ocfs2_dlm: fix cluster-wide refcounting of lock resources

This was previously broken and migration of some locks had to be temporarily
disabled. We use a new (and backward-incompatible) set of network messages
to account for all references to a lock resources held across the cluster.
once these are all freed, the master node may then free the lock resource
memory once its local references are dropped.

Signed-off-by: Kurt Hackel <kurt.hackel@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

1 files changed, 352 insertions, 42 deletions

diff --git a/fs/ocfs2/dlm/dlmmaster.c b/fs/ocfs2/dlm/dlmmaster.c
index 0ad872055cb3..4645ec2e0fc3 100644
--- a/fs/ocfs2/dlm/dlmmaster.c
+++ b/fs/ocfs2/dlm/dlmmaster.c
@@ -99,9 +99,9 @@ static void dlm_mle_node_up(struct dlm_ctxt *dlm,
                             int idx);
 
 static void dlm_assert_master_worker(struct dlm_work_item *item, void *data);
-static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
-                                unsigned int namelen, void *nodemap,
-                                u32 flags);
+static int dlm_do_assert_master(struct dlm_ctxt *dlm,
+                                struct dlm_lock_resource *res,
+                                void *nodemap, u32 flags);
 
 static inline int dlm_mle_equal(struct dlm_ctxt *dlm,
                                 struct dlm_master_list_entry *mle,
@@ -237,7 +237,8 @@ static int dlm_find_mle(struct dlm_ctxt *dlm,
                         struct dlm_master_list_entry **mle,
                         char *name, unsigned int namelen);
 
-static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to);
+static int dlm_do_master_request(struct dlm_lock_resource *res,
+                                 struct dlm_master_list_entry *mle, int to);
 
 
 static int dlm_wait_for_lock_mastery(struct dlm_ctxt *dlm,
@@ -687,6 +688,7 @@ static void dlm_init_lockres(struct dlm_ctxt *dlm,
         INIT_LIST_HEAD(&res->purge);
         atomic_set(&res->asts_reserved, 0);
         res->migration_pending = 0;
+        res->inflight_locks = 0;
 
         kref_init(&res->refs);
 
@@ -700,6 +702,7 @@ static void dlm_init_lockres(struct dlm_ctxt *dlm,
         res->last_used = 0;
 
         memset(res->lvb, 0, DLM_LVB_LEN);
+        memset(res->refmap, 0, sizeof(res->refmap));
 }
 
 struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
@@ -722,6 +725,42 @@ struct dlm_lock_resource *dlm_new_lockres(struct dlm_ctxt *dlm,
         return res;
 }
 
+void __dlm_lockres_grab_inflight_ref(struct dlm_ctxt *dlm,
+                                   struct dlm_lock_resource *res,
+                                   int new_lockres,
+                                   const char *file,
+                                   int line)
+{
+        if (!new_lockres)
+                assert_spin_locked(&res->spinlock);
+
+        if (!test_bit(dlm->node_num, res->refmap)) {
+                BUG_ON(res->inflight_locks != 0);
+                dlm_lockres_set_refmap_bit(dlm->node_num, res);
+        }
+        res->inflight_locks++;
+        mlog(0, "%s:%.*s: inflight++: now %u\n",
+             dlm->name, res->lockname.len, res->lockname.name,
+             res->inflight_locks);
+}
+
+void __dlm_lockres_drop_inflight_ref(struct dlm_ctxt *dlm,
+                                   struct dlm_lock_resource *res,
+                                   const char *file,
+                                   int line)
+{
+        assert_spin_locked(&res->spinlock);
+
+        BUG_ON(res->inflight_locks == 0);
+        res->inflight_locks--;
+        mlog(0, "%s:%.*s: inflight--: now %u\n",
+             dlm->name, res->lockname.len, res->lockname.name,
+             res->inflight_locks);
+        if (res->inflight_locks == 0)
+                dlm_lockres_clear_refmap_bit(dlm->node_num, res);
+        wake_up(&res->wq);
+}
+
 /*
  * lookup a lock resource by name.
  * may already exist in the hashtable.
@@ -752,6 +791,7 @@ struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
         unsigned int hash;
         int tries = 0;
         int bit, wait_on_recovery = 0;
+        int drop_inflight_if_nonlocal = 0;
 
         BUG_ON(!lockid);
 
