X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=5086c28e488d0308012585b926d05338f0a09ce0;hp=13e3e59d015f7f769896e754eb5f4ec5eea4731f;hb=5e30a2c06176f50f5e17aba68fdae7e38d922d33;hpb=2fe2d1e82005746180309d9b79057a418a729e54 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 13e3e59..5086c28e 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, 2015, Intel Corporation. + * Copyright (c) 2010, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -43,6 +39,9 @@ #define DEBUG_SUBSYSTEM S_LDLM +#include +#include +#include #include #include #include @@ -52,88 +51,93 @@ #include #include "ldlm_internal.h" -/* @priority: If non-zero, move the selected connection to the list head. +/* + * @priority: If non-zero, move the selected connection to the list head. * @create: If zero, only search in existing connections. */ 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; - - 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 (create) { - OBD_ALLOC(imp_conn, sizeof(*imp_conn)); - if (!imp_conn) { - GOTO(out_put, rc = -ENOMEM); - } - } + int priority, int create) +{ + 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 (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); + } 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) { - 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" : "")); + if (obd_uuid_equals(uuid, &item->oic_uuid)) { + if (priority) { + list_move(&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" : "")); 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) { - return import_set_conn(imp, uuid, 1, 0); + return import_set_conn(imp, uuid, 1, 0); } int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid, - int priority) + int priority) { - return import_set_conn(imp, uuid, priority, 1); + return import_set_conn(imp, uuid, priority, 1); } EXPORT_SYMBOL(client_import_add_conn); @@ -142,48 +146,52 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) struct obd_import_conn *imp_conn; struct obd_export *dlmexp; int rc = -ENOENT; + ENTRY; spin_lock(&imp->imp_lock); if (list_empty(&imp->imp_conn_list)) { - LASSERT(!imp->imp_connection); - GOTO(out, rc); - } + LASSERT(!imp->imp_connection); + GOTO(out, rc); + } list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) { - if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid)) - continue; - LASSERT(imp_conn->oic_conn); - - if (imp_conn == imp->imp_conn_current) { - LASSERT(imp_conn->oic_conn == imp->imp_connection); - - if (imp->imp_state != LUSTRE_IMP_CLOSED && - imp->imp_state != LUSTRE_IMP_DISCON) { - CERROR("can't remove current connection\n"); - GOTO(out, rc = -EBUSY); - } - - 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; - } - } + if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid)) + continue; + LASSERT(imp_conn->oic_conn); + + if (imp_conn == imp->imp_conn_current) { + LASSERT(imp_conn->oic_conn == imp->imp_connection); + + if (imp->imp_state != LUSTRE_IMP_CLOSED && + imp->imp_state != LUSTRE_IMP_DISCON) { + CERROR("can't remove current connection\n"); + GOTO(out, rc = -EBUSY); + } + + 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; + } + + if (dlmexp != NULL) + class_export_put(dlmexp); + } list_del(&imp_conn->oic_item); - ptlrpc_connection_put(imp_conn->oic_conn); - OBD_FREE(imp_conn, sizeof(*imp_conn)); - CDEBUG(D_HA, "imp %p@%s: remove connection %s\n", - imp, imp->imp_obd->obd_name, uuid->uuid); - rc = 0; - break; - } + ptlrpc_connection_put(imp_conn->oic_conn); + OBD_FREE(imp_conn, sizeof(*imp_conn)); + CDEBUG(D_HA, "imp %p@%s: remove connection %s\n", + imp, imp->imp_obd->obd_name, uuid->uuid); + rc = 0; + break; + } out: spin_unlock(&imp->imp_lock); if (rc == -ENOENT) @@ -201,17 +209,18 @@ int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, { struct obd_import_conn *conn; int rc = -ENOENT; + ENTRY; spin_lock(&imp->imp_lock); list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { /* Check if conn UUID does have this peer NID. */ - if (class_check_uuid(&conn->oic_uuid, peer)) { - *uuid = conn->oic_uuid; - rc = 0; - break; - } - } + if (class_check_uuid(&conn->oic_uuid, peer)) { + *uuid = conn->oic_uuid; + rc = 0; + break; + } + } spin_unlock(&imp->imp_lock); RETURN(rc); } @@ -219,9 +228,14 @@ EXPORT_SYMBOL(client_import_find_conn); void client_destroy_import(struct obd_import *imp) { - class_import_get(imp); - class_destroy_import(imp); - class_import_put(imp); + /* + * Drop security policy instance after all RPCs have finished/aborted + * to let all busy contexts be released. + */ + class_import_get(imp); + class_destroy_import(imp); + sptlrpc_import_sec_put(imp); + class_import_put(imp); } EXPORT_SYMBOL(client_destroy_import); @@ -248,12 +262,14 @@ static int osc_on_mdt(char *obdname) return 0; } -/* Configure an RPC client OBD device. +/* + * Configure an RPC client OBD device. * * lcfg parameters: * 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) { @@ -261,14 +277,19 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) struct obd_import *imp; struct obd_uuid server_uuid; int rq_portal, rp_portal, connect_op; - char *name = obddev->obd_type->typ_name; + const 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; @@ -305,52 +326,57 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) } rp_portal = OSC_REPLY_PORTAL; cli->cl_sp_me = LUSTRE_SP_MDT; - } else if (!strcmp(name, LUSTRE_MGC_NAME)) { - rq_portal = MGS_REQUEST_PORTAL; - rp_portal = MGC_REPLY_PORTAL; - connect_op = MGS_CONNECT; - cli->cl_sp_me = LUSTRE_SP_MGC; - cli->cl_sp_to = LUSTRE_SP_MGS; - cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID; - ns_type = LDLM_NS_TYPE_MGC; + } else if (!strcmp(name, LUSTRE_MGC_NAME)) { + rq_portal = MGS_REQUEST_PORTAL; + rp_portal = MGC_REPLY_PORTAL; + connect_op = MGS_CONNECT; + cli->cl_sp_me = LUSTRE_SP_MGC; + cli->cl_sp_to = LUSTRE_SP_MGS; + cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID; + ns_type = LDLM_NS_TYPE_MGC; } else { - CERROR("unknown client OBD type \"%s\", can't setup\n", - name); - RETURN(-EINVAL); - } - - if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) { - CERROR("requires a TARGET UUID\n"); - RETURN(-EINVAL); - } - - if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) { - CERROR("client UUID must be less than 38 characters\n"); - RETURN(-EINVAL); - } - - if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) { - CERROR("setup requires a SERVER UUID\n"); - RETURN(-EINVAL); - } - - if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) { - CERROR("target UUID must be less than 38 characters\n"); - RETURN(-EINVAL); - } + CERROR("unknown client OBD type \"%s\", can't setup\n", + name); + RETURN(-EINVAL); + } + + if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) { + CERROR("requires a TARGET UUID\n"); + RETURN(-EINVAL); + } + + if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) { + CERROR("client UUID must be less than 38 characters\n"); + RETURN(-EINVAL); + } + + if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) { + CERROR("setup requires a SERVER UUID\n"); + RETURN(-EINVAL); + } + + if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) { + CERROR("target UUID must be less than 38 characters\n"); + RETURN(-EINVAL); + } 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_dirty_max_pages = 0; cli->cl_avail_grant = 0; /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */ - /* cl_dirty_max_pages may be changed at connect time in - * ptlrpc_connect_interpret(). */ + /* + * cl_dirty_max_pages may be changed at connect time in + * ptlrpc_connect_interpret(). + */ client_adjust_max_dirty(cli); INIT_LIST_HEAD(&cli->cl_cache_waiters); INIT_LIST_HEAD(&cli->cl_loi_ready_list); @@ -379,9 +405,17 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) 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_grant_chain); + + 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); + + + cli->cl_supp_cksum_types = OBD_CKSUM_CRC32; + cli->cl_preferred_cksum_type = 0; #ifdef ENABLE_CHECKSUM /* Turn on checksumming by default. */ cli->cl_checksum = 1; @@ -390,35 +424,38 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) * Set cl_chksum* to CRC32 for now to avoid returning screwed info * through procfs. */ - cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32; + cli->cl_cksum_type = cli->cl_supp_cksum_types; #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_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES; - /* set cl_chunkbits default value to PAGE_CACHE_SHIFT, - * it will be updated at OSC connection time. */ - cli->cl_chunkbits = PAGE_CACHE_SHIFT; + /* + * set cl_chunkbits default value to PAGE_SHIFT, + * it will be updated at OSC connection time. + */ + 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 (cfs_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 (cfs_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 (cfs_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); @@ -428,6 +465,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) 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, @@ -436,67 +475,82 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err, rc = -ENOMEM); } - rc = ldlm_get_ref(); - if (rc) { - CERROR("ldlm_get_ref failed: %d\n", rc); - GOTO(err, rc); - } - - ptlrpc_init_client(rq_portal, rp_portal, name, - &obddev->obd_ldlm_client); + rc = ldlm_get_ref(); + if (rc) { + CERROR("ldlm_get_ref failed: %d\n", rc); + GOTO(err, rc); + } - imp = class_new_import(obddev); - if (imp == NULL) - GOTO(err_ldlm, rc = -ENOENT); - imp->imp_client = &obddev->obd_ldlm_client; - imp->imp_connect_op = connect_op; - memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1), - LUSTRE_CFG_BUFLEN(lcfg, 1)); - class_import_put(imp); + ptlrpc_init_client(rq_portal, rp_portal, name, + &obddev->obd_ldlm_client); + + imp = class_new_import(obddev); + if (imp == NULL) + GOTO(err_ldlm, rc = -ENOENT); + imp->imp_client = &obddev->obd_ldlm_client; + imp->imp_connect_op = connect_op; + memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1), + LUSTRE_CFG_BUFLEN(lcfg, 1)); + class_import_put(imp); + + 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); - } + 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 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")) { - CDEBUG(D_HA, "marking %s %s->%s as