Whamcloud - gitweb
b=23127 osc lock granted fix
[fs/lustre-release.git] / lustre / ldlm / ldlm_lockd.c
index cd8623d..81075ee 100644 (file)
@@ -133,6 +133,9 @@ struct ldlm_bl_work_item {
         struct ldlm_lock       *blwi_lock;
         cfs_list_t              blwi_head;
         int                     blwi_count;
+        cfs_completion_t        blwi_comp;
+        int                     blwi_mode;
+        int                     blwi_mem_pressure;
 };
 
 #ifdef __KERNEL__
@@ -252,7 +255,7 @@ static int ldlm_lock_busy(struct ldlm_lock *lock)
         if (lock->l_export == NULL)
                 return 0;
 
-        cfs_spin_lock(&lock->l_export->exp_lock);
+        cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
         cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
                                 rq_exp_list) {
                 if (req->rq_ops->hpreq_lock_match) {
@@ -261,7 +264,7 @@ static int ldlm_lock_busy(struct ldlm_lock *lock)
                                 break;
                 }
         }
-        cfs_spin_unlock(&lock->l_export->exp_lock);
+        cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
         RETURN(match);
 }
 
@@ -693,14 +696,14 @@ static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
                 RETURN_EXIT;
         }
 
-        cfs_spin_lock(&lock->l_export->exp_lock);
+        cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
         cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
                                 rq_exp_list) {
                 if (!req->rq_hp && req->rq_ops->hpreq_lock_match &&
                     req->rq_ops->hpreq_lock_match(req, lock))
                         ptlrpc_hpreq_reorder(req);
         }
-        cfs_spin_unlock(&lock->l_export->exp_lock);
+        cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
         EXIT;
 }
 
@@ -1436,10 +1439,7 @@ int ldlm_handle_cancel(struct ptlrpc_request *req)
         if (!ldlm_request_cancel(req, dlm_req, 0))
                 req->rq_status = ESTALE;
 
-        if (ptlrpc_reply(req) != 0)
-                LBUG();
-
-        RETURN(0);
+        RETURN(ptlrpc_reply(req));
 }
 
 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
@@ -1561,6 +1561,10 @@ static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 
         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 
+        /* Let Enqueue to call osc_lock_upcall() and initialize
+         * l_ast_data */
+        OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
+
         ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
 
         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
@@ -1622,22 +1626,46 @@ static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 }
 
 #ifdef __KERNEL__
-static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
-                             struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
-                             cfs_list_t *cancels, int count)
+static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
 {
         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
-        struct ldlm_bl_work_item *blwi;
         ENTRY;
 
-        if (cancels && count == 0)
-                RETURN(0);
+        cfs_spin_lock(&blp->blp_lock);
+        if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
+                /* add LDLM_FL_DISCARD_DATA requests to the priority list */
+                cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
+        } else {
+                /* other blocking callbacks are added to the regular list */
+                cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
+        }
+        cfs_spin_unlock(&blp->blp_lock);
 
-        OBD_ALLOC(blwi, sizeof(*blwi));
-        if (blwi == NULL)
-                RETURN(-ENOMEM);
+        cfs_waitq_signal(&blp->blp_waitq);
+
+        /* can not use blwi->blwi_mode as blwi could be already freed in
+           LDLM_ASYNC mode */
+        if (mode == LDLM_SYNC)
+                cfs_wait_for_completion(&blwi->blwi_comp);
+
+        RETURN(0);
+}
+
+static inline void init_blwi(struct ldlm_bl_work_item *blwi,
+                             struct ldlm_namespace *ns,
+                             struct ldlm_lock_desc *ld,
+                             cfs_list_t *cancels, int count,
+                             struct ldlm_lock *lock,
+                             int mode)
+{
+        cfs_init_completion(&blwi->blwi_comp);
+        CFS_INIT_LIST_HEAD(&blwi->blwi_head);
+
+        if (cfs_memory_pressure_get())
+                blwi->blwi_mem_pressure = 1;
 
