X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=1d4d6b78ff5c85509585fdfa0d9a9d75285324f8;hp=887979782554a86742bcf919e7db3b11b2771ce1;hb=892078e3b566c04471e7dcf2c28e66f2f3584f93;hpb=1b2547843817b4b7adbeb87ea9b070d9cac35c90 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 8879797..1d4d6b7 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -27,7 +27,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2010, 2013, Intel Corporation. + * Copyright (c) 2010, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -43,11 +43,8 @@ #define DEBUG_SUBSYSTEM S_LDLM -#ifdef __KERNEL__ -# include -#else -# include -#endif +#include +#include #include #include #include @@ -85,11 +82,11 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, } spin_lock(&imp->imp_lock); - cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) { + list_for_each_entry(item, &imp->imp_conn_list, oic_item) { if (obd_uuid_equals(uuid, &item->oic_uuid)) { if (priority) { - cfs_list_del(&item->oic_item); - cfs_list_add(&item->oic_item, + list_del(&item->oic_item); + list_add(&item->oic_item, &imp->imp_conn_list); item->oic_last_attempt = 0; } @@ -106,9 +103,9 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, imp_conn->oic_uuid = *uuid; imp_conn->oic_last_attempt = 0; if (priority) - cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list); + list_add(&imp_conn->oic_item, &imp->imp_conn_list); else - cfs_list_add_tail(&imp_conn->oic_item, + 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, @@ -148,12 +145,12 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) ENTRY; spin_lock(&imp->imp_lock); - if (cfs_list_empty(&imp->imp_conn_list)) { + if (list_empty(&imp->imp_conn_list)) { LASSERT(!imp->imp_connection); GOTO(out, rc); } - cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) { + 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); @@ -179,7 +176,7 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) } } - cfs_list_del(&imp_conn->oic_item); + 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", @@ -207,7 +204,7 @@ int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, ENTRY; spin_lock(&imp->imp_lock); - cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { + 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; @@ -263,41 +260,54 @@ static int osc_on_mdt(char *obdname) */ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) { - struct client_obd *cli = &obddev->u.cli; - struct obd_import *imp; - struct obd_uuid server_uuid; - int rq_portal, rp_portal, connect_op; - char *name = obddev->obd_type->typ_name; - ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN; - int rc; - char *cli_name = lustre_cfg_buf(lcfg, 0); - ENTRY; + struct client_obd *cli = &obddev->u.cli; + struct obd_import *imp; + struct obd_uuid server_uuid; + int rq_portal, rp_portal, connect_op; + char *name = obddev->obd_type->typ_name; + enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN; + char *cli_name = lustre_cfg_buf(lcfg, 0); + 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. */ - if (!strcmp(name, LUSTRE_OSC_NAME) || - (!(strcmp(name, LUSTRE_OSP_NAME)) && - (is_osp_on_mdt(cli_name) && - strstr(lustre_cfg_buf(lcfg, 1), "OST") != NULL))) { - /* OSC or OSP_on_MDT for OSTs */ - rq_portal = OST_REQUEST_PORTAL; - rp_portal = OSC_REPLY_PORTAL; - connect_op = OST_CONNECT; - cli->cl_sp_me = LUSTRE_SP_CLI; - cli->cl_sp_to = LUSTRE_SP_OST; - ns_type = LDLM_NS_TYPE_OSC; + /* 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; + connect_op = OST_CONNECT; + cli->cl_sp_me = LUSTRE_SP_CLI; + cli->cl_sp_to = LUSTRE_SP_OST; + ns_type = LDLM_NS_TYPE_OSC; } else if (!strcmp(name, LUSTRE_MDC_NAME) || - !strcmp(name, LUSTRE_LWP_NAME) || - (!strcmp(name, LUSTRE_OSP_NAME) && - (is_osp_on_mdt(cli_name) && - strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL))) { - /* MDC or OSP_on_MDT for other MDTs */ - rq_portal = MDS_REQUEST_PORTAL; - rp_portal = MDC_REPLY_PORTAL; - connect_op = MDS_CONNECT; - cli->cl_sp_me = LUSTRE_SP_CLI; - cli->cl_sp_to = LUSTRE_SP_MDT; - ns_type = LDLM_NS_TYPE_MDC; + !strcmp(name, LUSTRE_LWP_NAME)) { + rq_portal = MDS_REQUEST_PORTAL; + rp_portal = MDC_REPLY_PORTAL; + connect_op = MDS_CONNECT; + if (is_lwp_on_ost(cli_name)) + cli->cl_sp_me = LUSTRE_SP_OST; + else if (is_lwp_on_mdt(cli_name)) + cli->cl_sp_me = LUSTRE_SP_MDT; + else + cli->cl_sp_me = LUSTRE_SP_CLI; + cli->cl_sp_to = LUSTRE_SP_MDT; + ns_type = LDLM_NS_TYPE_MDC; + } else if (!strcmp(name, LUSTRE_OSP_NAME)) { + if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) { + /* OSP_on_MDT for other MDTs */ + connect_op = MDS_CONNECT; + cli->cl_sp_to = LUSTRE_SP_MDT; + ns_type = LDLM_NS_TYPE_MDC; + rq_portal = OUT_PORTAL; + } else { + /* OSP on MDT for OST */ + connect_op = OST_CONNECT; + cli->cl_sp_to = LUSTRE_SP_OST; + ns_type = LDLM_NS_TYPE_OSC; + rq_portal = OST_REQUEST_PORTAL; + } + 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; @@ -333,26 +343,26 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) } init_rwsem(&cli->cl_sem); - sema_init(&cli->cl_mgc_sem, 1); + 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_dirty = 0; - cli->cl_avail_grant = 0; - /* FIXME: Should limit this for the sum of all cl_dirty_max. */ - cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024; - if (cli->cl_dirty_max >> PAGE_CACHE_SHIFT > num_physpages / 8) - cli->cl_dirty_max = num_physpages << (PAGE_CACHE_SHIFT - 3); - CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters); - CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list); - CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list); - CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list); - CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list); - client_obd_list_lock_init(&cli->cl_loi_list_lock); - cfs_atomic_set(&cli->cl_pending_w_pages, 0); - cfs_atomic_set(&cli->cl_pending_r_pages, 0); + cli->cl_dirty_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(). */ + client_adjust_max_dirty(cli); + INIT_LIST_HEAD(&cli->cl_cache_waiters); + INIT_LIST_HEAD(&cli->cl_loi_ready_list); + INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list); + INIT_LIST_HEAD(&cli->cl_loi_write_list); + INIT_LIST_HEAD(&cli->cl_loi_read_list); + spin_lock_init(&cli->cl_loi_list_lock); + atomic_set(&cli->cl_pending_w_pages, 0); + atomic_set(&cli->cl_pending_r_pages, 0); cli->cl_r_in_flight = 0; cli->cl_w_in_flight = 0; @@ -364,26 +374,27 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) spin_lock_init(&cli->cl_write_offset_hist.oh_lock); /* lru for osc. */ - CFS_INIT_LIST_HEAD(&cli->cl_lru_osc); - cfs_atomic_set(&cli->cl_lru_shrinkers, 0); - cfs_atomic_set(&cli->cl_lru_busy, 0); - cfs_atomic_set(&cli->cl_lru_in_list, 0); - CFS_INIT_LIST_HEAD(&cli->cl_lru_list); - client_obd_list_lock_init(&cli->cl_lru_list_lock); - - cfs_waitq_init(&cli->cl_destroy_waitq); - cfs_atomic_set(&cli->cl_destroy_in_flight, 0); + INIT_LIST_HEAD(&cli->cl_lru_osc); + atomic_set(&cli->cl_lru_shrinkers, 0); + atomic_long_set(&cli->cl_lru_busy, 0); + atomic_long_set(&cli->cl_lru_in_list, 0); + INIT_LIST_HEAD(&cli->cl_lru_list); + spin_lock_init(&cli->cl_lru_list_lock); + atomic_long_set(&cli->cl_unstable_count, 0); + + init_waitqueue_head(&cli->cl_destroy_waitq); + atomic_set(&cli->cl_destroy_in_flight, 0); #ifdef ENABLE_CHECKSUM - /* Turn on checksumming by default. */ - cli->cl_checksum = 1; - /* - * The supported checksum types will be worked out at connect time - * 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; + /* Turn on checksumming by default. */ + cli->cl_checksum = 1; + /* + * The supported checksum types will be worked out at connect time + * 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; #endif - cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); + 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 @@ -392,20 +403,41 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES, LNET_MTU >> PAGE_CACHE_SHIFT); + /* set cl_chunkbits default value to PAGE_CACHE_SHIFT, + * it will be updated at OSC connection time. */ + cli->cl_chunkbits = PAGE_CACHE_SHIFT; + if (!strcmp(name, LUSTRE_MDC_NAME)) { - cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT; - } else if (num_physpages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) { + cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) { cli->cl_max_rpcs_in_flight = 2; - } else if (num_physpages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) { cli->cl_max_rpcs_in_flight = 3; - } else if (num_physpages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) { cli->cl_max_rpcs_in_flight = 4; } else { if (osc_on_mdt(obddev->obd_name)) - cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT; + cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX; else - cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT; - } + cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; + } + + spin_lock_init(&cli->cl_mod_rpcs_lock); + spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock); + cli->cl_max_mod_rpcs_in_flight = 0; + cli->cl_mod_rpcs_in_flight = 0; + cli->cl_close_rpcs_in_flight = 0; + init_waitqueue_head(&cli->cl_mod_rpcs_waitq); + cli->cl_mod_tag_bitmap = NULL; + + if (connect_op == MDS_CONNECT) { + cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1; + OBD_ALLOC(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + if (cli->cl_mod_tag_bitmap == NULL) + GOTO(err, rc = -ENOMEM); + } + rc = ldlm_get_ref(); if (rc) { CERROR("ldlm_get_ref failed: %d\n", rc); @@ -430,10 +462,9 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_import, rc); } - cli->cl_import = imp; - /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */ - cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3); - cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie); + cli->cl_import = imp; + /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */ + cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3); if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) { if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) { @@ -456,8 +487,6 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_import, rc = -ENOMEM); } - cli->cl_qchk_stat = CL_NOT_QUOTACHECKED; - RETURN(rc); err_import: @@ -465,6 +494,10 @@ err_import: err_ldlm: ldlm_put_ref(); err: + if (cli->cl_mod_tag_bitmap != NULL) + OBD_FREE(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + cli->cl_mod_tag_bitmap = NULL; RETURN(rc); } @@ -472,15 +505,23 @@ EXPORT_SYMBOL(client_obd_setup); int client_obd_cleanup(struct