X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=5c122fbbcf689ebbbc818a7792d7bff6151e2887;hp=69f35c1a577b515cf75143d7464a812a45937dfb;hb=c1d465de13ccf0eda8020c88661c3cc4d78538ca;hpb=f4b93dff9a8f4a59976ea864c4e3c2c42faa5770 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 69f35c1..5c122fb 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2010, 2014, Intel Corporation. + * Copyright (c) 2010, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -58,71 +54,75 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, int priority, int create) { - struct ptlrpc_connection *ptlrpc_conn; - struct obd_import_conn *imp_conn = NULL, *item; - int rc = 0; - ENTRY; + struct ptlrpc_connection *ptlrpc_conn; + struct obd_import_conn *imp_conn = NULL, *item; + lnet_nid_t nid4refnet = LNET_NID_ANY; + int rc = 0; + ENTRY; - if (!create && !priority) { - CDEBUG(D_HA, "Nothing to do\n"); - RETURN(-EINVAL); - } + if (!create && !priority) { + CDEBUG(D_HA, "Nothing to do\n"); + RETURN(-EINVAL); + } - ptlrpc_conn = ptlrpc_uuid_to_connection(uuid); - if (!ptlrpc_conn) { - CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid); - RETURN (-ENOENT); - } + if (imp->imp_connection && + imp->imp_connection->c_remote_uuid.uuid[0] == 0) + /* nid4refnet is used to restrict network connections */ + nid4refnet = imp->imp_connection->c_self; + ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet); + if (!ptlrpc_conn) { + CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid); + RETURN(-ENOENT); + } - if (create) { - OBD_ALLOC(imp_conn, sizeof(*imp_conn)); - if (!imp_conn) { - GOTO(out_put, rc = -ENOMEM); - } - } + if (create) { + OBD_ALLOC(imp_conn, sizeof(*imp_conn)); + if (!imp_conn) + GOTO(out_put, rc = -ENOMEM); + } spin_lock(&imp->imp_lock); list_for_each_entry(item, &imp->imp_conn_list, oic_item) { - if (obd_uuid_equals(uuid, &item->oic_uuid)) { - if (priority) { + if (obd_uuid_equals(uuid, &item->oic_uuid)) { + if (priority) { list_del(&item->oic_item); list_add(&item->oic_item, - &imp->imp_conn_list); - item->oic_last_attempt = 0; - } - CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n", - imp, imp->imp_obd->obd_name, uuid->uuid, - (priority ? ", moved to head" : "")); + &imp->imp_conn_list); + item->oic_last_attempt = 0; + } + CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n", + imp, imp->imp_obd->obd_name, uuid->uuid, + (priority ? ", moved to head" : "")); spin_unlock(&imp->imp_lock); - GOTO(out_free, rc = 0); - } - } + GOTO(out_free, rc = 0); + } + } /* No existing import connection found for \a uuid. */ - if (create) { - imp_conn->oic_conn = ptlrpc_conn; - imp_conn->oic_uuid = *uuid; - imp_conn->oic_last_attempt = 0; - if (priority) + if (create) { + imp_conn->oic_conn = ptlrpc_conn; + imp_conn->oic_uuid = *uuid; + imp_conn->oic_last_attempt = 0; + if (priority) list_add(&imp_conn->oic_item, &imp->imp_conn_list); - else + else list_add_tail(&imp_conn->oic_item, - &imp->imp_conn_list); - CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n", - imp, imp->imp_obd->obd_name, uuid->uuid, - (priority ? "head" : "tail")); - } else { + &imp->imp_conn_list); + CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n", + imp, imp->imp_obd->obd_name, uuid->uuid, + (priority ? "head" : "tail")); + } else { spin_unlock(&imp->imp_lock); GOTO(out_free, rc = -ENOENT); } spin_unlock(&imp->imp_lock); - RETURN(0); + RETURN(0); out_free: - if (imp_conn) - OBD_FREE(imp_conn, sizeof(*imp_conn)); + if (imp_conn) + OBD_FREE(imp_conn, sizeof(*imp_conn)); out_put: - ptlrpc_connection_put(ptlrpc_conn); - RETURN(rc); + ptlrpc_connection_put(ptlrpc_conn); + RETURN(rc); } int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid) @@ -167,14 +167,17 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) ptlrpc_connection_put(imp->imp_connection); imp->imp_connection = NULL; - dlmexp = class_conn2export(&imp->imp_dlm_handle); - if (dlmexp && dlmexp->exp_connection) { - LASSERT(dlmexp->exp_connection == - imp_conn->oic_conn); - ptlrpc_connection_put(dlmexp->exp_connection); - dlmexp->exp_connection = NULL; - } - } + dlmexp = class_conn2export(&imp->imp_dlm_handle); + if (dlmexp && dlmexp->exp_connection) { + LASSERT(dlmexp->exp_connection == + imp_conn->oic_conn); + ptlrpc_connection_put(dlmexp->exp_connection); + dlmexp->exp_connection = NULL; + } + + if (dlmexp != NULL) + class_export_put(dlmexp); + } list_del(&imp_conn->oic_item); ptlrpc_connection_put(imp_conn->oic_conn); @@ -257,20 +260,24 @@ static int osc_on_mdt(char *obdname) * 1 - client UUID * 2 - server UUID * 3 - inactive-on-startup + * 4 - restrictive net */ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) { - struct client_obd *cli = &obddev->u.cli; - struct obd_import *imp; - 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; + struct client_obd *cli = &obddev->u.cli; + struct obd_import *imp; + struct obd_uuid server_uuid; + int rq_portal, rp_portal, connect_op; + char *name = obddev->obd_type->typ_name; + enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN; + char *cli_name = lustre_cfg_buf(lcfg, 0); + struct ptlrpc_connection fake_conn = { .c_self = 0, + .c_remote_uuid.uuid[0] = 0 }; + int rc; + ENTRY; - /* In a more perfect world, we would hang a ptlrpc_client off of - * obd_type and just use the values from there. */ + /* In a more perfect world, we would hang a ptlrpc_client off of + * obd_type and just use the values from there. */ if (!strcmp(name, LUSTRE_OSC_NAME)) { rq_portal = OST_REQUEST_PORTAL; rp_portal = OSC_REPLY_PORTAL; @@ -283,7 +290,12 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) rq_portal = MDS_REQUEST_PORTAL; rp_portal = MDC_REPLY_PORTAL; connect_op = MDS_CONNECT; - cli->cl_sp_me = LUSTRE_SP_CLI; + if (is_lwp_on_ost(cli_name)) + cli->cl_sp_me = LUSTRE_SP_OST; + else if (is_lwp_on_mdt(cli_name)) + cli->cl_sp_me = LUSTRE_SP_MDT; + else + 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_OSP_NAME)) { @@ -301,7 +313,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) rq_portal = OST_REQUEST_PORTAL; } rp_portal = OSC_REPLY_PORTAL; - cli->cl_sp_me = LUSTRE_SP_CLI; + cli->cl_sp_me = LUSTRE_SP_MDT; } else if (!strcmp(name, LUSTRE_MGC_NAME)) { rq_portal = MGS_REQUEST_PORTAL; rp_portal = MGC_REPLY_PORTAL; @@ -338,10 +350,12 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) init_rwsem(&cli->cl_sem); mutex_init(&cli->cl_mgc_mutex); - cli->cl_conn_count = 0; - memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2), - min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2), - sizeof(server_uuid))); + cli->cl_seq = NULL; + init_rwsem(&cli->cl_seq_rwsem); + cli->cl_conn_count = 0; + memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2), + min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2), + sizeof(server_uuid))); cli->cl_dirty_pages = 0; cli->cl_avail_grant = 0; @@ -375,6 +389,10 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) INIT_LIST_HEAD(&cli->cl_lru_list); spin_lock_init(&cli->cl_lru_list_lock); atomic_long_set(&cli->cl_unstable_count, 0); + INIT_LIST_HEAD(&cli->cl_shrink_list); + + INIT_LIST_HEAD(&cli->cl_flight_waiters); + cli->cl_rpcs_in_flight = 0; init_waitqueue_head(&cli->cl_destroy_waitq); atomic_set(&cli->cl_destroy_in_flight, 0); @@ -390,31 +408,49 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) #endif atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); - /* This value may be reduced at connect time in - * ptlrpc_connect_interpret() . We initialize it to only - * 1MB until we know what the performance looks like. - * In the future this should likely be increased. LU-1431 */ - cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES, - LNET_MTU >> PAGE_CACHE_SHIFT); + /* Set it to possible maximum size. It may be reduced by ocd_brw_size + * from OFD after connecting. */ + cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES; + + cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES; - /* set cl_chunkbits default value to PAGE_CACHE_SHIFT, + /* set cl_chunkbits default value to PAGE_SHIFT, * it will be updated at OSC connection time. */ - cli->cl_chunkbits = PAGE_CACHE_SHIFT; + cli->cl_chunkbits = PAGE_SHIFT; if (!strcmp(name, LUSTRE_MDC_NAME)) { cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) { cli->cl_max_rpcs_in_flight = 2; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) { cli->cl_max_rpcs_in_flight = 3; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) { cli->cl_max_rpcs_in_flight = 4; } else { if (osc_on_mdt(obddev->obd_name)) cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX; else cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; - } + } + + spin_lock_init(&cli->cl_mod_rpcs_lock); + spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock); + cli->cl_max_mod_rpcs_in_flight = 0; + cli->cl_mod_rpcs_in_flight = 0; + cli->cl_close_rpcs_in_flight = 0; + init_waitqueue_head(&cli->cl_mod_rpcs_waitq); + cli->cl_mod_tag_bitmap = NULL; + + INIT_LIST_HEAD(&cli->cl_chg_dev_linkage); + + if (connect_op == MDS_CONNECT) { + cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1; + OBD_ALLOC(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + if (cli->cl_mod_tag_bitmap == NULL) + GOTO(err, rc = -ENOMEM); + } + rc = ldlm_get_ref(); if (rc) { CERROR("ldlm_get_ref failed: %d\n", rc); @@ -433,16 +469,30 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) LUSTRE_CFG_BUFLEN(lcfg, 1)); class_import_put(imp); - rc = client_import_add_conn(imp, &server_uuid, 1); - if (rc) { - CERROR("can't add initial connection\n"); - GOTO(err_import, rc); - } + if (lustre_cfg_buf(lcfg, 4)) { + __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4)); + + if (refnet == LNET_NIDNET(LNET_NID_ANY)) { + rc = -EINVAL; + CERROR("%s: bad mount option 'network=%s': rc = %d\n", + obddev->obd_name, lustre_cfg_string(lcfg, 4), + rc); + GOTO(err_import, rc); + } + fake_conn.c_self = LNET_MKNID(refnet, 0); + imp->imp_connection = &fake_conn; + } + + rc = client_import_add_conn(imp, &server_uuid, 1); + if (rc) { + CERROR("can't add initial connection\n"); + GOTO(err_import, rc); + } + imp->imp_connection = NULL; - 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")) { @@ -465,8 +515,6 @@ 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: @@ -474,6 +522,10 @@ err_import: err_ldlm: ldlm_put_ref(); err: + if (cli->cl_mod_tag_bitmap != NULL) + OBD_FREE(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + cli->cl_mod_tag_bitmap = NULL; RETURN(rc); } @@ -481,6 +533,7 @@ EXPORT_SYMBOL(client_obd_setup); int client_obd_cleanup(struct obd_device *obddev) { + struct client_obd *cli = &obddev->u.cli; ENTRY; ldlm_namespace_free_post(obddev->obd_namespace); @@ -490,6 +543,12 @@ int client_obd_cleanup(struct obd_device *obddev) LASSERT(obddev->u.cli.cl_import == NULL); ldlm_put_ref(); + + if (cli->cl_mod_tag_bitmap != NULL) + OBD_FREE(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + cli->cl_mod_tag_bitmap = NULL; + RETURN(0); } EXPORT_SYMBOL(client_obd_cleanup); @@ -507,55 +566,57 @@ int client_connect_import(const struct lu_env *env, int rc; ENTRY; - *exp = NULL; + *exp = NULL; down_write(&cli->cl_sem); if (cli->cl_conn_count > 0) GOTO(out_sem, rc = -EALREADY); - rc = class_connect(&conn, obd, cluuid); - if (rc) - GOTO(out_sem, rc); + rc = class_connect(&conn, obd, cluuid); + if (rc) + GOTO(out_sem, rc); - cli->cl_conn_count++; - *exp = class_conn2export(&conn); + cli->cl_conn_count++; + *exp = class_conn2export(&conn); - LASSERT(obd->obd_namespace); + LASSERT(obd->obd_namespace); - imp->imp_dlm_handle = conn; - rc = ptlrpc_init_import(imp); - if (rc != 0) - GOTO(out_ldlm, rc); + imp->imp_dlm_handle = conn; + rc = ptlrpc_init_import(imp); + if (rc != 0) + GOTO(out_ldlm, rc); - ocd = &imp->imp_connect_data; - if (data) { - *ocd = *data; - imp->imp_connect_flags_orig = data->ocd_connect_flags; - } + ocd = &imp->imp_connect_data; + if (data) { + *ocd = *data; + imp->imp_connect_flags_orig = data->ocd_connect_flags; + imp->imp_connect_flags2_orig = data->ocd_connect_flags2; + } - rc = ptlrpc_connect_import(imp); - if (rc != 0) { - LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); - GOTO(out_ldlm, rc); - } + rc = ptlrpc_connect_import(imp); + if (rc != 0) { + LASSERT(imp->imp_state == LUSTRE_IMP_DISCON); + GOTO(out_ldlm, rc); + } LASSERT(*exp != NULL && (*exp)->exp_connection); - if (data) { - LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) == - ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n", - data->ocd_connect_flags, ocd->ocd_connect_flags); - data->ocd_connect_flags = ocd->ocd_connect_flags; - } + if (data) { + LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) == + ocd->ocd_connect_flags, "old %#llx, new %#llx\n", + data->ocd_connect_flags, ocd->ocd_connect_flags); + data->ocd_connect_flags = ocd->ocd_connect_flags; + data->ocd_connect_flags2 = ocd->ocd_connect_flags2; + } - ptlrpc_pinger_add_import(imp); + ptlrpc_pinger_add_import(imp); - EXIT; + EXIT; - if (rc) { + if (rc) { out_ldlm: - cli->cl_conn_count--; - class_disconnect(*exp); - *exp = NULL; - } + cli->cl_conn_count--; + class_disconnect(*exp); + *exp = NULL; + } out_sem: up_write(&cli->cl_sem); @@ -572,7 +633,7 @@ int client_disconnect_export(struct obd_export *exp) ENTRY; if (!obd) { - CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n", + CERROR("invalid export for disconnect: exp %p cookie %#llx\n", exp, exp ? exp->exp_handle.h_cookie : -1); RETURN(-EINVAL); } @@ -648,8 +709,7 @@ int server_disconnect_export(struct obd_export *exp) if (exp->exp_imp_reverse) ptlrpc_cleanup_imp(exp->exp_imp_reverse); - if (exp->exp_obd->obd_namespace != NULL) - ldlm_cancel_locks_for_export(exp); + ldlm_bl_thread_wakeup(); /* complete all outstanding replies */ spin_lock(&exp->exp_lock); @@ -662,7 +722,10 @@ int server_disconnect_export(struct obd_export *exp) spin_lock(&svcpt->scp_rep_lock); list_del_init(&rs->rs_exp_list); + spin_lock(&rs->rs_lock); + /* clear rs_convert_lock to make sure rs is handled and put */ + rs->rs_convert_lock = 0; ptlrpc_schedule_difficult_reply(rs); spin_unlock(&rs->rs_lock); @@ -682,98 +745,225 @@ static int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp, struct obd_uuid *cluuid) { - ENTRY; - - if (exp->exp_connection && exp->exp_imp_reverse) { - struct lustre_handle *hdl; - struct obd_device *target; - - hdl = &exp->exp_imp_reverse->imp_remote_handle; - target = exp->exp_obd; - - /* Might be a re-connect after a partition. */ - if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) { - if (target->obd_recovering) { - int timeout = cfs_duration_sec(cfs_time_sub( - cfs_timer_deadline( - &target->obd_recovery_timer), - cfs_time_current())); - - LCONSOLE_WARN("%s: Client %s (at %s) reconnect" - "ing, waiting for %d clients in recov" - "ery for %d:%.02d\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - target->obd_max_recoverable_clients, - timeout / 60, timeout % 60); - } else { - LCONSOLE_WARN("%s: Client %s (at %s) " - "reconnecting\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp)); - } + struct obd_device *target; + struct lustre_handle *hdl; + time64_t deadline; + time64_t timeout; + time64_t now; + int rc = 0; - conn->cookie = exp->exp_handle.h_cookie; - /* target_handle_connect() treats EALREADY and - * -EALREADY