X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=2c288c236321f601b3a41731de04122a192063c4;hp=1b1c2a6ce3cc62029bf63afc884f2d2cfc1219f7;hb=4f53536d002c13886210b672b657795baa067144;hpb=77c40f46f1631212a8ad9f4ff2b3bfc5ca2d8aa0 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 1b1c2a6..2c288c2 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -27,7 +27,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2010, 2013, Intel Corporation. + * Copyright (c) 2010, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -43,6 +43,7 @@ #define DEBUG_SUBSYSTEM S_LDLM +#include #include #include #include @@ -353,7 +354,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list); INIT_LIST_HEAD(&cli->cl_loi_write_list); INIT_LIST_HEAD(&cli->cl_loi_read_list); - client_obd_list_lock_init(&cli->cl_loi_list_lock); + spin_lock_init(&cli->cl_loi_list_lock); atomic_set(&cli->cl_pending_w_pages, 0); atomic_set(&cli->cl_pending_r_pages, 0); cli->cl_r_in_flight = 0; @@ -369,11 +370,11 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) /* lru for osc. */ INIT_LIST_HEAD(&cli->cl_lru_osc); atomic_set(&cli->cl_lru_shrinkers, 0); - atomic_set(&cli->cl_lru_busy, 0); - atomic_set(&cli->cl_lru_in_list, 0); + atomic_long_set(&cli->cl_lru_busy, 0); + atomic_long_set(&cli->cl_lru_in_list, 0); INIT_LIST_HEAD(&cli->cl_lru_list); - client_obd_list_lock_init(&cli->cl_lru_list_lock); - atomic_set(&cli->cl_unstable_count, 0); + spin_lock_init(&cli->cl_lru_list_lock); + atomic_long_set(&cli->cl_unstable_count, 0); init_waitqueue_head(&cli->cl_destroy_waitq); atomic_set(&cli->cl_destroy_in_flight, 0); @@ -438,10 +439,9 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_import, rc); } - cli->cl_import = imp; - /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */ - cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3); - cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie); + cli->cl_import = imp; + /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */ + cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3); if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) { if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) { @@ -464,8 +464,6 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_import, rc = -ENOMEM); } - cli->cl_qchk_stat = CL_NOT_QUOTACHECKED; - RETURN(rc); err_import: @@ -508,7 +506,7 @@ int client_connect_import(const struct lu_env *env, *exp = NULL; down_write(&cli->cl_sem); - if (cli->cl_conn_count > 0 ) + if (cli->cl_conn_count > 0) GOTO(out_sem, rc = -EALREADY); rc = class_connect(&conn, obd, cluuid); @@ -580,17 +578,17 @@ int client_disconnect_export(struct obd_export *exp) imp = cli->cl_import; down_write(&cli->cl_sem); - CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name, - cli->cl_conn_count); + CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name, + cli->cl_conn_count); - if (!cli->cl_conn_count) { + if (cli->cl_conn_count == 0) { CERROR("disconnecting disconnected device (%s)\n", obd->obd_name); GOTO(out_disconnect, rc = -EINVAL); } cli->cl_conn_count--; - if (cli->cl_conn_count) + if (cli->cl_conn_count != 0) GOTO(out_disconnect, rc = 0); /* Mark import deactivated now, so we don't try to reconnect if any @@ -752,7 +750,6 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, spin_unlock(&exp->exp_lock); class_export_cb_put(exp); } -EXPORT_SYMBOL(target_client_add_cb); static void check_and_start_recovery_timer(struct obd_device *obd, @@ -774,6 +771,8 @@ int target_handle_connect(struct ptlrpc_request *req) char *target_start; int target_len; bool mds_conn = false, lw_client = false; + bool mds_mds_conn = false; + bool new_mds_mds_conn = false; struct obd_connect_data *data, *tmpdata; int size, tmpsize; lnet_nid_t *client_nid = NULL; @@ -871,6 +870,20 @@ int target_handle_connect(struct ptlrpc_request *req) if (rc) GOTO(out, rc); +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0) + /* Don't allow clients to connect that are using old 1.8 format + * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18, + * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The + * FULL20 flag should be set on all connections since 2.0, but no + * longer affects behaviour. + * + * Later this check will be disabled and the flag can be retired + * completely once interop with 3.0 is no longer needed. + */ + if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20)) + GOTO(out, rc = -EPROTO); +#endif + if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { if (data->ocd_version < LUSTRE_VERSION_CODE - LUSTRE_VERSION_ALLOWED_OFFSET || @@ -896,10 +909,47 @@ int target_handle_connect(struct ptlrpc_request *req) } } + /* Note: lw_client is needed in MDS-MDS failover during update log + * processing, so we needs to allow lw_client to be connected at + * anytime, instead of only the initial connection */ + lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0; + if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) { mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0; - lw_client = (data->ocd_connect_flags & - OBD_CONNECT_LIGHTWEIGHT) != 0; + mds_mds_conn = (data->ocd_connect_flags & + OBD_CONNECT_MDS_MDS) != 0; + + /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS + * for Imperative Recovery connection from MGC to MGS. + * + * Via check OBD_CONNECT_FID, we can distinguish whether + * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from + * MGC or MDT. */ + if (!lw_client && + (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && + (data->ocd_connect_flags & OBD_CONNECT_FID) && + (data->ocd_connect_flags & OBD_CONNECT_VERSION)) { + __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version); + __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version); + __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version); + + /* We do not support the MDT-MDT interoperations with + * different version MDT because of protocol changes. */ + if (unlikely(major != LUSTRE_MAJOR || + minor != LUSTRE_MINOR || + abs(patch - LUSTRE_PATCH) > 3)) { + LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the " + "connection from different version MDT " + "(%d.%d.%d.%d) %s %s\n", + target->obd_name, LUSTRE_MAJOR, + LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX, + major, minor, patch, + OBD_OCD_VERSION_FIX(data->ocd_version), + libcfs_nid2str(req->rq_peer.nid), str); + + GOTO(out, rc = -EPROTO); + } + } } /* lctl gets a backstage, all-access pass. */ @@ -1019,7 +1069,9 @@ no_export: if (export == NULL) { /* allow lightweight connections during recovery */ - if (target->obd_recovering && !lw_client) { + /* allow "new" MDT to be connected during recovery, since we + * need retrieve recovery update records from it */ + if (target->obd_recovering && !lw_client && !mds_mds_conn) { cfs_time_t t; int c; /* connected */ int i; /* in progress */ @@ -1033,47 +1085,38 @@ no_export: t = cfs_timer_deadline(&target->obd_recovery_timer); t = cfs_time_sub(t, cfs_time_current()); t = cfs_duration_sec(t); - LCONSOLE_WARN("%s: Denying connection for new client " - "%s (at %s), waiting for all %d known " - "clients (%d recovered, %d in progress, " - "and %d evicted) to recover in %d:%.02d\n", + LCONSOLE_WARN("%s: Denying connection for new client %s" + "(at %s), waiting for %d known clients " + "(%d recovered, %d in progress, and %d " + "evicted) to recover in %d:%.02d\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), k, c - i, i, s, (int)t / 60, (int)t % 60); - rc = -EBUSY; - } else { + rc = -EBUSY; + } else { dont_check_exports: - rc = obd_connect(req->rq_svc_thread->t_env, - &export, target, &cluuid, data, - client_nid); + rc = obd_connect(req->rq_svc_thread->t_env, + &export, target, &cluuid, data, + client_nid); if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG)) lustre_msg_add_op_flags(req->rq_repmsg, - MSG_CONNECT_RECOVERING); - if (rc == 0) - conn.cookie = export->exp_handle.h_cookie; - } - } else { - rc = obd_reconnect(req->rq_svc_thread->t_env, - export, target, &cluuid, data, client_nid); - } - if (rc) - GOTO(out, rc); + MSG_CONNECT_RECOVERING); + if (rc == 0) + conn.cookie = export->exp_handle.h_cookie; -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 53, 0) - /* 2.2.0 clients always swab nidtbl entries due to a bug, so server - * will do the swabbing for if the client is using the same endianness. - * - * This fixup is version-limited, because we don't want to carry the - * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need - * interop with unpatched 2.2 clients. For newer clients, servers - * will never do MNE swabbing, let the client handle that. LU-1644 */ - export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) && - !