obd_device *obddev) { - ENTRY; + struct client_obd *cli = &obddev->u.cli; + ENTRY; - ldlm_namespace_free_post(obddev->obd_namespace); - obddev->obd_namespace = NULL; + ldlm_namespace_free_post(obddev->obd_namespace); + obddev->obd_namespace = NULL; - LASSERT(obddev->u.cli.cl_import == NULL); + obd_cleanup_client_import(obddev); + LASSERT(obddev->u.cli.cl_import == NULL); - ldlm_put_ref(); - RETURN(0); + ldlm_put_ref(); + + if (cli->cl_mod_tag_bitmap != NULL) + OBD_FREE(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + cli->cl_mod_tag_bitmap = NULL; + + RETURN(0); } EXPORT_SYMBOL(client_obd_cleanup); @@ -495,11 +536,12 @@ int client_connect_import(const struct lu_env *env, struct obd_connect_data *ocd; struct lustre_handle conn = { 0 }; int rc; + bool is_mdc = false; ENTRY; *exp = NULL; down_write(&cli->cl_sem); - if (cli->cl_conn_count > 0 ) + if (cli->cl_conn_count > 0) GOTO(out_sem, rc = -EALREADY); rc = class_connect(&conn, obd, cluuid); @@ -519,6 +561,10 @@ int client_connect_import(const struct lu_env *env, 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; } @@ -527,13 +573,17 @@ int client_connect_import(const struct lu_env *env, LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); GOTO(out_ldlm, rc); } - LASSERT((*exp)->exp_connection); + 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; } ptlrpc_pinger_add_import(imp); @@ -571,17 +621,17 @@ int client_disconnect_export(struct obd_export *exp) imp = cli->cl_import; down_write(&cli->cl_sem); - CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name, - cli->cl_conn_count); + CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name, + cli->cl_conn_count); - if (!cli->cl_conn_count) { + if (cli->cl_conn_count == 0) { CERROR("disconnecting disconnected device (%s)\n", obd->obd_name); GOTO(out_disconnect, rc = -EINVAL); } cli->cl_conn_count--; - if (cli->cl_conn_count) + if (cli->cl_conn_count != 0) GOTO(out_disconnect, rc = 0); /* Mark import deactivated now, so we don't try to reconnect if any @@ -638,20 +688,19 @@ int server_disconnect_export(struct obd_export *exp) if (exp->exp_imp_reverse) ptlrpc_cleanup_imp(exp->exp_imp_reverse); - if (exp->exp_obd->obd_namespace != NULL) - ldlm_cancel_locks_for_export(exp); + ldlm_bl_thread_wakeup(); /* complete all outstanding replies */ spin_lock(&exp->exp_lock); - while (!cfs_list_empty(&exp->exp_outstanding_replies)) { + while (!list_empty(&exp->exp_outstanding_replies)) { struct ptlrpc_reply_state *rs = - cfs_list_entry(exp->exp_outstanding_replies.next, + list_entry(exp->exp_outstanding_replies.next, struct ptlrpc_reply_state, rs_exp_list); struct ptlrpc_service_part *svcpt = rs->rs_svcpt; spin_lock(&svcpt->scp_rep_lock); - cfs_list_del_init(&rs->rs_exp_list); + list_del_init(&rs->rs_exp_list); spin_lock(&rs->rs_lock); ptlrpc_schedule_difficult_reply(rs); spin_unlock(&rs->rs_lock); @@ -672,13 +721,13 @@ static int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp, struct obd_uuid *cluuid) { - ENTRY; + struct lustre_handle *hdl; + ENTRY; - if (exp->exp_connection && exp->exp_imp_reverse) { - struct lustre_handle *hdl; + hdl = &exp->exp_imp_reverse->imp_remote_handle; + if (exp->exp_connection && lustre_handle_is_used(hdl)) { struct obd_device *target; - hdl = &exp->exp_imp_reverse->imp_remote_handle; target = exp->exp_obd; /* Might be a re-connect after a partition. */ @@ -743,36 +792,136 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, spin_unlock(&exp->exp_lock); class_export_cb_put(exp); } -EXPORT_SYMBOL(target_client_add_cb); -#ifdef __KERNEL__ static void check_and_start_recovery_timer(struct obd_device *obd, struct ptlrpc_request *req, int new_client); -#else -static inline void -check_and_start_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, int new_client) + +/** + * update flags for import during reconnect process + */ +static int rev_import_flags_update(struct obd_import *revimp, + struct ptlrpc_request *req) { + int rc; + struct obd_connect_data *data; + + data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); + + if (data->ocd_connect_flags & OBD_CONNECT_AT) + revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; + else + revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + + revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; + + rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); + if (rc) { + CERROR("%s: cannot get reverse import %s security: rc = %d\n", + revimp->imp_client->cli_name, + libcfs_id2str(req->rq_peer), rc); + return rc; + } + + return 0; +} + +/** + * Allocate a new reverse import for an export. + * + * \retval -errno in case error hit + * \retval 0 if reverse import correctly init + **/ +int rev_import_init(struct obd_export *export) +{ + struct obd_device *obd = export->exp_obd; + struct obd_import *revimp; + + LASSERT(export->exp_imp_reverse == NULL); + + revimp = class_new_import(obd); + if (revimp == NULL) + return -ENOMEM; + + revimp->imp_remote_handle.cookie = 0ULL; + revimp->imp_client = &obd->obd_ldlm_client; + revimp->imp_dlm_fake = 1; + + /* it is safe to connect import in new state as no sends possible */ + spin_lock(&export->exp_lock); + export->exp_imp_reverse = revimp; + spin_unlock(&export->exp_lock); + class_import_put(revimp); + + return 0; +} +EXPORT_SYMBOL(rev_import_init); + +/** + * Handle reconnect for an export. + * + * \param exp export to handle reconnect process + * \param req client reconnect request + * + * \retval -rc in case securitfy flavor can't be changed + * \retval 0 in case none problems + */ +static int rev_import_reconnect(struct obd_export *exp, + struct ptlrpc_request *req) +{ + struct obd_import *revimp = exp->exp_imp_reverse; + struct lustre_handle *lh; + int rc; + + /* avoid sending a request until import flags are changed */ + ptlrpc_import_enter_resend(revimp); + + if (revimp->imp_connection != NULL) + ptlrpc_connection_put(revimp->imp_connection); + + /* + * client from recovery don't have a handle so we need to take from + * request. it may produce situation when wrong client connected + * to recovery as we trust a client uuid + */ + lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN); + revimp->imp_remote_handle = *lh; + + /* unknown versions will be caught in + * ptlrpc_handle_server_req_in->lustre_unpack_msg() */ + revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; + + revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection); + + rc = rev_import_flags_update(revimp, req); + if (rc != 0) { + /* it is safe to still be in RECOVERY phase as we are not able + * to setup correct security flavor so requests are not able to + * be delivered correctly */ + return rc; + } + + /* resend all rpc's via new connection */ + return ptlrpc_import_recovery_state_machine(revimp); } -#endif int target_handle_connect(struct ptlrpc_request *req) { - struct obd_device *target = NULL, *targref = NULL; - struct obd_export *export = NULL; - struct obd_import *revimp; - struct obd_import *tmp_imp = NULL; - struct lustre_handle conn; - struct lustre_handle *tmp; + struct obd_device *target = NULL; + struct obd_export *export = NULL; + /* connect handle - filled from target_handle_reconnect in + * reconnect case */ + struct lustre_handle conn; + struct lustre_handle *tmp; struct obd_uuid tgtuuid; struct obd_uuid cluuid; - struct obd_uuid remote_uuid; char *str; int rc = 0; char *target_start; int target_len; bool mds_conn = false, lw_client = false; + bool mds_mds_conn = false; + bool new_mds_mds_conn = false; struct obd_connect_data *data, *tmpdata; int size, tmpsize; lnet_nid_t *client_nid = NULL; @@ -794,12 +943,19 @@ int target_handle_connect(struct ptlrpc_request *req) if (!target) { deuuidify(str, NULL, &target_start, &target_len); LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect " - "from %s (no target)\n", str, + "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); + /* 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); + if (target->obd_stopping || !target->obd_set_up) { spin_unlock(&target->obd_dev_lock); @@ -821,11 +977,6 @@ int target_handle_connect(struct ptlrpc_request *req) GOTO(out, rc = -EAGAIN); } - /* Make sure the target isn't cleaned up while we're here. Yes, - * there's still a race between the above check and our incref here. - * Really, class_uuid2obd should take the ref. */ - targref = class_incref(target, __FUNCTION__, cfs_current()); - target->obd_conn_inprogress++; spin_unlock(&target->obd_dev_lock); @@ -837,21 +988,6 @@ int target_handle_connect(struct ptlrpc_request *req) obd_str2uuid(&cluuid, str); - /* XXX Extract a nettype and format accordingly. */ - switch (sizeof(lnet_nid_t)) { - /* NB the casts only avoid compiler warnings. */ - case 8: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid); - break; - case 4: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_%x_UUID", (__u32)req->rq_peer.nid); - break; - default: - LBUG(); - } - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); if (tmp == NULL) GOTO(out, rc = -EPROTO); @@ -868,41 +1004,87 @@ int target_handle_connect(struct ptlrpc_request *req) if (rc) GOTO(out, rc); - if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { - if (!data) { - DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) " - "libclient connection attempt"); - GOTO(out, rc = -EPROTO); - } else 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); - } - } +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0) + /* Don't allow clients to connect that are using old 1.8 format + * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18, + * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The + * FULL20 flag should be set on all connections since 2.0, but no + * longer affects behaviour. + * + * Later this check will be disabled and the flag can be retired + * completely once interop with 3.0 is no longer needed. + */ + if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20)) + GOTO(out, rc = -EPROTO); +#endif - if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) && - (data->ocd_connect_flags & OBD_CONNECT_MDS)) - mds_conn = true; + 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); + } + } - if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0) - lw_client = true; + /* Note: lw_client is needed in MDS-MDS failover during update log + * processing, so we needs to allow lw_client to be connected at + * anytime, instead of only the initial connection */ + lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0; + + if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) { + mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0; + mds_mds_conn = (data->ocd_connect_flags & + OBD_CONNECT_MDS_MDS) != 0; + + /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS + * for Imperative Recovery connection from MGC to MGS. + * + * Via check OBD_CONNECT_FID, we can distinguish whether + * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from + * MGC or MDT. */ + if (!lw_client && + (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && + (data->ocd_connect_flags & OBD_CONNECT_FID) && + (data->ocd_connect_flags & OBD_CONNECT_VERSION)) { + __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version); + __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version); + __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version); + + /* We do not support the MDT-MDT interoperations with + * different version MDT because of protocol changes. */ + if (unlikely(major != LUSTRE_MAJOR || + minor != LUSTRE_MINOR || + abs(patch - LUSTRE_PATCH) > 3)) { + LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the " + "connection from different version MDT " + "(%d.