Whamcloud - gitweb
LU-8306 ldlm: send blocking ASTs after lock replay
[fs/lustre-release.git] / lustre / ldlm / ldlm_lock.c
index c78c4bc..6e4663f 100644 (file)
@@ -1835,7 +1835,7 @@ enum ldlm_error ldlm_lock_enqueue(struct ldlm_namespace *ns,
         }
 
         policy = ldlm_processing_policy_table[res->lr_type];
-        policy(lock, flags, 1, &rc, NULL);
+       policy(lock, flags, LDLM_PROCESS_ENQUEUE, &rc, NULL);
         GOTO(out, rc);
 #else
         } else {
@@ -1860,7 +1860,8 @@ out:
  * Must be called with resource lock held.
  */
 int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
-                        struct list_head *work_list)
+                        struct list_head *work_list,
+                        enum ldlm_process_intention intention)
 {
        struct list_head *tmp, *pos;
        ldlm_processing_policy policy;
@@ -1873,6 +1874,8 @@ int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
 
        policy = ldlm_processing_policy_table[res->lr_type];
        LASSERT(policy);
+       LASSERT(intention == LDLM_PROCESS_RESCAN ||
+               intention == LDLM_PROCESS_RECOVERY);
 
        list_for_each_safe(tmp, pos, queue) {
                struct ldlm_lock *pending;
@@ -1882,13 +1885,116 @@ int ldlm_reprocess_queue(struct ldlm_resource *res, struct list_head *queue,
                 CDEBUG(D_INFO, "Reprocessing lock %p\n", pending);
 
                 flags = 0;
-                rc = policy(pending, &flags, 0, &err, work_list);
-                if (rc != LDLM_ITER_CONTINUE)
-                        break;
+               rc = policy(pending, &flags, intention, &err, work_list);
+               /*
+                * When this is called from recovery done, we always want
+                * to scan the whole list no matter what 'rc' is returned.
+                */
+               if (rc != LDLM_ITER_CONTINUE &&
+                   intention == LDLM_PROCESS_RESCAN)
+                       break;
         }
 
-        RETURN(rc);
+        RETURN(intention == LDLM_PROCESS_RESCAN ? rc : LDLM_ITER_CONTINUE);
 }
+
+/**
+ * Conflicting locks are detected for a lock to be enqueued, add the lock
+ * into waiting list and send blocking ASTs to the conflicting locks.
+ *
+ * \param[in] lock             The lock to be enqueued.
+ * \param[out] flags           Lock flags for the lock to be enqueued.
+ * \param[in] rpc_list         Conflicting locks list.
+ * \param[in] grant_flags      extra flags when granting a lock.
+ *
+ * \retval -ERESTART:  Some lock was instantly canceled while sending
+ *                     blocking ASTs, caller needs to re-check conflicting
+ *                     locks.
+ * \retval -EAGAIN:    Lock was destroyed, caller should return error.
+ * \reval 0:           Lock is successfully added in waiting list.
+ */
+int ldlm_handle_conflict_lock(struct ldlm_lock *lock, __u64 *flags,
+                             struct list_head *rpc_list, __u64 grant_flags)
+{
+       struct ldlm_resource *res = lock->l_resource;
+       int rc;
+       ENTRY;
+
+       check_res_locked(res);
+
+       /* If either of the compat_queue()s returned failure, then we
+        * have ASTs to send and must go onto the waiting list.
+        *
+        * bug 2322: we used to unlink and re-add here, which was a
+        * terrible folly -- if we goto restart, we could get
+        * re-ordered!  Causes deadlock, because ASTs aren't sent! */
+       if (list_empty(&lock->l_res_link))
+               ldlm_resource_add_lock(res, &res->lr_waiting, lock);
+       unlock_res(res);
+
+       rc = ldlm_run_ast_work(ldlm_res_to_ns(res), rpc_list,
+                              LDLM_WORK_BL_AST);
+
+       if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_OST_FAIL_RACE) &&
+           !ns_is_client(ldlm_res_to_ns(res)))
+               class_fail_export(lock->l_export);
+
+       lock_res(res);
+       if (rc == -ERESTART) {
+               /* 15715: The lock was granted and destroyed after
+                * resource lock was dropped. Interval node was freed
+                * in ldlm_lock_destroy. Anyway, this always happens
+                * when a client is being evicted. So it would be
+                * ok to return an error. -jay */
+               if (ldlm_is_destroyed(lock))
+                       RETURN(-EAGAIN);
+
+               /* lock was granted while resource was unlocked. */
+               if (lock->l_granted_mode == lock->l_req_mode) {
+                       /* bug 11300: if the lock has been granted,
+                        * break earlier because otherwise, we will go
+                        * to restart and ldlm_resource_unlink will be
+                        * called and it causes the interval node to be
+                        * freed. Then we will fail at
+                        * ldlm_extent_add_lock() */
+                       *flags &= ~LDLM_FL_BLOCKED_MASK;
+                       RETURN(0);
+               }
+
+               RETURN(rc);
+       }
+       *flags |= (LDLM_FL_BLOCK_GRANTED | grant_flags);
+
+       RETURN(0);
+}
+
+/**
+ * Discard all AST work items from list.
+ *
+ * If for whatever reason we do not want to send ASTs to conflicting locks
+ * anymore, disassemble the list with this function.
+ */
+void ldlm_discard_bl_list(struct list_head *bl_list)
+{
+       struct list_head *tmp, *pos;
+        ENTRY;
+
+       list_for_each_safe(pos, tmp, bl_list) {
+                struct ldlm_lock *lock =
+                       list_entry(pos, struct ldlm_lock, l_bl_ast);
+
+               list_del_init(&lock->l_bl_ast);
+               LASSERT(ldlm_is_ast_sent(lock));
+               ldlm_clear_ast_sent(lock);
+               LASSERT(lock->l_bl_ast_run == 0);
+               LASSERT(lock->l_blocking_lock);
+               LDLM_LOCK_RELEASE(lock->l_blocking_lock);
+               lock->l_blocking_lock = NULL;
+               LDLM_LOCK_RELEASE(lock);
+       }
+       EXIT;
+}
+
 #endif
 