@@ -761,9 +801,30 @@ struct dlm_lock_resource * dlm_get_lock_resource(struct dlm_ctxt *dlm,
 
 lookup:
         spin_lock(&dlm->spinlock);
-        tmpres = __dlm_lookup_lockres(dlm, lockid, namelen, hash);
+        tmpres = __dlm_lookup_lockres_full(dlm, lockid, namelen, hash);
         if (tmpres) {
+                int dropping_ref = 0;
+
+                spin_lock(&tmpres->spinlock);
+                if (tmpres->owner == dlm->node_num) {
+                        BUG_ON(tmpres->state & DLM_LOCK_RES_DROPPING_REF);
+                        dlm_lockres_grab_inflight_ref(dlm, tmpres);
+                } else if (tmpres->state & DLM_LOCK_RES_DROPPING_REF)
+                        dropping_ref = 1;
+                spin_unlock(&tmpres->spinlock);
                 spin_unlock(&dlm->spinlock);
+
+                /* wait until done messaging the master, drop our ref to allow
+                 * the lockres to be purged, start over. */
+                if (dropping_ref) {
+                        spin_lock(&tmpres->spinlock);
+                        __dlm_wait_on_lockres_flags(tmpres, DLM_LOCK_RES_DROPPING_REF);
+                        spin_unlock(&tmpres->spinlock);
+                        dlm_lockres_put(tmpres);
+                        tmpres = NULL;
+                        goto lookup;
+                }
+
                 mlog(0, "found in hash!\n");
                 if (res)
                         dlm_lockres_put(res);
@@ -793,6 +854,7 @@ lookup:
                 spin_lock(&res->spinlock);
                 dlm_change_lockres_owner(dlm, res, dlm->node_num);
                 __dlm_insert_lockres(dlm, res);
+                dlm_lockres_grab_inflight_ref(dlm, res);
                 spin_unlock(&res->spinlock);
                 spin_unlock(&dlm->spinlock);
                 /* lockres still marked IN_PROGRESS */
@@ -805,29 +867,40 @@ lookup:
         /* if we found a block, wait for lock to be mastered by another node */
         blocked = dlm_find_mle(dlm, &mle, (char *)lockid, namelen);
         if (blocked) {
+                int mig;
                 if (mle->type == DLM_MLE_MASTER) {
                         mlog(ML_ERROR, "master entry for nonexistent lock!\n");
                         BUG();
-                } else if (mle->type == DLM_MLE_MIGRATION) {
-                        /* migration is in progress! */
-                        /* the good news is that we now know the
-                         * "current" master (mle->master). */
-
+                }
+                mig = (mle->type == DLM_MLE_MIGRATION);
+                /* if there is a migration in progress, let the migration
+                 * finish before continuing.  we can wait for the absence
+                 * of the MIGRATION mle: either the migrate finished or
+                 * one of the nodes died and the mle was cleaned up.
+                 * if there is a BLOCK here, but it already has a master
+                 * set, we are too late.  the master does not have a ref
+                 * for us in the refmap.  detach the mle and drop it.
+                 * either way, go back to the top and start over. */
+                if (mig || mle->master != O2NM_MAX_NODES) {
+                        BUG_ON(mig && mle->master == dlm->node_num);
+                        /* we arrived too late.  the master does not
+                         * have a ref for us. retry. */
+                        mlog(0, "%s:%.*s: late on %s\n",
+                             dlm->name, namelen, lockid,
+                             mig ?  "MIGRATION" : "BLOCK");
                         spin_unlock(&dlm->master_lock);
-                        assert_spin_locked(&dlm->spinlock);
-
-                        /* set the lockres owner and hash it */
-                        spin_lock(&res->spinlock);
-                        dlm_set_lockres_owner(dlm, res, mle->master);
-                        __dlm_insert_lockres(dlm, res);
-                        spin_unlock(&res->spinlock);
                         spin_unlock(&dlm->spinlock);
 
                         /* master is known, detach */
-                        dlm_mle_detach_hb_events(dlm, mle);
+                        if (!mig)
+                                dlm_mle_detach_hb_events(dlm, mle);
                         dlm_put_mle(mle);
                         mle = NULL;
-                        goto wake_waiters;
+                        /* this is lame, but we cant wait on either
+                         * the mle or lockres waitqueue here */
+                        if (mig)
+                                msleep(100);
+                        goto lookup;
                 }
         } else {
                 /* go ahead and try to master lock on this node */
@@ -858,6 +931,13 @@ lookup:
 