%d.%d.%d) %s %s\n", + target->obd_name, LUSTRE_MAJOR, + LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX, + major, minor, patch, + OBD_OCD_VERSION_FIX(data->ocd_version), + libcfs_nid2str(req->rq_peer.nid), str); + + GOTO(out, rc = -EPROTO); + } + } + } /* lctl gets a backstage, all-access pass. */ if (obd_uuid_equals(&cluuid, &target->obd_uuid)) @@ -924,19 +1106,21 @@ int target_handle_connect(struct ptlrpc_request *req) class_export_put(export); export = NULL; rc = -EALREADY; - } else if (mds_conn && export->exp_connection) { + } else if ((mds_conn || lw_client) && export->exp_connection != NULL) { spin_unlock(&export->exp_lock); - if (req->rq_peer.nid != export->exp_connection->c_peer.nid) - /* MDS reconnected after failover. */ - LCONSOLE_WARN("%s: Received MDS connection from " + 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, libcfs_nid2str(req->rq_peer.nid), + 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 MDS connection from " - "%s, removing former export from same NID\n", - target->obd_name, libcfs_nid2str(req->rq_peer.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; @@ -969,28 +1153,13 @@ int target_handle_connect(struct ptlrpc_request *req) no_export: OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout); } else if (req->rq_export == NULL && - cfs_atomic_read(&export->exp_rpc_count) > 0) { + 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), - cfs_atomic_read(&export->exp_refcount)); + atomic_read(&export->exp_refcount)); GOTO(out, rc = -EBUSY); - } else if (req->rq_export != NULL && - (cfs_atomic_read(&export->exp_rpc_count) > 1)) { - /* The current connect RPC has increased exp_rpc_count. */ - LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, " - "still busy with %d active RPCs\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - cfs_atomic_read(&export->exp_rpc_count) - 1); - spin_lock(&export->exp_lock); - if (req->rq_export->exp_conn_cnt < - lustre_msg_get_conn_cnt(req->rq_reqmsg)) - /* try to abort active requests */ - req->rq_export->exp_abort_active_req = 1; - spin_unlock(&export->exp_lock); - 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 " @@ -1013,7 +1182,7 @@ no_export: 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 + /* 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); @@ -1034,60 +1203,56 @@ no_export: if (export == NULL) { /* allow lightweight connections during recovery */ - if (target->obd_recovering && !lw_client) { + /* allow "new" MDT to be connected during recovery, since we + * need retrieve recovery update records from it */ + if (target->obd_recovering && !lw_client && !mds_mds_conn) { cfs_time_t t; int c; /* connected */ int i; /* in progress */ int k; /* known */ int s; /* stale/evicted */ - c = cfs_atomic_read(&target->obd_connected_clients); - i = cfs_atomic_read(&target->obd_lock_replay_clients); + c = atomic_read(&target->obd_connected_clients); + i = atomic_read(&target->obd_lock_replay_clients); k = target->obd_max_recoverable_clients; s = target->obd_stale_clients; t = cfs_timer_deadline(&target->obd_recovery_timer); t = cfs_time_sub(t, cfs_time_current()); t = cfs_duration_sec(t); - LCONSOLE_WARN("%s: Denying connection for new client " - "%s (at %s), waiting for all %d known " - "clients (%d recovered, %d in progress, " - "and %d evicted) to recover in %d:%.02d\n", + LCONSOLE_WARN("%s: Denying connection for new client %s" + "(at %s), waiting for %d known clients " + "(%d recovered, %d in progress, and %d " + "evicted) to recover in %d:%.02d\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), k, c - i, i, s, (int)t / 60, (int)t % 60); - rc = -EBUSY; - } else { + rc = -EBUSY; + } else { dont_check_exports: - rc = obd_connect(req->rq_svc_thread->t_env, - &export, target, &cluuid, data, - client_nid); - if (rc == 0) - conn.cookie = export->exp_handle.h_cookie; - } - } else { - rc = obd_reconnect(req->rq_svc_thread->t_env, - export, target, &cluuid, data, client_nid); - } - if (rc) - GOTO(out, rc); - -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0) - /* 2.2.0 clients always swab nidtbl entries due to a bug, so server - * will do the swabbing for if the client is using the same endianness. - * - * This fixup is version-limited, because we don't want to carry the - * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need - * interop with unpatched 2.2 clients. For newer clients, servers - * will never do MNE swabbing, let the client handle that. LU-1644 */ - export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) && - !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB); -#else -#warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab" -#endif + rc = obd_connect(req->rq_svc_thread->t_env, + &export, target, &cluuid, data, + client_nid); + if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG)) + lustre_msg_add_op_flags(req->rq_repmsg, + MSG_CONNECT_RECOVERING); + if (rc == 0) { + conn.cookie = export->exp_handle.h_cookie; + rc = rev_import_init(export); + } + + if (mds_mds_conn) + new_mds_mds_conn = true; + } + } else { + rc = obd_reconnect(req->rq_svc_thread->t_env, + export, target, &cluuid, data, client_nid); + } + if (rc) + GOTO(out, rc); - LASSERT(target->u.obt.obt_magic == OBT_MAGIC); - data->ocd_instance = target->u.obt.obt_instance; + LASSERT(target->u.obt.obt_magic == OBT_MAGIC); + data->ocd_instance = target->u.obt.obt_instance; /* Return only the parts of obd_connect_data that we understand, so the * client knows that we don't understand the rest. */ @@ -1102,11 +1267,6 @@ dont_check_exports: memcpy(tmpdata, data, min(tmpsize, size)); } - /* If all else goes well, this is our RPC return code. */ - req->rq_status = 0; - - lustre_msg_set_handle(req->rq_repmsg, &conn); - /* If the client and the server are the same node, we will already * have an export that really points to the client's DLM export, * because we have a shared handles table. @@ -1129,7 +1289,6 @@ dont_check_exports: } LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0); export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg); - export->exp_abort_active_req = 0; /* Don't evict liblustre clients for not pinging. */ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { @@ -1137,7 +1296,7 @@ dont_check_exports: spin_unlock(&export->exp_lock); spin_lock(&target->obd_dev_lock); - cfs_list_del_init(&export->exp_obd_chain_timed); + list_del_init(&export->exp_obd_chain_timed); spin_unlock(&target->obd_dev_lock); } else { spin_unlock(&export->exp_lock); @@ -1146,7 +1305,7 @@ dont_check_exports: 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) && - !cfs_hlist_unhashed(&export->exp_nid_hash)) + !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); @@ -1154,14 +1313,19 @@ dont_check_exports: ptlrpc_connection_put(export->exp_connection); } - export->exp_connection = ptlrpc_connection_get(req->rq_peer, - req->rq_self, - &remote_uuid); - if (cfs_hlist_unhashed(&export->exp_nid_hash)) { - cfs_hash_add(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); - } + export->exp_connection = ptlrpc_connection_get(req->rq_peer, + req->rq_self, + &cluuid); + if (hlist_unhashed(&export->exp_nid_hash)) + cfs_hash_add(export->exp_obd->obd_nid_hash, + &export->exp_connection->c_peer.nid, + &export->exp_nid_hash); + + lustre_msg_set_handle(req->rq_repmsg, &conn); + + rc = rev_import_reconnect(export, req); + if (rc != 0) + GOTO(out, rc); if (target->obd_recovering && !export->exp_in_recovery && !lw_client) { int has_transno; @@ -1194,70 +1358,26 @@ dont_check_exports: spin_unlock(&target->obd_recovery_task_lock); } - cfs_atomic_inc(&target->obd_req_replay_clients); - cfs_atomic_inc(&target->obd_lock_replay_clients); - if (cfs_atomic_inc_return(&target->obd_connected_clients) == - target->obd_max_recoverable_clients) - cfs_waitq_signal(&target->obd_next_transno_waitq); - } + atomic_inc(&target->obd_req_replay_clients); + atomic_inc(&target->obd_lock_replay_clients); + /* Note: MDS-MDS connection is allowed to be connected during + * recovery, no matter if the exports needs to be recoveried. + * Because we need retrieve updates logs from all other MDTs. + * So if the MDS-MDS export is new, obd_max_recoverable_clients + * also needs to be increased to match other recovery checking + * condition. */ + if (new_mds_mds_conn) + target->obd_max_recoverable_clients++; + if (atomic_inc_return(&target->obd_connected_clients) == + target->obd_max_recoverable_clients) + wake_up(&target->obd_next_transno_waitq); + } /* 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); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); - conn = *tmp; - - /* Return -ENOTCONN in case of errors to let client reconnect. */ - revimp = class_new_import(target); - if (revimp == NULL) { - CERROR("fail to alloc new reverse import.