differently. EALREADY means we are - * doing a valid reconnect from the same client. */ - RETURN(EALREADY); - } else { - LCONSOLE_WARN("%s: already connected client %s (at %s) " - "with handle "LPX64". Rejecting client " - "with the same UUID trying to reconnect " - "with handle "LPX64"\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - hdl->cookie, conn->cookie); - memset(conn, 0, sizeof *conn); - /* target_handle_connect() treats EALREADY and - * -EALREADY differently. -EALREADY is an error - * (same UUID, different handle). */ - RETURN(-EALREADY); - } - } + ENTRY; + hdl = &exp->exp_imp_reverse->imp_remote_handle; + if (!exp->exp_connection || !lustre_handle_is_used(hdl)) { + conn->cookie = exp->exp_handle.h_cookie; + CDEBUG(D_HA, "connect export for UUID '%s' at %p," + " cookie %#llx\n", cluuid->uuid, exp, conn->cookie); + RETURN(0); + } - conn->cookie = exp->exp_handle.h_cookie; - CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n", - cluuid->uuid, exp, conn->cookie); - RETURN(0); -} + target = exp->exp_obd; + + /* Might be a re-connect after a partition. */ + if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) { + LCONSOLE_WARN("%s: already connected client %s (at %s) " + "with handle %#llx. Rejecting client " + "with the same UUID trying to reconnect " + "with handle %#llx\n", target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), + hdl->cookie, conn->cookie); + memset(conn, 0, sizeof *conn); + /* target_handle_connect() treats EALREADY and + * -EALREADY differently. -EALREADY is an error + * (same UUID, different handle). */ + RETURN(-EALREADY); + } -void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, - int error) -{ - struct obd_export *exp = cb_data; + if (!target->obd_recovering) { + LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n", + target->obd_name, obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + GOTO(out_already, rc); + } - CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n", - obd->obd_name, exp->exp_client_uuid.uuid); + now = ktime_get_seconds(); + deadline = cfs_duration_sec(target->obd_recovery_timer.expires); + if (now < deadline) { + struct target_distribute_txn_data *tdtd; + int size = 0; + int count = 0; + char *buf = NULL; + + timeout = deadline - now; + tdtd = class_exp2tgt(exp)->lut_tdtd; + if (tdtd && tdtd->tdtd_show_update_logs_retrievers) + buf = tdtd->tdtd_show_update_logs_retrievers( + tdtd->tdtd_show_retrievers_cbdata, + &size, &count); + + if (count > 0) + LCONSOLE_WARN("%s: Recovery already passed deadline " + "%lld:%.02lld. It is due to DNE recovery " + "failed/stuck on the %d MDT(s):%s. " + "Please wait until all MDTs recovered " + "or abort the recovery by force.\n", + target->obd_name, timeout / 60, + timeout % 60, count, + buf ? buf : "unknown (not enough RAM)"); + else + LCONSOLE_WARN("%s: Recovery already passed deadline " + "%lld:%.02lld. If you do not want to wait " + "more, please abort the recovery by " + "force.\n", target->obd_name, + timeout / 60, timeout % 60); + + if (buf != NULL) + OBD_FREE(buf, size); + } else { + timeout = now - deadline; + LCONSOLE_WARN("%s: Recovery already passed deadline" + " %lld:%.02lld, It is most likely due to DNE" + " recovery is failed or stuck, please wait a" + " few more minutes or abort the recovery.\n", + target->obd_name, timeout / 60, timeout % 60); + } - spin_lock(&exp->exp_lock); - exp->exp_need_sync = 0; - spin_unlock(&exp->exp_lock); - class_export_cb_put(exp); +out_already: + conn->cookie = exp->exp_handle.h_cookie; + /* target_handle_connect() treats EALREADY and + * -EALREADY differently. EALREADY means we are + * doing a valid reconnect from the same client. */ + RETURN(EALREADY); } static void check_and_start_recovery_timer(struct obd_device *obd, struct ptlrpc_request *req, int new_client); +/** + * update flags for import during reconnect process + */ +static int rev_import_flags_update(struct obd_import *revimp, + struct ptlrpc_request *req) +{ + int rc; + struct obd_connect_data *data; + + data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); + + if (data->ocd_connect_flags & OBD_CONNECT_AT) + revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; + else + revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + + revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; + + rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); + if (rc) { + CERROR("%s: cannot get reverse import %s security: rc = %d\n", + revimp->imp_client->cli_name, + libcfs_id2str(req->rq_peer), rc); + return rc; + } + + return 0; +} + +/** + * Allocate a new reverse import for an export. + * + * \retval -errno in case error hit + * \retval 0 if reverse import correctly init + **/ +int rev_import_init(struct obd_export *export) +{ + struct obd_device *obd = export->exp_obd; + struct obd_import *revimp; + + LASSERT(export->exp_imp_reverse == NULL); + + revimp = class_new_import(obd); + if (revimp == NULL) + return -ENOMEM; + + revimp->imp_remote_handle.cookie = 0ULL; + revimp->imp_client = &obd->obd_ldlm_client; + revimp->imp_dlm_fake = 1; + + /* it is safe to connect import in new state as no sends possible */ + spin_lock(&export->exp_lock); + export->exp_imp_reverse = revimp; + spin_unlock(&export->exp_lock); + class_import_put(revimp); + + return 0; +} +EXPORT_SYMBOL(rev_import_init); + +/** + * Handle reconnect for an export. + * + * \param exp export to handle reconnect process + * \param req client reconnect request + * + * \retval -rc in case securitfy flavor can't be changed + * \retval 0 in case none problems + */ +static int rev_import_reconnect(struct obd_export *exp, + struct ptlrpc_request *req) +{ + struct obd_import *revimp = exp->exp_imp_reverse; + struct lustre_handle *lh; + int rc; + + /* avoid sending a request until import flags are changed */ + ptlrpc_import_enter_resend(revimp); + + if (revimp->imp_connection != NULL) + ptlrpc_connection_put(revimp->imp_connection); + + /* + * client from recovery don't have a handle so we need to take from + * request. it may produce situation when wrong client connected + * to recovery as we trust a client uuid + */ + lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN); + revimp->imp_remote_handle = *lh; + + /* unknown versions will be caught in + * ptlrpc_handle_server_req_in->lustre_unpack_msg() */ + revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; + + revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection); + + rc = rev_import_flags_update(revimp, req); + if (rc != 0) { + /* it is safe to still be in RECOVERY phase as we are not able + * to setup correct security flavor so requests are not able to + * be delivered correctly */ + return rc; + } + + /* resend all rpc's via new connection */ + return ptlrpc_import_recovery_state_machine(revimp); +} + int target_handle_connect(struct ptlrpc_request *req) { - struct obd_device *target = NULL, *targref = NULL; - struct obd_export *export = NULL; - struct obd_import *revimp; - struct obd_import *tmp_imp = NULL; - struct lustre_handle conn; - struct lustre_handle *tmp; - struct obd_uuid tgtuuid; + struct obd_device *target = NULL; + struct obd_export *export = NULL; + /* connect handle - filled from target_handle_reconnect in + * reconnect case */ + struct lustre_handle conn; + struct lustre_handle *tmp; struct obd_uuid cluuid; - struct obd_uuid remote_uuid; char *str; int rc = 0; char *target_start; int target_len; - bool mds_conn = false, lw_client = false; + bool mds_conn = false, lw_client = false, initial_conn = 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; @@ -787,11 +977,7 @@ int target_handle_connect(struct ptlrpc_request *req) GOTO(out, rc = -EINVAL); } - obd_str2uuid(&tgtuuid, str); - target = class_uuid2obd(&tgtuuid); - if (!target) - target = class_name2obd(str); - + target = class_dev_by_str(str); if (!target) { deuuidify(str, NULL, &target_start, &target_len); LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect " @@ -803,6 +989,9 @@ int target_handle_connect(struct ptlrpc_request *req) } spin_lock(&target->obd_dev_lock); + + target->obd_conn_inprogress++; + if (target->obd_stopping || !target->obd_set_up) { spin_unlock(&target->obd_dev_lock); @@ -824,12 +1013,6 @@ int target_handle_connect(struct ptlrpc_request *req) GOTO(out, rc = -EAGAIN); } - /* Make sure the target isn't cleaned up while we're here. Yes, - * there's still a race between the above check and our incref here. - * Really, class_uuid2obd should take the ref. */ - targref = class_incref(target, __FUNCTION__, current); - - target->obd_conn_inprogress++; spin_unlock(&target->obd_dev_lock); str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID); @@ -840,21 +1023,6 @@ int target_handle_connect(struct ptlrpc_request *req) obd_str2uuid(&cluuid, str); - /* XXX Extract a nettype and format accordingly. */ - switch (sizeof(lnet_nid_t)) { - /* NB the casts only avoid compiler warnings. */ - case 8: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid); - break; - case 4: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_%x_UUID", (__u32)req->rq_peer.nid); - break; - default: - LBUG(); - } - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); if (tmp == NULL) GOTO(out, rc = -EPROTO); @@ -883,44 +1051,36 @@ int target_handle_connect(struct ptlrpc_request *req) */ if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20)) GOTO(out, rc = -EPROTO); -#endif + /* Don't allow liblustre clients to connect. + * - testing was disabled in v2_2_50_0-61-g6a75d65 + * - building was disabled in v2_5_58_0-28-g7277179 + * - client code was deleted in v2_6_50_0-101-gcdfbc72, + * - clients were refused connect for version difference > 0.0.1.32 */ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { - if (data->ocd_version < LUSTRE_VERSION_CODE - - LUSTRE_VERSION_ALLOWED_OFFSET || - data->ocd_version > LUSTRE_VERSION_CODE + - LUSTRE_VERSION_ALLOWED_OFFSET) { - DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) " - "libclient connection attempt", - data->ocd_version < LUSTRE_VERSION_CODE ? - "old" : "new", - OBD_OCD_VERSION_MAJOR(data->ocd_version), - OBD_OCD_VERSION_MINOR(data->ocd_version), - OBD_OCD_VERSION_PATCH(data->ocd_version), - OBD_OCD_VERSION_FIX(data->ocd_version)); - data = req_capsule_server_sized_get(&req->rq_pill, - &RMF_CONNECT_DATA, - offsetof(typeof(*data), ocd_version) + - sizeof(data->ocd_version)); - if (data) { - data->ocd_connect_flags = OBD_CONNECT_VERSION; - data->ocd_version = LUSTRE_VERSION_CODE; - } - GOTO(out, rc = -EPROTO); - } + DEBUG_REQ(D_WARNING, req, "Refusing libclient connection"); + GOTO(out, rc = -EPROTO); } +#endif + + /* 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) { + initial_conn = true; 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. */ + * MGC or MDT, since MGC does not use OBD_CONNECT_FID. + */ if (!lw_client && (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && (data->ocd_connect_flags & OBD_CONNECT_FID) && @@ -968,29 +1128,45 @@ int target_handle_connect(struct ptlrpc_request *req) class_export_put(export); export = NULL; rc = -EALREADY; - } else if ((mds_conn || lw_client) && export->exp_connection != NULL) { + } else if ((mds_conn || (lw_client && initial_conn) || + data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && + export->exp_connection != NULL) { spin_unlock(&export->exp_lock); - if (req->rq_peer.nid != export->exp_connection->c_peer.nid) + if (req->rq_peer.nid != export->exp_connection->c_peer.nid) { /* MDS or LWP reconnected after failover. */ LCONSOLE_WARN("%s: Received %s connection from " "%s, removing former export from %s\n", target->obd_name, mds_conn ? "MDS" : "LWP", libcfs_nid2str(req->rq_peer.nid), libcfs_nid2str(export->exp_connection->c_peer.nid)); - else + } else { /* New MDS connection from the same NID. */ LCONSOLE_WARN("%s: Received new %s connection from " "%s, removing former export from same NID\n", target->obd_name, mds_conn ? "MDS" : "LWP", libcfs_nid2str(req->rq_peer.nid)); - class_fail_export(export); - class_export_put(export); - export = NULL; - rc = 0; - } else if (export->exp_connection != NULL && - req->rq_peer.nid != export->exp_connection->c_peer.nid && - (lustre_msg_get_op_flags(req->rq_reqmsg) & - MSG_CONNECT_INITIAL)) { + } + + if (req->rq_peer.nid == export->exp_connection->c_peer.nid && + data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) { + /* Because exports between MDTs will always be + * kept, let's do not fail such export if they + * come from the same NID, otherwise it might + * cause eviction between MDTs, which might + * cause namespace inconsistency */ + spin_lock(&export->exp_lock); + export->exp_connecting = 1; + spin_unlock(&export->exp_lock); + conn.cookie = export->exp_handle.h_cookie; + rc = EALREADY; + } else { + class_fail_export(export); + class_export_put(export); + export = NULL; + rc = 0; + } + } else if (export->exp_connection != NULL && initial_conn && + req->rq_peer.nid != export->exp_connection->c_peer.nid) { spin_unlock(&export->exp_lock); /* In MDS failover we have static UUID but NID can change. */ LCONSOLE_WARN("%s: Client %s seen on new nid %s when " @@ -1038,11 +1214,11 @@ no_export: GOTO(out, rc); } - 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); + CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n", + target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), + target->obd_recovering ? "recovering/" : "", data->ocd_transno, + export, ktime_get_real_seconds(), + export ? export->exp_last_request_time : 0); /* If this is the first time a client connects, reset the recovery * timer. Discard lightweight connections which might be local. */ @@ -1065,44 +1241,50 @@ no_export: if (export == NULL) { /* allow lightweight connections during recovery */ - if (target->obd_recovering && !lw_client) { - cfs_time_t t; - int c; /* connected */ - int i; /* in progress */ - int k; /* known */ - int s; /* stale/evicted */ + /* 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) { + time64_t t; + int c; /* connected */ + int i; /* in progress */ + int k; /* known */ + int s; /* stale/evicted */ c = atomic_read(&target->obd_connected_clients); i = atomic_read(&target->obd_lock_replay_clients); k = target->obd_max_recoverable_clients; s = target->obd_stale_clients; - 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", + t = target->obd_recovery_timer.expires; + t = cfs_duration_sec(target->obd_recovery_timer.expires); + t -= ktime_get_seconds(); + 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 %lld:%.02lld\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 { + c - i, i, s, t / 60, t % 60); + 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); - } + MSG_CONNECT_RECOVERING); + if (rc == 0) { + conn.cookie = export->exp_handle.h_cookie; + rc = rev_import_init(export); + } + + 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); @@ -1122,11 +1304,6 @@ dont_check_exports: memcpy(tmpdata, data, min(tmpsize, size)); } - /* If all else goes well, this is our RPC return code. */ - req->rq_status = 0; - - lustre_msg_set_handle(req->rq_repmsg, &conn); - /* If the client and the server are the same node, we will already * have an export that really points to the client's DLM