X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=44bdcd46db0414f226950c3b55db3a62e5c1e311;hb=737da0331e8407a704cd11c04f18c2cd3d437800;hp=16578d8c506d2922932ca262e11aaad467149b08;hpb=f95393b0d0a59cf3dc2f29cffc35dcc4cc9d7728;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 16578d8..44bdcd4 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -28,6 +28,8 @@ /* * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -190,6 +192,31 @@ out: RETURN(rc); } +/** + * Find conn uuid by peer nid. @peer is a server nid. This function is used + * to find a conn uuid of @imp which can reach @peer. + */ +int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, + struct obd_uuid *uuid) +{ + struct obd_import_conn *conn; + int rc = -ENOENT; + ENTRY; + + cfs_spin_lock(&imp->imp_lock); + cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { + /* check if conn uuid does have this peer nid */ + if (class_check_uuid(&conn->oic_uuid, peer)) { + *uuid = conn->oic_uuid; + rc = 0; + break; + } + } + cfs_spin_unlock(&imp->imp_lock); + RETURN(rc); +} +EXPORT_SYMBOL(client_import_find_conn); + void client_destroy_import(struct obd_import *imp) { /* drop security policy instance after all rpc finished/aborted @@ -214,6 +241,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) struct obd_uuid server_uuid; int rq_portal, rp_portal, connect_op; char *name = obddev->obd_type->typ_name; + ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN; int rc; ENTRY; @@ -225,12 +253,16 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) connect_op = OST_CONNECT; cli->cl_sp_me = LUSTRE_SP_CLI; cli->cl_sp_to = LUSTRE_SP_OST; + ns_type = LDLM_NS_TYPE_OSC; + } else if (!strcmp(name, LUSTRE_MDC_NAME)) { rq_portal = MDS_REQUEST_PORTAL; rp_portal = MDC_REPLY_PORTAL; connect_op = MDS_CONNECT; cli->cl_sp_me = LUSTRE_SP_CLI; cli->cl_sp_to = LUSTRE_SP_MDT; + ns_type = LDLM_NS_TYPE_MDC; + } else if (!strcmp(name, LUSTRE_MGC_NAME)) { rq_portal = MGS_REQUEST_PORTAL; rp_portal = MGC_REPLY_PORTAL; @@ -238,6 +270,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) cli->cl_sp_me = LUSTRE_SP_MGC; cli->cl_sp_to = LUSTRE_SP_MGS; cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID; + ns_type = LDLM_NS_TYPE_MGC; + } else { CERROR("unknown client OBD type \"%s\", can't setup\n", name); @@ -366,7 +400,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name, LDLM_NAMESPACE_CLIENT, - LDLM_NAMESPACE_GREEDY); + LDLM_NAMESPACE_GREEDY, + ns_type); if (obddev->obd_namespace == NULL) { CERROR("Unable to create client namespace - %s\n", obddev->obd_name); @@ -435,7 +470,7 @@ int client_connect_import(const struct lu_env *env, imp->imp_connect_flags_orig = data->ocd_connect_flags; } - rc = ptlrpc_connect_import(imp, NULL); + rc = ptlrpc_connect_import(imp); if (rc != 0) { LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); GOTO(out_ldlm, rc); @@ -568,18 +603,15 @@ int server_disconnect_export(struct obd_export *exp) struct ptlrpc_reply_state, rs_exp_list); struct ptlrpc_service *svc = rs->rs_service; - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svc->srv_rs_lock); cfs_list_del_init(&rs->rs_exp_list); cfs_spin_lock(&rs->rs_lock); ptlrpc_schedule_difficult_reply(rs); cfs_spin_unlock(&rs->rs_lock); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&svc->srv_rs_lock); } cfs_spin_unlock(&exp->exp_lock); - /* release nid stat refererence */ - lprocfs_exp_cleanup(exp); - RETURN(rc); } @@ -639,10 +671,17 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, } EXPORT_SYMBOL(target_client_add_cb); +#ifdef __KERNEL__ static void -target_start_and_reset_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, - int new_client); +check_and_start_recovery_timer(struct obd_device *obd, + struct ptlrpc_request *req, int new_client); +#else +static inline void +check_and_start_recovery_timer(struct obd_device *obd, + struct ptlrpc_request *req, int new_client) +{ +} +#endif int target_handle_connect(struct ptlrpc_request *req) { @@ -658,6 +697,7 @@ int target_handle_connect(struct ptlrpc_request *req) int rc = 0; int mds_conn = 0; struct obd_connect_data *data, *tmpdata; + int size, tmpsize; lnet_nid_t *client_nid = NULL; ENTRY; @@ -676,7 +716,7 @@ int target_handle_connect(struct ptlrpc_request *req) if (!target || target->obd_stopping || !target->obd_set_up) { LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available " - " for connect (%s)\n", str, + "for connect (%s)\n", str, !target ? "no target" : (target->obd_stopping ? "stopping" : "not set up")); @@ -725,6 +765,8 @@ int target_handle_connect(struct ptlrpc_request *req) conn = *tmp; + size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_CLIENT); data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); if (!data) GOTO(out, rc = -EPROTO); @@ -815,7 +857,6 @@ int target_handle_connect(struct ptlrpc_request *req) cfs_spin_lock(&export->exp_lock); export->exp_connecting = 1; cfs_spin_unlock(&export->exp_lock); - class_export_put(export); LASSERT(export->exp_obd == target); rc = target_handle_reconnect(&conn, export, &cluuid); @@ -859,21 +900,16 @@ no_export: GOTO(out, rc); } - CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n", + CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), target->obd_recovering ? "recovering/" : "", data->ocd_transno, export, (long)cfs_time_current_sec(), export ? (long)export->exp_last_request_time : 0); - /* Tell the client if we're in recovery. */ - if (target->obd_recovering) { - lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); - /* If this is the first time a client connects, - reset the recovery timer */ - if (rc == 0) - target_start_and_reset_recovery_timer(target, req, - !export); - } + /* If this is the first time a client connects, reset the recovery + * timer */ + if (rc == 0 && target->obd_recovering) + check_and_start_recovery_timer(target, req, export == NULL); /* We want to handle EALREADY but *not* -EALREADY from * target_handle_reconnect(), return reconnection state in a flag */ @@ -908,8 +944,12 @@ dont_check_exports: rc = obd_connect(req->rq_svc_thread->t_env, &export, target, &cluuid, data, client_nid); - if (rc == 0) + if (rc == 0) { conn.cookie = export->exp_handle.h_cookie; + /* LU-1092 reconnect put export refcount in the + * end, connect needs take one here too. */ + class_export_get(export); + } } } else { rc = obd_reconnect(req->rq_svc_thread->t_env, @@ -920,13 +960,21 @@ dont_check_exports: } if (rc) GOTO(out, rc); + + LASSERT(target->u.obt.obt_magic == OBT_MAGIC); + data->ocd_instance = target->u.obt.obt_instance; + /* Return only the parts of obd_connect_data that we understand, so the * client knows that we don't understand the rest. */ if (data) { - tmpdata = req_capsule_server_get(&req->rq_pill, - &RMF_CONNECT_DATA); - //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED; - *tmpdata = *data; + tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_SERVER); + tmpdata = req_capsule_server_get(&req->rq_pill, + &RMF_CONNECT_DATA); + /* Don't use struct assignment here, because the client reply + * buffer may be smaller/larger than the local struct + * obd_connect_data. */ + memcpy(tmpdata, data, min(tmpsize, size)); } /* If all else goes well, this is our RPC return code. */ @@ -987,33 +1035,50 @@ dont_check_exports: req->rq_self, &remote_uuid); if (cfs_hlist_unhashed(&export->exp_nid_hash)) { - cfs_hash_add_unique(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); + cfs_hash_add(export->exp_obd->obd_nid_hash, + &export->exp_connection->c_peer.nid, + &export->exp_nid_hash); } - - cfs_spin_lock_bh(&target->obd_processing_task_lock); + /** + class_disconnect->class_export_recovery_cleanup() race + */ if (target->obd_recovering && !export->exp_in_recovery) { + int has_transno; + __u64 transno = data->ocd_transno; + cfs_spin_lock(&export->exp_lock); export->exp_in_recovery = 1; export->exp_req_replay_needed = 1; export->exp_lock_replay_needed = 1; cfs_spin_unlock(&export->exp_lock); - if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO) - && (data->ocd_transno == 0)) + + has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) & + MSG_CONNECT_TRANSNO); + if (has_transno && transno == 0) CWARN("Connect with zero transno!\n"); - if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO) - && data->ocd_transno < target->obd_next_recovery_transno) - target->obd_next_recovery_transno = data->ocd_transno; - target->obd_connected_clients++; + if (has_transno && transno > 0 && + transno < target->obd_next_recovery_transno && + transno > target->obd_last_committed) { + /* another way is to use cmpxchg() so it will be + * lock free */ + cfs_spin_lock(&target->obd_recovery_task_lock); + if (transno < target->obd_next_recovery_transno) + target->obd_next_recovery_transno = transno; + cfs_spin_unlock(&target->obd_recovery_task_lock); + } + cfs_atomic_inc(&target->obd_req_replay_clients); cfs_atomic_inc(&target->obd_lock_replay_clients); - if (target->obd_connected_clients == + if (cfs_atomic_inc_return(&target->obd_connected_clients) == target->obd_max_recoverable_clients) cfs_waitq_signal(&target->obd_next_transno_waitq); } - cfs_spin_unlock_bh(&target->obd_processing_task_lock); + + /* Tell the client we're in recovery, when client is involved in it. */ + if (target->obd_recovering) + lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); + tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); conn = *tmp; @@ -1069,6 +1134,8 @@ out: cfs_spin_lock(&export->exp_lock); export->exp_connecting = 0; cfs_spin_unlock(&export->exp_lock); + + class_export_put(export); } if (targref) class_decref(targref, __FUNCTION__, cfs_current()); @@ -1099,10 +1166,10 @@ void target_destroy_export(struct obd_export *exp) if (exp->exp_imp_reverse != NULL) client_destroy_import(exp->exp_imp_reverse); - LASSERT(cfs_atomic_read(&exp->exp_locks_count) == 0); - LASSERT(cfs_atomic_read(&exp->exp_rpc_count) == 0); - LASSERT(cfs_atomic_read(&exp->exp_cb_count) == 0); - LASSERT(cfs_atomic_read(&exp->exp_replay_count) == 0); + LASSERT_ATOMIC_ZERO(&exp->exp_locks_count); + LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count); + LASSERT_ATOMIC_ZERO(&exp->exp_cb_count); + LASSERT_ATOMIC_ZERO(&exp->exp_replay_count); } /* @@ -1113,25 +1180,20 @@ static void target_request_copy_get(struct ptlrpc_request *req) class_export_rpc_get(req->rq_export); LASSERT(cfs_list_empty(&req->rq_list)); CFS_INIT_LIST_HEAD(&req->rq_replay_list); + /* increase refcount to keep request in queue */ - LASSERT(cfs_atomic_read(&req->rq_refcount)); cfs_atomic_inc(&req->rq_refcount); /** let export know it has replays