\n"); - GOTO(out, rc = -ENOTCONN); - } - - spin_lock(&export->exp_lock); - if (export->exp_imp_reverse != NULL) - /* destroyed import can be still referenced in ctxt */ - tmp_imp = export->exp_imp_reverse; - export->exp_imp_reverse = revimp; - spin_unlock(&export->exp_lock); - - revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection); - revimp->imp_client = &export->exp_obd->obd_ldlm_client; - revimp->imp_remote_handle = conn; - revimp->imp_dlm_fake = 1; - revimp->imp_state = LUSTRE_IMP_FULL; - - /* Unknown versions will be caught in - * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */ - revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; - - if ((data->ocd_connect_flags & OBD_CONNECT_AT) && - (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) - revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT; - else - revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; - - if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) && - (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) - revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; - else - revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; - - rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); - if (rc) { - CERROR("Failed to get sec for reverse import: %d\n", rc); - spin_lock(&export->exp_lock); - export->exp_imp_reverse = NULL; - spin_unlock(&export->exp_lock); - class_destroy_import(revimp); - } - - class_import_put(revimp); - out: - if (tmp_imp != NULL) - client_destroy_import(tmp_imp); if (export) { spin_lock(&export->exp_lock); export->exp_connecting = 0; @@ -1265,18 +1385,16 @@ out: class_export_put(export); } - if (targref) { + if (target != NULL) { spin_lock(&target->obd_dev_lock); target->obd_conn_inprogress--; spin_unlock(&target->obd_dev_lock); - class_decref(targref, __func__, cfs_current()); + class_decref(target, __func__, current); } - if (rc) - req->rq_status = rc; + req->rq_status = rc; RETURN(rc); } -EXPORT_SYMBOL(target_handle_connect); int target_handle_disconnect(struct ptlrpc_request *req) { @@ -1292,7 +1410,6 @@ int target_handle_disconnect(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(target_handle_disconnect); void target_destroy_export(struct obd_export *exp) { @@ -1321,23 +1438,23 @@ EXPORT_SYMBOL(target_destroy_export); static void target_request_copy_get(struct ptlrpc_request *req) { class_export_rpc_inc(req->rq_export); - LASSERT(cfs_list_empty(&req->rq_list)); - CFS_INIT_LIST_HEAD(&req->rq_replay_list); + LASSERT(list_empty(&req->rq_list)); + INIT_LIST_HEAD(&req->rq_replay_list); /* Increase refcount to keep request in queue. */ - cfs_atomic_inc(&req->rq_refcount); + atomic_inc(&req->rq_refcount); /* Let export know it has replays to be handled. */ - cfs_atomic_inc(&req->rq_export->exp_replay_count); + atomic_inc(&req->rq_export->exp_replay_count); } static void target_request_copy_put(struct ptlrpc_request *req) { - LASSERT(cfs_list_empty(&req->rq_replay_list)); - LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count); + LASSERT(list_empty(&req->rq_replay_list)); + LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count); - cfs_atomic_dec(&req->rq_export->exp_replay_count); + atomic_dec(&req->rq_export->exp_replay_count); class_export_rpc_dec(req->rq_export); - ptlrpc_server_drop_request(req); + ptlrpc_server_drop_request(req); } static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) @@ -1345,14 +1462,16 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); struct obd_export *exp = req->rq_export; struct ptlrpc_request *reqiter; + struct ptlrpc_request *dup_req = NULL; int dup = 0; LASSERT(exp); spin_lock(&exp->exp_lock); - cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue, + list_for_each_entry(reqiter, &exp->exp_req_replay_queue, rq_replay_list) { if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { + dup_req = reqiter; dup = 1; break; } @@ -1364,8 +1483,18 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY)) CERROR("invalid flags %x of resent replay\n", lustre_msg_get_flags(req->rq_reqmsg)); + + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + __u32 new_conn; + + new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg); + if (new_conn > + lustre_msg_get_conn_cnt(dup_req->rq_reqmsg)) + lustre_msg_set_conn_cnt(dup_req->rq_reqmsg, + new_conn); + } } else { - cfs_list_add_tail(&req->rq_replay_list, + list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue); } @@ -1375,17 +1504,17 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) { - LASSERT(!cfs_list_empty(&req->rq_replay_list)); + LASSERT(!list_empty(&req->rq_replay_list)); LASSERT(req->rq_export); spin_lock(&req->rq_export->exp_lock); - cfs_list_del_init(&req->rq_replay_list); + list_del_init(&req->rq_replay_list); spin_unlock(&req->rq_export->exp_lock); } -#ifdef __KERNEL__ -static void target_finish_recovery(struct obd_device *obd) +static void target_finish_recovery(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; ENTRY; /* Only log a recovery message when recovery has occurred. */ @@ -1396,28 +1525,35 @@ static void target_finish_recovery(struct obd_device *obd) "%d recovered and %d %s evicted.\n", obd->obd_name, (int)elapsed_time / 60, (int)elapsed_time % 60, obd->obd_max_recoverable_clients, - cfs_atomic_read(&obd->obd_connected_clients), + atomic_read(&obd->obd_connected_clients), obd->obd_stale_clients, obd->obd_stale_clients == 1 ? "was" : "were"); } ldlm_reprocess_all_ns(obd->obd_namespace); spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_req_replay_queue) || - !cfs_list_empty(&obd->obd_lock_replay_queue) || - !cfs_list_empty(&obd->obd_final_req_queue)) { + 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, - cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ", - cfs_list_empty(&obd->obd_lock_replay_queue) ? \ + list_empty(&obd->obd_req_replay_queue) ? "" : "req ", + list_empty(&obd->obd_lock_replay_queue) ? \ "" : "lock ", - cfs_list_empty(&obd->obd_final_req_queue) ? \ + list_empty(&obd->obd_final_req_queue) ? \ "" : "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(); /* When recovery finished, cleanup orphans on MDS and OST. */ @@ -1433,13 +1569,13 @@ static void target_finish_recovery(struct obd_device *obd) static void abort_req_replay_queue(struct obd_device *obd) { struct ptlrpc_request *req, *n; - cfs_list_t abort_list; + struct list_head abort_list; - CFS_INIT_LIST_HEAD(&abort_list); + INIT_LIST_HEAD(&abort_list); spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list); + list_splice_init(&obd->obd_req_replay_queue, &abort_list); spin_unlock(&obd->obd_recovery_task_lock); - cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) { + 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)) { @@ -1454,13 +1590,13 @@ static void abort_req_replay_queue(struct obd_device *obd) static void abort_lock_replay_queue(struct obd_device *obd) { struct ptlrpc_request *req, *n; - cfs_list_t abort_list; + struct list_head abort_list; - CFS_INIT_LIST_HEAD(&abort_list); + INIT_LIST_HEAD(&abort_list); spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list); + list_splice_init(&obd->obd_lock_replay_queue, &abort_list); spin_unlock(&obd->obd_recovery_task_lock); - cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){ + 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)) { @@ -1483,10 +1619,9 @@ static void abort_lock_replay_queue(struct obd_device *obd) void target_cleanup_recovery(struct obd_device *obd) { struct ptlrpc_request *req, *n; - cfs_list_t clean_list; - ENTRY; + struct list_head clean_list; - CFS_INIT_LIST_HEAD(&clean_list); + INIT_LIST_HEAD(&clean_list); spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering) { spin_unlock(&obd->obd_dev_lock); @@ -1494,26 +1629,27 @@ 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); target_cancel_recovery_timer(obd); - cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list); + list_splice_init(&obd->obd_req_replay_queue, &clean_list); spin_unlock(&obd->obd_recovery_task_lock); - cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); + list_for_each_entry_safe(req, n, &clean_list, rq_list) { + LASSERT(req->rq_reply_state == NULL); target_exp_dequeue_req_replay(req); target_request_copy_put(req); } spin_lock(&obd->obd_recovery_task_lock); - cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list); - cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list); + list_splice_init(&obd->obd_lock_replay_queue, &clean_list); + list_splice_init(&obd->obd_final_req_queue, &clean_list); spin_unlock(&obd->obd_recovery_task_lock); - cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){ - LASSERT(req->rq_reply_state == 0); + list_for_each_entry_safe(req, n, &clean_list, rq_list) { + LASSERT(req->rq_reply_state == NULL); target_request_copy_put(req); } @@ -1527,7 +1663,6 @@ void target_cancel_recovery_timer(struct obd_device *obd) CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name); cfs_timer_disarm(&obd->obd_recovery_timer); } -EXPORT_SYMBOL(target_cancel_recovery_timer); static void target_start_recovery_timer(struct obd_device *obd) { @@ -1535,7 +1670,8 @@ static void target_start_recovery_timer(struct obd_device *obd) return; spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery) { + if (!obd->obd_recovering || obd->obd_abort_recovery || + obd->obd_force_abort_recovery) { spin_unlock(&obd->obd_dev_lock); return; } @@ -1576,7 +1712,8 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) int to; spin_lock(&obd->obd_dev_lock); - if (!obd->obd_recovering || obd->obd_abort_recovery) { + if (!obd->obd_recovering || obd->obd_abort_recovery || + obd->obd_force_abort_recovery) { spin_unlock(&obd->obd_dev_lock); return; } @@ -1591,22 +1728,20 @@ static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) to += drt - left; } else if (!extend && (drt > to)) { to = drt; - /* reduce drt by already passed time */ - drt -= obd->obd_recovery_timeout - left; } if (to > obd->obd_recovery_time_hard) to = obd->obd_recovery_time_hard; - if (obd->obd_recovery_timeout < to || - obd->obd_recovery_timeout == obd->obd_recovery_time_hard) { + if (obd->obd_recovery_timeout < to) { obd->obd_recovery_timeout = to; - cfs_timer_arm(&obd->obd_recovery_timer, - cfs_time_shift(drt)); + end = obd->obd_recovery_start + to; + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(end - now)); } spin_unlock(&obd->obd_dev_lock); CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)drt); + obd->obd_name, (unsigned)cfs_time_sub(end, now)); } /* Reset the timer with each new client connection */ @@ -1627,7 +1762,6 @@ check_and_start_recovery_timer(struct obd_device *obd, { int service_time = lustre_msg_get_service_time(req->rq_reqmsg); struct obd_device_target *obt = &obd->u.obt; - struct lustre_sb_info *lsi; if (!new_client && service_time) /* Teach server about old server's estimates, as first guess @@ -1647,12 +1781,9 @@ check_and_start_recovery_timer(struct obd_device *obd, service_time += 2 * INITIAL_CONNECT_TIMEOUT; LASSERT(obt->obt_magic == OBT_MAGIC); - lsi = s2lsi(obt->obt_sb); - if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE)) - 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); + 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); } /** Health checking routines */ @@ -1664,14 +1795,14 @@ static inline int exp_connect_healthy(struct obd_export *exp) /** if export done req_replay or has replay in queue */ static inline int exp_req_replay_healthy(struct obd_export *exp) { - return (!exp->exp_req_replay_needed || - cfs_atomic_read(&exp->exp_replay_count) > 0); + return (!exp->exp_req_replay_needed || + atomic_read(&exp->exp_replay_count) > 0); } /** if export done lock_replay or has replay in queue */ static inline int exp_lock_replay_healthy(struct obd_export *exp) { - return (!exp->exp_lock_replay_needed || - cfs_atomic_read(&exp->exp_replay_count) > 0); + return (!exp->exp_lock_replay_needed || + atomic_read(&exp->exp_replay_count) > 0); } static inline int exp_vbr_healthy(struct obd_export *exp) @@ -1684,103 +1815,105 @@ static inline int exp_finished(struct obd_export *exp) return (exp->exp_in_recovery && !exp->exp_lock_replay_needed); } -/** Checking routines for recovery */ -static int check_for_clients(struct obd_device *obd) -{ - unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients); - - if (obd->obd_abort_recovery || obd->obd_recovery_expired) - return 1; - LASSERT(clnts <= obd->obd_max_recoverable_clients); - return (clnts + obd->obd_stale_clients == - obd->obd_max_recoverable_clients); -} - -static int check_for_next_transno(struct obd_device *obd) +static int check_for_next_transno(struct lu_target *lut) { struct ptlrpc_request *req = NULL; + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; int wake_up = 0, connected, completed, queue_len; - __u64 next_transno, req_transno; + __u64 req_transno = 0; + __u64 update_transno = 0; + __u64 next_transno = 0; ENTRY; spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_req_replay_queue)) { - req = cfs_list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); - req_transno = lustre_msg_get_transno(req->rq_reqmsg); - } else { - req_transno = 0; - } + if (!list_empty(&obd->obd_req_replay_queue)) { + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + req_transno = lustre_msg_get_transno(req->rq_reqmsg); + } + + if (tdtd != NULL) + update_transno = distribute_txn_get_next_transno(tdtd); + + connected = atomic_read(&obd->obd_connected_clients); + completed = connected - atomic_read(&obd->obd_req_replay_clients); + queue_len = obd->obd_requests_queued_for_recovery; + next_transno = obd->obd_next_recovery_transno; - connected = cfs_atomic_read(&obd->obd_connected_clients); - completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients); - queue_len = obd->obd_requests_queued_for_recovery; - next_transno = obd->obd_next_recovery_transno; - - 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, - queue_len, req_transno, next_transno); - - if (obd->obd_abort_recovery) { - CDEBUG(D_HA, "waking for aborted recovery\n"); - wake_up = 1; - } else if (obd->obd_recovery_expired) { - CDEBUG(D_HA, "waking for expired recovery\n"); - wake_up = 1; - } else if (req_transno == next_transno) { - CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno); - wake_up = 1; + 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, + queue_len, req_transno, next_transno); + + if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { + CDEBUG(D_HA, "waking for aborted recovery\n"); + wake_up = 1; + } else if (obd->obd_recovery_expired) { + CDEBUG(D_HA, "waking for expired recovery\n"); + wake_up = 1; + } else if (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", + 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); + wake_up = 1; } else if (queue_len > 0 && - queue_len == cfs_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, 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 - ", last_committed: "LPD64")\n", - obd->obd_name, obd->obd_version_recov ? "ON" : "OFF", - next_transno, queue_len, completed, connected, - req_transno, obd->obd_last_committed); - obd->obd_next_recovery_transno = req_transno; - wake_up = 1; - } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 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, 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", + obd->obd_name, obd->obd_version_recov ? "ON" : "OFF", + next_transno, queue_len, completed, connected, + req_transno, update_transno, obd->obd_last_committed); + obd->obd_next_recovery_transno = req_transno; + wake_up = 1; + } else if (atomic_read(&obd->obd_req_replay_clients) == 0) { 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); - obd->obd_next_recovery_transno = req_transno; - 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); + obd->obd_next_recovery_transno = req_transno; + wake_up = 1; + } spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } -static int check_for_next_lock(struct obd_device *obd) +static int check_for_next_lock(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; int wake_up = 0; spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { - CDEBUG(D_HA, "waking for next lock\n"); - wake_up = 1; - } else if (cfs_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) { - CDEBUG(D_HA, "waking for aborted recovery\n"); - wake_up = 1; - } else if (obd->obd_recovery_expired) { - CDEBUG(D_HA, "waking for expired recovery\n"); - wake_up = 1; - } + if (!list_empty(&obd->obd_lock_replay_queue)) { + CDEBUG(D_HA, "waking for next lock\n"); + wake_up = 1; + } 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) { + CDEBUG(D_HA, "waking for aborted recovery\n"); + wake_up = 1; + } else if (obd->obd_recovery_expired) { + CDEBUG(D_HA, "waking for expired recovery\n"); + wake_up = 1; + } spin_unlock(&obd->obd_recovery_task_lock); return wake_up; @@ -1791,93 +1924,74 @@ static int check_for_next_lock(struct obd_device *obd) * check its status with help of check_routine * evict dead clients via health_check */ -static int target_recovery_overseer(struct obd_device *obd, - int (*check_routine)(struct obd_device *), - int (*health_check)(struct obd_export *)) +static int target_recovery_overseer(struct lu_target *lut, + int (*check_routine)(struct lu_target *), + int (*health_check)(struct obd_export *)) { + struct obd_device *obd = lut->lut_obd; repeat: - cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); - if (obd->obd_abort_recovery) { - CWARN("recovery is aborted, evict exports in recovery\n"); - /** evict exports which didn't finish recovery yet */ - class_disconnect_stale_exports(obd, exp_finished); - return 1; - } else if (obd->obd_recovery_expired) { - obd->obd_recovery_expired = 0; - /** If some clients died being recovered, evict them */ - 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 */ + if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >= + (obd->obd_recovery_start + obd->obd_recovery_time_hard))) { + CWARN("recovery is aborted by hard timeout\n"); + obd->obd_abort_recovery = 1; + } + + while (wait_event_timeout(obd->obd_next_transno_waitq, + check_routine(lut), + msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0) + /* wait indefinitely for event, but don't trigger watchdog */; + + if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) { + CWARN("recovery is aborted, evict exports in recovery\n"); + /** evict exports which didn't finish recovery yet */ + class_disconnect_stale_exports(obd, exp_finished); + return 1; + } else if (obd->obd_recovery_expired) { + 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); + /** 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; spin_unlock(&obd->obd_dev_lock); - /** - * reset timer, recovery will proceed with versions now, - * timeout is set just to handle reconnection delays - */ - extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); - /** Wait for recovery events again, after evicting bad clients */ - goto repeat; - } - return 0; + /** + * reset timer, recovery will proceed with versions now, + * timeout is set just to handle reconnection delays + */ + extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true); + /** Wait for recovery events again, after evicting bad clients */ + goto repeat; + } + return 0; } -static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) +static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut) { - struct ptlrpc_request *req = NULL; - ENTRY; - - CDEBUG(D_HA, "Waiting for transno "LPD64"\n", - obd->obd_next_recovery_transno); + struct obd_device *obd = lut->lut_obd; + struct ptlrpc_request *req = NULL; - if (target_recovery_overseer(obd, check_for_next_transno, - exp_req_replay_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } + CDEBUG(D_HA, "Waiting for lock\n"); + if (target_recovery_overseer(lut, check_for_next_lock, + exp_lock_replay_healthy)) + abort_lock_replay_queue(obd); spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_req_replay_queue)) { - req = cfs_list_entry(obd->obd_req_replay_queue.next, + if (!list_empty(&obd->obd_lock_replay_queue)) { + req = list_entry(obd->obd_lock_replay_queue.next, struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); - obd->obd_requests_queued_for_recovery--; + list_del_init(&req->rq_list); spin_unlock(&obd->obd_recovery_task_lock); } else { spin_unlock(&obd->obd_recovery_task_lock); - LASSERT(cfs_list_empty(&obd->obd_req_replay_queue)); - LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0); + LASSERT(list_empty(&obd->obd_lock_replay_queue)); + LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0); /** evict exports failed VBR */ class_disconnect_stale_exports(obd, exp_vbr_healthy); } - RETURN(req); -} - -static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) -{ - struct ptlrpc_request *req = NULL; - - CDEBUG(D_HA, "Waiting for lock\n"); - if (target_recovery_overseer(obd, check_for_next_lock, - exp_lock_replay_healthy)) - abort_lock_replay_queue(obd); - - spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_lock_replay_queue)) { - req = cfs_list_entry(obd->obd_lock_replay_queue.next, - struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); - spin_unlock(&obd->obd_recovery_task_lock); - } else { - spin_unlock(&obd->obd_recovery_task_lock); - LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue)); - LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0); - /** evict exports failed VBR */ - class_disconnect_stale_exports(obd, exp_vbr_healthy); - } - return