ocfs2/dlm: continue to purge recovery lockres when recovery master goes down
authorpiaojun <piaojun@huawei.com>
Tue, 2 Aug 2016 21:02:19 +0000 (14:02 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 2 Aug 2016 21:31:41 +0000 (17:31 -0400)
We found a dlm-blocked situation caused by continuous breakdown of
recovery masters described below.  To solve this problem, we should
purge recovery lock once detecting recovery master goes down.

N3                      N2                   N1(reco master)
                        go down
                                             pick up recovery lock and
                                             begin recoverying for N2

                                             go down

pick up recovery
lock failed, then
purge it:
dlm_purge_lockres
  ->DROPPING_REF is set

send deref to N1 failed,
recovery lock is not purged

find N1 go down, begin
recoverying for N1, but
blocked in dlm_do_recovery
as DROPPING_REF is set:
dlm_do_recovery
  ->dlm_pick_recovery_master
    ->dlmlock
      ->dlm_get_lock_resource
        ->__dlm_wait_on_lockres_flags(tmpres,
   DLM_LOCK_RES_DROPPING_REF);

Fixes: 8c0343968163 ("ocfs2/dlm: clear DROPPING_REF flag when the master goes down")
Link: http://lkml.kernel.org/r/578453AF.8030404@huawei.com
Signed-off-by: Jun Piao <piaojun@huawei.com>
Reviewed-by: Joseph Qi <joseph.qi@huawei.com>
Reviewed-by: Jiufei Xue <xuejiufei@huawei.com>
Reviewed-by: Mark Fasheh <mfasheh@suse.de>
Cc: Joel Becker <jlbec@evilplan.org>
Cc: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/ocfs2/dlm/dlmcommon.h
fs/ocfs2/dlm/dlmmaster.c
fs/ocfs2/dlm/dlmrecovery.c
fs/ocfs2/dlm/dlmthread.c

index 8107d0d0c3f6e8857f4cbbfd8f14f3573b0267c2..e9f3705c4c9fef7e9614eb3f6c48c98f1f42550e 100644 (file)
@@ -1004,6 +1004,8 @@ int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data,
 int dlm_do_master_requery(struct dlm_ctxt *dlm, struct dlm_lock_resource *res,
                          u8 nodenum, u8 *real_master);
 
+void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
+               struct dlm_lock_resource *res);
 
 int dlm_dispatch_assert_master(struct dlm_ctxt *dlm,
                               struct dlm_lock_resource *res,
index 553d220df4063a9caf7651b45891adf9fa175801..6ea06f8a7d295acd45e24f36a8fd50046897d78a 100644 (file)
@@ -2425,51 +2425,20 @@ int dlm_deref_lockres_done_handler(struct o2net_msg *msg, u32 len, void *data,
                mlog(ML_NOTICE, "%s:%.*s: node %u sends deref done "
                        "but it is already derefed!\n", dlm->name,
                        res->lockname.len, res->lockname.name, node);
-               dlm_lockres_put(res);
                ret = 0;
                goto done;
        }
 
-       if (!list_empty(&res->purge)) {
-               mlog(0, "%s: Removing res %.*s from purgelist\n",
-                       dlm->name, res->lockname.len, res->lockname.name);
-               list_del_init(&res->purge);
-               dlm_lockres_put(res);
-               dlm->purge_count--;
-       }
-
-       if (!__dlm_lockres_unused(res)) {
-               mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
-                       dlm->name, res->lockname.len, res->lockname.name);
-               __dlm_print_one_lock_resource(res);
-               BUG();
-       }
-
-       __dlm_unhash_lockres(dlm, res);
-
-       spin_lock(&dlm->track_lock);
-       if (!list_empty(&res->tracking))
-               list_del_init(&res->tracking);
-       else {
-               mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
-                    dlm->name, res->lockname.len, res->lockname.name);
-               __dlm_print_one_lock_resource(res);
-       }
-       spin_unlock(&dlm->track_lock);
-
-       /* lockres is not in the hash now. drop the flag and wake up
-        * any processes waiting in dlm_get_lock_resource.
-        */
-       res->state &= ~DLM_LOCK_RES_DROPPING_REF;
+       __dlm_do_purge_lockres(dlm, res);
        spin_unlock(&res->spinlock);
        wake_up(&res->wq);
 
-       dlm_lockres_put(res);
-
        spin_unlock(&dlm->spinlock);
 
        ret = 0;
 done:
+       if (res)
+               dlm_lockres_put(res);
        dlm_put(dlm);
        return ret;
 }
index f6b313898763a33173200eab37f1ab8d61f23c44..dd5cb8bcefd13820ae0d242dc1e02da5798736c6 100644 (file)
@@ -2343,6 +2343,7 @@ static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
        struct dlm_lock_resource *res;
        int i;
        struct hlist_head *bucket;
+       struct hlist_node *tmp;
        struct dlm_lock *lock;
 