inactive\n", - name, obddev->obd_name, - cli->cl_target_uuid.uuid); + if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) { + if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) { + CDEBUG(D_HA, "marking %s %s->%s as inactive\n", + name, obddev->obd_name, + cli->cl_target_uuid.uuid); spin_lock(&imp->imp_lock); imp->imp_deactive = 1; spin_unlock(&imp->imp_lock); - } - } + } + } - obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name, - LDLM_NAMESPACE_CLIENT, - LDLM_NAMESPACE_GREEDY, - ns_type); - if (obddev->obd_namespace == NULL) { - CERROR("Unable to create client namespace - %s\n", - obddev->obd_name); - GOTO(err_import, rc = -ENOMEM); - } + obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name, + LDLM_NAMESPACE_CLIENT, + LDLM_NAMESPACE_GREEDY, + ns_type); + if (obddev->obd_namespace == NULL) { + CERROR("Unable to create client namespace - %s\n", + obddev->obd_name); + GOTO(err_import, rc = -ENOMEM); + } - RETURN(rc); + RETURN(rc); err_import: - class_destroy_import(imp); + class_destroy_import(imp); err_ldlm: - ldlm_put_ref(); + 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); + RETURN(rc); } EXPORT_SYMBOL(client_obd_setup); @@ -504,6 +558,7 @@ 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); @@ -525,160 +580,178 @@ EXPORT_SYMBOL(client_obd_cleanup); /* ->o_connect() method for client side (OSC and MDC and MGC) */ int client_connect_import(const struct lu_env *env, - struct obd_export **exp, - struct obd_device *obd, struct obd_uuid *cluuid, - struct obd_connect_data *data, void *localdata) + struct obd_export **exp, + struct obd_device *obd, struct obd_uuid *cluuid, + struct obd_connect_data *data, void *localdata) { - struct client_obd *cli = &obd->u.cli; - struct obd_import *imp = cli->cl_import; + struct client_obd *cli = &obd->u.cli; + struct obd_import *imp = cli->cl_import; struct obd_connect_data *ocd; - struct lustre_handle conn = { 0 }; - int rc; - bool is_mdc = false; + struct lustre_handle conn = { 0 }; + 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); - - cli->cl_conn_count++; - *exp = class_conn2export(&conn); - - LASSERT(obd->obd_namespace); - - 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; - is_mdc = strncmp(imp->imp_obd->obd_type->typ_name, - LUSTRE_MDC_NAME, 3) == 0; - if (is_mdc) - data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS; - imp->imp_connect_flags_orig = data->ocd_connect_flags; - } - - rc = ptlrpc_connect_import(imp); - if (rc != 0) { - LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); - GOTO(out_ldlm, rc); - } + GOTO(out_sem, rc = -EALREADY); + + rc = class_connect(&conn, obd, cluuid); + if (rc) + GOTO(out_sem, rc); + + cli->cl_conn_count++; + *exp = class_conn2export(&conn); + + LASSERT(obd->obd_namespace); + + 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; + 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); + } 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; - /* clear the flag as it was not set and is not known - * by upper layers */ - if (is_mdc) - data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS; - } + 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); + if (!rc && localdata) { + LASSERT(cli->cl_cache == NULL); /* only once */ + cli->cl_cache = (struct cl_client_cache *)localdata; + cl_cache_incref(cli->cl_cache); + cli->cl_lru_left = &cli->cl_cache->ccc_lru_left; + + /* add this osc into entity list */ + LASSERT(list_empty(&cli->cl_lru_osc)); + spin_lock(&cli->cl_cache->ccc_lru_lock); + list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru); + spin_unlock(&cli->cl_cache->ccc_lru_lock); + } + return rc; } EXPORT_SYMBOL(client_connect_import); int client_disconnect_export(struct obd_export *exp) { - struct obd_device *obd = class_exp2obd(exp); - struct client_obd *cli; - struct obd_import *imp; - int rc = 0, err; - ENTRY; + struct obd_device *obd = class_exp2obd(exp); + struct client_obd *cli; + struct obd_import *imp; + int rc = 0, err; - if (!obd) { - CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n", - exp, exp ? exp->exp_handle.h_cookie : -1); - RETURN(-EINVAL); - } + ENTRY; - cli = &obd->u.cli; - imp = cli->cl_import; + if (!obd) { + CERROR("invalid export for disconnect: exp %p cookie %#llx\n", + exp, exp ? exp->exp_handle.h_cookie : -1); + RETURN(-EINVAL); + } + + cli = &obd->u.cli; + imp = cli->cl_import; down_write(&cli->cl_sem); CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name, cli->cl_conn_count); if (cli->cl_conn_count == 0) { - CERROR("disconnecting disconnected device (%s)\n", - obd->obd_name); - GOTO(out_disconnect, rc = -EINVAL); - } + CERROR("disconnecting disconnected device (%s)\n", + obd->obd_name); + GOTO(out_disconnect, rc = -EINVAL); + } - cli->cl_conn_count--; + cli->cl_conn_count--; if (cli->cl_conn_count != 0) - GOTO(out_disconnect, rc = 0); + GOTO(out_disconnect, rc = 0); - /* Mark import deactivated now, so we don't try to reconnect if any + /* + * Mark import deactivated now, so we don't try to reconnect if any * of the cleanup RPCs fails (e.g. LDLM cancel, etc). We don't - * fully deactivate the import, or that would drop all requests. */ + * fully deactivate the import, or that would drop all requests. + */ spin_lock(&imp->imp_lock); imp->imp_deactive = 1; spin_unlock(&imp->imp_lock); - /* Some non-replayable imports (MDS's OSCs) are pinged, so just - * delete it regardless. (It's safe to delete an import that was - * never added.) */ - (void)ptlrpc_pinger_del_import(imp); - - if (obd->obd_namespace != NULL) { - /* obd_force == local only */ - ldlm_cli_cancel_unused(obd->obd_namespace, NULL, - obd->obd_force ? LCF_LOCAL : 0, NULL); - ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force); - } + /* + * Some non-replayable imports (MDS's OSCs) are pinged, so just + * delete it regardless. (It's safe to delete an import that was + * never added.) + */ + (void)ptlrpc_pinger_del_import(imp); + + if (obd->obd_namespace != NULL) { + /* obd_force == local only */ + ldlm_cli_cancel_unused(obd->obd_namespace, NULL, + obd->obd_force ? LCF_LOCAL : 0, NULL); + ldlm_namespace_free_prior(obd->obd_namespace, imp, + obd->obd_force); + } - /* There's no need to hold sem while disconnecting an import, - * and it may actually cause deadlock in GSS. */ + /* + * There's no need to hold sem while disconnecting an import, + * and it may actually cause deadlock in GSS. + */ up_write(&cli->cl_sem); rc = ptlrpc_disconnect_import(imp, 0); down_write(&cli->cl_sem); - ptlrpc_invalidate_import(imp); + ptlrpc_invalidate_import(imp); - EXIT; + EXIT; out_disconnect: - /* Use server style - class_disconnect should be always called for - * o_disconnect. */ - err = class_disconnect(exp); - if (!rc && err) - rc = err; + /* + * Use server style - class_disconnect should be always called for + * o_disconnect. + */ + err = class_disconnect(exp); + if (!rc && err) + rc = err; up_write(&cli->cl_sem); - RETURN(rc); + RETURN(rc); } EXPORT_SYMBOL(client_disconnect_export); #ifdef HAVE_SERVER_SUPPORT int server_disconnect_export(struct obd_export *exp) { - int rc; - ENTRY; + int rc; + + ENTRY; /* Disconnect early so that clients can't keep using export. */ rc = class_disconnect(exp); @@ -688,7 +761,7 @@ int server_disconnect_export(struct obd_export *exp) ldlm_bl_thread_wakeup(); - /* complete all outstanding replies */ + /* complete all outstanding replies */ spin_lock(&exp->exp_lock); while (!list_empty(&exp->exp_outstanding_replies)) { struct ptlrpc_reply_state *rs = @@ -699,7 +772,10 @@ 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); @@ -711,89 +787,109 @@ int server_disconnect_export(struct obd_export *exp) } EXPORT_SYMBOL(server_disconnect_export); -/* -------------------------------------------------------------------------- +/* + * -------------------------------------------------------------------------- * from old lib/target.c - * -------------------------------------------------------------------------- */ + * -------------------------------------------------------------------------- + */ static int target_handle_reconnect(struct lustre_handle *conn, - struct obd_export *exp, - struct obd_uuid *cluuid) + struct obd_export *exp, + struct obd_uuid *cluuid) { + struct obd_device *target; struct lustre_handle *hdl; - ENTRY; + ktime_t remaining; + s64 timeout; + int rc = 0; + ENTRY; hdl = &exp->exp_imp_reverse->imp_remote_handle; - if (exp->exp_connection && lustre_handle_is_used(hdl)) { - struct obd_device *target; - - 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)); - } - - 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); - } - } + 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); + remaining = hrtimer_expires_remaining(&target->obd_recovery_timer); + timeout = ktime_divns(remaining, NSEC_PER_SEC); + if (timeout > 0) { + LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n", + target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), + atomic_read(&target->obd_max_recoverable_clients), + timeout / 60, timeout % 60); + } else { + struct target_distribute_txn_data *tdtd; + int size = 0; + int count = 0; + char *buf = NULL; + + 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: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n", + target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), count, + buf ? buf : "unknown (not enough RAM)", + (abs(timeout) + target->obd_recovery_timeout) / 60, + (abs(timeout) + target->obd_recovery_timeout) % 60, + target->obd_name); + else + LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n", + target->obd_name, abs(timeout) / 60, + abs(timeout) % 60, target->obd_name); - spin_lock(&exp->exp_lock); - exp->exp_need_sync = 0; - spin_unlock(&exp->exp_lock); - class_export_cb_put(exp); + if (buf != NULL) + OBD_FREE(buf, size); + } + +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); + struct ptlrpc_request *req, int new_client); /** * update flags for import during reconnect process @@ -885,17 +981,21 @@ static int rev_import_reconnect(struct obd_export *exp, 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() */ + /* + * 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 + /* + * 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 */ + * be delivered correctly + */ return rc; } @@ -907,49 +1007,47 @@ int target_handle_connect(struct ptlrpc_request *req) { struct obd_device *target = NULL; struct obd_export *export = NULL; - /* connect handle - filled from target_handle_reconnect in - * reconnect case */ + /* + * connect handle - filled from target_handle_reconnect in + * reconnect case + */ struct lustre_handle conn; struct lustre_handle *tmp; - struct obd_uuid tgtuuid; - struct obd_uuid cluuid; - char *str; - int rc = 0; - char *target_start; - int target_len; - bool mds_conn = false, lw_client = false; - bool mds_mds_conn = false; - bool new_mds_mds_conn = false; - bool target_referenced = false; - struct obd_connect_data *data, *tmpdata; - int size, tmpsize; - lnet_nid_t *client_nid = NULL; - ENTRY; + struct obd_uuid cluuid; + char *str; + int rc = 0; + char *target_start; + int target_len; + 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; - OBD_RACE(OBD_FAIL_TGT_CONN_RACE); + ENTRY; - str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID); - if (str == NULL) { - DEBUG_REQ(D_ERROR, req, "bad target UUID for connect"); - GOTO(out, rc = -EINVAL); - } + OBD_RACE(OBD_FAIL_TGT_CONN_RACE); - obd_str2uuid(&tgtuuid, str); - target = class_uuid2obd(&tgtuuid); - if (!target) - target = class_name2obd(str); + str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID); + if (str == NULL) { + DEBUG_REQ(D_ERROR, req, "bad target UUID for connect"); + GOTO(out, rc = -EINVAL); + } + target = class_dev_by_str(str); if (!target) { deuuidify(str, NULL, &target_start, &target_len); - LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect " - "from %s (no target). If you are running " - "an HA pair check that the target is " - "mounted on the other server.\n", str, - libcfs_nid2str(req->rq_peer.nid)); + LCONSOLE_ERROR_MSG(0x137, + "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n", + str, libcfs_nid2str(req->rq_peer.nid)); GOTO(out, rc = -ENODEV); } spin_lock(&target->obd_dev_lock); + + target->obd_conn_inprogress++; + if (target->obd_stopping || !target->obd_set_up) { spin_unlock(&target->obd_dev_lock); @@ -962,50 +1060,46 @@ int target_handle_connect(struct ptlrpc_request *req) GOTO(out, rc = -ENODEV); } - if (target->obd_no_conn) { + if (target->obd_no_conn) { spin_unlock(&target->obd_dev_lock); - CDEBUG(D_INFO, "%s: Temporarily refusing client connection " - "from %s\n", target->obd_name, - libcfs_nid2str(req->rq_peer.nid)); + CDEBUG(D_INFO, + "%s: Temporarily refusing client connection from %s\n", + target->obd_name, libcfs_nid2str(req->rq_peer.nid)); 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. */ - class_incref(target, __func__, current); - target_referenced = true; - - target->obd_conn_inprogress++; spin_unlock(&target->obd_dev_lock); - str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID); - if (str == NULL) { - DEBUG_REQ(D_ERROR, req, "bad client UUID for connect"); - GOTO(out, rc = -EINVAL); - } + str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID); + if (str == NULL) { + DEBUG_REQ(D_ERROR, req, "bad client UUID for connect"); + GOTO(out, rc = -EINVAL); + } - obd_str2uuid(&cluuid, str); + obd_str2uuid(&cluuid, str); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); - if (tmp == NULL) - GOTO(out, rc = -EPROTO); + tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); + if (tmp == NULL) + GOTO(out, rc = -EPROTO); - conn = *tmp; + conn = *tmp; - size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, - RCL_CLIENT); - data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); - if (!data) - GOTO(out, rc = -EPROTO); + size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_CLIENT); + if (size < 0 || size > 8 * sizeof(struct obd_connect_data)) + GOTO(out, rc = -EPROTO); + data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); + if (!data) + GOTO(out, rc = -EPROTO); - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - GOTO(out, rc); + rc = req_capsule_server_pack(&req->rq_pill); + 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 + /* + * 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 @@ -1016,49 +1110,41 @@ 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 + /* + * 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 */ + * 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; mds_mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) != 0; - /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS + /* + * 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) && @@ -1067,38 +1153,38 @@ int target_handle_connect(struct ptlrpc_request *req) __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. */ + /* + * 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); - + 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. */ - if (obd_uuid_equals(&cluuid, &target->obd_uuid)) - goto dont_check_exports; + /* lctl gets a backstage, all-access pass. */ + if (obd_uuid_equals(&cluuid, &target->obd_uuid)) + goto dont_check_exports; - export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid); - if (!export) - goto no_export; + export = obd_uuid_lookup(target, &cluuid); + if (!export) + goto no_export; /* We've found an export in the hash. */ spin_lock(&export->exp_lock); - if (export->exp_connecting) { /* bug 9635, et. al. */ + if (export->exp_connecting) { /* b=9635, et. al. */ spin_unlock(&export->exp_lock); LCONSOLE_WARN("%s: Export %p already connecting from %s\n", export->exp_obd->obd_name, export, @@ -1106,41 +1192,63 @@ 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 + 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 { /* 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)) { + 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)); + } + + 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; + export->exp_conn_cnt = 0; + 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 " - "existing nid %s is already connected\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - libcfs_nid2str( - export->exp_connection->c_peer.nid)); - rc = -EALREADY; - class_export_put(export); - export = NULL; - } else { + LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid), + libcfs_nid2str( + export->exp_connection->c_peer.nid)); + rc = -EALREADY; + class_export_put(export); + export = NULL; + } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) && + !lw_client) { + spin_unlock(&export->exp_lock); + rc = -EAGAIN; + } else { export->exp_connecting = 1; spin_unlock(&export->exp_lock); LASSERT(export->exp_obd == target); @@ -1148,85 +1256,97 @@ int target_handle_connect(struct ptlrpc_request *req) rc = target_handle_reconnect(&conn, export, &cluuid); } - /* If we found an export, we already unlocked. */ - if (!export) { + /* If we found an export, we already unlocked. */ + if (!export) { no_export: - OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout); - } else if (req->rq_export == NULL && + OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout); + } else if (req->rq_export == NULL && atomic_read(&export->exp_rpc_count) > 0) { - LCONSOLE_WARN("%s: Client %s (at %s) refused connection, " - "still busy with %d references\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - atomic_read(&export->exp_refcount)); - GOTO(out, rc = -EBUSY); - } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) { - if (!strstr(cluuid.uuid, "mdt")) - LCONSOLE_WARN("%s: Rejecting reconnect from the " - "known client %s (at %s) because it " - "is indicating it is a new client", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid)); - GOTO(out, rc = -EALREADY); - } else { - OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout); - } - - if (rc < 0) { - 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); - - /* If this is the first time a client connects, reset the recovery - * timer. Discard lightweight connections which might be local. */ + LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid), + refcount_read(&export->exp_handle.h_ref)); + GOTO(out, rc = -EBUSY); + } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 && + rc != EALREADY) { + if (!strstr(cluuid.uuid, "mdt")) + LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid)); + GOTO(out, rc = -EALREADY); + } else { + OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout); + } + + if (rc < 0) + GOTO(out, rc); + + 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_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. + */ if (!lw_client && rc == 0 && target->obd_recovering) check_and_start_recovery_timer(target, req, export == NULL); - /* We want to handle EALREADY but *not* -EALREADY from - * target_handle_reconnect(), return reconnection state in a flag. */ - if (rc == EALREADY) { - lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT); - rc = 0; - } else { - LASSERT(rc == 0); - } + /* + * We want to handle EALREADY but *not* -EALREADY from + * target_handle_reconnect(), return reconnection state in a flag. + */ + if (rc == EALREADY) { + lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT); + rc = 0; + } else { + LASSERT(rc == 0); + } /* Tell the client if we support replayable requests. */ - if (target->obd_replayable) - lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE); - client_nid = &req->rq_peer.nid; + if (target->obd_replayable) + lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE); + client_nid = &req->rq_peer.nid; - if (export == NULL) { + if (export == NULL) { /* allow lightweight connections during recovery */ - /* allow "new" MDT to be connected during recovery, since we - * need retrieve recovery update records from it */ + /* + * 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 */ - 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 %d known clients " - "(%d recovered, %d in progress, and %d " - "evicted) to recover in %d:%.02d\n", + struct hrtimer *timer = &target->obd_recovery_timer; + ktime_t remaining; + s64 timeout, left; + int in_progress; + int connected; + int known; + int stale; + char *msg; + + connected = atomic_read(&target->obd_connected_clients); + in_progress = atomic_read(&target->obd_lock_replay_clients); + known = + atomic_read(&target->obd_max_recoverable_clients); + stale = target->obd_stale_clients; + remaining = hrtimer_expires_remaining(timer); + left = ktime_divns(remaining, NSEC_PER_SEC); + + if (ktime_to_ns(remaining) > 