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB); -#endif + if (mds_mds_conn) + new_mds_mds_conn = true; + } + } else { + rc = obd_reconnect(req->rq_svc_thread->t_env, + export, target, &cluuid, data, client_nid); + } + if (rc) + GOTO(out, rc); - LASSERT(target->u.obt.obt_magic == OBT_MAGIC); - data->ocd_instance = target->u.obt.obt_instance; + 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. */ @@ -1181,6 +1224,14 @@ dont_check_exports: atomic_inc(&target->obd_req_replay_clients); atomic_inc(&target->obd_lock_replay_clients); + /* Note: MDS-MDS connection is allowed to be connected during + * recovery, no matter if the exports needs to be recoveried. + * Because we need retrieve updates logs from all other MDTs. + * So if the MDS-MDS export is new, obd_max_recoverable_clients + * also needs to be increased to match other recovery checking + * condition. */ + if (new_mds_mds_conn) + target->obd_max_recoverable_clients++; if (atomic_inc_return(&target->obd_connected_clients) == target->obd_max_recoverable_clients) wake_up(&target->obd_next_transno_waitq); @@ -1217,17 +1268,12 @@ dont_check_exports: * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */ revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; - if ((data->ocd_connect_flags & OBD_CONNECT_AT) && - (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) + if (data->ocd_connect_flags & OBD_CONNECT_AT) revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; else revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) && - (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) - revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; - else - revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; + revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); if (rc) { @@ -1261,7 +1307,6 @@ out: req->rq_status = rc; RETURN(rc); } -EXPORT_SYMBOL(target_handle_connect); int target_handle_disconnect(struct ptlrpc_request *req) { @@ -1277,7 +1322,6 @@ int target_handle_disconnect(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(target_handle_disconnect); void target_destroy_export(struct obd_export *exp) { @@ -1368,8 +1412,9 @@ static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) spin_unlock(&req->rq_export->exp_lock); } -static void target_finish_recovery(struct obd_device *obd) +static void target_finish_recovery(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; ENTRY; /* Only log a recovery message when recovery has occurred. */ @@ -1402,6 +1447,10 @@ static void target_finish_recovery(struct obd_device *obd) } spin_unlock(&obd->obd_recovery_task_lock); + if (lut->lut_tdtd != NULL && + !list_empty(&lut->lut_tdtd->tdtd_replay_list)) + dtrq_list_dump(lut->lut_tdtd, D_ERROR); + obd->obd_recovery_end = cfs_time_current_sec(); /* When recovery finished, cleanup orphans on MDS and OST. */ @@ -1478,6 +1527,7 @@ void target_cleanup_recovery(struct obd_device *obd) return; } obd->obd_recovering = obd->obd_abort_recovery = 0; + obd->obd_force_abort_recovery = 0; spin_unlock(&obd->obd_dev_lock); spin_lock(&obd->obd_recovery_task_lock); @@ -1486,7 +1536,7 @@ void target_cleanup_recovery(struct obd_device *obd) spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); + LASSERT(req->rq_reply_state == NULL); target_exp_dequeue_req_replay(req); target_request_copy_put(req); } @@ -1497,7 +1547,7 @@ void target_cleanup_recovery(struct obd_device *obd) spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); + LASSERT(req->rq_reply_state == NULL); target_request_copy_put(req); } @@ -1511,7 +1561,6 @@ 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); } -EXPORT_SYMBOL(target_cancel_recovery_timer); static void target_start_recovery_timer(struct obd_device *obd) { @@ -1519,7 +1568,8 @@ static void target_start_recovery_timer(struct obd_device *obd) return; spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery) { + if (!obd->obd_recovering || obd->obd_abort_recovery || + obd->obd_force_abort_recovery) { spin_unlock(&obd->obd_dev_lock); return; } @@ -1560,7 +1610,8 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) int to; spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery) { + if (!obd->obd_recovering || obd->obd_abort_recovery || + obd->obd_force_abort_recovery) { spin_unlock(&obd->obd_dev_lock); return; } @@ -1575,22 +1626,20 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) 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 == 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)); + end = obd->obd_recovery_start + to; + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(end - now)); } spin_unlock(&obd->obd_dev_lock); CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)drt); + obd->obd_name, (unsigned)cfs_time_sub(end, now)); } /* Reset the timer with each new client connection */ @@ -1664,23 +1713,14 @@ static inline int exp_finished(struct obd_export *exp) return (exp->exp_in_recovery && !exp->exp_lock_replay_needed); } -/** Checking routines for recovery */ -static int check_for_clients(struct obd_device *obd) -{ - unsigned int clnts = atomic_read(&obd->obd_connected_clients); - - if (obd->obd_abort_recovery || obd->obd_recovery_expired) - return 1; - LASSERT(clnts <= obd->obd_max_recoverable_clients); - return (clnts + obd->obd_stale_clients == - obd->obd_max_recoverable_clients); -} - -static int check_for_next_transno(struct obd_device *obd) +static int check_for_next_transno(struct lu_target *lut) { struct ptlrpc_request *req = NULL; + struct obd_device *obd = lut->lut_obd; int wake_up = 0, connected, completed, queue_len; - __u64 next_transno, req_transno; + __u64 req_transno = 0; + __u64 update_transno = 0; + __u64 next_transno = 0; ENTRY; spin_lock(&obd->obd_recovery_task_lock); @@ -1688,8 +1728,14 @@ static int check_for_next_transno(struct obd_device *obd) req = list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); req_transno = lustre_msg_get_transno(req->rq_reqmsg); - } else { - req_transno = 0; + } + + if (lut->lut_tdtd != NULL) { + struct target_distribute_txn_data *tdtd; + __u64 update_transno; + + tdtd = lut->lut_tdtd; + update_transno = distribute_txn_get_next_transno(lut->lut_tdtd); } connected = atomic_read(&obd->obd_connected_clients); @@ -1702,13 +1748,14 @@ static int check_for_next_transno(struct obd_device *obd) obd->obd_max_recoverable_clients, connected, completed, queue_len, req_transno, next_transno); - if (obd->obd_abort_recovery) { + if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; } else if (obd->obd_recovery_expired) { CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; - } else if (req_transno == next_transno) { + } else if (req_transno == next_transno || (update_transno != 0 && + update_transno <= next_transno)) { CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno); wake_up = 1; } else if (queue_len > 0 && @@ -1724,10 +1771,10 @@ static int check_for_next_transno(struct obd_device *obd) CDEBUG(d_lvl, "%s: waking for gap in transno, VBR is %s (skip: " LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64 - ", last_committed: "LPD64")\n", + ", next_update "LPD64" last_committed: "LPD64")\n", obd->obd_name, obd->obd_version_recov ? "ON" : "OFF", next_transno, queue_len, completed, connected, - req_transno, obd->obd_last_committed); + req_transno, update_transno, obd->obd_last_committed); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } else if (atomic_read(&obd->obd_req_replay_clients) == 0) { @@ -1743,8 +1790,9 @@ static int check_for_next_transno(struct obd_device *obd) return wake_up; } -static int check_for_next_lock(struct obd_device *obd) +static int check_for_next_lock(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; int wake_up = 0; spin_lock(&obd->obd_recovery_task_lock); @@ -1754,7 +1802,7 @@ static int check_for_next_lock(struct obd_device *obd) } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) { CDEBUG(D_HA, "waking for completed lock replay\n"); wake_up = 1; - } else if (obd->obd_abort_recovery) { + } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; } else if (obd->obd_recovery_expired) { @@ -1771,13 +1819,24 @@ static int check_for_next_lock(struct obd_device *obd) * check its status with help of check_routine * evict dead clients via health_check */ -static int target_recovery_overseer(struct obd_device *obd, - int (*check_routine)(struct obd_device *), +static int target_recovery_overseer(struct lu_target *lut, + int (*check_routine)(struct lu_target *), int (*health_check)(struct obd_export *)) { + struct obd_device *obd = lut->lut_obd; repeat: - wait_event(obd->obd_next_transno_waitq, check_routine(obd)); - if (obd->obd_abort_recovery) { + if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >= + (obd->obd_recovery_start + obd->obd_recovery_time_hard))) { + CWARN("recovery is aborted by hard timeout\n"); + obd->obd_abort_recovery = 1; + } + + while (wait_event_timeout(obd->obd_next_transno_waitq, + check_routine(lut), + msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0) + /* wait indefinitely for event, but don't trigger watchdog */; + + if (obd->obd_abort_recovery || obd->obd_force_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); @@ -1804,43 +1863,13 @@ repeat: return 0; } -static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) -{ - struct ptlrpc_request *req = NULL; - ENTRY; - - CDEBUG(D_HA, "Waiting for transno "LPD64"\n", - obd->obd_next_recovery_transno); - - if (target_recovery_overseer(obd, check_for_next_transno, - exp_req_replay_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } - - spin_lock(&obd->obd_recovery_task_lock); - if (!list_empty(&obd->obd_req_replay_queue)) { - req = list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); - list_del_init(&req->rq_list); - obd->obd_requests_queued_for_recovery--; - spin_unlock(&obd->obd_recovery_task_lock); - } else { - spin_unlock(&obd->obd_recovery_task_lock); - LASSERT(list_empty(&obd->obd_req_replay_queue)); - LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); - /** evict exports failed VBR */ - class_disconnect_stale_exports(obd, exp_vbr_healthy); - } - RETURN(req); -} - -static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) +static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; struct ptlrpc_request *req = NULL; CDEBUG(D_HA, "Waiting for lock\n"); - if (target_recovery_overseer(obd, check_for_next_lock, + if (target_recovery_overseer(lut, check_for_next_lock, exp_lock_replay_healthy)) abort_lock_replay_queue(obd); @@ -1934,6 +1963,318 @@ static void handle_recovery_req(struct ptlrpc_thread *thread, EXIT; } +/** Checking routines for recovery */ +static int check_for_recovery_ready(struct lu_target *lut) +{ + struct obd_device *obd = lut->lut_obd; + unsigned int clnts = atomic_read(&obd->obd_connected_clients); + + CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d" + " abort %d expired %d\n", clnts, obd->obd_stale_clients, + obd->obd_max_recoverable_clients, obd->obd_abort_recovery, + obd->obd_recovery_expired); + + if (obd->obd_force_abort_recovery) + return 1; + + if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) { + LASSERT(clnts <= obd->obd_max_recoverable_clients); + if (clnts + obd->obd_stale_clients < + obd->obd_max_recoverable_clients) + return 0; + } + + if (lut->lut_tdtd != NULL) { + if (!lut->lut_tdtd->tdtd_replay_ready) { + /* Let's extend recovery timer, in case the recovery + * timer expired, and some clients got evicted */ + extend_recovery_timer(obd, obd->obd_recovery_timeout, + true); + return 0; + } else { + dtrq_list_dump(lut->lut_tdtd, D_HA); + } + } + + return 1; +} + +enum { + REQUEST_RECOVERY = 1, + UPDATE_RECOVERY = 2, +}; + +static __u64 get_next_replay_req_transno(struct obd_device *obd) +{ + __u64 transno = 0; + + if (!list_empty(&obd->obd_req_replay_queue)) { + struct ptlrpc_request *req; + + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + transno = lustre_msg_get_transno(req->rq_reqmsg); + } + + return transno; +} +__u64 get_next_transno(struct lu_target *lut, int *type) +{ + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + __u64 transno = 0; + __u64 update_transno; + ENTRY; + + transno = get_next_replay_req_transno(obd); + if (type != NULL) + *type = REQUEST_RECOVERY; + + if (tdtd == NULL) + RETURN(transno); + + update_transno = distribute_txn_get_next_transno(tdtd); + if (transno == 0 || (transno >= update_transno && + update_transno != 0)) { + transno = update_transno; + if (type != NULL) + *type = UPDATE_RECOVERY; + } + + RETURN(transno); +} + +/** + * drop duplicate replay request + * + * Because the operation has been replayed by update recovery, the request + * with the same transno will be dropped and also notify the client to send + * next replay request. + * + * \param[in] env execution environment + * \param[in] obd failover obd device + * \param[in] req request to be dropped + */ +static void drop_duplicate_replay_req(struct lu_env *env, + struct obd_device *obd, + struct ptlrpc_request *req) +{ + DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate" + " update records are found.\n", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + /* Right now, only for MDS reint operation update replay and + * normal request replay can have the same transno */ + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) { + req_capsule_set(&req->rq_pill, &RQF_MDS_REINT); + req->rq_status = req_capsule_server_pack(&req->rq_pill); + if (likely(req->rq_export)) + target_committed_to_req(req); + lustre_msg_set_transno(req->rq_repmsg, req->rq_transno); + target_send_reply(req, req->rq_status, 0); + } else { + DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n", + libcfs_nid2str(req->rq_peer.nid)); + } + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; +} + +/** + * Update last_rcvd of the update + * + * Because update recovery might update the last_rcvd by updates, i.e. + * it will not update the last_rcvd information in memory, so we need + * refresh these information in