Whamcloud - gitweb
LU-8753 osp: add rpc generation
[fs/lustre-release.git] / lustre / osp / osp_trans.c
index 281478d..ebecd8a 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2014, 2015, Intel Corporation.
+ * Copyright (c) 2014, 2016, Intel Corporation.
  */
 /*
  * lustre/osp/osp_trans.c
@@ -77,6 +77,7 @@ struct osp_update_args {
        atomic_t                 *oaua_count;
        wait_queue_head_t        *oaua_waitq;
        bool                      oaua_flow_control;
+       const struct lu_env      *oaua_update_env;
 };
 
 /**
@@ -186,6 +187,7 @@ osp_current_object_update_request(struct osp_update_request *our)
 struct osp_update_request *osp_update_request_create(struct dt_device *dt)
 {
        struct osp_update_request *our;
+       int rc;
 
        OBD_ALLOC_PTR(our);
        if (our == NULL)
@@ -194,13 +196,19 @@ struct osp_update_request *osp_update_request_create(struct dt_device *dt)
        INIT_LIST_HEAD(&our->our_req_list);
        INIT_LIST_HEAD(&our->our_cb_items);
        INIT_LIST_HEAD(&our->our_list);
+       INIT_LIST_HEAD(&our->our_invalidate_cb_list);
        spin_lock_init(&our->our_list_lock);
 
-       osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
+       rc = osp_object_update_request_create(our, OUT_UPDATE_INIT_BUFFER_SIZE);
+       if (rc != 0) {
+               OBD_FREE_PTR(our);
+               return ERR_PTR(rc);
+       }
        return our;
 }
 
-void osp_update_request_destroy(struct osp_update_request *our)
+void osp_update_request_destroy(const struct lu_env *env,
+                               struct osp_update_request *our)
 {
        struct osp_update_request_sub *ours;
        struct osp_update_request_sub *tmp;
@@ -214,6 +222,31 @@ void osp_update_request_destroy(struct osp_update_request *our)
                        OBD_FREE_LARGE(ours->ours_req, ours->ours_req_size);
                OBD_FREE_PTR(ours);
        }
+
+       if (!list_empty(&our->our_invalidate_cb_list)) {
+               struct lu_env lenv;
+               struct osp_object *obj;
+               struct osp_object *next;
+
+               if (env == NULL) {
+                       lu_env_init(&lenv, LCT_MD_THREAD | LCT_DT_THREAD);
+                       env = &lenv;
+               }
+
+               list_for_each_entry_safe(obj, next,
+                                        &our->our_invalidate_cb_list,
+                                        opo_invalidate_cb_list) {
+                       spin_lock(&obj->opo_lock);
+                       list_del_init(&obj->opo_invalidate_cb_list);
+                       spin_unlock(&obj->opo_lock);
+
+                       lu_object_put(env, &obj->opo_obj.do_lu);
+               }
+
+               if (env == &lenv)
+                       lu_env_fini(&lenv);
+       }
+
        OBD_FREE_PTR(our);
 }
 
@@ -231,7 +264,7 @@ object_update_request_dump(const struct object_update_request *ourq,
                update = object_update_request_get(ourq, i, &size);
                LASSERT(update != NULL);
                CDEBUG(mask, "i = %u fid = "DFID" op = %s "
-                      "params = %d batchid = "LPU64" size = %zu repsize %u\n",
+                      "params = %d batchid = %llu size = %zu repsize %u\n",
                       i, PFID(&update->ou_fid),
                       update_op_str(update->ou_type),
                       update->ou_params_count,
@@ -451,10 +484,7 @@ int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
        if (rc != 0)
                RETURN(rc);
 
-       /* This will only be called with read-only update, and these updates
-        * might be used to retrieve update log during recovery process, so
-        * it will be allowed to send during recovery process */
-       req->rq_allow_replay = 1;
+       osp_set_req_replay(osp, req);
        req->rq_allow_intr = 1;
 
        /* Note: some dt index api might return non-zero result here, like
@@ -478,48 +508,26 @@ int osp_remote_sync(const struct lu_env *env, struct osp_device *osp,
  * \param[in] oth      osp thandle.
  */
 static void osp_thandle_invalidate_object(const struct lu_env *env,
-                                         struct osp_thandle *oth)
+                                         struct osp_thandle *oth,
+                                         int result)
 {
        struct osp_update_request *our = oth->ot_our;
-       struct osp_update_request_sub *ours;
+       struct osp_object *obj;
+       struct osp_object *next;
 
        if (our == NULL)
                return;
 