0) { + msg = "to recover in"; + timeout = left; + } else { + msg = "already passed deadline"; + timeout = -left; + } + + LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %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); + libcfs_nid2str(req->rq_peer.nid), known, + connected - in_progress, in_progress, + stale, msg, timeout / 60, timeout % 60); rc = -EBUSY; } else { dont_check_exports: @@ -1254,64 +1374,58 @@ dont_check_exports: LASSERT(target->u.obt.obt_magic == OBT_MAGIC); data->ocd_instance = target->u.obt.obt_instance; - /* Return only the parts of obd_connect_data that we understand, so the - * client knows that we don't understand the rest. */ - if (data) { - tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, - RCL_SERVER); - tmpdata = req_capsule_server_get(&req->rq_pill, - &RMF_CONNECT_DATA); - /* Don't use struct assignment here, because the client reply - * buffer may be smaller/larger than the local struct - * obd_connect_data. */ - memcpy(tmpdata, data, min(tmpsize, size)); - } - - /* If 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. - * - * XXX this will go away when shaver stops sending the "connect" handle - * in the real "remote handle" field of the request --phik 24 Apr 2003 - */ + /* + * Return only the parts of obd_connect_data that we understand, so the + * client knows that we don't understand the rest. + */ + if (data) { + tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_SERVER); + tmpdata = req_capsule_server_get(&req->rq_pill, + &RMF_CONNECT_DATA); + /* + * Don't use struct assignment here, because the client reply + * buffer may be smaller/larger than the local struct + * obd_connect_data. + */ + memcpy(tmpdata, data, min(tmpsize, size)); + } + + /* + * If 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. + * + * XXX this will go away when shaver stops sending the "connect" handle + * in the real "remote handle" field of the request --phik 24 Apr 2003 + */ ptlrpc_request_change_export(req, export); spin_lock(&export->exp_lock); if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) { 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); + 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)); - 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, @@ -1328,12 +1442,14 @@ dont_check_exports: GOTO(out, rc); if (target->obd_recovering && !export->exp_in_recovery && !lw_client) { - int has_transno; - __u64 transno = data->ocd_transno; + int has_transno; + __u64 transno = data->ocd_transno; spin_lock(&export->exp_lock); - /* possible race with class_disconnect_stale_exports, - * export may be already in the eviction process */ + /* + * possible race with class_disconnect_stale_exports, + * export may be already in the eviction process + */ if (export->exp_failed) { spin_unlock(&export->exp_lock); GOTO(out, rc = -ENODEV); @@ -1343,39 +1459,42 @@ dont_check_exports: export->exp_lock_replay_needed = 1; spin_unlock(&export->exp_lock); - has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) & - MSG_CONNECT_TRANSNO); - if (has_transno && transno == 0) - CWARN("Connect with zero transno!\n"); + has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) & + MSG_CONNECT_TRANSNO); + if (has_transno && transno == 0) + CWARN("Connect with zero transno!\n"); - if (has_transno && transno > 0 && - transno < target->obd_next_recovery_transno && - transno > target->obd_last_committed) { + if (has_transno && transno > 0 && + transno < target->obd_next_recovery_transno && + transno > target->obd_last_committed) { /* Another way is to use cmpxchg() to be lock-free. */ spin_lock(&target->obd_recovery_task_lock); if (transno < target->obd_next_recovery_transno) target->obd_next_recovery_transno = transno; spin_unlock(&target->obd_recovery_task_lock); - } + } atomic_inc(&target->obd_req_replay_clients); atomic_inc(&target->obd_lock_replay_clients); - /* Note: MDS-MDS connection is allowed to be connected during + /* + * 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. */ + * condition. + */ if (new_mds_mds_conn) - target->obd_max_recoverable_clients++; + atomic_inc(&target->obd_max_recoverable_clients); + if (atomic_inc_return(&target->obd_connected_clients) == - target->obd_max_recoverable_clients) + atomic_read(&target->obd_max_recoverable_clients)) wake_up(&target->obd_next_transno_waitq); } - /* Tell the client we're in recovery, when client is involved in it. */ + /* Tell the client we're in recovery, when client is involved in it. */ if (target->obd_recovering && !lw_client) - lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); + lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); out: if (export) { @@ -1385,12 +1504,11 @@ out: class_export_put(export); } - if (target_referenced == true && target != NULL) { + if (target != NULL) { spin_lock(&target->obd_dev_lock); target->obd_conn_inprogress--; spin_unlock(&target->obd_dev_lock); - - class_decref(target, __func__, current); + class_decref(target, "find", current); } req->rq_status = rc; RETURN(rc); @@ -1398,24 +1516,27 @@ out: int target_handle_disconnect(struct ptlrpc_request *req) { - int rc; - ENTRY; + int rc; + + ENTRY; - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - RETURN(rc); + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); /* Keep the rq_export around so we can send the reply. */ - req->rq_status = obd_disconnect(class_export_get(req->rq_export)); + req->rq_status = obd_disconnect(class_export_get(req->rq_export)); - RETURN(0); + RETURN(0); } void target_destroy_export(struct obd_export *exp) { - struct obd_import *imp = NULL; - /* exports created from last_rcvd data, and "fake" - exports created by lctl don't have an import */ + struct obd_import *imp = NULL; + /* + * exports created from last_rcvd data, and "fake" + * exports created by lctl don't have an import + */ spin_lock(&exp->exp_lock); if (exp->exp_imp_reverse != NULL) { imp = exp->exp_imp_reverse; @@ -1459,30 +1580,30 @@ static void target_request_copy_put(struct ptlrpc_request *req) 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; + __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; + int dup = 0; - LASSERT(exp); + LASSERT(exp); spin_lock(&exp->exp_lock); list_for_each_entry(reqiter, &exp->exp_req_replay_queue, - rq_replay_list) { - if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { + rq_replay_list) { + if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { dup_req = reqiter; - dup = 1; - break; - } - } + dup = 1; + break; + } + } - if (dup) { + if (dup) { /* We expect it with RESENT and REPLAY flags. */ - if ((lustre_msg_get_flags(req->rq_reqmsg) & - (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_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; @@ -1493,13 +1614,13 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) lustre_msg_set_conn_cnt(dup_req->rq_reqmsg, new_conn); } - } else { + } else { list_add_tail(&req->rq_replay_list, - &exp->exp_req_replay_queue); - } + &exp->exp_req_replay_queue); + } spin_unlock(&exp->exp_lock); - return dup; + return dup; } static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) @@ -1515,55 +1636,53 @@ static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) static void target_finish_recovery(struct lu_target *lut) { struct obd_device *obd = lut->lut_obd; - ENTRY; + + 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 " - "%d recovered and %d %s evicted.\n", obd->obd_name, - (int)elapsed_time / 60, (int)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); + time64_t now = ktime_get_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, (s64)elapsed_time / 60, + (s64)elapsed_time % 60, + atomic_read(&obd->obd_max_recoverable_clients), + atomic_read(&obd->obd_connected_clients), + obd->obd_stale_clients, + obd->obd_stale_clients == 1 ? "was" : "were"); + } + + 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) || !list_empty(&obd->obd_final_req_queue)) { - CERROR("%s: Recovery queues ( %s%s%s) are not empty\n", - obd->obd_name, + CERROR("%s: Recovery queues ( %s%s%s) are not empty\n", + obd->obd_name, list_empty(&obd->obd_req_replay_queue) ? "" : "req ", list_empty(&obd->obd_lock_replay_queue) ? \ - "" : "lock ", + "" : "lock ", list_empty(&obd->obd_final_req_queue) ? \ - "" : "final "); + "" : "final "); spin_unlock(&obd->obd_recovery_task_lock); LBUG(); } spin_unlock(&obd->obd_recovery_task_lock); - if (lut->lut_tdtd != NULL && - (!list_empty(&lut->lut_tdtd->tdtd_replay_list) || - !list_empty(&lut->lut_tdtd->tdtd_replay_finish_list))) { - dtrq_list_dump(lut->lut_tdtd, D_ERROR); - dtrq_list_destroy(lut->lut_tdtd); - } - - obd->obd_recovery_end = cfs_time_current_sec(); + obd->obd_recovery_end = ktime_get_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); - } - EXIT; + 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; } static void abort_req_replay_queue(struct obd_device *obd) @@ -1576,15 +1695,15 @@ static void abort_req_replay_queue(struct obd_device *obd) list_splice_init(&obd->obd_req_replay_queue, &abort_list); spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &abort_list, rq_list) { - DEBUG_REQ(D_WARNING, req, "aborted:"); - req->rq_status = -ENOTCONN; - if (ptlrpc_error(req)) { - DEBUG_REQ(D_ERROR, req, - "failed abort_req_reply; skipping"); - } - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - } + DEBUG_REQ(D_WARNING, req, "aborted:"); + req->rq_status = -ENOTCONN; + if (ptlrpc_error(req)) { + DEBUG_REQ(D_ERROR, req, + "failed abort_req_reply; skipping"); + } + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + } } static void abort_lock_replay_queue(struct obd_device *obd) @@ -1597,28 +1716,28 @@ static void abort_lock_replay_queue(struct obd_device *obd) list_splice_init(&obd->obd_lock_replay_queue, &abort_list); spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &abort_list, rq_list) { - DEBUG_REQ(D_ERROR, req, "aborted:"); - req->rq_status = -ENOTCONN; - if (ptlrpc_error(req)) { - DEBUG_REQ(D_ERROR, req, - "failed abort_lock_reply; skipping"); - } - target_request_copy_put(req); - } -} - -/* Called from a cleanup function if the device is being cleaned up - forcefully. The exports should all have been disconnected already, - the only thing left to do is - - clear the recovery flags - - cancel the timer - - free