to be handled */ cfs_atomic_inc(&req->rq_export->exp_replay_count); - /* release service thread while request is queued - * we are moving the request from active processing - * to waiting on the replay queue */ - ptlrpc_server_active_request_dec(req); } static void target_request_copy_put(struct ptlrpc_request *req) { LASSERT(cfs_list_empty(&req->rq_replay_list)); - LASSERT(cfs_atomic_read(&req->rq_export->exp_replay_count) > 0); + LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count); + cfs_atomic_dec(&req->rq_export->exp_replay_count); class_export_rpc_put(req->rq_export); - /* ptlrpc_server_drop_request() assumes the request is active */ - ptlrpc_server_active_request_inc(req); ptlrpc_server_drop_request(req); } @@ -1181,12 +1243,20 @@ static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) #ifdef __KERNEL__ static void target_finish_recovery(struct obd_device *obd) { + time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() - + obd->obd_recovery_start); ENTRY; - LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n", - obd->obd_name); + + LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients " + "%d recovered and %d %s evicted.\n", obd->obd_name, + (int)elapsed_time / 60, (int)elapsed_time % 60, + obd->obd_max_recoverable_clients, + cfs_atomic_read(&obd->obd_connected_clients), + obd->obd_stale_clients, + obd->obd_stale_clients == 1 ? "was" : "were"); ldlm_reprocess_all_ns(obd->obd_namespace); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_req_replay_queue) || !cfs_list_empty(&obd->obd_lock_replay_queue) || !cfs_list_empty(&obd->obd_final_req_queue)) { @@ -1197,10 +1267,10 @@ static void target_finish_recovery(struct obd_device *obd) "" : "lock ", cfs_list_empty(&obd->obd_final_req_queue) ? \ "" : "final "); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); LBUG(); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); obd->obd_recovery_end = cfs_time_current_sec(); @@ -1220,9 +1290,9 @@ static void abort_req_replay_queue(struct obd_device *obd) cfs_list_t abort_list; CFS_INIT_LIST_HEAD(&abort_list); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) { DEBUG_REQ(D_WARNING, req, "aborted:"); req->rq_status = -ENOTCONN; @@ -1241,9 +1311,9 @@ static void abort_lock_replay_queue(struct obd_device *obd) cfs_list_t abort_list; CFS_INIT_LIST_HEAD(&abort_list); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){ DEBUG_REQ(D_ERROR, req, "aborted:"); req->rq_status = -ENOTCONN; @@ -1254,7 +1324,6 @@ static void abort_lock_replay_queue(struct obd_device *obd) target_request_copy_put(req); } } -#endif /* Called from a cleanup function if the device is being cleaned up forcefully. The exports should all have been disconnected already, @@ -1272,17 +1341,19 @@ void target_cleanup_recovery(struct obd_device *obd) ENTRY; CFS_INIT_LIST_HEAD(&clean_list); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); EXIT; return; } obd->obd_recovering = obd->obd_abort_recovery = 0; - target_cancel_recovery_timer(obd); + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); + target_cancel_recovery_timer(obd); cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) { LASSERT(req->rq_reply_state == 0); @@ -1290,10 +1361,10 @@ void target_cleanup_recovery(struct obd_device *obd) target_request_copy_put(req); } - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list); cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){ LASSERT(req->rq_reply_state == 0); @@ -1303,63 +1374,91 @@ void target_cleanup_recovery(struct obd_device *obd) EXIT; } -/* obd_processing_task_lock should be held */ +/* obd_recovery_task_lock should be held */ void target_cancel_recovery_timer(struct obd_device *obd) { CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name); cfs_timer_disarm(&obd->obd_recovery_timer); } -/* extend = 1 means require at least "duration" seconds left in the timer, - extend = 0 means set the total duration (start_recovery_timer) */ -static void reset_recovery_timer(struct obd_device *obd, int duration, - int extend) +static void target_start_recovery_timer(struct obd_device *obd) { - cfs_time_t now = cfs_time_current_sec(); - cfs_duration_t left; + if (obd->obd_recovery_start != 0) + return; - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering || obd->obd_abort_recovery) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); return; } - left = cfs_time_sub(obd->obd_recovery_end, now); - - if (extend && (duration > left)) - obd->obd_recovery_timeout += duration - left; - else if (!extend && (duration > obd->obd_recovery_timeout)) - /* Track the client's largest expected replay time */ - obd->obd_recovery_timeout = duration; + LASSERT(obd->obd_recovery_timeout != 0); - /* Hard limit of obd_recovery_time_hard which should not happen */ - if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard) - obd->obd_recovery_timeout = obd->obd_recovery_time_hard; - - obd->obd_recovery_end = obd->obd_recovery_start + - obd->obd_recovery_timeout; - if (!cfs_timer_is_armed(&obd->obd_recovery_timer) || - cfs_time_before(now, obd->obd_recovery_end)) { - left = cfs_time_sub(obd->obd_recovery_end, now); - cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left)); + if (obd->obd_recovery_start != 