export, * because we have a shared handles table. @@ -1141,46 +1318,40 @@ dont_check_exports: spin_unlock(&export->exp_lock); CDEBUG(D_RPCTRACE, "%s: %s already connected at greater " "or equal conn_cnt: %d >= %d\n", - cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), - export->exp_conn_cnt, - lustre_msg_get_conn_cnt(req->rq_reqmsg)); - - GOTO(out, rc = -EALREADY); - } - LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0); - export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg); - - /* Don't evict liblustre clients for not pinging. */ - if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { - export->exp_libclient = 1; - spin_unlock(&export->exp_lock); + cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), + export->exp_conn_cnt, + lustre_msg_get_conn_cnt(req->rq_reqmsg)); - spin_lock(&target->obd_dev_lock); - list_del_init(&export->exp_obd_chain_timed); - spin_unlock(&target->obd_dev_lock); - } else { - spin_unlock(&export->exp_lock); + GOTO(out, rc = -EALREADY); } + LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0); + export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg); + spin_unlock(&export->exp_lock); - if (export->exp_connection != NULL) { + if (export->exp_connection != NULL) { /* Check to see if connection came from another NID. */ - if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) && + if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) && !hlist_unhashed(&export->exp_nid_hash)) - cfs_hash_del(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); + cfs_hash_del(export->exp_obd->obd_nid_hash, + &export->exp_connection->c_peer.nid, + &export->exp_nid_hash); - ptlrpc_connection_put(export->exp_connection); - } + ptlrpc_connection_put(export->exp_connection); + } - export->exp_connection = ptlrpc_connection_get(req->rq_peer, - req->rq_self, - &remote_uuid); - if (hlist_unhashed(&export->exp_nid_hash)) { - cfs_hash_add(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); - } + export->exp_connection = ptlrpc_connection_get(req->rq_peer, + req->rq_self, + &cluuid); + if (hlist_unhashed(&export->exp_nid_hash)) + cfs_hash_add(export->exp_obd->obd_nid_hash, + &export->exp_connection->c_peer.nid, + &export->exp_nid_hash); + + lustre_msg_set_handle(req->rq_repmsg, &conn); + + rc = rev_import_reconnect(export, req); + if (rc != 0) + GOTO(out, rc); if (target->obd_recovering && !export->exp_in_recovery && !lw_client) { int has_transno; @@ -1215,6 +1386,14 @@ 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); @@ -1224,54 +1403,7 @@ dont_check_exports: if (target->obd_recovering && !lw_client) lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); - conn = *tmp; - - /* Return -ENOTCONN in case of errors to let client reconnect. */ - revimp = class_new_import(target); - if (revimp == NULL) { - CERROR("fail to alloc new reverse import.\n"); - GOTO(out, rc = -ENOTCONN); - } - - spin_lock(&export->exp_lock); - if (export->exp_imp_reverse != NULL) - /* destroyed import can be still referenced in ctxt */ - tmp_imp = export->exp_imp_reverse; - export->exp_imp_reverse = revimp; - spin_unlock(&export->exp_lock); - - revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection); - revimp->imp_client = &export->exp_obd->obd_ldlm_client; - revimp->imp_remote_handle = conn; - revimp->imp_dlm_fake = 1; - revimp->imp_state = LUSTRE_IMP_FULL; - - /* Unknown versions will be caught in - * 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_msghdr_flags |= MSGHDR_AT_SUPPORT; - else - revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - - revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; - - rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); - if (rc) { - CERROR("Failed to get sec for reverse import: %d\n", rc); - spin_lock(&export->exp_lock); - export->exp_imp_reverse = NULL; - spin_unlock(&export->exp_lock); - class_destroy_import(revimp); - } - - class_import_put(revimp); - out: - if (tmp_imp != NULL) - client_destroy_import(tmp_imp); if (export) { spin_lock(&export->exp_lock); export->exp_connecting = 0; @@ -1279,15 +1411,13 @@ out: class_export_put(export); } - if (targref) { + if (target != NULL) { spin_lock(&target->obd_dev_lock); target->obd_conn_inprogress--; spin_unlock(&target->obd_dev_lock); - - class_decref(targref, __func__, current); + class_decref(target, "find", current); } - if (rc) - req->rq_status = rc; + req->rq_status = rc; RETURN(rc); } @@ -1357,6 +1487,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); struct obd_export *exp = req->rq_export; struct ptlrpc_request *reqiter; + struct ptlrpc_request *dup_req = NULL; int dup = 0; LASSERT(exp); @@ -1365,6 +1496,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) list_for_each_entry(reqiter, &exp->exp_req_replay_queue, rq_replay_list) { if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { + dup_req = reqiter; dup = 1; break; } @@ -1376,6 +1508,16 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY)) CERROR("invalid flags %x of resent replay\n", lustre_msg_get_flags(req->rq_reqmsg)); + + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + __u32 new_conn; + + new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg); + if (new_conn > + lustre_msg_get_conn_cnt(dup_req->rq_reqmsg)) + lustre_msg_set_conn_cnt(dup_req->rq_reqmsg, + new_conn); + } } else { list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue); @@ -1395,24 +1537,27 @@ 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. */ if (obd->obd_recovery_start) { - time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() - - obd->obd_recovery_start); - LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients " + time64_t now = ktime_get_real_seconds(); + time64_t elapsed_time; + + elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1); + LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients " "%d recovered and %d %s evicted.\n", obd->obd_name, - (int)elapsed_time / 60, (int)elapsed_time % 60, + (s64)elapsed_time / 60, (s64)elapsed_time % 60, obd->obd_max_recoverable_clients, atomic_read(&obd->obd_connected_clients), obd->obd_stale_clients, obd->obd_stale_clients == 1 ? "was" : "were"); } - ldlm_reprocess_all_ns(obd->obd_namespace); + ldlm_reprocess_recovery_done(obd->obd_namespace); spin_lock(&obd->obd_recovery_task_lock); if (!list_empty(&obd->obd_req_replay_queue) || !list_empty(&obd->obd_lock_replay_queue) || @@ -1429,15 +1574,16 @@ static void target_finish_recovery(struct obd_device *obd) } spin_unlock(&obd->obd_recovery_task_lock); - obd->obd_recovery_end = cfs_time_current_sec(); + obd->obd_recovery_end = ktime_get_real_seconds(); /* When recovery finished, cleanup orphans on MDS and OST. */ - if (OBT(obd) && OBP(obd, postrecov)) { - int rc = OBP(obd, postrecov)(obd); - if (rc < 0) - LCONSOLE_WARN("%s: Post recovery failed, rc %d\n", - obd->obd_name, rc); - } + if (obd->obd_type && OBP(obd, postrecov)) { + int rc = OBP(obd, postrecov)(obd); + + if (rc < 0) + LCONSOLE_WARN("%s: Post recovery failed, rc %d\n", + obd->obd_name, rc); + } EXIT; } @@ -1495,7 +1641,6 @@ void target_cleanup_recovery(struct obd_device *obd) { struct ptlrpc_request *req, *n; struct list_head clean_list; - ENTRY; INIT_LIST_HEAD(&clean_list); spin_lock(&obd->obd_dev_lock); @@ -1536,7 +1681,7 @@ EXPORT_SYMBOL(target_cleanup_recovery); 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); + del_timer(&obd->obd_recovery_timer); } static void target_start_recovery_timer(struct obd_device *obd) @@ -1557,19 +1702,18 @@ static void target_start_recovery_timer(struct obd_device *obd) return; } - cfs_timer_arm(&obd->obd_recovery_timer, - cfs_time_shift(obd->obd_recovery_timeout)); - obd->obd_recovery_start = cfs_time_current_sec(); + mod_timer(&obd->obd_recovery_timer, + jiffies + cfs_time_seconds(obd->obd_recovery_timeout)); + obd->obd_recovery_start = ktime_get_real_seconds(); 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": ""); + LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, 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": ""); } /** @@ -1578,24 +1722,25 @@ static void target_start_recovery_timer(struct obd_device *obd) * 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) +static void extend_recovery_timer(struct obd_device *obd, time64_t drt, + bool extend) { - cfs_time_t now; - cfs_time_t end; - cfs_duration_t left; - int to; + time64_t now; + time64_t end; + time64_t left; + time64_t to; spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering || obd->obd_abort_recovery) { spin_unlock(&obd->obd_dev_lock); - return; - } - LASSERT(obd->obd_recovery_start != 0); + return; + } + 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); + now = ktime_get_real_seconds(); + to = obd->obd_recovery_timeout; + end = obd->obd_recovery_start + to; + left = end - now; if (extend && (drt > left)) { to += drt - left; @@ -1603,18 +1748,22 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) to = drt; } - if (to > obd->obd_recovery_time_hard) - to = obd->obd_recovery_time_hard; + if (to > obd->obd_recovery_time_hard) { + to = obd->obd_recovery_time_hard; + CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n", + obd->obd_name, to, extend); + } + if (obd->obd_recovery_timeout < to) { obd->obd_recovery_timeout = to; end = obd->obd_recovery_start + to; - cfs_timer_arm(&obd->obd_recovery_timer, - cfs_time_shift(end - now)); + mod_timer(&obd->obd_recovery_timer, + jiffies + cfs_time_seconds(end - now)); } spin_unlock(&obd->obd_dev_lock); - CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)cfs_time_sub(end, now)); + CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n", + obd->obd_name, (s64)(end - now)); } /* Reset the timer with each new client connection */ @@ -1633,7 +1782,7 @@ 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); + time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg); struct obd_device_target *obt = &obd->u.obt; if (!new_client && service_time) @@ -1645,9 +1794,14 @@ check_and_start_recovery_timer(struct obd_device *obd, target_start_recovery_timer(obd); /* Convert the service time to RPC timeout, - * and reuse service_time to limit stack usage. */ + * and reuse service_time to limit stack usage. + */ service_time = at_est2timeout(service_time); + if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) && + service_time < at_extra) + service_time = at_extra; + /* 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. */ @@ -1671,6 +1825,14 @@ static inline int exp_req_replay_healthy(struct obd_export *exp) return (!exp->exp_req_replay_needed || atomic_read(&exp->exp_replay_count) > 0); } + + +static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp) +{ + return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) || + exp_req_replay_healthy(exp); +} + /** if export done lock_replay or has replay in queue */ static inline int exp_lock_replay_healthy(struct obd_export *exp) { @@ -1688,23 +1850,21 @@ 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) +static inline int exp_finished_or_from_mdt(struct obd_export *exp) { - 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); + return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) || + exp_finished(exp); } -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; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; 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); @@ -1712,17 +1872,18 @@ 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 (tdtd != NULL) + update_transno = distribute_txn_get_next_transno(tdtd); + connected = atomic_read(&obd->obd_connected_clients); completed = connected - atomic_read(&obd->obd_req_replay_clients); queue_len = obd->obd_requests_queued_for_recovery; next_transno = obd->obd_next_recovery_transno; CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, " - "req_transno: "LPU64", next_transno: "LPU64"\n", + "req_transno: %llu, next_transno: %llu\n", obd->obd_max_recoverable_clients, connected, completed, queue_len, req_transno, next_transno); @@ -1732,26 +1893,30 @@ static int check_for_next_transno(struct obd_device *obd) } else if (obd->obd_recovery_expired) { CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; - } else if (req_transno == next_transno) { - CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno); + } else if (tdtd != NULL && req != NULL && + is_req_replayed_by_update(req)) { + LASSERTF(req_transno < next_transno, "req_transno %llu" + "next_transno%llu\n", req_transno, next_transno); + CDEBUG(D_HA, "waking for duplicate req (%llu)\n", + req_transno); + wake_up = 1; + } else if (req_transno == next_transno || + (update_transno != 0 && update_transno <= next_transno)) { + CDEBUG(D_HA, "waking for next (%lld)\n", next_transno); wake_up = 1; } else if (queue_len > 0 && queue_len == atomic_read(&obd->obd_req_replay_clients)) { - int d_lvl = D_HA; /** handle gaps occured due to lost reply or VBR */ LASSERTF(req_transno >= next_transno, - "req_transno: "LPU64", next_transno: "LPU64"\n", + "req_transno: %llu, next_transno: %llu\n", req_transno, next_transno); - if (req_transno > obd->obd_last_committed && - !obd->obd_version_recov) - d_lvl = D_ERROR; - CDEBUG(d_lvl, + CDEBUG(D_HA, "%s: waking for gap in transno, VBR is %s (skip: " - LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64 - ", last_committed: "LPD64")\n", + "%lld, ql: %d, comp: %d, conn: %d, next: %lld" + ", next_update %lld last_committed: %lld)\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) { @@ -1759,7 +1924,7 @@ static int check_for_next_transno(struct obd_device *obd) wake_up = 1; } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) { CDEBUG(D_HA, "accepting transno gaps is explicitly allowed" - " by fail_lock, waking up ("LPD64")\n", next_transno); + " by fail_lock, waking up (%lld)\n", next_transno); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } @@ -1767,8 +1932,9 @@ 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); @@ -1795,24 +1961,97 @@ 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; + struct target_distribute_txn_data *tdtd; + time64_t last = 0; + time64_t now; repeat: - 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; + if (obd->obd_recovering && obd->obd_recovery_start == 0) { + now = ktime_get_seconds(); + if (now - last > 600) { + LCONSOLE_INFO("%s: in recovery but waiting for " + "the first client to connect\n", + obd->obd_name); + last = now; + } + } + if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >= + (obd->obd_recovery_start + obd->obd_recovery_time_hard)) { + __u64 next_update_transno = 0; + + /* Only abort the recovery if there are no update recovery + * left in the queue */ + spin_lock(&obd->obd_recovery_task_lock); + if (lut->lut_tdtd != NULL) { + next_update_transno = + distribute_txn_get_next_transno(lut->lut_tdtd); + + tdtd = lut->lut_tdtd; + /* If next_update_transno == 0, it probably because + * updatelog retrieve threads did not get any records + * yet, let's wait those threads stopped */ + if (next_update_transno == 0) { + struct l_wait_info lwi = { 0 }; + + l_wait_event(tdtd->tdtd_recovery_threads_waitq, + atomic_read( + &tdtd->tdtd_recovery_threads_count) == 0, + &lwi); + + next_update_transno = + distribute_txn_get_next_transno( + lut->lut_tdtd); + } + } + + if (next_update_transno != 0 && !obd->obd_abort_recovery) { + obd->obd_next_recovery_transno = next_update_transno; + spin_unlock(&obd->obd_recovery_task_lock); + /* Disconnect unfinished exports from clients, and + * keep connection from MDT to make sure the update + * recovery will still keep trying until some one + * manually abort the recovery */ + class_disconnect_stale_exports(obd, + exp_finished_or_from_mdt); + /* Abort all of replay and replay lock req from + * clients */ + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + CDEBUG(D_HA, "%s: there are still update replay (%#llx" + ")in the queue.