-       list_for_each_entry(ours, &our->our_req_list, ours_list) {
-               struct object_update_request *our_req = ours->ours_req;
-               unsigned int i;
-               struct lu_object *obj;
-               struct osp_object *osp_obj;
-
-               for (i = 0; i < our_req->ourq_count; i++) {
-                       struct object_update *update;
-
-                       update = object_update_request_get(our_req, i, NULL);
-                       if (update == NULL)
-                               break;
-
-                       if (update->ou_type != OUT_WRITE)
-                               continue;
-
-                       if (!fid_is_sane(&update->ou_fid))
-                               continue;
-
-                       obj = lu_object_find_slice(env,
-                                       &oth->ot_super.th_dev->dd_lu_dev,
-                                       &update->ou_fid, NULL);
-                       if (IS_ERR(obj))
-                               break;
-
-                       osp_obj = lu2osp_obj(obj);
-                       if (osp_obj->opo_ooa != NULL) {
-                               spin_lock(&osp_obj->opo_lock);
-                               osp_obj->opo_ooa->ooa_attr.la_valid = 0;
-                               osp_obj->opo_stale = 1;
-                               spin_unlock(&osp_obj->opo_lock);
-                       }
-                       lu_object_put(env, obj);
-               }
+       list_for_each_entry_safe(obj, next, &our->our_invalidate_cb_list,
+                                opo_invalidate_cb_list) {
+               if (result < 0)
+                       osp_invalidate(env, &obj->opo_obj);
+
+               spin_lock(&obj->opo_lock);
+               list_del_init(&obj->opo_invalidate_cb_list);
+               spin_unlock(&obj->opo_lock);
+
+               lu_object_put(env, &obj->opo_obj.do_lu);
        }
 }
 
@@ -539,8 +547,7 @@ static void osp_trans_stop_cb(const struct lu_env *env,
                dcb->dcb_func(NULL, &oth->ot_super, dcb, result);
        }
 
-       if (result < 0)
-               osp_thandle_invalidate_object(env, oth);
+       osp_thandle_invalidate_object(env, oth, result);
 }
 
 /**
@@ -619,6 +626,17 @@ static int osp_update_interpret(const struct lu_env *env,
        if (our == NULL)
                RETURN(0);
 
+       /* Sigh env might be NULL in some cases, see
+        * this calling path.
+        * osp_send_update_thread()
+        *  ptlrpc_set_wait() ----> null env.
+        *   ptlrpc_check_set()
+        *    osp_update_interpret()
+        * Let's use env in oaua for this case.
+        */
+       if (env == NULL)
+               env = oaua->oaua_update_env;
+
        oaua->oaua_update = NULL;
        oth = our->our_th;
        if (oaua->oaua_flow_control) {
@@ -680,9 +698,9 @@ static int osp_update_interpret(const struct lu_env *env,
                /* oth and osp_update_requests will be destoryed in
                 * osp_thandle_put */
                osp_trans_stop_cb(env, oth, rc);
-               osp_thandle_put(oth);
+               osp_thandle_put(env, oth);
        } else {
-               osp_update_request_destroy(our);
+               osp_update_request_destroy(env, our);
        }
 
        RETURN(rc);
@@ -722,12 +740,16 @@ int osp_unplug_async_request(const struct lu_env *env,
                                                     ouc->ouc_data, 0, rc);
                        osp_update_callback_fini(env, ouc);
                }
-               osp_update_request_destroy(our);
+               osp_update_request_destroy(env, our);
        } else {
                args = ptlrpc_req_async_args(req);
                args->oaua_update = our;
                args->oaua_count = NULL;
                args->oaua_waitq = NULL;
+               /* Note: this is asynchronous call for the request, so the
+                * interrupte cb and current function will be different
+                * thread, so we need use different env */
+               args->oaua_update_env = NULL;
                args->oaua_flow_control = false;
                req->rq_interpret_reply = osp_update_interpret;
                ptlrpcd_add_req(req);
@@ -900,13 +922,14 @@ int osp_trans_update_request_create(struct thandle *th)
        return 0;
 }
 
-void osp_thandle_destroy(struct osp_thandle *oth)
+void osp_thandle_destroy(const struct lu_env *env,
+                        struct osp_thandle *oth)
 {
        LASSERT(oth->ot_magic == OSP_THANDLE_MAGIC);
        LASSERT(list_empty(&oth->ot_commit_dcb_list));
        LASSERT(list_empty(&oth->ot_stop_dcb_list));
        if (oth->ot_our != NULL)
-               osp_update_request_destroy(oth->ot_our);
+               osp_update_request_destroy(env, oth->ot_our);
        OBD_FREE_PTR(oth);
 }
 