         /* finally add the lockres to its hash bucket */
         __dlm_insert_lockres(dlm, res);
+        /* since this lockres is new it doesnt not require the spinlock */
+        dlm_lockres_grab_inflight_ref_new(dlm, res);
+
+        /* if this node does not become the master make sure to drop
+         * this inflight reference below */
+        drop_inflight_if_nonlocal = 1;
+
         /* get an extra ref on the mle in case this is a BLOCK
          * if so, the creator of the BLOCK may try to put the last
          * ref at this time in the assert master handler, so we
@@ -910,7 +990,7 @@ redo_request:
         ret = -EINVAL;
         dlm_node_iter_init(mle->vote_map, &iter);
         while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
-                ret = dlm_do_master_request(mle, nodenum);
+                ret = dlm_do_master_request(res, mle, nodenum);
                 if (ret < 0)
                         mlog_errno(ret);
                 if (mle->master != O2NM_MAX_NODES) {
@@ -960,6 +1040,8 @@ wait:
 
 wake_waiters:
         spin_lock(&res->spinlock);
+        if (res->owner != dlm->node_num && drop_inflight_if_nonlocal)
+                dlm_lockres_drop_inflight_ref(dlm, res);
         res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
         spin_unlock(&res->spinlock);
         wake_up(&res->wq);
@@ -998,7 +1080,7 @@ recheck:
                 /* this will cause the master to re-assert across
                  * the whole cluster, freeing up mles */
                 if (res->owner != dlm->node_num) {
-                        ret = dlm_do_master_request(mle, res->owner);
+                        ret = dlm_do_master_request(res, mle, res->owner);
                         if (ret < 0) {
                                 /* give recovery a chance to run */
                                 mlog(ML_ERROR, "link to %u went down?: %d\n", res->owner, ret);
@@ -1062,6 +1144,8 @@ recheck:
                                  * now tell other nodes that I am
                                  * mastering this. */
                                 mle->master = dlm->node_num;
+                                /* ref was grabbed in get_lock_resource
+                                 * will be dropped in dlmlock_master */
                                 assert = 1;
                                 sleep = 0;
                         }
@@ -1087,7 +1171,8 @@ recheck:
                                          (atomic_read(&mle->woken) == 1),
                                          timeo);
                 if (res->owner == O2NM_MAX_NODES) {
-                        mlog(0, "waiting again\n");
+                        mlog(0, "%s:%.*s: waiting again\n", dlm->name,
+                             res->lockname.len, res->lockname.name);
                         goto recheck;
                 }
                 mlog(0, "done waiting, master is %u\n", res->owner);
@@ -1100,8 +1185,7 @@ recheck:
                 m = dlm->node_num;
                 mlog(0, "about to master %.*s here, this=%u\n",
                      res->lockname.len, res->lockname.name, m);
-                ret = dlm_do_assert_master(dlm, res->lockname.name,
-                                           res->lockname.len, mle->vote_map, 0);
+                ret = dlm_do_assert_master(dlm, res, mle->vote_map, 0);
                 if (ret) {
                         /* This is a failure in the network path,
                          * not in the response to the assert_master
@@ -1117,6 +1201,8 @@ recheck:
 
         /* set the lockres owner */
         spin_lock(&res->spinlock);
+        /* mastery reference obtained either during
+         * assert_master_handler or in get_lock_resource */
         dlm_change_lockres_owner(dlm, res, m);
         spin_unlock(&res->spinlock);
 
@@ -1283,7 +1369,8 @@ static int dlm_restart_lock_mastery(struct dlm_ctxt *dlm,
  *
  */
 