queued requests and replies, but don't send replies - Because the obd_stopping flag is set, no new requests should be received. - -*/ + DEBUG_REQ(D_ERROR, req, "aborted:"); + req->rq_status = -ENOTCONN; + if (ptlrpc_error(req)) { + DEBUG_REQ(D_ERROR, req, + "failed abort_lock_reply; skipping"); + } + target_request_copy_put(req); + } +} + +/* + * Called from a cleanup function if the device is being cleaned up + * forcefully. The exports should all have been disconnected already, + * the only thing left to do is + * - clear the recovery flags + * - cancel the timer + * - free queued requests and replies, but don't send replies + * Because the obd_stopping flag is set, no new requests should be received. + */ void target_cleanup_recovery(struct obd_device *obd) { - struct ptlrpc_request *req, *n; + struct ptlrpc_request *req, *n; struct list_head clean_list; INIT_LIST_HEAD(&clean_list); @@ -1629,7 +1748,6 @@ 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); @@ -1650,28 +1768,29 @@ void target_cleanup_recovery(struct obd_device *obd) list_for_each_entry_safe(req, n, &clean_list, rq_list) { LASSERT(req->rq_reply_state == NULL); - target_request_copy_put(req); - } + target_request_copy_put(req); + } - EXIT; + EXIT; } EXPORT_SYMBOL(target_cleanup_recovery); /* obd_recovery_task_lock should be held */ void target_cancel_recovery_timer(struct obd_device *obd) { - CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name); - cfs_timer_disarm(&obd->obd_recovery_timer); + CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name); + hrtimer_cancel(&obd->obd_recovery_timer); } static void target_start_recovery_timer(struct obd_device *obd) { + ktime_t delay; + if (obd->obd_recovery_start != 0) return; spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery || - obd->obd_force_abort_recovery) { + if (!obd->obd_recovering || obd->obd_abort_recovery) { spin_unlock(&obd->obd_dev_lock); return; } @@ -1683,65 +1802,82 @@ 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(); + obd->obd_recovery_start = ktime_get_seconds(); + delay = ktime_set(obd->obd_recovery_start + + obd->obd_recovery_timeout, 0); + hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS); 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 %lu:%02lu, or until %d client%s reconnect%s\n", + obd->obd_name, + obd->obd_recovery_timeout / 60, + obd->obd_recovery_timeout % 60, + atomic_read(&obd->obd_max_recoverable_clients), + (atomic_read(&obd->obd_max_recoverable_clients) == 1) ? + "" : "s", + (atomic_read(&obd->obd_max_recoverable_clients) == 1) ? + "s" : ""); } /** * extend recovery window. * - * if @extend is true, extend recovery window to have @drt remaining at least; - * otherwise, make sure the recovery timeout value is not less than @drt. + * if @extend is true, extend recovery window to have @dr_timeout remaining + * at least; otherwise, make sure the recovery timeout value is not less + * than @dr_timeout. */ -static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) +static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout, + bool extend) { - cfs_time_t now; - cfs_time_t end; - cfs_duration_t left; - int to; + ktime_t left_ns; + time_t timeout; + time_t left; spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering || obd->obd_abort_recovery || - obd->obd_force_abort_recovery) { + obd->obd_stopping) { spin_unlock(&obd->obd_dev_lock); - 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); - - if (extend && (drt > left)) { - to += drt - left; - } else if (!extend && (drt > to)) { - to = drt; - } - - if (to > obd->obd_recovery_time_hard) - to = obd->obd_recovery_time_hard; - 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)); - } + return; + } + LASSERT(obd->obd_recovery_start != 0); + + left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer); + left = ktime_divns(left_ns, NSEC_PER_SEC); + + if (extend) { + timeout = obd->obd_recovery_timeout; + /* dr_timeout will happen after the hrtimer has expired. + * Add the excess time to the soft recovery timeout without + * exceeding the hard recovery timeout. + */ + if (dr_timeout > left) { + timeout += dr_timeout - left; + timeout = min_t(time_t, obd->obd_recovery_time_hard, + timeout); + } + } else { + timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout, + obd->obd_recovery_time_hard); + } + + if (timeout == obd->obd_recovery_time_hard) + CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n", + obd->obd_name, timeout, extend); + + if (obd->obd_recovery_timeout < timeout) { + ktime_t end, now; + + obd->obd_recovery_timeout = timeout; + end = ktime_set(obd->obd_recovery_start + timeout, 0); + now = ktime_set(ktime_get_seconds(), 0); + left_ns = ktime_sub(end, now); + hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS); + left = ktime_divns(left_ns, NSEC_PER_SEC); + } 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 %ld seconds\n", + obd->obd_name, left); } /* Reset the timer with each new client connection */ @@ -1754,33 +1890,42 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) * be extended to make sure the client could be reconnected, in the * process, the timeout from the new client should be ignored. */ - static void check_and_start_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, - int new_client) + struct ptlrpc_request *req, + int new_client) { - int service_time = lustre_msg_get_service_time(req->rq_reqmsg); - struct obd_device_target *obt = &obd->u.obt; + time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg); + struct obd_device_target *obt = &obd->u.obt; - if (!new_client && service_time) - /* Teach server about old server's estimates, as first guess - * at how long new requests will take. */ + if (!new_client && service_time) + /* + * Teach server about old server's estimates, as first guess + * at how long new requests will take. + */ at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate, - service_time); + service_time); - target_start_recovery_timer(obd); + target_start_recovery_timer(obd); - /* Convert the service time to RPC timeout, - * and reuse service_time to limit stack usage. */ + /* + * Convert the service time to RPC timeout, + * and reuse service_time to limit stack usage. + */ service_time = at_est2timeout(service_time); - /* We expect other clients to timeout within service_time, then try + 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. */ - service_time += 2 * INITIAL_CONNECT_TIMEOUT; + * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. + */ + service_time += 2 * INITIAL_CONNECT_TIMEOUT; - LASSERT(obt->obt_magic == OBT_MAGIC); + LASSERT(obt->obt_magic == OBT_MAGIC); service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC); if (service_time > obd->obd_recovery_timeout && !new_client) extend_recovery_timer(obd, service_time, false); @@ -1789,7 +1934,7 @@ check_and_start_recovery_timer(struct obd_device *obd, /** Health checking routines */ static inline int exp_connect_healthy(struct obd_export *exp) { - return (exp->exp_in_recovery); + return exp->exp_in_recovery; } /** if export done req_replay or has replay in queue */ @@ -1798,6 +1943,14 @@ 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) { @@ -1807,12 +1960,18 @@ static inline int exp_lock_replay_healthy(struct obd_export *exp) static inline int exp_vbr_healthy(struct obd_export *exp) { - return (!exp->exp_vbr_failed); + return !exp->exp_vbr_failed; } static inline int exp_finished(struct obd_export *exp) { - return (exp->exp_in_recovery && !exp->exp_lock_replay_needed); + return exp->exp_in_recovery && !exp->exp_lock_replay_needed; +} + +static inline int exp_finished_or_from_mdt(struct obd_export *exp) +{ + return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) || + exp_finished(exp); } static int check_for_next_transno(struct lu_target *lut) @@ -1824,6 +1983,7 @@ static int check_for_next_transno(struct lu_target *lut) __u64 req_transno = 0; __u64 update_transno = 0; __u64 next_transno = 0; + ENTRY; spin_lock(&obd->obd_recovery_task_lock); @@ -1841,12 +2001,13 @@ static int check_for_next_transno(struct lu_target *lut) 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", - obd->obd_max_recoverable_clients, connected, completed, + CDEBUG(D_HA, + "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n", + atomic_read(&obd->obd_max_recoverable_clients), + connected, completed, queue_len, req_transno, next_transno); - if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { + if (obd->obd_abort_recovery) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; } else if (obd->obd_recovery_expired) { @@ -1854,29 +2015,24 @@ static int check_for_next_transno(struct lu_target *lut) wake_up = 1; } else if (tdtd != NULL && req != NULL && is_req_replayed_by_update(req)) { - LASSERTF(req_transno < next_transno, "req_transno "LPU64 - "next_transno"LPU64"\n", req_transno, next_transno); - CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n", + 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 ("LPD64")\n", 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, - "%s: waking for gap in transno, VBR is %s (skip: " - LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64 - ", next_update "LPD64" last_committed: "LPD64")\n", + CDEBUG(D_HA, + "%s: waking for gap in transno, VBR is %s (skip: %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, update_transno, obd->obd_last_committed); @@ -1886,8 +2042,9 @@ static int check_for_next_transno(struct lu_target *lut) CDEBUG(D_HA, "waking for completed recovery\n"); 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); + CDEBUG(D_HA, + "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n", + next_transno); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } @@ -1907,7 +2064,7 @@ static int check_for_next_lock(struct lu_target *lut) } 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 || obd->obd_force_abort_recovery) { + } else if (obd->obd_abort_recovery) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; } else if (obd->obd_recovery_expired) { @@ -1928,31 +2085,111 @@ 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 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_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) { + spin_unlock(&obd->obd_recovery_task_lock); + wait_event_idle( + tdtd->tdtd_recovery_threads_waitq, + atomic_read(&tdtd->tdtd_recovery_threads_count) + == 