 
@@ -2365,7 +2366,7 @@ static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
         */
        for (i = 0; i < DLM_HASH_BUCKETS; i++) {
                bucket = dlm_lockres_hash(dlm, i);
-               hlist_for_each_entry(res, bucket, hash_node) {
+               hlist_for_each_entry_safe(res, tmp, bucket, hash_node) {
                        /* always prune any $RECOVERY entries for dead nodes,
                         * otherwise hangs can occur during later recovery */
                        if (dlm_is_recovery_lock(res->lockname.name,
@@ -2386,8 +2387,17 @@ static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
                                                break;
                                        }
                                }
-                               dlm_lockres_clear_refmap_bit(dlm, res,
-                                               dead_node);
+
+                               if ((res->owner == dead_node) &&
+                                                       (res->state & DLM_LOCK_RES_DROPPING_REF)) {
+                                       dlm_lockres_get(res);
+                                       __dlm_do_purge_lockres(dlm, res);
+                                       spin_unlock(&res->spinlock);
+                                       wake_up(&res->wq);
+                                       dlm_lockres_put(res);
+                                       continue;
+                               } else if (res->owner == dlm->node_num)
+                                       dlm_lockres_clear_refmap_bit(dlm, res, dead_node);
                                spin_unlock(&res->spinlock);
                                continue;
                        }
@@ -2398,14 +2408,17 @@ static void dlm_do_local_recovery_cleanup(struct dlm_ctxt *dlm, u8 dead_node)
                                if (res->state & DLM_LOCK_RES_DROPPING_REF) {
                                        mlog(0, "%s:%.*s: owned by "
                                                "dead node %u, this node was "
-                                               "dropping its ref when it died. "
-                                               "continue, dropping the flag.\n",
+                                               "dropping its ref when master died. "
+                                               "continue, purging the lockres.\n",
                                                dlm->name, res->lockname.len,
                                                res->lockname.name, dead_node);
+                                       dlm_lockres_get(res);
+                                       __dlm_do_purge_lockres(dlm, res);
+                                       spin_unlock(&res->spinlock);
+                                       wake_up(&res->wq);
+                                       dlm_lockres_put(res);
+                                       continue;
                                }
-                               res->state &= ~DLM_LOCK_RES_DROPPING_REF;
-                               dlm_move_lockres_to_recovery_list(dlm,
-                                               res);
                        } else if (res->owner == dlm->node_num) {
                                dlm_free_dead_locks(dlm, res, dead_node);
                                __dlm_lockres_calc_usage(dlm, res);
index ce397229acc04e28bad738d69bf143b42ca5f59d..838a06d4066a6e36d820a1350d45b7656a26f31f 100644 (file)
@@ -160,6 +160,52 @@ void dlm_lockres_calc_usage(struct dlm_ctxt *dlm,
        spin_unlock(&dlm->spinlock);
 }
 
+/*
+ * Do the real purge work:
+ *     unhash the lockres, and
+ *     clear flag DLM_LOCK_RES_DROPPING_REF.
+ * It requires dlm and lockres spinlock to be taken.
+ */
+void __dlm_do_purge_lockres(struct dlm_ctxt *dlm,
+               struct dlm_lock_resource *res)
+{
+       assert_spin_locked(&dlm->spinlock);
+       assert_spin_locked(&res->spinlock);
+
+       if (!list_empty(&res->purge)) {
+               mlog(0, "%s: Removing res %.*s from purgelist\n",
+                    dlm->name, res->lockname.len, res->lockname.name);
+               list_del_init(&res->purge);
+               dlm_lockres_put(res);
+               dlm->purge_count--;
+       }
+
+       if (!__dlm_lockres_unused(res)) {
+               mlog(ML_ERROR, "%s: res %.*s in use after deref\n",
+                    dlm->name, res->lockname.len, res->lockname.name);
+               __dlm_print_one_lock_resource(res);
+               BUG();
+       }
+
+       __dlm_unhash_lockres(dlm, res);
+
+       spin_lock(&dlm->track_lock);
+       if (!list_empty(&res->tracking))
+               list_del_init(&res->tracking);
+       else {
+               mlog(ML_ERROR, "%s: Resource %.*s not on the Tracking list\n",
+                    dlm->name, res->lockname.len, res->lockname.name);
+               __dlm_print_one_lock_resource(res);
+       }
+       spin_unlock(&dlm->track_lock);
+
+       /*
+        * lockres is not in the hash now. drop the flag and wake up
+        * any processes waiting in dlm_get_lock_resource.
+        */
+       res->state &= ~DLM_LOCK_RES_DROPPING_REF;
+}
+
 static void dlm_purge_lockres(struct dlm_ctxt *dlm,
                             struct dlm_lock_resource *res)
 {
@@ -176,10 +222,8 @@ static void dlm_purge_lockres(struct dlm_ctxt *dlm,
 
        if (!master) {
                if (res->state & DLM_LOCK_RES_DROPPING_REF) {
-                       mlog(ML_NOTICE, "%s: res %.*s already in "
-                               "DLM_LOCK_RES_DROPPING_REF state\n",
-                               dlm->name, res->lockname.len,
-                               res->lockname.name);
+                       mlog(ML_NOTICE, "%s: res %.*s already in DLM_LOCK_RES_DROPPING_REF state\n",
+                               dlm->name, res->lockname.len, res->lockname.name);
                        spin_unlock(&res->spinlock);
                        return;
                }