-static int dlm_do_master_request(struct dlm_master_list_entry *mle, int to)
+static int dlm_do_master_request(struct dlm_lock_resource *res,
+                                 struct dlm_master_list_entry *mle, int to)
 {
         struct dlm_ctxt *dlm = mle->dlm;
         struct dlm_master_request request;
@@ -1339,6 +1426,9 @@ again:
                 case DLM_MASTER_RESP_YES:
                         set_bit(to, mle->response_map);
                         mlog(0, "node %u is the master, response=YES\n", to);
+                        mlog(0, "%s:%.*s: master node %u now knows I have a "
+                             "reference\n", dlm->name, res->lockname.len,
+                             res->lockname.name, to);
                         mle->master = to;
                         break;
                 case DLM_MASTER_RESP_NO:
@@ -1428,8 +1518,10 @@ way_up_top:
                 }
 
                 if (res->owner == dlm->node_num) {
+                        mlog(0, "%s:%.*s: setting bit %u in refmap\n",
+                             dlm->name, namelen, name, request->node_idx);
+                        dlm_lockres_set_refmap_bit(request->node_idx, res);
                         spin_unlock(&res->spinlock);
-                        // mlog(0, "this node is the master\n");
                         response = DLM_MASTER_RESP_YES;
                         if (mle)
                                 kmem_cache_free(dlm_mle_cache, mle);
@@ -1477,7 +1569,6 @@ way_up_top:
                         mlog(0, "node %u is master, but trying to migrate to "
                              "node %u.\n", tmpmle->master, tmpmle->new_master);
                         if (tmpmle->master == dlm->node_num) {
-                                response = DLM_MASTER_RESP_YES;
                                 mlog(ML_ERROR, "no owner on lockres, but this "
                                      "node is trying to migrate it to %u?!\n",
                                      tmpmle->new_master);
@@ -1494,6 +1585,10 @@ way_up_top:
                                  * go back and clean the mles on any
                                  * other nodes */
                                 dispatch_assert = 1;
+                                dlm_lockres_set_refmap_bit(request->node_idx, res);
+                                mlog(0, "%s:%.*s: setting bit %u in refmap\n",
+                                     dlm->name, namelen, name,
+                                     request->node_idx);
                         } else
                                 response = DLM_MASTER_RESP_NO;
                 } else {
@@ -1607,15 +1702,17 @@ send_response:
  * can periodically run all locks owned by this node
  * and re-assert across the cluster...
  */
-static int dlm_do_assert_master(struct dlm_ctxt *dlm, const char *lockname,
-                                unsigned int namelen, void *nodemap,
-                                u32 flags)
+int dlm_do_assert_master(struct dlm_ctxt *dlm,
+                         struct dlm_lock_resource *res,
+                         void *nodemap, u32 flags)
 {
         struct dlm_assert_master assert;
         int to, tmpret;
         struct dlm_node_iter iter;
         int ret = 0;
         int reassert;
+        const char *lockname = res->lockname.name;
+        unsigned int namelen = res->lockname.len;
 
         BUG_ON(namelen > O2NM_MAX_NAME_LEN);
 again:
@@ -1647,6 +1744,7 @@ again:
                         mlog(0, "link to %d went down!\n", to);
                         /* any nonzero status return will do */
                         ret = tmpret;
+                        r = 0;
                 } else if (r < 0) {
                         /* ok, something horribly messed.  kill thyself. */
                         mlog(ML_ERROR,"during assert master of %.*s to %u, "
@@ -1661,12 +1759,29 @@ again:
                         spin_unlock(&dlm->master_lock);
                         spin_unlock(&dlm->spinlock);
                         BUG();
-                } else if (r == EAGAIN) {
+                }
+
+                if (r & DLM_ASSERT_RESPONSE_REASSERT &&
+                    !(r & DLM_ASSERT_RESPONSE_MASTERY_REF)) {
+                                mlog(ML_ERROR, "%.*s: very strange, "
+                                     "master MLE but no lockres on %u\n",
+                                     namelen, lockname, to);
+                }
+
+                if (r & DLM_ASSERT_RESPONSE_REASSERT) {
                         mlog(0, "%.*s: node %u create mles on other "
                              "nodes and requests a re-assert\n", 
                              namelen, lockname, to);
                         reassert = 1;
                 }
+                if (r & DLM_ASSERT_RESPONSE_MASTERY_REF) {
+                        mlog(0, "%.*s: node %u has a reference to this "
+                             "lockres, set the bit in the refmap\n",
+                             namelen, lockname, to);
+                        spin_lock(&res->spinlock);
+                        dlm_lockres_set_refmap_bit(to, res);
+                        spin_unlock(&res->spinlock);
+                }
         }
 
         if (reassert)
@@ -1693,7 +1808,7 @@ int dlm_assert_master_handler(struct o2net_msg *msg, u32 len, void *data)
         char *name;
         unsigned int namelen, hash;
         u32 flags;
-        int master_request = 0;
+        int master_request = 0, have_lockres_ref = 0;
         int ret = 0;
 
         if (!dlm_grab(dlm))
@@ -1864,6 +1979,7 @@ ok:
                                 dlm_change_lockres_owner(dlm, res, mle->master);
                         }
                         spin_unlock(&res->spinlock);
+                        have_lockres_ref = 1;
                 }
 