req; + return req; } static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) @@ -1885,10 +1999,10 @@ static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) struct ptlrpc_request *req = NULL; spin_lock(&obd->obd_recovery_task_lock); - if (!cfs_list_empty(&obd->obd_final_req_queue)) { - req = cfs_list_entry(obd->obd_final_req_queue.next, + if (!list_empty(&obd->obd_final_req_queue)) { + req = list_entry(obd->obd_final_req_queue.next, struct ptlrpc_request, rq_list); - cfs_list_del_init(&req->rq_list); + list_del_init(&req->rq_list); spin_unlock(&obd->obd_recovery_task_lock); if (req->rq_export->exp_in_recovery) { spin_lock(&req->rq_export->exp_lock); @@ -1901,38 +2015,28 @@ static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) return req; } -static int handle_recovery_req(struct ptlrpc_thread *thread, - struct ptlrpc_request *req, - svc_handler_t handler) +static void handle_recovery_req(struct ptlrpc_thread *thread, + struct ptlrpc_request *req, + svc_handler_t handler) { - int rc; - ENTRY; - - /** - * export can be evicted during recovery, no need to handle replays for - * it after that, discard such request silently - */ - if (req->rq_export->exp_disconnected) - GOTO(reqcopy_put, rc = 0); + ENTRY; - rc = lu_context_init(&req->rq_recov_session, LCT_SESSION); - if (rc) { - CERROR("Failure to initialize session: %d\n", rc); - GOTO(reqcopy_put, rc); - } + /** + * export can be evicted during recovery, no need to handle replays for + * it after that, discard such request silently + */ + if (req->rq_export->exp_disconnected) + RETURN_EXIT; - req->rq_recov_session.lc_thread = thread; - lu_context_enter(&req->rq_recov_session); - req->rq_svc_thread = thread; - req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session; + req->rq_session.lc_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); - lu_context_exit(&req->rq_recov_session); - lu_context_fini(&req->rq_recov_session); /* don't reset timer for final stage */ if (!exp_finished(req->rq_export)) { int to = obd_timeout; @@ -1961,8 +2065,250 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, } extend_recovery_timer(class_exp2obd(req->rq_export), to, true); } -reqcopy_put: - RETURN(rc); + EXIT; +} + +/** Checking routines for recovery */ +static int check_for_recovery_ready(struct lu_target *lut) +{ + struct obd_device *obd = lut->lut_obd; + unsigned int clnts = atomic_read(&obd->obd_connected_clients); + + CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d" + " abort %d expired %d\n", clnts, obd->obd_stale_clients, + obd->obd_max_recoverable_clients, obd->obd_abort_recovery, + obd->obd_recovery_expired); + + if (obd->obd_force_abort_recovery) + return 1; + + if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) { + LASSERT(clnts <= obd->obd_max_recoverable_clients); + if (clnts + obd->obd_stale_clients < + obd->obd_max_recoverable_clients) + return 0; + } + + if (lut->lut_tdtd != NULL) { + if (!lut->lut_tdtd->tdtd_replay_ready) { + /* Let's extend recovery timer, in case the recovery + * timer expired, and some clients got evicted */ + extend_recovery_timer(obd, obd->obd_recovery_timeout, + true); + return 0; + } else { + dtrq_list_dump(lut->lut_tdtd, D_HA); + } + } + + return 1; +} + +enum { + REQUEST_RECOVERY = 1, + UPDATE_RECOVERY = 2, +}; + +static __u64 get_next_replay_req_transno(struct obd_device *obd) +{ + __u64 transno = 0; + + if (!list_empty(&obd->obd_req_replay_queue)) { + struct ptlrpc_request *req; + + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + transno = lustre_msg_get_transno(req->rq_reqmsg); + } + + return transno; +} +__u64 get_next_transno(struct lu_target *lut, int *type) +{ + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + __u64 transno = 0; + __u64 update_transno; + ENTRY; + + transno = get_next_replay_req_transno(obd); + if (type != NULL) + *type = REQUEST_RECOVERY; + + if (tdtd == NULL) + RETURN(transno); + + update_transno = distribute_txn_get_next_transno(tdtd); + if (transno == 0 || (transno >= update_transno && + update_transno != 0)) { + transno = update_transno; + if (type != NULL) + *type = UPDATE_RECOVERY; + } + + RETURN(transno); +} + +/** + * drop duplicate replay request + * + * Because the operation has been replayed by update recovery, the request + * with the same transno will be dropped and also notify the client to send + * next replay request. + * + * \param[in] env execution environment + * \param[in] obd failover obd device + * \param[in] req request to be dropped + */ +static void drop_duplicate_replay_req(struct lu_env *env, + struct obd_device *obd, + struct ptlrpc_request *req) +{ + DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate" + " update records are found.\n", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + /* Right now, only for MDS reint operation update replay and + * normal request replay can have the same transno */ + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) { + req_capsule_set(&req->rq_pill, &RQF_MDS_REINT); + req->rq_status = req_capsule_server_pack(&req->rq_pill); + if (likely(req->rq_export)) + target_committed_to_req(req); + lustre_msg_set_transno(req->rq_repmsg, req->rq_transno); + target_send_reply(req, req->rq_status, 0); + } else { + DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n", + libcfs_nid2str(req->rq_peer.nid)); + } + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; +} + +static void replay_request_or_update(struct lu_env *env, + struct lu_target *lut, + struct target_recovery_data *trd, + struct ptlrpc_thread *thread) +{ + struct obd_device *obd = lut->lut_obd; + struct ptlrpc_request *req = NULL; + int type; + __u64 transno; + ENTRY; + + CDEBUG(D_HA, "Waiting for transno "LPD64"\n", + obd->obd_next_recovery_transno); + + /* Replay all of request and update by transno */ + do { + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + + CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val); + + /** It is needed to extend recovery window above + * recovery_time_soft. Extending is possible only in the + * end of recovery window (see more details in + * handle_recovery_req()). + */ + CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300); + + if (target_recovery_overseer(lut, check_for_next_transno, + exp_req_replay_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + } + + spin_lock(&obd->obd_recovery_task_lock); + transno = get_next_transno(lut, &type); + if (type == REQUEST_RECOVERY && transno != 0) { + /* Drop replay request from client side, if the + * replay has been executed by update with the + * same transno */ + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + + list_del_init(&req->rq_list); + obd->obd_requests_queued_for_recovery--; + spin_unlock(&obd->obd_recovery_task_lock); + + /* Let's check if the request has been redone by + * update replay */ + if (is_req_replayed_by_update(req)) { + struct distribute_txn_replay_req *dtrq; + + dtrq = distribute_txn_lookup_finish_list(tdtd, + req->rq_xid); + LASSERT(dtrq != NULL); + spin_lock(&tdtd->tdtd_replay_list_lock); + list_del_init(&dtrq->dtrq_list); + spin_unlock(&tdtd->tdtd_replay_list_lock); + dtrq_destroy(dtrq); + + drop_duplicate_replay_req(env, obd, req); + + continue; + } + + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + handle_recovery_req(thread, req, + trd->trd_recovery_handler); + /** + * bz18031: increase next_recovery_transno before + * target_request_copy_put() will drop exp_rpc reference + */ + spin_lock(&obd->obd_recovery_task_lock); + obd->obd_next_recovery_transno++; + spin_unlock(&obd->obd_recovery_task_lock); + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; + } else if (type == UPDATE_RECOVERY && transno != 0) { + struct distribute_txn_replay_req *dtrq; + + spin_unlock(&obd->obd_recovery_task_lock); + + LASSERT(tdtd != NULL); + dtrq = distribute_txn_get_next_req(tdtd); + lu_context_enter(&thread->t_env->le_ctx); + tdtd->tdtd_replay_handler(env, tdtd, dtrq); + lu_context_exit(&thread->t_env->le_ctx); + extend_recovery_timer(obd, obd_timeout, true); + + /* 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, + dtrq->dtrq_master_transno); + list_add(&dtrq->dtrq_list, + &tdtd->tdtd_replay_finish_list); + } else { + dtrq_destroy(dtrq); + } + spin_unlock(&tdtd->tdtd_replay_list_lock); + + spin_lock(&obd->obd_recovery_task_lock); + if (transno == obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno++; + else if (transno > obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno = transno + 1; + spin_unlock(&obd->obd_recovery_task_lock); + + } else { + spin_unlock(&obd->obd_recovery_task_lock); + LASSERT(list_empty(&obd->obd_req_replay_queue)); + LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); + /** evict exports failed VBR */ + class_disconnect_stale_exports(obd, exp_vbr_healthy); + break; + } + } while (1); } static int target_recovery_thread(void *arg) @@ -1998,68 +2344,54 @@ static int target_recovery_thread(void *arg) thread->t_env = env; thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */ env->le_ctx.lc_thread = thread; - thread->t_data = NULL; - thread->t_watchdog = NULL; + 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, - cfs_curproc_pid()); - trd->trd_processing_task = cfs_curproc_pid(); + CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, + current_pid()); + trd->trd_processing_task = current_pid(); spin_lock(&obd->obd_dev_lock); obd->obd_recovering = 1; spin_unlock(&obd->obd_dev_lock); complete(&trd->trd_starting); - /* first of all, we have to know the first transno to replay */ - if (target_recovery_overseer(obd, check_for_clients, - exp_connect_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } - - /* next stage: replay requests */ - delta = jiffies; - CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", - cfs_atomic_read(&obd->obd_req_replay_clients), - obd->obd_next_recovery_transno); - while ((req = target_next_replay_req(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", - lustre_msg_get_transno(req->rq_reqmsg), - libcfs_nid2str(req->rq_peer.nid)); - handle_recovery_req(thread, req, - trd->trd_recovery_handler); - /** - * bz18031: increase next_recovery_transno before - * target_request_copy_put() will drop exp_rpc reference - */ - spin_lock(&obd->obd_recovery_task_lock); - obd->obd_next_recovery_transno++; - spin_unlock(&obd->obd_recovery_task_lock); - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - obd->obd_replayed_requests++; - } + /* first of