 /**
@@ -2102,38 +2208,6 @@ out:
        return rc;
 }
 
-static int reprocess_one_queue(struct ldlm_resource *res, void *closure)
-{
-        ldlm_reprocess_all(res);
-        return LDLM_ITER_CONTINUE;
-}
-
-static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
-                             struct hlist_node *hnode, void *arg)
-{
-        struct ldlm_resource *res = cfs_hash_object(hs, hnode);
-        int    rc;
-
-        rc = reprocess_one_queue(res, arg);
-
-        return rc == LDLM_ITER_STOP;
-}
-
-/**
- * Iterate through all resources on a namespace attempting to grant waiting
- * locks.
- */
-void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
-{
-       ENTRY;
-
-       if (ns != NULL) {
-               cfs_hash_for_each_nolock(ns->ns_rs_hash,
-                                        ldlm_reprocess_res, NULL, 0);
-       }
-       EXIT;
-}
-
 /**
  * Try to grant all waiting locks on a resource.
  *
@@ -2142,7 +2216,8 @@ void ldlm_reprocess_all_ns(struct ldlm_namespace *ns)
  * Typically called after some resource locks are cancelled to see
  * if anything could be granted as a result of the cancellation.
  */
-void ldlm_reprocess_all(struct ldlm_resource *res)
+static void __ldlm_reprocess_all(struct ldlm_resource *res,
+                                enum ldlm_process_intention intention)
 {
        struct list_head rpc_list;
 #ifdef HAVE_SERVER_SUPPORT
@@ -2165,11 +2240,13 @@ void ldlm_reprocess_all(struct ldlm_resource *res)
            atomic_read(&obd->obd_req_replay_clients) == 0)
                RETURN_EXIT;
 restart:
-        lock_res(res);
-        rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list);
-        if (rc == LDLM_ITER_CONTINUE)
-                ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list);
-        unlock_res(res);
+       lock_res(res);
+       rc = ldlm_reprocess_queue(res, &res->lr_converting, &rpc_list,
+                                 intention);
+       if (rc == LDLM_ITER_CONTINUE)
+               ldlm_reprocess_queue(res, &res->lr_waiting, &rpc_list,
+                                    intention);
+       unlock_res(res);
 
         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), &rpc_list,
                                LDLM_WORK_CP_AST);
@@ -2189,8 +2266,38 @@ restart:
 #endif
         EXIT;
 }
+
+void ldlm_reprocess_all(struct ldlm_resource *res)
+{
+       __ldlm_reprocess_all(res, LDLM_PROCESS_RESCAN);
+}
 EXPORT_SYMBOL(ldlm_reprocess_all);
 
+static int ldlm_reprocess_res(struct cfs_hash *hs, struct cfs_hash_bd *bd,
+                             struct hlist_node *hnode, void *arg)
+{
+       struct ldlm_resource *res = cfs_hash_object(hs, hnode);
+
+       /* This is only called once after recovery done. LU-8306. */
+       __ldlm_reprocess_all(res, LDLM_PROCESS_RECOVERY);
+       return 0;
+}
+
+/**
+ * Iterate through all resources on a namespace attempting to grant waiting
+ * locks.
+ */
+void ldlm_reprocess_recovery_done(struct ldlm_namespace *ns)
+{
+       ENTRY;
+
+       if (ns != NULL) {
+               cfs_hash_for_each_nolock(ns->ns_rs_hash,
+                                        ldlm_reprocess_res, NULL, 0);
+       }
+       EXIT;
+}
+
 static bool is_bl_done(struct ldlm_lock *lock)
 {
        bool bl_done = true;
@@ -2557,7 +2664,8 @@ struct ldlm_resource *ldlm_lock_convert(struct ldlm_lock *lock,
                ldlm_processing_policy policy;
 
                 policy = ldlm_processing_policy_table[res->lr_type];
-                rc = policy(lock, &pflags, 0, &err, &rpc_list);
+               rc = policy(lock, &pflags, LDLM_PROCESS_RESCAN, &err,
+                           &rpc_list);
                 if (rc == LDLM_ITER_STOP) {
                         lock->l_req_mode = old_mode;
                         if (res->lr_type == LDLM_EXTENT)