name, dlm_errname(ret),
+ dlm_errname(lksb.status));
+ res = dlm_lookup_lockres(dlm, DLM_RECOVERY_LOCK_NAME,
+ DLM_RECOVERY_LOCK_NAME_LEN);
+ if (res) {
+ dlm_print_one_lock_resource(res);
+ dlm_lockres_put(res);
+ } else {
+ mlog(ML_ERROR, "recovery lock not found\n");
+ }
+ BUG();
}
return status;
mlog(0, "not sending begin reco to self\n");
continue;
}
-
+retry:
ret = -EINVAL;
mlog(0, "attempting to send begin reco msg to %d\n",
nodenum);
/* negative status is handled ok by caller here */
if (ret >= 0)
ret = status;
+ if (dlm_is_host_down(ret)) {
+ /* node is down. not involved in recovery
+ * so just keep going */
+ mlog(0, "%s: node %u was down when sending "
+ "begin reco msg (%d)\n", dlm->name, nodenum, ret);
+ ret = 0;
+ }
if (ret < 0) {
struct dlm_lock_resource *res;
+ /* this is now a serious problem, possibly ENOMEM
+ * in the network stack. must retry */
mlog_errno(ret);
mlog(ML_ERROR, "begin reco of dlm %s to node %u "
" returned %d\n", dlm->name, nodenum, ret);
} else {
mlog(ML_ERROR, "recovery lock not found\n");
}
- break;
+ /* sleep for a bit in hopes that we can avoid
+ * another ENOMEM */
+ msleep(100);
+ goto retry;
}
}
spin_lock(&dlm->spinlock);
if (dlm->reco.new_master != O2NM_INVALID_NODE_NUM) {
- mlog(0, "new_master already set to %u!\n",
- dlm->reco.new_master);
+ if (test_bit(dlm->reco.new_master, dlm->recovery_map)) {
+ mlog(0, "%s: new_master %u died, changing "
+ "to %u\n", dlm->name, dlm->reco.new_master,
+ br->node_idx);
+ } else {
+ mlog(0, "%s: new_master %u NOT DEAD, changing "
+ "to %u\n", dlm->name, dlm->reco.new_master,
+ br->node_idx);
+ /* may not have seen the new master as dead yet */
+ }
}
if (dlm->reco.dead_node != O2NM_INVALID_NODE_NUM) {
- mlog(0, "dead_node already set to %u!\n",
- dlm->reco.dead_node);
+ mlog(ML_NOTICE, "%s: dead_node previously set to %u, "
+ "node %u changing it to %u\n", dlm->name,
+ dlm->reco.dead_node, br->node_idx, br->dead_node);
}
dlm->reco.new_master = br->node_idx;
dlm->reco.dead_node = br->dead_node;
if (!test_bit(br->dead_node, dlm->recovery_map)) {
- mlog(ML_ERROR, "recovery master %u sees %u as dead, but this "
+ mlog(0, "recovery master %u sees %u as dead, but this "
"node has not yet. marking %u as dead\n",
br->node_idx, br->dead_node, br->dead_node);
+ if (!test_bit(br->dead_node, dlm->domain_map) ||
+ !test_bit(br->dead_node, dlm->live_nodes_map))
+ mlog(0, "%u not in domain/live_nodes map "
+ "so setting it in reco map manually\n",
+ br->dead_node);
+ set_bit(br->dead_node, dlm->recovery_map);
__dlm_hb_node_down(dlm, br->dead_node);
}
spin_unlock(&dlm->spinlock);
projects / linux-2.6 / commitdiff