\n", obd->obd_name, + next_update_transno); + } else { + obd->obd_abort_recovery = 1; + spin_unlock(&obd->obd_recovery_task_lock); + CWARN("%s recovery is aborted by hard timeout\n", + obd->obd_name); + } } while (wait_event_timeout(obd->obd_next_transno_waitq, - check_routine(obd), + 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) { CWARN("recovery is aborted, evict exports in recovery\n"); + if (lut->lut_tdtd != NULL) { + struct l_wait_info lwi = { 0 }; + + tdtd = lut->lut_tdtd; + /* Let's wait all of the update log recovery thread + * finished */ + l_wait_event(tdtd->tdtd_recovery_threads_waitq, + atomic_read(&tdtd->tdtd_recovery_threads_count) == 0, + &lwi); + /* Then abort the update recovery list */ + dtrq_list_destroy(lut->lut_tdtd); + } + /** evict exports which didn't finish recovery yet */ class_disconnect_stale_exports(obd, exp_finished); return 1; @@ -1823,6 +2062,7 @@ repeat: "evict stale exports\n", obd->obd_name); /** evict cexports with no replay in queue, they are stalled */ class_disconnect_stale_exports(obd, health_check); + /** continue with VBR */ spin_lock(&obd->obd_dev_lock); obd->obd_version_recov = 1; @@ -1838,50 +2078,13 @@ 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); - - 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(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); @@ -1946,7 +2149,7 @@ static void handle_recovery_req(struct ptlrpc_thread *thread, /* don't reset timer for final stage */ if (!exp_finished(req->rq_export)) { - int to = obd_timeout; + time64_t to = obd_timeout; /** * Add request timeout to the recovery time so next request from @@ -1975,6 +2178,253 @@ 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_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 && + !obd->obd_abort_recovery) { + /* 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); + CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n", + obd->obd_name, obd->obd_recovery_timeout); + return 0; + } + } + + 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; +} + +static __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%lld 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++; +} + +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 %lld\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_or_from_mdt)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + goto abort; + } + + spin_lock(&obd->obd_recovery_task_lock); + transno = get_next_transno(lut, &type); + if (type == REQUEST_RECOVERY && transno != 0) { + /* 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); + + /* Let's check if the request has been redone by + * update replay */ + if (is_req_replayed_by_update(req)) { + struct distribute_txn_replay_req *dtrq; + + dtrq = distribute_txn_lookup_finish_list(tdtd, + req->rq_xid); + LASSERT(dtrq != NULL); + spin_lock(&tdtd->tdtd_replay_list_lock); + list_del_init(&dtrq->dtrq_list); + spin_unlock(&tdtd->tdtd_replay_list_lock); + dtrq_destroy(dtrq); + + drop_duplicate_replay_req(env, obd, req); + + continue; + } + + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing t%lld 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; + int rc; + + 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); + rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq); + lu_context_exit(&thread->t_env->le_ctx); + extend_recovery_timer(obd, obd_timeout, true); + + if (rc == 0 && dtrq->dtrq_xid != 0) { + CDEBUG(D_HA, "Move x%llu t%llu" + " to finish list\n", dtrq->dtrq_xid, + dtrq->dtrq_master_transno); + + /* Add it to the replay finish list */ + spin_lock(&tdtd->tdtd_replay_list_lock); + list_add(&dtrq->dtrq_list, + &tdtd->tdtd_replay_finish_list); + spin_unlock(&tdtd->tdtd_replay_list_lock); + + spin_lock(&obd->obd_recovery_task_lock); + if (transno == obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno++; + else if (transno > + obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno = + transno + 1; + spin_unlock(&obd->obd_recovery_task_lock); + } else { + dtrq_destroy(dtrq); + } + } else { + spin_unlock(&obd->obd_recovery_task_lock); +abort: + 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; @@ -2020,43 +2470,28 @@ 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", + CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\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)); @@ -2103,7 +2538,7 @@ 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; @@ -2121,14 +2556,19 @@ static int target_start_recovery_thread(struct lu_target *lut, struct obd_device *obd = lut->lut_obd; int rc = 0; struct target_recovery_data *trd = &obd->obd_recovery_data; + int index; memset(trd, 0, sizeof(*trd)); init_completion(&trd->trd_starting); init_completion(&trd->trd_finishing); trd->trd_recovery_handler = handler; + rc = server_name2index(obd->obd_name, &index, NULL); + if (rc < 0) + return rc; + if (!IS_ERR(kthread_run(target_recovery_thread, - lut, "tgt_recov"))) { + lut, "tgt_recover_%d", index))) { wait_for_completion(&trd->trd_starting); LASSERT(obd->obd_recovering != 0); } else { @@ -2167,9 +2607,9 @@ static void target_recovery_expired(unsigned long castmeharder) { struct obd_device *obd = (struct obd_device *)castmeharder; CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery" - " after %lds (%d clients connected)\n", + " after %llus (%d clients connected)\n", obd->obd_name, atomic_read(&obd->obd_lock_replay_clients), - cfs_time_current_sec()- obd->obd_recovery_start, + (s64)(ktime_get_real_seconds() - obd->obd_recovery_start), atomic_read(&obd->obd_connected_clients)); obd->obd_recovery_expired = 1; @@ -2178,27 +2618,31 @@ 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); - return; - } + struct obd_device *obd = lut->lut_obd; + + if (lut->lut_bottom->dd_rdonly) + return; + + if (obd->obd_max_recoverable_clients == 0) { + /** Update server last boot epoch */ + tgt_boot_epoch_update(lut); + return; + } CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, " - "last_transno "LPU64"\n", obd->obd_name, + "last_transno %llu\n", obd->obd_name, obd->obd_max_recoverable_clients, obd->obd_last_committed); LASSERT(obd->obd_stopping == 0); obd->obd_next_recovery_transno = obd->obd_last_committed + 1; obd->obd_recovery_start = 0; obd->obd_recovery_end = 0; - cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd); - target_start_recovery_thread(lut, handler); + setup_timer(&obd->obd_recovery_timer, target_recovery_expired, + (unsigned long)obd); + target_start_recovery_thread(lut, handler); } EXPORT_SYMBOL(target_recovery_init); - static int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req) { @@ -2250,6 +2694,17 @@ int target_queue_recovery_request(struct ptlrpc_request *req, target_process_req_flags(obd, req); if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) { + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) { + if (cfs_fail_val == 1) { + cfs_race_state = 1; + cfs_fail_val = 0; + wake_up(&cfs_race_waitq); + + set_current_state(TASK_INTERRUPTIBLE); + schedule_timeout(cfs_time_seconds(1)); + } + } + /* client