memory after update recovery. + * + * \param[in] obd obd_device under recoverying. + * \param[in] dtrq the update replay requests being replayed. + */ +static void target_update_lcd(struct lu_env *env, struct lu_target *lut, + struct distribute_txn_replay_req *dtrq) +{ + struct obd_device *obd = lut->lut_obd; + struct obd_export *export; + struct tg_export_data *ted; + struct distribute_txn_replay_req_sub *dtrqs; + struct seq_server_site *site; + struct update_records *ur; + const struct lu_fid *fid; + struct update_ops *ops; + struct update_params *params; + struct update_op *op; + __u32 mdt_index; + unsigned int i; + struct lsd_client_data *lcd = NULL; + + /* if Updates has been executed(committed) on the recovery target, + * i.e. the updates is not being executed on the target, so we do + * not need update it in memory */ + site = lu_site2seq(obd->obd_lu_dev->ld_site); + mdt_index = site->ss_node_id; + dtrqs = dtrq_sub_lookup(dtrq, mdt_index); + if (dtrqs != NULL) + return; + + if (dtrq->dtrq_lur == NULL) + return; + + /* Find the update last_rcvd record */ + fid = lu_object_fid(&lut->lut_last_rcvd->do_lu); + ur = &dtrq->dtrq_lur->lur_update_rec; + ops = &ur->ur_ops; + params = update_records_get_params(ur); + for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count; + i++, op = update_op_next_op(op)) { + __u64 pos; + __u16 size; + void *buf; + + if (!lu_fid_eq(&op->uop_fid, fid)) + continue; + + if (op->uop_type != OUT_WRITE) + continue; + + buf = update_params_get_param_buf(params, op->uop_params_off[1], + ur->ur_param_count, NULL); + if (buf == NULL) + continue; + + pos = le64_to_cpu(*(__u64 *)buf); + if (pos == 0) + continue; + + buf = update_params_get_param_buf(params, op->uop_params_off[0], + ur->ur_param_count, &size); + if (buf == NULL) + continue; + + if (size != sizeof(*lcd)) + continue; + lcd = buf; + } + + if (lcd == NULL || lcd->lcd_uuid[0] == '\0') + return; + + /* locate the export then update the exp_target_data if needed */ + export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid); + if (export == NULL) + return; + + ted = &export->exp_target_data; + if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) { + CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n", + lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid, + lcd->lcd_last_xid); + ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid; + ted->ted_lcd->lcd_last_result = lcd->lcd_last_result; + } + class_export_put(export); +} + +static void replay_request_or_update(struct lu_env *env, + struct lu_target *lut, + struct target_recovery_data *trd, + struct ptlrpc_thread *thread) +{ + struct obd_device *obd = lut->lut_obd; + struct ptlrpc_request *req = NULL; + int type; + __u64 transno; + ENTRY; + + CDEBUG(D_HA, "Waiting for transno "LPD64"\n", + obd->obd_next_recovery_transno); + + /* Replay all of request and update by transno */ + do { + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + + CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val); + + /** It is needed to extend recovery window above + * recovery_time_soft. Extending is possible only in the + * end of recovery window (see more details in + * handle_recovery_req()). + */ + CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300); + + if (target_recovery_overseer(lut, check_for_next_transno, + exp_req_replay_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + } + + spin_lock(&obd->obd_recovery_task_lock); + transno = get_next_transno(lut, &type); + if (type == REQUEST_RECOVERY && tdtd != NULL && + transno == tdtd->tdtd_last_update_transno) { + /* Drop replay request from client side, if the + * replay has been executed by update with the + * same transno */ + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + list_del_init(&req->rq_list); + obd->obd_requests_queued_for_recovery--; + spin_unlock(&obd->obd_recovery_task_lock); + drop_duplicate_replay_req(env, obd, req); + } else if (type == REQUEST_RECOVERY && transno != 0) { + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + list_del_init(&req->rq_list); + obd->obd_requests_queued_for_recovery--; + spin_unlock(&obd->obd_recovery_task_lock); + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + handle_recovery_req(thread, req, + trd->trd_recovery_handler); + /** + * bz18031: increase next_recovery_transno before + * target_request_copy_put() will drop exp_rpc reference + */ + spin_lock(&obd->obd_recovery_task_lock); + obd->obd_next_recovery_transno++; + spin_unlock(&obd->obd_recovery_task_lock); + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; + } else