all, we have to know the first transno to replay */ + if (target_recovery_overseer(lut, check_for_recovery_ready, + exp_connect_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + if (lut->lut_tdtd != NULL) + dtrq_list_destroy(lut->lut_tdtd); + } - /** - * The second stage: replay locks - */ - CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", - cfs_atomic_read(&obd->obd_lock_replay_clients)); - while ((req = target_next_replay_lock(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing lock from %s: ", - libcfs_nid2str(req->rq_peer.nid)); - handle_recovery_req(thread, req, - trd->trd_recovery_handler); - target_request_copy_put(req); - obd->obd_replayed_locks++; - } + /* next stage: replay requests or update */ + delta = jiffies; + CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", + atomic_read(&obd->obd_req_replay_clients), + obd->obd_next_recovery_transno); + replay_request_or_update(env, lut, trd, thread); + + /** + * The second stage: replay locks + */ + CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", + atomic_read(&obd->obd_lock_replay_clients)); + while ((req = target_next_replay_lock(lut))) { + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing lock from %s: ", + libcfs_nid2str(req->rq_peer.nid)); + handle_recovery_req(thread, req, + trd->trd_recovery_handler); + target_request_copy_put(req); + obd->obd_replayed_locks++; + } /** * 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); @@ -2071,32 +2403,37 @@ static int target_recovery_thread(void *arg) spin_lock(&obd->obd_recovery_task_lock); target_cancel_recovery_timer(obd); spin_unlock(&obd->obd_recovery_task_lock); - while ((req = target_next_final_ping(obd))) { - LASSERT(trd->trd_processing_task == cfs_curproc_pid()); - DEBUG_REQ(D_HA, req, "processing final ping from %s: ", - libcfs_nid2str(req->rq_peer.nid)); + while ((req = target_next_final_ping(obd))) { + LASSERT(trd->trd_processing_task == current_pid()); + 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); - target_request_copy_put(req); - } + /* 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 */ + ptlrpc_update_export_timer(req->rq_export, 0); + target_request_copy_put(req); + } - delta = (jiffies - delta) / CFS_HZ; - 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(); - } + delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC; + CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n", + delta, obd->obd_replayed_requests, obd->obd_replayed_locks); + if (delta > OBD_RECOVERY_TIME_SOFT) { + CWARN("too long recovery - read logs\n"); + libcfs_debug_dumplog(); + } - target_finish_recovery(obd); + target_finish_recovery(lut); lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; complete(&trd->trd_finishing); - OBD_FREE_PTR(thread); - OBD_FREE_PTR(env); - RETURN(rc); + tgt_io_thread_done(thread); + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); + RETURN(rc); } static int target_start_recovery_thread(struct lu_target *lut, @@ -2131,7 +2468,7 @@ void target_stop_recovery_thread(struct obd_device *obd) if (obd->obd_recovering) { CERROR("%s: Aborting recovery\n", obd->obd_name); obd->obd_abort_recovery = 1; - cfs_waitq_signal(&obd->obd_next_transno_waitq); + wake_up(&obd->obd_next_transno_waitq); } spin_unlock(&obd->obd_dev_lock); wait_for_completion(&trd->trd_finishing); @@ -2149,20 +2486,21 @@ EXPORT_SYMBOL(target_recovery_fini); static void target_recovery_expired(unsigned long castmeharder) { - struct obd_device *obd = (struct obd_device *)castmeharder; - CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery" - " after %lds (%d clients connected)\n", - obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients), - cfs_time_current_sec()- obd->obd_recovery_start, - cfs_atomic_read(&obd->obd_connected_clients)); - - obd->obd_recovery_expired = 1; - cfs_waitq_signal(&obd->obd_next_transno_waitq); + 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", + obd->obd_name, atomic_read(&obd->obd_lock_replay_clients), + cfs_time_current_sec()- obd->obd_recovery_start, + atomic_read(&obd->obd_connected_clients)); + + obd->obd_recovery_expired = 1; + wake_up(&obd->obd_next_transno_waitq); } void target_recovery_init(struct lu_target *lut, svc_handler_t handler) { struct obd_device *obd = lut->lut_obd; + if (obd->obd_max_recoverable_clients == 0) { /** Update server last boot epoch */ tgt_boot_epoch_update(lut); @@ -2182,8 +2520,6 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) } EXPORT_SYMBOL(target_recovery_init); -#endif /* __KERNEL__ */ - static int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req) { @@ -2197,7 +2533,7 @@ static int target_process_req_flags(struct obd_device *obd, spin_unlock(&exp->exp_lock); LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients); - cfs_atomic_dec(&obd->obd_req_replay_clients); + atomic_dec(&obd->obd_req_replay_clients); } else { spin_unlock(&exp->exp_lock); } @@ -2211,7 +2547,7 @@ static int target_process_req_flags(struct obd_device *obd, spin_unlock(&exp->exp_lock); LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients); - cfs_atomic_dec(&obd->obd_lock_replay_clients); + atomic_dec(&obd->obd_lock_replay_clients); } else { spin_unlock(&exp->exp_lock); } @@ -2222,27 +2558,27 @@ static int target_process_req_flags(struct obd_device *obd, int target_queue_recovery_request(struct ptlrpc_request *req, struct obd_device *obd) { - cfs_list_t *tmp; - int inserted = 0; __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); - ENTRY; + struct ptlrpc_request *reqiter; + int inserted = 0; + ENTRY; - if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) { - /* Processing the queue right now, don't re-add. */ - RETURN(1); - } + if (obd->obd_recovery_data.trd_processing_task == current_pid()) { + /* Processing the queue right now, don't re-add. */ + RETURN(1); + } target_process_req_flags(obd, req); 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 */ - target_request_copy_get(req); - DEBUG_REQ(D_HA, req, "queue final req"); - cfs_waitq_signal(&obd->obd_next_transno_waitq); + 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) { - cfs_list_add_tail(&req->rq_list, + list_add_tail(&req->rq_list, &obd->obd_final_req_queue); } else { spin_unlock(&obd->obd_recovery_task_lock); @@ -2256,7 +2592,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, /* client declares he's ready to replay locks */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue lock replay req"); - cfs_waitq_signal(&obd->obd_next_transno_waitq); + wake_up(&obd->obd_next_transno_waitq); spin_lock(&obd->obd_recovery_task_lock); LASSERT(obd->obd_recovering); /* usually due to recovery abort */ @@ -2266,7 +2602,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(-ENOTCONN); } LASSERT(req->rq_export->exp_lock_replay_needed); - cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); + list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); spin_unlock(&obd->obd_recovery_task_lock); RETURN(0); } @@ -2276,7 +2612,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */ if (!transno) { - CFS_INIT_LIST_HEAD(&req->rq_list); + INIT_LIST_HEAD(&req->rq_list); DEBUG_REQ(D_HA, req, "not queueing"); RETURN(1); } @@ -2294,10 +2630,15 @@ int target_queue_recovery_request(struct ptlrpc_request *req, 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 + * request to the recovery thread (replay_request_or_update()), where + * it will be handled */ spin_lock(&obd->obd_recovery_task_lock); - if (transno < obd->obd_next_recovery_transno) { + if (transno < obd->obd_next_recovery_transno && + !is_req_replayed_by_update(req)) { /* Processing the queue right now, don't re-add. */ - LASSERT(cfs_list_empty(&req->rq_list)); + LASSERT(list_empty(&req->rq_list)); spin_unlock(&obd->obd_recovery_task_lock); RETURN(1); } @@ -2319,18 +2660,15 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(0); } - /* XXX O(n^2) */ + /* XXX O(n^2) */ spin_lock(&obd->obd_recovery_task_lock); - LASSERT(obd->obd_recovering); - cfs_list_for_each(tmp, &obd->obd_req_replay_queue) { - struct ptlrpc_request *reqiter = - cfs_list_entry(tmp, struct ptlrpc_request, rq_list); - - if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) { - cfs_list_add_tail(&req->rq_list, &reqiter->rq_list); - inserted = 1; - break; - } + LASSERT(obd->obd_recovering); + list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) { + if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) { + list_add_tail(&req->rq_list, &reqiter->rq_list); + inserted = 1; + goto added; + } if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) == transno)) { @@ -2342,23 +2680,21 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(0); } } - +added: if (!inserted) - cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); + list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); obd->obd_requests_queued_for_recovery++; spin_unlock(&obd->obd_recovery_task_lock); - cfs_waitq_signal(&obd->obd_next_transno_waitq); + wake_up(&obd->obd_next_transno_waitq); RETURN(0); } -EXPORT_SYMBOL(target_queue_recovery_request); int target_handle_ping(struct ptlrpc_request *req) { obd_ping(req->rq_svc_thread->t_env, req->rq_export); return req_capsule_server_pack(&req->rq_pill); } -EXPORT_SYMBOL(target_handle_ping); void target_committed_to_req(struct ptlrpc_request *req) { @@ -2375,7 +2711,6 @@ void target_committed_to_req(struct ptlrpc_request *req) CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", exp->exp_last_committed, req->rq_transno, req->rq_xid); } -EXPORT_SYMBOL(target_committed_to_req); #endif /* HAVE_SERVER_SUPPORT */ @@ -2406,24 +2741,27 @@ int target_pack_pool_reply(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(target_pack_pool_reply); -int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) +static int target_send_reply_msg(struct ptlrpc_request *req, + int rc, int fail_id) { - if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) { - DEBUG_REQ(D_ERROR, req, "dropping reply"); - return (-ECOMM); - } + if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) { + DEBUG_REQ(D_ERROR, req, "dropping reply"); + return -ECOMM; + } + if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT && + OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP))) + return -ECOMM; - if (unlikely(rc)) { - DEBUG_REQ(D_NET, req, "processing error (%d)", rc); - req->rq_status = rc; - return (ptlrpc_send_error(req, 1)); - } else { - DEBUG_REQ(D_NET, req, "sending reply"); - } + if (unlikely(rc)) { + DEBUG_REQ(D_NET, req, "processing