0) { + cfs_spin_unlock(&obd->obd_dev_lock); + return; } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)left); + + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(obd->obd_recovery_timeout)); + obd->obd_recovery_start = cfs_time_current_sec(); + cfs_spin_unlock(&obd->obd_dev_lock); + + LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, " + "or until %d client%s reconnect%s\n", + obd->obd_name, + obd->obd_recovery_timeout / 60, + obd->obd_recovery_timeout % 60, + obd->obd_max_recoverable_clients, + (obd->obd_max_recoverable_clients == 1) ? "" : "s", + (obd->obd_max_recoverable_clients == 1) ? "s": ""); } -static void check_and_start_recovery_timer(struct obd_device *obd) +/** + * extend recovery window. + * + * if @extend is true, extend recovery window to have @drt remaining at least; + * otherwise, make sure the recovery timeout value is not less than @drt. + */ +static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) { - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - if (cfs_timer_is_armed(&obd->obd_recovery_timer)) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_time_t now; + cfs_time_t end; + cfs_duration_t left; + int to; + + cfs_spin_lock(&obd->obd_dev_lock); + if (!obd->obd_recovering || obd->obd_abort_recovery) { + cfs_spin_unlock(&obd->obd_dev_lock); return; } - CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name); - obd->obd_recovery_start = cfs_time_current_sec(); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - reset_recovery_timer(obd, obd->obd_recovery_timeout, 0); + LASSERT(obd->obd_recovery_start != 0); + + now = cfs_time_current_sec(); + to = obd->obd_recovery_timeout; + end = obd->obd_recovery_start + to; + left = cfs_time_sub(end, now); + + if (extend && (drt > left)) { + to += drt - left; + } else if (!extend && (drt > to)) { + to = drt; + /* reduce drt by already passed time */ + drt -= obd->obd_recovery_timeout - left; + } + + if (to > obd->obd_recovery_time_hard) + to = obd->obd_recovery_time_hard; + if (obd->obd_recovery_timeout < to) { + obd->obd_recovery_timeout = to; + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(drt)); + } + cfs_spin_unlock(&obd->obd_dev_lock); + + CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", + obd->obd_name, (unsigned)drt); } /* Reset the timer with each new client connection */ @@ -1374,11 +1473,13 @@ static void check_and_start_recovery_timer(struct obd_device *obd) */ static void -target_start_and_reset_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, - int new_client) +check_and_start_recovery_timer(struct obd_device *obd, + struct ptlrpc_request *req, + int new_client) { int service_time = lustre_msg_get_service_time(req->rq_reqmsg); + struct obd_device_target *obt = &obd->u.obt; + struct lustre_sb_info *lsi; if (!new_client && service_time) /* Teach server about old server's estimates, as first guess @@ -1386,7 +1487,7 @@ target_start_and_reset_recovery_timer(struct obd_device *obd, at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate, service_time); - check_and_start_recovery_timer(obd); + target_start_recovery_timer(obd); /* convert the service time to rpc timeout, * reuse service_time to limit stack usage */ @@ -1395,14 +1496,17 @@ target_start_and_reset_recovery_timer(struct obd_device *obd, /* We expect other clients to timeout within service_time, then try * to reconnect, then try the failover server. The max delay between * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */ - service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC + - INITIAL_CONNECT_TIMEOUT); + service_time += 2 * INITIAL_CONNECT_TIMEOUT; + + LASSERT(obt->obt_magic == OBT_MAGIC); + lsi = s2lsi(obt->obt_sb); + if (!(lsi->lsi_flags | LSI_IR_CAPABLE)) + service_time += 2 * (CONNECTION_SWITCH_MAX + + CONNECTION_SWITCH_INC); if (service_time > obd->obd_recovery_timeout && !new_client) - reset_recovery_timer(obd, service_time, 0); + extend_recovery_timer(obd, service_time, false); } -#ifdef __KERNEL__ - /** Health checking routines */ static inline int exp_connect_healthy(struct obd_export *exp) { @@ -1435,12 +1539,13 @@ static inline int exp_finished(struct obd_export *exp) /** Checking routines for recovery */ static int check_for_clients(struct obd_device *obd) { + unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients); + if (obd->obd_abort_recovery || obd->obd_recovery_expired) return 1; - LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients); + LASSERT(clnts <= obd->obd_max_recoverable_clients); if (obd->obd_no_conn == 0 && - obd->obd_connected_clients + obd->obd_stale_clients == - obd->obd_max_recoverable_clients) + clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients) return 1; return 0; } @@ -1451,8 +1556,8 @@ static int check_for_next_transno(struct obd_device *obd) int wake_up = 0, connected, completed, queue_len; __u64 next_transno, req_transno; ENTRY; - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_req_replay_queue)) { req = cfs_list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); @@ -1461,7 +1566,7 @@ static int check_for_next_transno(struct obd_device *obd) req_transno = 0; } - connected = obd->obd_connected_clients; + connected = cfs_atomic_read(&obd->obd_connected_clients); completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients); queue_len = obd->obd_requests_queued_for_recovery; next_transno = obd->obd_next_recovery_transno; @@ -1507,7 +1612,7 @@ static int check_for_next_transno(struct obd_device *obd) obd->obd_next_recovery_transno = req_transno; wake_up = 1; } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } @@ -1515,7 +1620,7 @@ static int check_for_next_lock(struct obd_device *obd) { int wake_up = 0; - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { CDEBUG(D_HA, "waking for next lock\n"); wake_up = 1; @@ -1529,7 +1634,7 @@ static int check_for_next_lock(struct obd_device *obd) CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } @@ -1543,38 +1648,33 @@ static int target_recovery_overseer(struct obd_device *obd, int (*check_routine)(struct obd_device *), int (*health_check)(struct obd_export *)) { - int abort = 0, expired = 1; - - do { - cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - abort = obd->obd_abort_recovery; - expired = obd->obd_recovery_expired; +repeat: + cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); + if (obd->obd_abort_recovery) { + CWARN("recovery is aborted, evict exports in recovery\n"); + /** evict exports which didn't finish recovery yet */ + class_disconnect_stale_exports(obd, exp_finished); + return 1; + } else if (obd->obd_recovery_expired) { obd->obd_recovery_expired = 0; - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - if (abort) { - CWARN("recovery is aborted, evict exports in recovery\n"); - /** evict exports which didn't finish recovery yet */ - class_disconnect_stale_exports(obd, exp_finished); - } else if (expired) { - /** If some clients died being recovered, evict them */ - CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n"); - /** evict cexports with no replay in queue, they are stalled */ - class_disconnect_stale_exports(obd, health_check); - /** continue with VBR */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - obd->obd_version_recov = 1; - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - /** - * reset timer, recovery will proceed with versions now, - * timeout is set just to handle reconnection delays - */ - reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1); - /** Wait for recovery events again, after evicting bad clients */ - } - } while (!abort && expired); - - return abort; + /** If some clients died being recovered, evict them */ + CDEBUG(D_WARNING, + "recovery is timed out, evict stale exports\n"); + /** evict cexports with no replay in queue, they are stalled */ + class_disconnect_stale_exports(obd, health_check); + /** continue with VBR */ + cfs_spin_lock(&obd->obd_dev_lock); + obd->obd_version_recov = 1; + cfs_spin_unlock(&obd->obd_dev_lock); + /** + * reset timer, recovery will proceed with versions now, + * timeout is set just to handle reconnection delays + */ + extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); + /** Wait for recovery events again, after evicting bad clients */ + goto repeat; + } + return 0; } static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) @@ -1591,15 +1691,15 @@ static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) abort_lock_replay_queue(obd); } - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_req_replay_queue)) { req = cfs_list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); cfs_list_del_init(&req->rq_list); obd->obd_requests_queued_for_recovery--; - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); } else { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); LASSERT(cfs_list_empty(&obd->obd_req_replay_queue)); LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0); /** evict exports failed VBR */ @@ -1617,14 +1717,14 @@ static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) exp_lock_replay_healthy)) abort_lock_replay_queue(obd); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { req = cfs_list_entry(obd->obd_lock_replay_queue.next, struct ptlrpc_request, rq_list); cfs_list_del_init(&req->rq_list); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); } else { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue)); LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0); /** evict exports failed VBR */ @@ -1637,18 +1737,20 @@ static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) { struct ptlrpc_request *req = NULL; - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (!cfs_list_empty(&obd->obd_final_req_queue)) { req = cfs_list_entry(obd->obd_final_req_queue.next, struct ptlrpc_request, rq_list); cfs_list_del_init(&req->rq_list); + cfs_spin_unlock(&obd->obd_recovery_task_lock); if (req->rq_export->exp_in_recovery) { cfs_spin_lock(&req->rq_export->exp_lock); req->rq_export->exp_in_recovery = 0; cfs_spin_unlock(&req->rq_export->exp_lock); } + } else { + cfs_spin_unlock(&obd->obd_recovery_task_lock); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); return req; } @@ -1659,17 +1761,18 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, int rc; ENTRY; - rc = lu_context_init(&req->rq_recov_session, LCT_SESSION); - if (rc) { - CERROR("Failure to initialize session: %d\n", rc); - GOTO(reqcopy_put, rc); - } /** * export can be evicted during recovery, no need to handle replays for * it after that, discard such request silently */ if (req->rq_export->exp_disconnected) + GOTO(reqcopy_put, rc = 0); + + rc = lu_context_init(&req->rq_recov_session, LCT_SESSION); + if (rc) { + CERROR("Failure to initialize session: %d\n", rc); GOTO(reqcopy_put, rc); + } req->rq_recov_session.lc_thread = thread; lu_context_enter(&req->rq_recov_session); @@ -1685,13 +1788,26 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, lu_context_fini(&req->rq_recov_session); /* don't reset timer for final stage */ if (!exp_finished(req->rq_export)) { + int to = obd_timeout; + /** * Add request timeout to the recovery time so next request from * this client may come in recovery time */ - reset_recovery_timer(class_exp2obd(req->rq_export), - AT_OFF ? obd_timeout : - lustre_msg_get_timeout(req->rq_reqmsg), 1); + if (!AT_OFF) { + struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + /* If the server sent early reply for this request, + * the client will recalculate the timeout according to + * current server estimate service time, so we will + * use the maxium timeout here for waiting the client + * sending the next req */ + to = max((int)at_est2timeout( + at_get(&svc->srv_at_estimate)), + (int)lustre_msg_get_timeout(req->rq_reqmsg)); + /* Add net_latency (see ptlrpc_replay_req) */ + to += lustre_msg_get_service_time(req->rq_reqmsg); + } + extend_recovery_timer(class_exp2obd(req->rq_export), to, true); } reqcopy_put: RETURN(rc); @@ -1705,8 +1821,8 @@ static int target_recovery_thread(void *arg) struct target_recovery_data *trd = &obd->obd_recovery_data; unsigned long delta; unsigned long flags; - struct lu_env env; - struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread; + struct lu_env *env; + struct ptlrpc_thread *thread = NULL; int rc = 0; ENTRY; @@ -1717,20 +1833,36 @@ static int target_recovery_thread(void *arg) RECALC_SIGPENDING; SIGNAL_MASK_UNLOCK(current, flags); - rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD); - if (rc) + OBD_ALLOC_PTR(thread); + if (thread == NULL) + RETURN(-ENOMEM); + + OBD_ALLOC_PTR(env); + if (env == NULL) { + OBD_FREE_PTR(thread); + RETURN(-ENOMEM); + } + + rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD); + if (rc) { + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); RETURN(rc); + } - thread->t_env = &env; + thread->t_env = env; thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */ - env.le_ctx.lc_thread = thread; + env->le_ctx.lc_thread = thread; thread->t_data = NULL; + thread->t_watchdog = NULL; - CERROR("%s: started recovery thread pid %d\n", obd->obd_name, + CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, cfs_curproc_pid()); trd->trd_processing_task = cfs_curproc_pid(); + cfs_spin_lock(&obd->obd_dev_lock); obd->obd_recovering = 1; + cfs_spin_unlock(&obd->obd_dev_lock); cfs_complete(&trd->trd_starting); /* first of all, we have to know the first transno to replay */ @@ -1742,7 +1874,6 @@ static int target_recovery_thread(void *arg) /* next stage: replay requests */ delta = jiffies; - obd->obd_req_replaying = 1; CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", cfs_atomic_read(&obd->obd_req_replay_clients), obd->obd_next_recovery_transno); @@ -1757,9 +1888,9 @@ static int target_recovery_thread(void *arg) * bz18031: increase next_recovery_transno before * target_request_copy_put() will drop exp_rpc reference */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); obd->obd_next_recovery_transno++; - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); obd->obd_replayed_requests++; @@ -1789,10 +1920,12 @@ static int target_recovery_thread(void *arg) lut_boot_epoch_update(lut); /* We drop recoverying flag to forward all new requests * to regular mds_handle() since now */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); obd->obd_recovering = obd->obd_abort_recovery = 0; + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); target_cancel_recovery_timer(obd); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); while ((req = target_next_final_ping(obd))) { LASSERT(trd->trd_processing_task == cfs_curproc_pid()); DEBUG_REQ(D_HA, req, "processing final ping from %s: ", @@ -1812,9 +1945,12 @@ static int target_recovery_thread(void *arg) target_finish_recovery(obd); - lu_context_fini(&env.le_ctx); + lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; cfs_complete(&trd->trd_finishing); + + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); RETURN(rc); } @@ -1830,7 +1966,7 @@ static int target_start_recovery_thread(struct lu_target *lut, cfs_init_completion(&trd->trd_finishing); trd->trd_recovery_handler = handler; - if (cfs_kernel_thread(target_recovery_thread, lut, 0) > 0) { + if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) { cfs_wait_for_completion(&trd->trd_starting); LASSERT(obd->obd_recovering != 0); } else @@ -1841,19 +1977,17 @@ static int target_start_recovery_thread(struct lu_target *lut, void target_stop_recovery_thread(struct obd_device *obd) { - cfs_spin_lock_bh(&obd->obd_processing_task_lock); if (obd->obd_recovery_data.trd_processing_task > 0) { struct target_recovery_data *trd = &obd->obd_recovery_data; /** recovery can be done but postrecovery is not yet */ + cfs_spin_lock(&obd->obd_dev_lock); if (obd->obd_recovering) { CERROR("%s: Aborting recovery\n", obd->obd_name); obd->obd_abort_recovery = 1; cfs_waitq_signal(&obd->obd_next_transno_waitq); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); cfs_wait_for_completion(&trd->trd_finishing); - } else { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); } } @@ -1872,12 +2006,10 @@ static void target_recovery_expired(unsigned long castmeharder) " after %lds (%d clients connected)\n", obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients), cfs_time_current_sec()- obd->obd_recovery_start, - obd->obd_connected_clients); + cfs_atomic_read(&obd->obd_connected_clients)); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); obd->obd_recovery_expired = 1; cfs_waitq_signal(&obd->obd_next_transno_waitq); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); } void target_recovery_init(struct lu_target *lut, svc_handler_t handler) @@ -1897,11 +2029,6 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) obd->obd_recovery_start = 0; obd->obd_recovery_end = 0; - /* both values can be get from mount data already */ - if (obd->obd_recovery_timeout == 0) - obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT; - if (obd->obd_recovery_time_hard == 0) - obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD; cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd); target_start_recovery_thread(lut, handler); } @@ -1916,32 +2043,31 @@ static int target_process_req_flags(struct obd_device *obd, LASSERT(exp != NULL); if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { /* client declares he's ready to replay locks */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&exp->exp_lock); if (exp->exp_req_replay_needed) { - LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) > - 0); - cfs_spin_lock(&exp->exp_lock); exp->exp_req_replay_needed = 0; cfs_spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients); cfs_atomic_dec(&obd->obd_req_replay_clients); + } else { + cfs_spin_unlock(&exp->exp_lock); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); } if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) { /* client declares he's ready to complete recovery * so, we put the request on th final queue */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&exp->exp_lock); if (exp->exp_lock_replay_needed) { - LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) > - 0); - cfs_spin_lock(&exp->exp_lock); exp->exp_lock_replay_needed = 0; cfs_spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients); cfs_atomic_dec(&obd->obd_lock_replay_clients); + } else { + cfs_spin_unlock(&exp->exp_lock); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); } - return 0; } @@ -1965,41 +2091,41 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * so, we put the request on th final queue */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue final req"); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (obd->obd_recovering) { cfs_list_add_tail(&req->rq_list, &obd->obd_final_req_queue); } else { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(obd->obd_stopping ? -ENOTCONN : 1); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); RETURN(0); } if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { /* client declares he's ready to replay locks */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue lock replay req"); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); + cfs_spin_lock(&obd->obd_recovery_task_lock); LASSERT(obd->obd_recovering); /* usually due to recovery abort */ if (!req->rq_export->exp_in_recovery) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(-ENOTCONN); } LASSERT(req->rq_export->exp_lock_replay_needed); cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); RETURN(0); } /* CAVEAT EMPTOR: The incoming request message has been swabbed * (i.e. buflens etc are in my own byte order), but type-dependent - * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */ + * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */ if (!transno) { CFS_INIT_LIST_HEAD(&req->rq_list); @@ -2007,8 +2133,6 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(1); } - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - /* If we're processing the queue, we want don't want to queue this * message. * @@ -2019,37 +2143,36 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * Also, a resent, replayed request that has already been * handled will pass through here and be processed immediately. */ - CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n", - obd->obd_next_recovery_transno, transno, obd->obd_req_replaying); - if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) { + CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n", + obd->obd_next_recovery_transno, transno); + cfs_spin_lock(&obd->obd_recovery_task_lock); + if (transno < obd->obd_next_recovery_transno) { /* Processing the queue right now, don't re-add. */ LASSERT(cfs_list_empty(&req->rq_list)); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); RETURN(1); } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP)) RETURN(0); target_request_copy_get(req); - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - LASSERT(obd->obd_recovering); if (!req->rq_export->exp_in_recovery) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); target_request_copy_put(req); RETURN(-ENOTCONN); } LASSERT(req->rq_export->exp_req_replay_needed); if (target_exp_enqueue_req_replay(req)) { - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); DEBUG_REQ(D_ERROR, req, "dropping