                 /* master is known, detach if not already detached.
@@ -1918,7 +2034,19 @@ done:
         dlm_put(dlm);
         if (master_request) {
                 mlog(0, "need to tell master to reassert\n");
-                ret = EAGAIN;  // positive. negative would shoot down the node.
+                /* positive. negative would shoot down the node. */
+                ret |= DLM_ASSERT_RESPONSE_REASSERT;
+                if (!have_lockres_ref) {
+                        mlog(ML_ERROR, "strange, got assert from %u, MASTER "
+                             "mle present here for %s:%.*s, but no lockres!\n",
+                             assert->node_idx, dlm->name, namelen, name);
+                }
+        }
+        if (have_lockres_ref) {
+                /* let the master know we have a reference to the lockres */
+                ret |= DLM_ASSERT_RESPONSE_MASTERY_REF;
+                mlog(0, "%s:%.*s: got assert from %u, need a ref\n",
+                     dlm->name, namelen, name, assert->node_idx);
         }
         return ret;
 
@@ -2023,9 +2151,7 @@ static void dlm_assert_master_worker(struct dlm_work_item *item, void *data)
          * even if one or more nodes die */
         mlog(0, "worker about to master %.*s here, this=%u\n",
                      res->lockname.len, res->lockname.name, dlm->node_num);
-        ret = dlm_do_assert_master(dlm, res->lockname.name,
-                                   res->lockname.len,
-                                   nodemap, flags);
+        ret = dlm_do_assert_master(dlm, res, nodemap, flags);
         if (ret < 0) {
                 /* no need to restart, we are done */
                 if (!dlm_is_host_down(ret))
@@ -2097,6 +2223,104 @@ static int dlm_pre_master_reco_lockres(struct dlm_ctxt *dlm,
         return ret;
 }
 
+/*
+ * DLM_DEREF_LOCKRES_MSG
+ */
+
+int dlm_drop_lockres_ref(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
+{
+        struct dlm_deref_lockres deref;
+        int ret = 0, r;
+        const char *lockname;
+        unsigned int namelen;
+
+        lockname = res->lockname.name;
+        namelen = res->lockname.len;
+        BUG_ON(namelen > O2NM_MAX_NAME_LEN);
+
+        mlog(0, "%s:%.*s: sending deref to %d\n",
+             dlm->name, namelen, lockname, res->owner);
+        memset(&deref, 0, sizeof(deref));
+        deref.node_idx = dlm->node_num;
+        deref.namelen = namelen;
+        memcpy(deref.name, lockname, namelen);
+
+        ret = o2net_send_message(DLM_DEREF_LOCKRES_MSG, dlm->key,
+                                 &deref, sizeof(deref), res->owner, &r);
+        if (ret < 0)
+                mlog_errno(ret);
+        else if (r < 0) {
+                /* BAD.  other node says I did not have a ref. */
+                mlog(ML_ERROR,"while dropping ref on %s:%.*s "
+                    "(master=%u) got %d.\n", dlm->name, namelen,
+                    lockname, res->owner, r);
+                dlm_print_one_lock_resource(res);
+                BUG();
+        }
+        return ret;
+}
+
+int dlm_deref_lockres_handler(struct o2net_msg *msg, u32 len, void *data)
+{
+        struct dlm_ctxt *dlm = data;
+        struct dlm_deref_lockres *deref = (struct dlm_deref_lockres *)msg->buf;
+        struct dlm_lock_resource *res = NULL;
+        char *name;
+        unsigned int namelen;
+        int ret = -EINVAL;
+        u8 node;
+        unsigned int hash;
+
+        if (!dlm_grab(dlm))
+                return 0;
+
+        name = deref->name;
+        namelen = deref->namelen;
+        node = deref->node_idx;
+
+        if (namelen > DLM_LOCKID_NAME_MAX) {
+                mlog(ML_ERROR, "Invalid name length!");
+                goto done;
+        }
+        if (deref->node_idx >= O2NM_MAX_NODES) {
+                mlog(ML_ERROR, "Invalid node number: %u\n", node);
+                goto done;
+        }
+
+        hash = dlm_lockid_hash(name, namelen);
+
+        spin_lock(&dlm->spinlock);
+        res = __dlm_lookup_lockres_full(dlm, name, namelen, hash);
+        if (!res) {
+                spin_unlock(&dlm->spinlock);
+                mlog(ML_ERROR, "%s:%.*s: bad lockres name\n",
+                     dlm->name, namelen, name);
+                goto done;
+        }
+        spin_unlock(&dlm->spinlock);
+
+        spin_lock(&res->spinlock);
+        BUG_ON(res->state & DLM_LOCK_RES_DROPPING_REF);
+        if (test_bit(node, res->refmap)) {
+                ret = 0;
+                dlm_lockres_clear_refmap_bit(node, res);
+        } else {
+                mlog(ML_ERROR, "%s:%.*s: node %u trying to drop ref "
+                     "but it is already dropped!\n", dlm->name, namelen,
+                     name, node);
+                __dlm_print_one_lock_resource(res);
+        }
+        spin_unlock(&res->spinlock);
+
+        if (!ret)
+                dlm_lockres_calc_usage(dlm, res);
+done:
+        if (res)
+                dlm_lockres_put(res);
+        dlm_put(dlm);
+        return ret;
+}
+
 