0); + + spin_lock(&obd->obd_recovery_task_lock); + 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 & 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(lut), - msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0) - /* wait indefinitely for event, but don't trigger watchdog */; + cfs_time_seconds(60)) == 0) + ; /* wait indefinitely for event, but don't trigger watchdog */ - if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { + if (obd->obd_abort_recovery) { CWARN("recovery is aborted, evict exports in recovery\n"); + if (lut->lut_tdtd != NULL) { + tdtd = lut->lut_tdtd; + /* + * Let's wait all of the update log recovery thread + * finished + */ + wait_event_idle( + tdtd->tdtd_recovery_threads_waitq, + atomic_read(&tdtd->tdtd_recovery_threads_count) + == 0); + /* 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; - } else if (obd->obd_recovery_expired) { + } else if (obd->obd_recovery_expired && + obd->obd_recovery_timeout < obd->obd_recovery_time_hard) { obd->obd_recovery_expired = 0; + /** If some clients died being recovered, evict them */ - LCONSOLE_WARN("%s: recovery is timed out, " - "evict stale exports\n", obd->obd_name); + LCONSOLE_WARN("%s: recovery is timed out, 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; @@ -1962,7 +2199,9 @@ repeat: * timeout is set just to handle reconnection delays */ extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); - /** Wait for recovery events again, after evicting bad clients */ + /** + * Wait for recovery events again, after evicting bad clients + */ goto repeat; } return 0; @@ -1970,7 +2209,7 @@ repeat: static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut) { - struct obd_device *obd = lut->lut_obd; + struct obd_device *obd = lut->lut_obd; struct ptlrpc_request *req = NULL; CDEBUG(D_HA, "Waiting for lock\n"); @@ -2032,39 +2271,45 @@ static void handle_recovery_req(struct ptlrpc_thread *thread, req->rq_svc_thread = thread; req->rq_svc_thread->t_env->le_ses = &req->rq_session; - /* thread context */ - lu_context_enter(&thread->t_env->le_ctx); - (void)handler(req); - lu_context_exit(&thread->t_env->le_ctx); + /* thread context */ + lu_context_enter(&thread->t_env->le_ctx); + (void)handler(req); + lu_context_exit(&thread->t_env->le_ctx); + + req->rq_svc_thread->t_env->le_ses = NULL; - /* don't reset timer for final stage */ - if (!exp_finished(req->rq_export)) { - int to = obd_timeout; + /* don't reset timer for final stage */ + if (!exp_finished(req->rq_export)) { + time_t to = obd_timeout; - /** - * Add request timeout to the recovery time so next request from - * this client may come in recovery time - */ - if (!AT_OFF) { + /** + * Add request timeout to the recovery time so next request from + * this client may come in recovery time + */ + if (!AT_OFF) { struct ptlrpc_service_part *svcpt; svcpt = req->rq_rqbd->rqbd_svcpt; - /* If the server sent early reply for this request, + /* + * If the server sent early reply for this request, * the client will recalculate the timeout according to * current server estimate service time, so we will * use the maxium timeout here for waiting the client - * sending the next req */ - to = max((int)at_est2timeout( - at_get(&svcpt->scp_at_estimate)), - (int)lustre_msg_get_timeout(req->rq_reqmsg)); - /* Add 2 net_latency, one for balance rq_deadline + * sending the next req + */ + to = max_t(time_t, + at_est2timeout(at_get(&svcpt->scp_at_estimate)), + lustre_msg_get_timeout(req->rq_reqmsg)); + /* + * Add 2 net_latency, one for balance rq_deadline * (see ptl_send_rpc), one for resend the req to server, * Note: client will pack net_latency in replay req - * (see ptlrpc_replay_req) */ + * (see ptlrpc_replay_req) + */ to += 2 * lustre_msg_get_service_time(req->rq_reqmsg); - } - extend_recovery_timer(class_exp2obd(req->rq_export), to, true); - } + } + extend_recovery_timer(class_exp2obd(req->rq_export), to, true); + } EXIT; } @@ -2074,30 +2319,33 @@ 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; + CDEBUG(D_HA, + "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n", + clnts, obd->obd_stale_clients, + atomic_read(&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); + LASSERT(clnts <= + atomic_read(&obd->obd_max_recoverable_clients)); if (clnts + obd->obd_stale_clients < - obd->obd_max_recoverable_clients) + atomic_read(&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 */ + if (!lut->lut_tdtd->tdtd_replay_ready && + !obd->obd_abort_recovery && !obd->obd_stopping) { + /* + * 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 %lu\n", + obd->obd_name, obd->obd_recovery_timeout); return 0; - } else { - dtrq_list_dump(lut->lut_tdtd, D_HA); } } @@ -2123,12 +2371,14 @@ static __u64 get_next_replay_req_transno(struct obd_device *obd) return transno; } -__u64 get_next_transno(struct lu_target *lut, int *type) + +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); @@ -2164,13 +2414,15 @@ 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", + DEBUG_REQ(D_HA, req, + "remove t%lld from %s because duplicate update records found", 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 */ + /* + * 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); @@ -2179,7 +2431,7 @@ static void drop_duplicate_replay_req(struct lu_env *env, 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", + DEBUG_REQ(D_ERROR, req, "wrong opc from %s", libcfs_nid2str(req->rq_peer.nid)); } target_exp_dequeue_req_replay(req); @@ -2194,11 +2446,12 @@ static void replay_request_or_update(struct lu_env *env, { struct obd_device *obd = lut->lut_obd; struct ptlrpc_request *req = NULL; - int type; - __u64 transno; + int type; + __u64 transno; + ENTRY; - CDEBUG(D_HA, "Waiting for transno "LPD64"\n", + CDEBUG(D_HA, "Waiting for transno %lld\n", obd->obd_next_recovery_transno); /* Replay all of request and update by transno */ @@ -2207,7 +2460,8 @@ static void replay_request_or_update(struct lu_env *env, CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val); - /** It is needed to extend recovery window above + /** + * 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()). @@ -2215,17 +2469,20 @@ static void replay_request_or_update(struct lu_env *env, CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300); if (target_recovery_overseer(lut, check_for_next_transno, - exp_req_replay_healthy)) { + 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 + /* + * Drop replay request from client side, if the * replay has been executed by update with the - * same transno */ + * same transno + */ req = list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); @@ -2233,13 +2490,15 @@ static void replay_request_or_update(struct lu_env *env, 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 */ + /* + * 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); + transno); LASSERT(dtrq != NULL); spin_lock(&tdtd->tdtd_replay_list_lock); list_del_init(&dtrq->dtrq_list); @@ -2252,7 +2511,8 @@ static void replay_request_or_update(struct lu_env *env, } LASSERT(trd->trd_processing_task == current_pid()); - DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", + DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s", + req->rq_xid, lustre_msg_get_transno(req->rq_reqmsg), libcfs_nid2str(req->rq_peer.nid)); @@ -2270,32 +2530,29 @@ static void replay_request_or_update(struct lu_env *env, obd->obd_replayed_requests++; } else if (type == UPDATE_RECOVERY && transno != 0) { struct distribute_txn_replay_req *dtrq; - bool update_transno = false; + 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); - tdtd->tdtd_replay_handler(env, tdtd, dtrq); + rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq); lu_context_exit(&thread->t_env->le_ctx); extend_recovery_timer(obd, obd_timeout, true); - /* Add it to the replay finish list */ - spin_lock(&tdtd->tdtd_replay_list_lock); - if (dtrq->dtrq_xid != 0) { - CDEBUG(D_HA, "Move x"LPU64" t"LPU64 - " to finish list\n", dtrq->dtrq_xid, + 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); - update_transno = true; - } else { - dtrq_destroy(dtrq); - } - spin_unlock(&tdtd->tdtd_replay_list_lock); + spin_unlock(&tdtd->tdtd_replay_list_lock); - if (update_transno) { spin_lock(&obd->obd_recovery_task_lock); if (transno == obd->obd_next_recovery_transno) obd->obd_next_recovery_transno++; @@ -2304,9 +2561,12 @@ static void replay_request_or_update(struct lu_env *env, 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 */ @@ -2318,39 +2578,37 @@ static void replay_request_or_update(struct lu_env *env, static int target_recovery_thread(void *arg) { - struct lu_target *lut = arg; - struct obd_device *obd = lut->lut_obd; - struct ptlrpc_request *req; - struct target_recovery_data *trd = &obd->obd_recovery_data; - unsigned long delta; - struct lu_env *env; - struct ptlrpc_thread *thread = NULL; - int rc = 0; - ENTRY; + struct lu_target *lut = arg; + struct obd_device *obd = lut->lut_obd; + struct ptlrpc_request *req; + struct target_recovery_data *trd = &obd->obd_recovery_data; + unsigned long delta; + struct lu_env *env; + struct ptlrpc_thread *thread = NULL; + int rc = 0; + + ENTRY; unshare_fs_struct(); - OBD_ALLOC_PTR(thread); - if (thread == NULL) - RETURN(-ENOMEM); - - OBD_ALLOC_PTR(env); - if (env == NULL) { - OBD_FREE_PTR(thread); - RETURN(-ENOMEM); - } - - rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD); - if (rc) { - OBD_FREE_PTR(thread); - OBD_FREE_PTR(env); - RETURN(rc); - } - - thread->t_env = env; - thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */ - env->le_ctx.lc_thread = thread; + OBD_ALLOC_PTR(thread); + if (thread == NULL) + RETURN(-ENOMEM); + + OBD_ALLOC_PTR(env); + if (env == NULL) + GOTO(out_thread, rc = -ENOMEM); + rc = lu_env_add(env); + if (rc) + GOTO(out_env, rc); + + rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD); + if (rc) + GOTO(out_env_remove, rc); + + thread->t_env = env; + thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */ + env->le_ctx.lc_thread = thread; tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */ - thread->t_watchdog = NULL; CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, current_pid()); @@ -2372,7 +2630,7 @@ static int target_recovery_thread(void *arg) /* 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); replay_request_or_update(env, lut, trd, thread); @@ -2384,7 +2642,7 @@ static int target_recovery_thread(void *arg) atomic_read(&obd->obd_lock_replay_clients)); while ((req = target_next_replay_lock(lut))) { LASSERT(trd->trd_processing_task == current_pid()); - DEBUG_REQ(D_HA, req, "processing lock from %s: ", + DEBUG_REQ(D_HA, req, "processing lock from %s:", libcfs_nid2str(req->rq_peer.nid)); handle_recovery_req(thread, req, trd->trd_recovery_handler); @@ -2392,16 +2650,18 @@ static int target_recovery_thread(void *arg) obd->obd_replayed_locks++; } - /** - * The third stage: reply on final pings, at this moment all clients - * must have request in final queue - */ + /** + * 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); - /* We drop recoverying flag to forward all new requests - * to regular mds_handle() since now */ + CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); + /** Update server last boot epoch */ + tgt_boot_epoch_update(lut); + /* + * We drop recoverying flag to forward all new requests + * to regular mds_handle() since now + */ spin_lock(&obd->obd_dev_lock); obd->obd_recovering = obd->obd_abort_recovery = 0; spin_unlock(&obd->obd_dev_lock); @@ -2410,39 +2670,43 @@ static int target_recovery_thread(void *arg) spin_unlock(&obd->obd_recovery_task_lock); while ((req = target_next_final_ping(obd))) { LASSERT(trd->trd_processing_task == current_pid()); - DEBUG_REQ(D_HA, req, "processing final ping from %s: ", + DEBUG_REQ(D_HA, req, "processing final ping from %s:", libcfs_nid2str(req->rq_peer.nid)); - handle_recovery_req(thread, req, - trd->trd_recovery_handler); - /* Because the waiting client can not send ping to server, + handle_recovery_req(thread, req, + trd->trd_recovery_handler); + /* + * Because the waiting client can not send ping to server, * so we need refresh the last_request_time, to avoid the - * export is being evicted */ + * export is being evicted + */ ptlrpc_update_export_timer(req->rq_export, 0); target_request_copy_put(req); } 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); + 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) { CWARN("too long recovery - read logs\n"); libcfs_debug_dumplog(); } target_finish_recovery(lut); - - lu_context_fini(&env->le_ctx); - trd->trd_processing_task = 0; - complete(&trd->trd_finishing); - + lu_context_fini(&env->le_ctx); + trd->trd_processing_task = 0; + complete_all(&trd->trd_finishing); tgt_io_thread_done(thread); - OBD_FREE_PTR(thread); +out_env_remove: + lu_env_remove(env); +out_env: OBD_FREE_PTR(env); +out_thread: + OBD_FREE_PTR(thread); RETURN(rc); } static int target_start_recovery_thread(struct lu_target *lut, - svc_handler_t handler) + svc_handler_t handler) { struct obd_device *obd = lut->lut_obd; int rc = 0; @@ -2488,52 +2752,62 @@ EXPORT_SYMBOL(target_stop_recovery_thread); void target_recovery_fini(struct obd_device *obd) { - class_disconnect_exports(obd); - target_stop_recovery_thread(obd); - target_cleanup_recovery(obd); + class_disconnect_exports(obd); + target_stop_recovery_thread(obd); + target_cleanup_recovery(obd); } EXPORT_SYMBOL(target_recovery_fini); -static void target_recovery_expired(unsigned long castmeharder) +static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer) { - 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", + struct obd_device *obd = container_of(timer, struct obd_device, + obd_recovery_timer); + + CDEBUG(D_HA, + "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n", obd->obd_name, atomic_read(&obd->obd_lock_replay_clients), - cfs_time_current_sec()- obd->obd_recovery_start, + ktime_get_real_seconds() - obd->obd_recovery_start, atomic_read(&obd->obd_connected_clients)); obd->obd_recovery_expired = 1; wake_up(&obd->obd_next_transno_waitq); + return HRTIMER_NORESTART; } void target_recovery_init(struct lu_target *lut, svc_handler_t handler) { - struct obd_device *obd = lut->lut_obd; + 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; - } + if (lut->lut_bottom->dd_rdonly) + return; - CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, " - "last_transno "LPU64"\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; + if (atomic_read(&obd->obd_max_recoverable_clients) == 0) { + /** Update server last boot epoch */ + tgt_boot_epoch_update(lut); + return; + } - cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd); - target_start_recovery_thread(lut, handler); + CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, " + "last_transno %llu\n", obd->obd_name, + atomic_read(&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; + + hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC, + HRTIMER_MODE_ABS); + obd->obd_recovery_timer.function = &target_recovery_expired; + 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) + struct ptlrpc_request *req) { struct obd_export *exp = req->rq_export; + LASSERT(exp != NULL); if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { /* client declares he's ready to replay locks */ @@ -2549,8 +2823,10 @@ static int target_process_req_flags(struct obd_device *obd, } } if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) { - /* client declares he's ready to complete recovery - * so, we put the request on th final queue */ + /* + * client declares he's ready to complete recovery + * so, we put the request on th final queue + */ spin_lock(&exp->exp_lock); if (exp->exp_lock_replay_needed) { exp->exp_lock_replay_needed = 0; @@ -2566,11 +2842,12 @@ static int target_process_req_flags(struct obd_device *obd, } int target_queue_recovery_request(struct ptlrpc_request *req, - struct obd_device *obd) + struct obd_device *obd) { - __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); + __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); struct ptlrpc_request *reqiter; int inserted = 0; + ENTRY; if (obd->obd_recovery_data.trd_processing_task == current_pid()) { @@ -2578,25 +2855,71 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(1); } - target_process_req_flags(obd, 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)); + } + } - if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) { - /* client declares he's ready to complete recovery - * so, we put the request on th final queue */ + /* + * client declares he's ready to complete recovery + * so, we put the request on th final queue + */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue final 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"); + 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 */ @@ -2617,33 +2940,38 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(0); } - /* CAVEAT EMPTOR: The incoming request message has been swabbed - * (i.e. buflens etc are in my own byte order), but type-dependent - * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */ + /* + * CAVEAT EMPTOR: The incoming request message has been swabbed + * (i.e. buflens etc are in my own byte order), but type-dependent + * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. + */ - if (!transno) { + if (!transno) { INIT_LIST_HEAD(&req->rq_list); - DEBUG_REQ(D_HA, req, "not queueing"); - RETURN(1); - } - - /* If we're processing the queue, we want don't want to queue this - * message. - * - * Also, if this request has a transno less than the one we're waiting - * for, we should process it now. It could (and currently always will) - * be an open request for a descriptor that was opened some time ago. - * - * 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", - obd->obd_next_recovery_transno, transno); - - /* If the request has been replayed by update replay, then sends this + DEBUG_REQ(D_HA, req, "not queueing"); + RETURN(1); + } + + /* + * If we're processing the queue, we want don't want to queue this + * message. + * + * Also, if this request has a transno less than the one we're waiting + * for, we should process it now. It could (and currently always will) + * be an open request for a descriptor that was opened some time ago. + * + * Also, a resent, replayed request that has already been + * handled will pass through here and be processed immediately. + */ + 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 */ + * it will be handled + */ spin_lock(&obd->obd_recovery_task_lock); if (transno < obd->obd_next_recovery_transno && !is_req_replayed_by_update(req)) { @@ -2654,21 +2982,21 @@ int target_queue_recovery_request(struct ptlrpc_request *req, } spin_unlock(&obd->obd_recovery_task_lock); - if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP)) - RETURN(0); + if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP)) + RETURN(0); - target_request_copy_get(req); - if (!req->rq_export->exp_in_recovery) { - target_request_copy_put(req); - RETURN(-ENOTCONN); - } - LASSERT(req->rq_export->exp_req_replay_needed); + target_request_copy_get(req); + if (!req->rq_export->exp_in_recovery) { + target_request_copy_put(req); + RETURN(-ENOTCONN); + } + LASSERT(req->rq_export->exp_req_replay_needed); - if (target_exp_enqueue_req_replay(req)) { - DEBUG_REQ(D_ERROR, req, "dropping resent queued req"); - target_request_copy_put(req); - RETURN(0); - } + if (target_exp_enqueue_req_replay(req)) { + DEBUG_REQ(D_ERROR, req, "dropping resent queued req"); + target_request_copy_put(req); + RETURN(0); + } /* XXX O(n^2) */ spin_lock(&obd->obd_recovery_task_lock); @@ -2680,46 +3008,41 @@ int target_queue_recovery_request(struct ptlrpc_request *req, goto added; } - if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) == - transno)) { - DEBUG_REQ(D_ERROR, req, "dropping replay: transno " - "has been claimed by another client"); + if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) == + transno)) { + DEBUG_REQ(D_ERROR, req, + "dropping replay: transno has been claimed by another client"); spin_unlock(&obd->obd_recovery_task_lock); - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - RETURN(0); - } - } + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + RETURN(0); + } + } added: - if (!inserted) + if (!inserted) list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); - obd->obd_requests_queued_for_recovery++; + obd->obd_requests_queued_for_recovery++; spin_unlock(&obd->obd_recovery_task_lock); wake_up(&obd->obd_next_transno_waitq); RETURN(0); } -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); -} - void