declares he's ready to complete recovery * so, we put the request on th final queue */ target_request_copy_get(req); @@ -2257,15 +2712,46 @@ int target_queue_recovery_request(struct ptlrpc_request *req, wake_up(&obd->obd_next_transno_waitq); spin_lock(&obd->obd_recovery_task_lock); if (obd->obd_recovering) { + struct ptlrpc_request *tmp; + struct ptlrpc_request *duplicate = NULL; + + if (likely(!req->rq_export->exp_replay_done)) { + req->rq_export->exp_replay_done = 1; + list_add_tail(&req->rq_list, + &obd->obd_final_req_queue); + spin_unlock(&obd->obd_recovery_task_lock); + RETURN(0); + } + + /* XXX O(n), but only happens if final ping is + * timed out, probably reorganize the list as + * a hash list later */ + list_for_each_entry_safe(reqiter, tmp, + &obd->obd_final_req_queue, + rq_list) { + if (reqiter->rq_export == req->rq_export) { + list_del_init(&reqiter->rq_list); + duplicate = reqiter; + break; + } + } + list_add_tail(&req->rq_list, - &obd->obd_final_req_queue); + &obd->obd_final_req_queue); + req->rq_export->exp_replay_done = 1; + spin_unlock(&obd->obd_recovery_task_lock); + + if (duplicate != NULL) { + DEBUG_REQ(D_HA, duplicate, + "put prev final req\n"); + target_request_copy_put(duplicate); + } + RETURN(0); } else { spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(obd->obd_stopping ? -ENOTCONN : 1); } - 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 */ @@ -2306,11 +2792,16 @@ 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. */ - CDEBUG(D_HA, "Next recovery transno: "LPU64 - ", current: "LPU64", replaying\n", + CDEBUG(D_HA, "Next recovery transno: %llu" + ", current: %llu, replaying\n", obd->obd_next_recovery_transno, transno); + + /* If the request has been replayed by update replay, then sends this + * request to the recovery thread (replay_request_or_update()), where + * it will be handled */ spin_lock(&obd->obd_recovery_task_lock); - if (transno < obd->obd_next_recovery_transno) { + if (transno < obd->obd_next_recovery_transno && + !is_req_replayed_by_update(req)) { /* Processing the queue right now, don't re-add. */ LASSERT(list_empty(&req->rq_list)); spin_unlock(&obd->obd_recovery_task_lock); @@ -2382,7 +2873,7 @@ void target_committed_to_req(struct ptlrpc_request *req) "%d)", exp->exp_obd->obd_no_transno, req->rq_repmsg == NULL); - CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", + CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n", exp->exp_last_committed, req->rq_transno, req->rq_xid); } @@ -2419,20 +2910,26 @@ int target_pack_pool_reply(struct ptlrpc_request *req) 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"); - return (-ECOMM); - } - - if (unlikely(rc)) { - DEBUG_REQ(D_NET, req, "processing error (%d)", rc); - req->rq_status = rc; - return (ptlrpc_send_error(req, 1)); - } else { - DEBUG_REQ(D_NET, req, "sending reply"); - } + if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) { + DEBUG_REQ(D_ERROR, req, "dropping reply"); + return -ECOMM; + } + /* We can have a null rq_reqmsg in the event of bad signature or + * no context when unwrapping */ + if (req->rq_reqmsg && + unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT && + OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP))) + return -ECOMM; + + if (unlikely(rc)) { + DEBUG_REQ(D_NET, req, "processing error (%d)", rc); + req->rq_status = rc; + return ptlrpc_send_error(req, 1); + } else { + DEBUG_REQ(D_NET, req, "sending reply"); + } - return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT)); + return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT); } void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) @@ -2482,7 +2979,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg); spin_lock(&exp->exp_uncommitted_replies_lock); - CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n", + CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n", rs->rs_transno, exp->exp_last_committed); if (rs->rs_transno > exp->exp_last_committed) { /* not committed already */ @@ -2527,7 +3024,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) EXIT; } -ldlm_mode_t lck_compat_array[] = { +enum ldlm_mode lck_compat_array[] = { [LCK_EX] = LCK_COMPAT_EX, [LCK_PW] = LCK_COMPAT_PW, [LCK_PR] = LCK_COMPAT_PR, @@ -2542,12 +3039,12 @@ ldlm_mode_t lck_compat_array[] = { * Rather arbitrary mapping from LDLM error codes to errno values. This should * not escape to the user level. */ -int ldlm_error2errno(ldlm_error_t error) +int ldlm_error2errno(enum ldlm_error error) { - int result; + int result; - switch (error) { - case ELDLM_OK: + switch (error) { + case ELDLM_OK: case ELDLM_LOCK_MATCHED: result = 0; break; @@ -2569,22 +3066,22 @@ int ldlm_error2errno(ldlm_error_t error) case ELDLM_BAD_NAMESPACE: result = -EBADF; break; - default: - if (((int)error) < 0) /* cast to signed type */ - result = error; /* as ldlm_error_t can be unsigned */ - else { - CERROR("Invalid DLM result code: %d\n", error); - result = -EPROTO; - } - } - return result; + default: + if (((int)error) < 0) { /* cast to signed type */ + result = error; /* as ldlm_error can be unsigned */ + } else { + CERROR("Invalid DLM result code: %d\n", error); + result = -EPROTO; + } + } + return result; } EXPORT_SYMBOL(ldlm_error2errno); /** - * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t. + * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error. */ -ldlm_error_t ldlm_errno2error(int err_no) +enum ldlm_error ldlm_errno2error(int err_no) { int error; @@ -2655,10 +3152,10 @@ static inline const char *bulk2type(struct ptlrpc_request *req) 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; - time_t start = cfs_time_current_sec(); - time_t deadline; - int rc = 0; + struct ptlrpc_request *req = desc->bd_req; + time64_t start = ktime_get_real_seconds(); + time64_t deadline; + int rc = 0; ENTRY; @@ -2678,6 +3175,12 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, } else { if (req->rq_bulk_read) rc = sptlrpc_svc_wrap_bulk(req, desc); + + if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS)) + req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg); + else /* old version, bulk matchbits is rq_xid */ + req->rq_mbits = req->rq_xid; + if (rc == 0) rc = ptlrpc_start_bulk_transfer(desc); } @@ -2699,12 +3202,13 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, deadline = req->rq_deadline; do { - long timeoutl = deadline - cfs_time_current_sec(); - cfs_duration_t timeout = timeoutl <= 0 ? - CFS_TICK : cfs_time_seconds(timeoutl); - time_t rq_deadline; + time64_t timeoutl = deadline - ktime_get_real_seconds(); + long timeout_jiffies = timeoutl <= 0 ? + 1 : cfs_time_seconds(timeoutl); + time64_t rq_deadline; - *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), + *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies, + cfs_time_seconds(1), target_bulk_timeout, desc); rc = l_wait_event(desc->bd_waitq, !ptlrpc_server_bulk_active(desc) || @@ -2718,13 +3222,13 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, deadline = start + bulk_timeout; if (deadline > rq_deadline) deadline = rq_deadline; - } while ((rc == -ETIMEDOUT) && - (deadline > cfs_time_current_sec())); + } while (rc == -ETIMEDOUT && + deadline > ktime_get_real_seconds()); if (rc == -ETIMEDOUT) { - DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds", + DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds", bulk2type(req), deadline - start, - cfs_time_current_sec() - deadline); + ktime_get_real_seconds() - deadline); ptlrpc_abort_bulk(desc); } else if (exp->exp_failed) { DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",