         blwi->blwi_ns = ns;
+        blwi->blwi_mode = mode;
         if (ld != NULL)
                 blwi->blwi_ld = *ld;
         if (count) {
@@ -1647,36 +1675,53 @@ static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
         } else {
                 blwi->blwi_lock = lock;
         }
-        cfs_spin_lock(&blp->blp_lock);
-        if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
-                /* add LDLM_FL_DISCARD_DATA requests to the priority list */
-                cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
+}
+
+static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
+                             struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
+                             cfs_list_t *cancels, int count, int mode)
+{
+        ENTRY;
+
+        if (cancels && count == 0)
+                RETURN(0);
+
+        if (mode == LDLM_SYNC) {
+                /* if it is synchronous call do minimum mem alloc, as it could
+                 * be triggered from kernel shrinker
+                 */
+                struct ldlm_bl_work_item blwi;
+                memset(&blwi, 0, sizeof(blwi));
+                init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
+                RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
         } else {
-                /* other blocking callbacks are added to the regular list */
-                cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
-        }
-        cfs_waitq_signal(&blp->blp_waitq);
-        cfs_spin_unlock(&blp->blp_lock);
+                struct ldlm_bl_work_item *blwi;
+                OBD_ALLOC(blwi, sizeof(*blwi));
+                if (blwi == NULL)
+                        RETURN(-ENOMEM);
+                init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
 
-        RETURN(0);
+                RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
+        }
 }
+
 #endif
 
 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
                            struct ldlm_lock *lock)
 {
 #ifdef __KERNEL__
-        RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
+        RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
 #else
         RETURN(-ENOSYS);
 #endif
 }
 
 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
-                           cfs_list_t *cancels, int count)
+                           cfs_list_t *cancels, int count, int mode)
 {
 #ifdef __KERNEL__
-        RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
+        RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
 #else
         RETURN(-ENOSYS);
 #endif
@@ -1964,9 +2009,11 @@ static int ldlm_cancel_handler(struct ptlrpc_request *req)
         if (req->rq_export == NULL) {
                 struct ldlm_request *dlm_req;
 
-                CERROR("operation %d from %s with bad export cookie "LPU64"\n",
-                       lustre_msg_get_opc(req->rq_reqmsg),
-                       libcfs_id2str(req->rq_peer),
+                CERROR("%s from %s arrived at %lu with bad export cookie "
+                       LPU64"\n",
+                       ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
+                       libcfs_nid2str(req->rq_peer.nid),
+                       req->rq_arrival_time.tv_sec,
                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
 
                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
@@ -2169,19 +2216,30 @@ static int ldlm_bl_thread_main(void *arg)
                                 /* added by ldlm_cleanup() */
                                 break;
                 }
+                if (blwi->blwi_mem_pressure)
+                        cfs_memory_pressure_set();
 
                 if (blwi->blwi_count) {
+                        int count;
                         /* The special case when we cancel locks in lru
                          * asynchronously, we pass the list of locks here.
-                         * Thus lock is marked LDLM_FL_CANCELING, and already
-                         * canceled locally. */
-                        ldlm_cli_cancel_list(&blwi->blwi_head,
-                                             blwi->blwi_count, NULL, 0);
+                         * Thus locks are marked LDLM_FL_CANCELING, but NOT
+                         * canceled locally yet. */
+                        count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
+                                                           blwi->blwi_count,
+                                                           LCF_BL_AST);
+                        ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
                 } else {
                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
                                                 blwi->blwi_lock);
                 }
-                OBD_FREE(blwi, sizeof(*blwi));
+                if (blwi->blwi_mem_pressure)
+                        cfs_memory_pressure_clr();
+
+                if (blwi->blwi_mode == LDLM_ASYNC)
+                        OBD_FREE(blwi, sizeof(*blwi));
+                else
+                        cfs_complete(&blwi->blwi_comp);
         }
 
         cfs_atomic_dec(&blp->blp_busy_threads);
@@ -2604,6 +2662,7 @@ EXPORT_SYMBOL(ldlm_namespace_foreach);
 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
 EXPORT_SYMBOL(ldlm_resource_iterate);
 EXPORT_SYMBOL(ldlm_cancel_resource_local);
+EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
 EXPORT_SYMBOL(ldlm_cli_cancel_list);
 
 /* ldlm_lockd.c */