target_committed_to_req(struct ptlrpc_request *req) { - struct obd_export *exp = req->rq_export; + struct obd_export *exp = req->rq_export; - if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) - lustre_msg_set_last_committed(req->rq_repmsg, - exp->exp_last_committed); - else - DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/" - "%d)", exp->exp_obd->obd_no_transno, - req->rq_repmsg == NULL); + if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) + lustre_msg_set_last_committed(req->rq_repmsg, + exp->exp_last_committed); + else + DEBUG_REQ(D_IOCTL, req, + "not sending last_committed update (%d/%d)", + exp->exp_obd->obd_no_transno, + req->rq_repmsg == NULL); - CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", - exp->exp_last_committed, req->rq_transno, req->rq_xid); + CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n", + exp->exp_last_committed, req->rq_transno, req->rq_xid); } #endif /* HAVE_SERVER_SUPPORT */ @@ -2729,27 +3052,30 @@ void target_committed_to_req(struct ptlrpc_request *req) */ int target_pack_pool_reply(struct ptlrpc_request *req) { - struct obd_device *obd; - ENTRY; + struct obd_device *obd; - /* Check that we still have all structures alive as this may - * be some late RPC at shutdown time. */ - if (unlikely(!req->rq_export || !req->rq_export->exp_obd || - !exp_connect_lru_resize(req->rq_export))) { - lustre_msg_set_slv(req->rq_repmsg, 0); - lustre_msg_set_limit(req->rq_repmsg, 0); - RETURN(0); - } + ENTRY; + + /* + * Check that we still have all structures alive as this may + * be some late RPC at shutdown time. + */ + if (unlikely(!req->rq_export || !req->rq_export->exp_obd || + !exp_connect_lru_resize(req->rq_export))) { + lustre_msg_set_slv(req->rq_repmsg, 0); + lustre_msg_set_limit(req->rq_repmsg, 0); + RETURN(0); + } /* OBD is alive here as export is alive, which we checked above. */ - obd = req->rq_export->exp_obd; + obd = req->rq_export->exp_obd; read_lock(&obd->obd_pool_lock); - lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv); - lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit); + lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv); + lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit); read_unlock(&obd->obd_pool_lock); - RETURN(0); + RETURN(0); } static int target_send_reply_msg(struct ptlrpc_request *req, @@ -2759,8 +3085,10 @@ static int target_send_reply_msg(struct ptlrpc_request *req, 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 */ + /* + * 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))) @@ -2770,9 +3098,8 @@ static int target_send_reply_msg(struct ptlrpc_request *req, 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"); } + DEBUG_REQ(D_NET, req, "sending reply"); return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT); } @@ -2780,31 +3107,32 @@ static int target_send_reply_msg(struct ptlrpc_request *req, void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) { struct ptlrpc_service_part *svcpt; - int netrc; - struct ptlrpc_reply_state *rs; - struct obd_export *exp; - ENTRY; + int netrc; + struct ptlrpc_reply_state *rs; + struct obd_export *exp; - if (req->rq_no_reply) { - EXIT; - return; - } + ENTRY; + + if (req->rq_no_reply) { + EXIT; + return; + } svcpt = req->rq_rqbd->rqbd_svcpt; - rs = req->rq_reply_state; - if (rs == NULL || !rs->rs_difficult) { - /* no notifiers */ - target_send_reply_msg (req, rc, fail_id); - EXIT; - return; - } - - /* must be an export if locks saved */ + rs = req->rq_reply_state; + if (rs == NULL || !rs->rs_difficult) { + /* no notifiers */ + target_send_reply_msg(req, rc, fail_id); + EXIT; + return; + } + + /* must be an export if locks saved */ LASSERT(req->rq_export != NULL); - /* req/reply consistent */ + /* req/reply consistent */ LASSERT(rs->rs_svcpt == svcpt); - /* "fresh" reply */ + /* "fresh" reply */ LASSERT(!rs->rs_scheduled); LASSERT(!rs->rs_scheduled_ever); LASSERT(!rs->rs_handled); @@ -2815,16 +3143,16 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) exp = class_export_get(req->rq_export); - /* disable reply scheduling while I'm setting up */ - rs->rs_scheduled = 1; - rs->rs_on_net = 1; - rs->rs_xid = req->rq_xid; - rs->rs_transno = req->rq_transno; - rs->rs_export = exp; - rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg); + /* disable reply scheduling while I'm setting up */ + rs->rs_scheduled = 1; + rs->rs_on_net = 1; + rs->rs_xid = req->rq_xid; + rs->rs_transno = req->rq_transno; + rs->rs_export = exp; + 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 */ @@ -2844,11 +3172,13 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) atomic_inc(&svcpt->scp_nreps_difficult); if (netrc != 0) { - /* error sending: reply is off the net. Also we need +1 + /* + * error sending: reply is off the net. Also we need +1 * reply ref until ptlrpc_handle_rs() is done * with the reply state (if the send was successful, there * would have been +1 ref for the net, which - * reply_out_callback leaves alone) */ + * reply_out_callback leaves alone) + */ rs->rs_on_net = 0; ptlrpc_rs_addref(rs); } @@ -2891,26 +3221,26 @@ int ldlm_error2errno(enum ldlm_error error) switch (error) { case ELDLM_OK: case ELDLM_LOCK_MATCHED: - result = 0; - break; - case ELDLM_LOCK_CHANGED: - result = -ESTALE; - break; - case ELDLM_LOCK_ABORTED: - result = -ENAVAIL; - break; - case ELDLM_LOCK_REPLACED: - result = -ESRCH; - break; - case ELDLM_NO_LOCK_DATA: - result = -ENOENT; - break; - case ELDLM_NAMESPACE_EXISTS: - result = -EEXIST; - break; - case ELDLM_BAD_NAMESPACE: - result = -EBADF; - break; + result = 0; + break; + case ELDLM_LOCK_CHANGED: + result = -ESTALE; + break; + case ELDLM_LOCK_ABORTED: + result = -ENAVAIL; + break; + case ELDLM_LOCK_REPLACED: + result = -ESRCH; + break; + case ELDLM_NO_LOCK_DATA: + result = -ENOENT; + break; + case ELDLM_NAMESPACE_EXISTS: + result = -EEXIST; + break; + case ELDLM_BAD_NAMESPACE: + result = -EBADF; + break; default: if (((int)error) < 0) { /* cast to signed type */ result = error; /* as ldlm_error can be unsigned */ @@ -2928,34 +3258,34 @@ EXPORT_SYMBOL(ldlm_error2errno); */ enum ldlm_error ldlm_errno2error(int err_no) { - int error; - - switch (err_no) { - case 0: - error = ELDLM_OK; - break; - case -ESTALE: - error = ELDLM_LOCK_CHANGED; - break; - case -ENAVAIL: - error = ELDLM_LOCK_ABORTED; - break; - case -ESRCH: - error = ELDLM_LOCK_REPLACED; - break; - case -ENOENT: - error = ELDLM_NO_LOCK_DATA; - break; - case -EEXIST: - error = ELDLM_NAMESPACE_EXISTS; - break; - case -EBADF: - error = ELDLM_BAD_NAMESPACE; - break; - default: - error = err_no; - } - return error; + int error; + + switch (err_no) { + case 0: + error = ELDLM_OK; + break; + case -ESTALE: + error = ELDLM_LOCK_CHANGED; + break; + case -ENAVAIL: + error = ELDLM_LOCK_ABORTED; + break; + case -ESRCH: + error = ELDLM_LOCK_REPLACED; + break; + case -ENOENT: + error = ELDLM_NO_LOCK_DATA; + break; + case -EEXIST: + error = ELDLM_NAMESPACE_EXISTS; + break; + case -EBADF: + error = ELDLM_BAD_NAMESPACE; + break; + default: + error = err_no; + } + return error; } #if LUSTRE_TRACKS_LOCK_EXP_REFS @@ -2965,8 +3295,8 @@ void ldlm_dump_export_locks(struct obd_export *exp) if (!list_empty(&exp->exp_locks_list)) { struct ldlm_lock *lock; - CERROR("dumping locks for export %p," - "ignore if the unmount doesn't hang\n", exp); + CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n", + exp); list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link) LDLM_ERROR(lock, "lock:"); @@ -2976,15 +3306,6 @@ void ldlm_dump_export_locks(struct obd_export *exp) #endif #ifdef HAVE_SERVER_SUPPORT -static int target_bulk_timeout(void *data) -{ - ENTRY; - /* We don't fail the connection here, because having the export - * killed makes the (vital) call to commitrw very sad. - */ - RETURN(1); -} - static inline const char *bulk2type(struct ptlrpc_request *req) { if (req->rq_bulk_read) @@ -2994,24 +3315,19 @@ static inline const char *bulk2type(struct ptlrpc_request *req) return "UNKNOWN"; } -int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, - struct l_wait_info *lwi) +int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc) { - 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_seconds(); + time64_t deadline; + int rc = 0; ENTRY; /* If there is eviction in progress, wait for it to finish. */ - if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) { - *lwi = LWI_INTR(NULL, NULL); - rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, - !atomic_read(&exp->exp_obd-> - obd_evict_inprogress), - lwi); - } + wait_event_idle( + exp->exp_obd->obd_evict_inprogress_waitq, + !atomic_read(&exp->exp_obd->obd_evict_inprogress)); /* Check if client was evicted or reconnected already. */ if (exp->exp_failed || @@ -3021,8 +3337,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, if (req->rq_bulk_read) rc = sptlrpc_svc_wrap_bulk(req, desc); - if ((exp->exp_connect_data.ocd_connect_flags & - OBD_CONNECT_BULK_MBITS) != 0) + 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; @@ -3032,7 +3347,7 @@ 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", + DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d", bulk2type(req), rc); RETURN(rc); } @@ -3048,32 +3363,32 @@ 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; - - *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), - target_bulk_timeout, desc); - rc = l_wait_event(desc->bd_waitq, - !ptlrpc_server_bulk_active(desc) || - exp->exp_failed || - exp->exp_conn_cnt > - lustre_msg_get_conn_cnt(req->rq_reqmsg), - lwi); - LASSERT(rc == 0 || rc == -ETIMEDOUT); + time64_t timeoutl = deadline - ktime_get_seconds(); + time64_t rq_deadline; + + while (timeoutl >= 0 && + wait_event_idle_timeout( + desc->bd_waitq, + !ptlrpc_server_bulk_active(desc) || + exp->exp_failed || + exp->exp_conn_cnt > + lustre_msg_get_conn_cnt(req->rq_reqmsg), + timeoutl ? cfs_time_seconds(1) : 1) == 0) + timeoutl -= 1; + rc = timeoutl < 0 ? -ETIMEDOUT : 0; + /* Wait again if we changed rq_deadline. */ - rq_deadline = ACCESS_ONCE(req->rq_deadline); + rq_deadline = READ_ONCE(req->rq_deadline); 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_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",