@@ -1022,7 +1045,7 @@ static void osp_request_commit_cb(struct ptlrpc_request *req)
                last_committed_transno =
                        req->rq_import->imp_peer_committed_transno;
 
-       CDEBUG(D_HA, "trans no "LPU64" committed transno "LPU64"\n",
+       CDEBUG(D_HA, "trans no %llu committed transno %llu\n",
               req->rq_transno, last_committed_transno);
 
        /* If the transaction is not really committed, mark result = 1 */
@@ -1032,7 +1055,7 @@ static void osp_request_commit_cb(struct ptlrpc_request *req)
 
        osp_trans_commit_cb(oth, result);
        req->rq_committed = 1;
-       osp_thandle_put(oth);
+       osp_thandle_put(NULL, oth);
        EXIT;
 }
 
@@ -1091,7 +1114,6 @@ static int osp_send_update_req(const struct lu_env *env,
 {
        struct osp_update_args  *args;
        struct ptlrpc_request   *req;
-       struct lu_device *top_device;
        struct osp_thandle      *oth = our->our_th;
        int     rc = 0;
        ENTRY;
@@ -1106,19 +1128,22 @@ static int osp_send_update_req(const struct lu_env *env,
 
        args = ptlrpc_req_async_args(req);
        args->oaua_update = our;
+       /* set env to NULL, in case the interrupt cb and current function
+        * are in different thread */
+       args->oaua_update_env = NULL;
        osp_thandle_get(oth); /* hold for update interpret */
        req->rq_interpret_reply = osp_update_interpret;
        if (!oth->ot_super.th_wait_submit && !oth->ot_super.th_sync) {
                if (!osp->opd_imp_active || !osp->opd_imp_connected) {
                        osp_trans_callback(env, oth, rc);
-                       osp_thandle_put(oth);
+                       osp_thandle_put(env, oth);
                        GOTO(out, rc = -ENOTCONN);
                }
 
                rc = obd_get_request_slot(&osp->opd_obd->u.cli);
                if (rc != 0) {
                        osp_trans_callback(env, oth, rc);
-                       osp_thandle_put(oth);
+                       osp_thandle_put(env, oth);
                        GOTO(out, rc = -ENOTCONN);
                }
                args->oaua_flow_control = true;
@@ -1145,10 +1170,12 @@ static int osp_send_update_req(const struct lu_env *env,
                 * status, in case the other target is being recoveried
                 * at the same time, and if we wait here for the import
                 * to be recoveryed, it might cause deadlock */
-               top_device = osp->opd_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
-               if (top_device->ld_obd->obd_recovering)
-                       req->rq_allow_replay = 1;
+               osp_set_req_replay(osp, req);
 
+               /* Because this req will be synchronus, i.e. it will be called
+                * in the same thread, so it will be safe to use current
+                * env */
+               args->oaua_update_env = env;
                if (osp->opd_connect_mdt)
                        osp_get_rpc_lock(osp);
                rc = ptlrpc_queue_wait(req);
@@ -1160,13 +1187,13 @@ static int osp_send_update_req(const struct lu_env *env,
                                /* If osp_update_interpret is not being called,
                                 * release the osp_thandle */
                                args->oaua_update = NULL;
-                               osp_thandle_put(oth);
+                               osp_thandle_put(env, oth);
                        }
 
                        req->rq_cb_data = NULL;
                        rc = rc == 0 ? req->rq_status : rc;
                        osp_trans_callback(env, oth, rc);
-                       osp_thandle_put(oth);
+                       osp_thandle_put(env, oth);
                        GOTO(out, rc);
                }
        }
@@ -1231,7 +1258,7 @@ struct thandle *osp_get_storage_thandle(const struct lu_env *env,
  *
  * Set the version for the transaction and add the request to
  * the sending list, then after transaction stop, the request
- * will be picked in the order of version, by sending thread.
+ * will be sent in the order of version by the sending thread.
  *
  * \param [in] oth     osp thandle to be set version.
  *
@@ -1261,6 +1288,7 @@ int osp_check_and_set_rpc_version(struct osp_thandle *oth,
        /* Assign the version and add it to the sending list */
        osp_thandle_get(oth);
        oth->ot_our->our_version = ou->ou_version++;
+       oth->ot_our->our_generation = ou->ou_generation;
        list_add_tail(&oth->ot_our->our_list,
                      &osp->opd_update->ou_list);
        oth->ot_our->our_req_ready = 0;
@@ -1268,8 +1296,8 @@ int osp_check_and_set_rpc_version(struct osp_thandle *oth,
        spin_unlock(&ou->ou_lock);
 