 /*
  * DLM_MIGRATE_LOCKRES
@@ -2376,6 +2600,53 @@ leave:
         return ret;
 }
 
+#define DLM_MIGRATION_RETRY_MS  100
+
+/* Should be called only after beginning the domain leave process.
+ * There should not be any remaining locks on nonlocal lock resources,
+ * and there should be no local locks left on locally mastered resources.
+ *
+ * Called with the dlm spinlock held, may drop it to do migration, but
+ * will re-acquire before exit.
+ *
+ * Returns: 1 if dlm->spinlock was dropped/retaken, 0 if never dropped */
+int dlm_empty_lockres(struct dlm_ctxt *dlm, struct dlm_lock_resource *res)
+{
+        int ret;
+        int lock_dropped = 0;
+
+        if (res->owner != dlm->node_num) {
+                if (!__dlm_lockres_unused(res)) {
+                        mlog(ML_ERROR, "%s:%.*s: this node is not master, "
+                             "trying to free this but locks remain\n",
+                             dlm->name, res->lockname.len, res->lockname.name);
+                }
+                goto leave;
+        }
+
+        /* Wheee! Migrate lockres here! Will sleep so drop spinlock. */
+        spin_unlock(&dlm->spinlock);
+        lock_dropped = 1;
+        while (1) {
+                ret = dlm_migrate_lockres(dlm, res, O2NM_MAX_NODES);
+                if (ret >= 0)
+                        break;
+                if (ret == -ENOTEMPTY) {
+                        mlog(ML_ERROR, "lockres %.*s still has local locks!\n",
+                                res->lockname.len, res->lockname.name);
+                        BUG();
+                }
+
+                mlog(0, "lockres %.*s: migrate failed, "
+                     "retrying\n", res->lockname.len,
+                     res->lockname.name);
+                msleep(DLM_MIGRATION_RETRY_MS);
+        }
+        spin_lock(&dlm->spinlock);
+leave:
+        return lock_dropped;
+}
+
 int dlm_lock_basts_flushed(struct dlm_ctxt *dlm, struct dlm_lock *lock)
 {
         int ret;
@@ -2490,7 +2761,7 @@ static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
 {
         struct list_head *iter, *iter2;
         struct list_head *queue = &res->granted;
-        int i;
+        int i, bit;
         struct dlm_lock *lock;
 
         assert_spin_locked(&res->spinlock);
@@ -2508,12 +2779,28 @@ static void dlm_remove_nonlocal_locks(struct dlm_ctxt *dlm,
                                 BUG_ON(!list_empty(&lock->bast_list));
                                 BUG_ON(lock->ast_pending);
                                 BUG_ON(lock->bast_pending);
+                                dlm_lockres_clear_refmap_bit(lock->ml.node, res);
                                 list_del_init(&lock->list);
                                 dlm_lock_put(lock);
                         }
                 }
                 queue++;
         }
+        bit = 0;
+        while (1) {
+                bit = find_next_bit(res->refmap, O2NM_MAX_NODES, bit);
+                if (bit >= O2NM_MAX_NODES)
+                        break;
+                /* do not clear the local node reference, if there is a
+                 * process holding this, let it drop the ref itself */
+                if (bit != dlm->node_num) {
+                        mlog(0, "%s:%.*s: node %u had a ref to this "
+                             "migrating lockres, clearing\n", dlm->name,
+                             res->lockname.len, res->lockname.name, bit);
+                        dlm_lockres_clear_refmap_bit(bit, res);
+                }
+                bit++;
+        }
 }
 