if (type == UPDATE_RECOVERY && transno != 0) { + struct distribute_txn_replay_req *dtrq; + + spin_unlock(&obd->obd_recovery_task_lock); + + LASSERT(tdtd != NULL); + dtrq = distribute_txn_get_next_req(tdtd); + lu_context_enter(&thread->t_env->le_ctx); + tdtd->tdtd_replay_handler(env, tdtd, dtrq); + lu_context_exit(&thread->t_env->le_ctx); + extend_recovery_timer(obd, obd_timeout, true); + LASSERT(tdtd->tdtd_last_update_transno <= transno); + tdtd->tdtd_last_update_transno = transno; + spin_lock(&obd->obd_recovery_task_lock); + if (transno > obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno = transno; + spin_unlock(&obd->obd_recovery_task_lock); + target_update_lcd(env, lut, dtrq); + dtrq_destory(dtrq); + } else { + spin_unlock(&obd->obd_recovery_task_lock); + LASSERT(list_empty(&obd->obd_req_replay_queue)); + LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); + /** evict exports failed VBR */ + class_disconnect_stale_exports(obd, exp_vbr_healthy); + break; + } + } while (1); +} + static int target_recovery_thread(void *arg) { struct lu_target *lut = arg; @@ -1979,43 +2320,28 @@ static int target_recovery_thread(void *arg) spin_unlock(&obd->obd_dev_lock); complete(&trd->trd_starting); - /* first of all, we have to know the first transno to replay */ - if (target_recovery_overseer(obd, check_for_clients, - exp_connect_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } + /* first of all, we have to know the first transno to replay */ + if (target_recovery_overseer(lut, check_for_recovery_ready, + exp_connect_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + if (lut->lut_tdtd != NULL) + dtrq_list_destroy(lut->lut_tdtd); + } - /* next stage: replay requests */ + /* next stage: replay requests or update */ delta = jiffies; CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", atomic_read(&obd->obd_req_replay_clients), obd->obd_next_recovery_transno); - while ((req = target_next_replay_req(obd))) { - LASSERT(trd->trd_processing_task == current_pid()); - DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", - lustre_msg_get_transno(req->rq_reqmsg), - libcfs_nid2str(req->rq_peer.nid)); - handle_recovery_req(thread, req, - trd->trd_recovery_handler); - /** - * bz18031: increase next_recovery_transno before - * target_request_copy_put() will drop exp_rpc reference - */ - spin_lock(&obd->obd_recovery_task_lock); - obd->obd_next_recovery_transno++; - spin_unlock(&obd->obd_recovery_task_lock); - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - obd->obd_replayed_requests++; - } + replay_request_or_update(env, lut, trd, thread); /** * The second stage: replay locks */ CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", atomic_read(&obd->obd_lock_replay_clients)); - while ((req = target_next_replay_lock(obd))) { + while ((req = target_next_replay_lock(lut))) { LASSERT(trd->trd_processing_task == current_pid()); DEBUG_REQ(D_HA, req, "processing lock from %s: ", libcfs_nid2str(req->rq_peer.nid)); @@ -2029,6 +2355,7 @@ static int target_recovery_thread(void *arg) * The third stage: reply on final pings, at this moment all clients * must have request in final queue */ + CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val); CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); /** Update server last boot epoch */ tgt_boot_epoch_update(lut); @@ -2051,9 +2378,9 @@ static int target_recovery_thread(void *arg) * export is being evicted */ ptlrpc_update_export_timer(req->rq_export, 0); target_request_copy_put(req); - } + } - delta = (jiffies - delta) / HZ; + delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC; CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n", delta, obd->obd_replayed_requests, obd->obd_replayed_locks); if (delta > OBD_RECOVERY_TIME_SOFT) { @@ -2061,7 +2388,7 @@ static int target_recovery_thread(void *arg) libcfs_debug_dumplog(); } - target_finish_recovery(obd); + target_finish_recovery(lut); lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; @@ -2137,6 +2464,7 @@ static void target_recovery_expired(unsigned long castmeharder) void target_recovery_init(struct lu_target *lut, svc_handler_t handler) { struct obd_device *obd = lut->lut_obd; + if (obd->obd_max_recoverable_clients == 0) { /** Update server last boot epoch */ tgt_boot_epoch_update(lut); @@ -2156,7 +2484,6 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) } EXPORT_SYMBOL(target_recovery_init); - static int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req) { @@ -2321,14 +2648,12 @@ added: wake_up(&obd->obd_next_transno_waitq); RETURN(0); } -EXPORT_SYMBOL(target_queue_recovery_request); int target_handle_ping(struct ptlrpc_request *req) { obd_ping(req->rq_svc_thread->t_env, req->rq_export); return req_capsule_server_pack(&req->rq_pill); } -EXPORT_SYMBOL(target_handle_ping); void target_committed_to_req(struct ptlrpc_request *req) { @@ -2345,7 +2670,6 @@ void target_committed_to_req(struct ptlrpc_request *req) CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", exp->exp_last_committed, req->rq_transno, req->rq_xid); } -EXPORT_SYMBOL(target_committed_to_req); #endif /* HAVE_SERVER_SUPPORT */ @@ -2376,9 +2700,9 @@ int target_pack_pool_reply(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(target_pack_pool_reply); -int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) +static int target_send_reply_msg(struct ptlrpc_request *req, + int rc, int fail_id) { if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) { DEBUG_REQ(D_ERROR, req, "dropping reply"); @@ -2487,7 +2811,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) spin_unlock(&svcpt->scp_rep_lock); EXIT; } -EXPORT_SYMBOL(target_send_reply); ldlm_mode_t lck_compat_array[] = { [LCK_EX] = LCK_COMPAT_EX, @@ -2510,6 +2833,7 @@ int ldlm_error2errno(ldlm_error_t error) switch (error) { case ELDLM_OK: + case ELDLM_LOCK_MATCHED: result = 0; break; case ELDLM_LOCK_CHANGED: @@ -2576,7 +2900,6 @@ ldlm_error_t ldlm_errno2error(int err_no) } return error; } -EXPORT_SYMBOL(ldlm_errno2error); #if LUSTRE_TRACKS_LOCK_EXP_REFS void ldlm_dump_export_locks(struct obd_export *exp) @@ -2605,9 +2928,13 @@ static int target_bulk_timeout(void *data) RETURN(1); } -static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) +static inline const char *bulk2type(struct ptlrpc_request *req) { - return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT"; + if (req->rq_bulk_read) + return "READ"; + if (req->rq_bulk_write) + return "WRITE"; + return "UNKNOWN"; } int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, @@ -2634,7 +2961,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { rc = -ENOTCONN; } else { - if (desc->bd_type == BULK_PUT_SINK) + if (req->rq_bulk_read) rc = sptlrpc_svc_wrap_bulk(req, desc); if (rc == 0) rc = ptlrpc_start_bulk_transfer(desc); @@ -2642,7 +2969,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, if (rc < 0) { DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", - bulk2type(desc), rc); + bulk2type(req), rc); RETURN(rc); } @@ -2660,6 +2987,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, long timeoutl = deadline - cfs_time_current_sec(); cfs_duration_t timeout = timeoutl <= 0 ? CFS_TICK : cfs_time_seconds(timeoutl); + time_t rq_deadline; *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), target_bulk_timeout, desc); @@ -2671,39 +2999,45 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); /* Wait again if we changed rq_deadline. */ + rq_deadline = ACCESS_ONCE(req->rq_deadline); deadline = start + bulk_timeout; - if (deadline > req->rq_deadline) - deadline = req->rq_deadline; + if (deadline > rq_deadline) + deadline = rq_deadline; } while ((rc == -ETIMEDOUT) && (deadline > cfs_time_current_sec())); if (rc == -ETIMEDOUT) { DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds", - bulk2type(desc), deadline - start, + bulk2type(req), deadline - start, cfs_time_current_sec() - deadline); ptlrpc_abort_bulk(desc); } else if (exp->exp_failed) { DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s", - bulk2type(desc)); + bulk2type(req)); rc = -ENOTCONN; ptlrpc_abort_bulk(desc); } else if (exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s", - bulk2type(desc)); + bulk2type(req)); /* We don't reply anyway. */ rc = -ETIMEDOUT; ptlrpc_abort_bulk(desc); - } else if (desc->bd_failure || - desc->bd_nob_transferred != desc->bd_nob) { - DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", - desc->bd_failure ? "network error on" : "truncated", - bulk2type(desc), desc->bd_nob_transferred, - desc->bd_nob); - /* XXX Should this be a different errno? */ + } else if (desc->bd_failure) { + DEBUG_REQ(D_ERROR, req, "network error on bulk %s", + bulk2type(req)); + /* 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 { + if (req->rq_bulk_write) + rc = sptlrpc_svc_unwrap_bulk(req, desc); + if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)", + bulk2type(req), desc->bd_nob_transferred, + desc->bd_nob); + /* XXX should this be a different errno? */ + rc = -ETIMEDOUT; + } } RETURN(rc);