        LASSERT(oth->ot_super.th_wait_submit == 1);
-       CDEBUG(D_INFO, "%s: version "LPU64" oth:version %p:"LPU64"\n",
-              osp->opd_obd->obd_name, ou->ou_version, oth,
+       CDEBUG(D_INFO, "%s: version %llu gen %llu oth:version %p:%llu\n",
+              osp->opd_obd->obd_name, ou->ou_version, ou->ou_generation, oth,
               oth->ot_our->our_version);
 
        return 0;
@@ -1298,7 +1326,7 @@ osp_get_next_request(struct osp_updates *ou, struct osp_update_request **ourp)
        spin_lock(&ou->ou_lock);
        list_for_each_entry_safe(our, tmp, &ou->ou_list, our_list) {
                LASSERT(our->our_th != NULL);
-               CDEBUG(D_HA, "ou %p version "LPU64" rpc_version "LPU64"\n",
+               CDEBUG(D_HA, "ou %p version %llu rpc_version %llu\n",
                       ou, our->our_version, ou->ou_rpc_version);
                spin_lock(&our->our_list_lock);
                /* Find next osp_update_request in the list */
@@ -1345,6 +1373,11 @@ void osp_invalidate_request(struct osp_device *osp)
        if (rc < 0) {
                CERROR("%s: init env error: rc = %d\n", osp->opd_obd->obd_name,
                       rc);
+
+               spin_lock(&ou->ou_lock);
+               ou->ou_generation++;
+               spin_unlock(&ou->ou_lock);
+
                return;
        }
 
@@ -1370,6 +1403,9 @@ void osp_invalidate_request(struct osp_device *osp)
                       our);
        }
 
+       /* Increase the generation, then the update request with old generation
+        * will fail with -EIO. */
+       ou->ou_generation++;
        spin_unlock(&ou->ou_lock);
 
        /* invalidate all of request in the sending list */
@@ -1379,7 +1415,7 @@ void osp_invalidate_request(struct osp_device *osp)
                spin_unlock(&our->our_list_lock);
                osp_trans_callback(&env, our->our_th,
                                   our->our_th->ot_super.th_result);
-               osp_thandle_put(our->our_th);
+               osp_thandle_put(&env, our->our_th);
        }
        lu_env_fini(&env);
 }
@@ -1427,7 +1463,7 @@ int osp_send_update_thread(void *arg)
                if (!osp_send_update_thread_running(osp)) {
                        if (our != NULL) {
                                osp_trans_callback(&env, our->our_th, -EINTR);
-                               osp_thandle_put(our->our_th);
+                               osp_thandle_put(&env, our->our_th);
                        }
                        break;
                }
@@ -1437,7 +1473,8 @@ int osp_send_update_thread(void *arg)
                        osp_trans_callback(&env, our->our_th,
                                our->our_th->ot_super.th_result);
                        rc = our->our_th->ot_super.th_result;
-               } else if (OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
+               } else if (ou->ou_generation != our->our_generation ||
+                          OBD_FAIL_CHECK(OBD_FAIL_INVALIDATE_UPDATE)) {
                        rc = -EIO;
                        osp_trans_callback(&env, our->our_th, rc);
                } else {
@@ -1458,7 +1495,7 @@ int osp_send_update_thread(void *arg)
                        osp_invalidate_request(osp);
 
                /* Balanced for thandle_get in osp_check_and_set_rpc_version */
-               osp_thandle_put(our->our_th);
+               osp_thandle_put(&env, our->our_th);
        }
 
        thread->t_flags = SVC_STOPPED;
@@ -1550,7 +1587,7 @@ int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
                GOTO(out, rc = -EIO);
        }
 
-       CDEBUG(D_HA, "%s: add oth %p with version "LPU64"\n",
+       CDEBUG(D_HA, "%s: add oth %p with version %llu\n",
               osp->opd_obd->obd_name, oth, our->our_version);
 
        LASSERT(our->our_req_ready == 0);
@@ -1570,7 +1607,7 @@ int osp_trans_stop(const struct lu_env *env, struct dt_device *dt,
                osp_trans_callback(env, oth, th->th_result);
        }
 out:
-       osp_thandle_put(oth);
+       osp_thandle_put(env, oth);
 
        RETURN(rc);
 }