 /* for now this is not too intelligent.  we will
@@ -2601,6 +2888,16 @@ static int dlm_do_migrate_request(struct dlm_ctxt *dlm,
                         mlog(0, "migrate request (node %u) returned %d!\n",
                              nodenum, status);
                         ret = status;
+                } else if (status == DLM_MIGRATE_RESPONSE_MASTERY_REF) {
+                        /* during the migration request we short-circuited
+                         * the mastery of the lockres.  make sure we have
+                         * a mastery ref for nodenum */
+                        mlog(0, "%s:%.*s: need ref for node %u\n",
+                             dlm->name, res->lockname.len, res->lockname.name,
+                             nodenum);
+                        spin_lock(&res->spinlock);
+                        dlm_lockres_set_refmap_bit(nodenum, res);
+                        spin_unlock(&res->spinlock);
                 }
         }
 
@@ -2745,7 +3042,13 @@ static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
                         /* remove it from the list so that only one
                          * mle will be found */
                         list_del_init(&tmp->list);
-                        __dlm_mle_detach_hb_events(dlm, mle);
+                        /* this was obviously WRONG.  mle is uninited here.  should be tmp. */
+                        __dlm_mle_detach_hb_events(dlm, tmp);
+                        ret = DLM_MIGRATE_RESPONSE_MASTERY_REF;
+                        mlog(0, "%s:%.*s: master=%u, newmaster=%u, "
+                            "telling master to get ref for cleared out mle "
+                            "during migration\n", dlm->name, namelen, name,
+                            master, new_master);
                 }
                 spin_unlock(&tmp->spinlock);
         }
@@ -2753,6 +3056,8 @@ static int dlm_add_migration_mle(struct dlm_ctxt *dlm,
         /* now add a migration mle to the tail of the list */
         dlm_init_mle(mle, DLM_MLE_MIGRATION, dlm, res, name, namelen);
         mle->new_master = new_master;
+        /* the new master will be sending an assert master for this.
+         * at that point we will get the refmap reference */
         mle->master = master;
         /* do this for consistency with other mle types */
         set_bit(new_master, mle->maybe_map);
@@ -2902,6 +3207,13 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
         clear_bit(dlm->node_num, iter.node_map);
         spin_unlock(&dlm->spinlock);
 
+        /* ownership of the lockres is changing.  account for the
+         * mastery reference here since old_master will briefly have
+         * a reference after the migration completes */
+        spin_lock(&res->spinlock);
+        dlm_lockres_set_refmap_bit(old_master, res);
+        spin_unlock(&res->spinlock);
+
         mlog(0, "now time to do a migrate request to other nodes\n");
         ret = dlm_do_migrate_request(dlm, res, old_master,
                                      dlm->node_num, &iter);
@@ -2914,8 +3226,7 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
              res->lockname.len, res->lockname.name);
         /* this call now finishes out the nodemap
          * even if one or more nodes die */
-        ret = dlm_do_assert_master(dlm, res->lockname.name,
-                                   res->lockname.len, iter.node_map,
+        ret = dlm_do_assert_master(dlm, res, iter.node_map,
                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
         if (ret < 0) {
                 /* no longer need to retry.  all living nodes contacted. */
@@ -2927,8 +3238,7 @@ int dlm_finish_migration(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
         set_bit(old_master, iter.node_map);
         mlog(0, "doing assert master of %.*s back to %u\n",
              res->lockname.len, res->lockname.name, old_master);
-        ret = dlm_do_assert_master(dlm, res->lockname.name,
-                                   res->lockname.len, iter.node_map,
+        ret = dlm_do_assert_master(dlm, res, iter.node_map,
                                    DLM_ASSERT_MASTER_FINISH_MIGRATION);
         if (ret < 0) {
                 mlog(0, "assert master to original master failed "