error (%d)", rc); + req->rq_status = rc; + return ptlrpc_send_error(req, 1); + } else { + DEBUG_REQ(D_NET, req, "sending reply"); + } - return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT)); + return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT); } void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) @@ -2449,20 +2787,20 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) } /* must be an export if locks saved */ - LASSERT (req->rq_export != NULL); + LASSERT(req->rq_export != NULL); /* req/reply consistent */ LASSERT(rs->rs_svcpt == svcpt); /* "fresh" reply */ - LASSERT (!rs->rs_scheduled); - LASSERT (!rs->rs_scheduled_ever); - LASSERT (!rs->rs_handled); - LASSERT (!rs->rs_on_net); - LASSERT (rs->rs_export == NULL); - LASSERT (cfs_list_empty(&rs->rs_obd_list)); - LASSERT (cfs_list_empty(&rs->rs_exp_list)); + LASSERT(!rs->rs_scheduled); + LASSERT(!rs->rs_scheduled_ever); + LASSERT(!rs->rs_handled); + LASSERT(!rs->rs_on_net); + LASSERT(rs->rs_export == NULL); + LASSERT(list_empty(&rs->rs_obd_list)); + LASSERT(list_empty(&rs->rs_exp_list)); - exp = class_export_get (req->rq_export); + exp = class_export_get(req->rq_export); /* disable reply scheduling while I'm setting up */ rs->rs_scheduled = 1; @@ -2477,69 +2815,69 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) rs->rs_transno, exp->exp_last_committed); if (rs->rs_transno > exp->exp_last_committed) { /* not committed already */ - cfs_list_add_tail(&rs->rs_obd_list, + list_add_tail(&rs->rs_obd_list, &exp->exp_uncommitted_replies); } spin_unlock(&exp->exp_uncommitted_replies_lock); spin_lock(&exp->exp_lock); - cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies); + list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies); spin_unlock(&exp->exp_lock); netrc = target_send_reply_msg(req, rc, fail_id); spin_lock(&svcpt->scp_rep_lock); - cfs_atomic_inc(&svcpt->scp_nreps_difficult); + atomic_inc(&svcpt->scp_nreps_difficult); - if (netrc != 0) { - /* 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) */ - rs->rs_on_net = 0; - ptlrpc_rs_addref(rs); - } + if (netrc != 0) { + /* 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) */ + rs->rs_on_net = 0; + ptlrpc_rs_addref(rs); + } spin_lock(&rs->rs_lock); if (rs->rs_transno <= exp->exp_last_committed || (!rs->rs_on_net && !rs->rs_no_ack) || - cfs_list_empty(&rs->rs_exp_list) || /* completed already */ - cfs_list_empty(&rs->rs_obd_list)) { + list_empty(&rs->rs_exp_list) || /* completed already */ + list_empty(&rs->rs_obd_list)) { CDEBUG(D_HA, "Schedule reply immediately\n"); ptlrpc_dispatch_difficult_reply(rs); } else { - cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active); + list_add(&rs->rs_list, &svcpt->scp_rep_active); rs->rs_scheduled = 0; /* allow notifier to schedule */ } spin_unlock(&rs->rs_lock); spin_unlock(&svcpt->scp_rep_lock); EXIT; } -EXPORT_SYMBOL(target_send_reply); - -ldlm_mode_t lck_compat_array[] = { - [LCK_EX] LCK_COMPAT_EX, - [LCK_PW] LCK_COMPAT_PW, - [LCK_PR] LCK_COMPAT_PR, - [LCK_CW] LCK_COMPAT_CW, - [LCK_CR] LCK_COMPAT_CR, - [LCK_NL] LCK_COMPAT_NL, - [LCK_GROUP] LCK_COMPAT_GROUP, - [LCK_COS] LCK_COMPAT_COS, + +enum ldlm_mode lck_compat_array[] = { + [LCK_EX] = LCK_COMPAT_EX, + [LCK_PW] = LCK_COMPAT_PW, + [LCK_PR] = LCK_COMPAT_PR, + [LCK_CW] = LCK_COMPAT_CW, + [LCK_CR] = LCK_COMPAT_CR, + [LCK_NL] = LCK_COMPAT_NL, + [LCK_GROUP] = LCK_COMPAT_GROUP, + [LCK_COS] = LCK_COMPAT_COS, }; /** * Rather arbitrary mapping from LDLM error codes to errno values. This should * not escape to the user level. */ -int ldlm_error2errno(ldlm_error_t error) +int ldlm_error2errno(enum ldlm_error error) { - int result; + int result; - switch (error) { - case ELDLM_OK: + switch (error) { + case ELDLM_OK: + case ELDLM_LOCK_MATCHED: result = 0; break; case ELDLM_LOCK_CHANGED: @@ -2560,22 +2898,22 @@ int ldlm_error2errno(ldlm_error_t error) case ELDLM_BAD_NAMESPACE: result = -EBADF; break; - default: - if (((int)error) < 0) /* cast to signed type */ - result = error; /* as ldlm_error_t can be unsigned */ - else { - CERROR("Invalid DLM result code: %d\n", error); - result = -EPROTO; - } - } - return result; + default: + if (((int)error) < 0) { /* cast to signed type */ + result = error; /* as ldlm_error can be unsigned */ + } else { + CERROR("Invalid DLM result code: %d\n", error); + result = -EPROTO; + } + } + return result; } EXPORT_SYMBOL(ldlm_error2errno); /** - * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t. + * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error. */ -ldlm_error_t ldlm_errno2error(int err_no) +enum ldlm_error ldlm_errno2error(int err_no) { int error; @@ -2606,18 +2944,17 @@ ldlm_error_t ldlm_errno2error(int err_no) } return error; } -EXPORT_SYMBOL(ldlm_errno2error); #if LUSTRE_TRACKS_LOCK_EXP_REFS void ldlm_dump_export_locks(struct obd_export *exp) { spin_lock(&exp->exp_locks_list_guard); - if (!cfs_list_empty(&exp->exp_locks_list)) { + 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); - cfs_list_for_each_entry(lock, &exp->exp_locks_list, + list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link) LDLM_ERROR(lock, "lock:"); } @@ -2635,97 +2972,126 @@ static int target_bulk_timeout(void *data) RETURN(1); } -static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) +static inline const char *bulk2type(struct ptlrpc_request *req) { - return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT"; + if (req->rq_bulk_read) + return "READ"; + if (req->rq_bulk_write) + return "WRITE"; + return "UNKNOWN"; } int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, struct l_wait_info *lwi) { - struct ptlrpc_request *req = desc->bd_req; - int rc = 0; - ENTRY; + struct ptlrpc_request *req = desc->bd_req; + time_t start = cfs_time_current_sec(); + time_t deadline; + int rc = 0; + + ENTRY; /* If there is eviction in progress, wait for it to finish. */ - if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) { - *lwi = LWI_INTR(NULL, NULL); - rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, - !cfs_atomic_read(&exp->exp_obd-> - obd_evict_inprogress), - lwi); - } + 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); + } - /* Check if client was evicted or tried to reconnect already. */ - if (exp->exp_failed || exp->exp_abort_active_req) { - rc = -ENOTCONN; - } else { - if (desc->bd_type == BULK_PUT_SINK) - rc = sptlrpc_svc_wrap_bulk(req, desc); - if (rc == 0) - rc = ptlrpc_start_bulk_transfer(desc); - } + /* Check if client was evicted or reconnected already. */ + if (exp->exp_failed || + exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { + rc = -ENOTCONN; + } else { + if (req->rq_bulk_read) + rc = sptlrpc_svc_wrap_bulk(req, desc); - if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) { - ptlrpc_abort_bulk(desc); - } else if (rc == 0) { - time_t start = cfs_time_current_sec(); - do { - long timeoutl = req->rq_deadline - cfs_time_current_sec(); - cfs_duration_t timeout = timeoutl <= 0 ? - CFS_TICK : cfs_time_seconds(timeoutl); - *lwi = LWI_TIMEOUT_INTERVAL(timeout, - cfs_time_seconds(1), - target_bulk_timeout, - desc); - rc = l_wait_event(desc->bd_waitq, - !ptlrpc_server_bulk_active(desc) || - exp->exp_failed || - exp->exp_abort_active_req, - lwi); - LASSERT(rc == 0 || rc == -ETIMEDOUT); - /* Wait again if we changed deadline. */ - } while ((rc == -ETIMEDOUT) && - (req->rq_deadline > cfs_time_current_sec())); - - if (rc == -ETIMEDOUT) { - DEBUG_REQ(D_ERROR, req, - "timeout on bulk %s after %ld%+lds", - bulk2type(desc), - req->rq_deadline - start, - cfs_time_current_sec() - - req->rq_deadline); - ptlrpc_abort_bulk(desc); - } else if (exp->exp_failed) { - DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s", - bulk2type(desc)); - rc = -ENOTCONN; - ptlrpc_abort_bulk(desc); - } else if (exp->exp_abort_active_req) { - DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s", - bulk2type(desc)); - /* We don't reply anyway. */ - rc = -ETIMEDOUT; - ptlrpc_abort_bulk(desc); - } else if (desc->bd_failure || - desc->bd_nob_transferred != desc->bd_nob) { - DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", - desc->bd_failure ? - "network error on" : "truncated", - bulk2type(desc), - desc->bd_nob_transferred, + if ((exp->exp_connect_data.ocd_connect_flags & + OBD_CONNECT_BULK_MBITS) != 0) + req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg); + else /* old version, bulk matchbits is rq_xid */ + req->rq_mbits = req->rq_xid; + + if (rc == 0) + rc = ptlrpc_start_bulk_transfer(desc); + } + + if (rc < 0) { + DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", + bulk2type(req), rc); + RETURN(rc); + } + + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) { + ptlrpc_abort_bulk(desc); + RETURN(0); + } + + /* limit actual bulk transfer to bulk_timeout seconds */ + deadline = start + bulk_timeout; + if (deadline > req->rq_deadline) + 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); + /* Wait again if we changed rq_deadline. */ + rq_deadline = ACCESS_ONCE(req->rq_deadline); + deadline = start + bulk_timeout; + if (deadline > rq_deadline) + deadline = rq_deadline; + } while ((rc == -ETIMEDOUT) && + (deadline > cfs_time_current_sec())); + + if (rc == -ETIMEDOUT) { + DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds", + bulk2type(req), deadline - start, + cfs_time_current_sec() - deadline); + ptlrpc_abort_bulk(desc); + } else if (exp->exp_failed) { + DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s", + bulk2type(req)); + rc = -ENOTCONN; + ptlrpc_abort_bulk(desc); + } else if (exp->exp_conn_cnt > + lustre_msg_get_conn_cnt(req->rq_reqmsg)) { + DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s", + bulk2type(req)); + /* We don't reply anyway. */ + rc = -ETIMEDOUT; + ptlrpc_abort_bulk(desc); + } else if (desc->bd_failure) { + DEBUG_REQ(D_ERROR, req, "network error on bulk %s", + bulk2type(req)); + /* XXX should this be a different errno? */ + rc = -ETIMEDOUT; + } else { + if (req->rq_bulk_write) + rc = sptlrpc_svc_unwrap_bulk(req, desc); + if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)", + bulk2type(req), desc->bd_nob_transferred, desc->bd_nob); - /* XXX Should this be a different errno? */ + /* XXX should this be a different errno? */ rc = -ETIMEDOUT; - } else if (desc->bd_type == BULK_GET_SINK) { - rc = sptlrpc_svc_unwrap_bulk(req, desc); - } - } else { - DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", - bulk2type(desc), rc); - } + } + } - RETURN(rc); + RETURN(rc); } EXPORT_SYMBOL(target_bulk_io);