resent queued req"); target_request_copy_put(req); RETURN(0); } /* XXX O(n^2) */ + cfs_spin_lock(&obd->obd_recovery_task_lock); + LASSERT(obd->obd_recovering); cfs_list_for_each(tmp, &obd->obd_req_replay_queue) { struct ptlrpc_request *reqiter = cfs_list_entry(tmp, struct ptlrpc_request, rq_list); @@ -2064,7 +2187,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, transno)) { DEBUG_REQ(D_ERROR, req, "dropping replay: transno " "has been claimed by another client"); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); RETURN(0); @@ -2075,8 +2198,8 @@ int target_queue_recovery_request(struct ptlrpc_request *req, cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); obd->obd_requests_queued_for_recovery++; + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); RETURN(0); } @@ -2134,7 +2257,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) { int netrc; struct ptlrpc_reply_state *rs; - struct obd_device *obd; struct obd_export *exp; struct ptlrpc_service *svc; ENTRY; @@ -2168,7 +2290,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) LASSERT (cfs_list_empty(&rs->rs_exp_list)); exp = class_export_get (req->rq_export); - obd = exp->exp_obd; /* disable reply scheduling while I'm setting up */ rs->rs_scheduled = 1; @@ -2194,7 +2315,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) netrc = target_send_reply_msg (req, rc, fail_id); - cfs_spin_lock(&svc->srv_lock); + cfs_spin_lock(&svc->srv_rs_lock); cfs_atomic_inc(&svc->srv_n_difficult_replies); @@ -2206,7 +2327,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) * reply_out_callback leaves alone) */ rs->rs_on_net = 0; ptlrpc_rs_addref(rs); - cfs_atomic_inc (&svc->srv_outstanding_replies); } cfs_spin_lock(&rs->rs_lock); @@ -2221,7 +2341,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) rs->rs_scheduled = 0; /* allow notifier to schedule */ } cfs_spin_unlock(&rs->rs_lock); - cfs_spin_unlock(&svc->srv_lock); + cfs_spin_unlock(&svc->srv_rs_lock); EXIT; } @@ -2406,7 +2526,7 @@ int ldlm_error2errno(ldlm_error_t error) if (((int)error) < 0) /* cast to signed type */ result = error; /* as ldlm_error_t can be unsigned */ else { - CERROR("Invalid DLM result code: %i\n", error); + CERROR("Invalid DLM result code: %d\n", error); result = -EPROTO; } } @@ -2465,3 +2585,107 @@ void ldlm_dump_export_locks(struct obd_export *exp) cfs_spin_unlock(&exp->exp_locks_list_guard); } #endif + +static int target_bulk_timeout(void *data) +{ + ENTRY; + /* We don't fail the connection here, because having the export + * killed makes the (vital) call to commitrw very sad. + */ + RETURN(1); +} + +static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) +{ + return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT"; +} + +int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, + struct l_wait_info *lwi) +{ + struct ptlrpc_request *req = desc->bd_req; + int rc = 0; + ENTRY; + + /* Check if there is eviction in progress, and if so, wait for + * it to finish */ + if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) { + *lwi = LWI_INTR(NULL, NULL); + rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, + !cfs_atomic_read(&exp->exp_obd-> + obd_evict_inprogress), + lwi); + } + + /* Check if client was evicted or tried to reconnect already */ + if (exp->exp_failed || exp->exp_abort_active_req) { + rc = -ENOTCONN; + } else { + if (desc->bd_type == BULK_PUT_SINK) + rc = sptlrpc_svc_wrap_bulk(req, desc); + if (rc == 0) + rc = ptlrpc_start_bulk_transfer(desc); + } + + if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) { + ptlrpc_abort_bulk(desc); + } else if (rc == 0) { + time_t start = cfs_time_current_sec(); + do { + long timeoutl = req->rq_deadline - cfs_time_current_sec(); + cfs_duration_t timeout = timeoutl <= 0 ? + CFS_TICK : cfs_time_seconds(timeoutl); + *lwi = LWI_TIMEOUT_INTERVAL(timeout, + cfs_time_seconds(1), + target_bulk_timeout, + desc); + rc = l_wait_event(desc->bd_waitq, + !ptlrpc_server_bulk_active(desc) || + exp->exp_failed || + exp->exp_abort_active_req, + lwi); + LASSERT(rc == 0 || rc == -ETIMEDOUT); + /* Wait again if we changed deadline */ + } while ((rc == -ETIMEDOUT) && + (req->rq_deadline > cfs_time_current_sec())); + + if (rc == -ETIMEDOUT) { + DEBUG_REQ(D_ERROR, req, + "timeout on bulk %s after %ld%+lds", + bulk2type(desc), + req->rq_deadline - start, + cfs_time_current_sec() - + req->rq_deadline); + ptlrpc_abort_bulk(desc); + } else if (exp->exp_failed) { + DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s", + bulk2type(desc)); + rc = -ENOTCONN; + ptlrpc_abort_bulk(desc); + } else if (exp->exp_abort_active_req) { + DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s", + bulk2type(desc)); + /* we don't reply anyway */ + rc = -ETIMEDOUT; + ptlrpc_abort_bulk(desc); + } else if (!desc->bd_success || + desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", + desc->bd_success ? + "truncated" : "network error on", + bulk2type(desc), + desc->bd_nob_transferred, + desc->bd_nob); + /* XXX should this be a different errno? */ + rc = -ETIMEDOUT; + } else if (desc->bd_type == BULK_GET_SINK) { + rc = sptlrpc_svc_unwrap_bulk(req, desc); + } + } else { + DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", + bulk2type(desc), rc); + } + + RETURN(rc); +} +EXPORT_SYMBOL(target_bulk_io);