X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=95e566d44a78aa3a3f18d494d1f2452651de3b2e;hp=1c370e6e68d426ae4bd8c6ac4123f47976255f0d;hb=bb75072cb679bf52e00537c19e42f8e4e95255b6;hpb=bcec0e8d347a5f01552fd966e1bbc167e33ac1a4 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 1c370e6..95e566d 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, 2015, 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 @@ -170,14 +167,17 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) ptlrpc_connection_put(imp->imp_connection); imp->imp_connection = NULL; - dlmexp = class_conn2export(&imp->imp_dlm_handle); - if (dlmexp && dlmexp->exp_connection) { - LASSERT(dlmexp->exp_connection == - imp_conn->oic_conn); - ptlrpc_connection_put(dlmexp->exp_connection); - dlmexp->exp_connection = NULL; - } - } + dlmexp = class_conn2export(&imp->imp_dlm_handle); + if (dlmexp && dlmexp->exp_connection) { + LASSERT(dlmexp->exp_connection == + imp_conn->oic_conn); + ptlrpc_connection_put(dlmexp->exp_connection); + dlmexp->exp_connection = NULL; + } + + if (dlmexp != NULL) + class_export_put(dlmexp); + } list_del(&imp_conn->oic_item); ptlrpc_connection_put(imp_conn->oic_conn); @@ -263,17 +263,18 @@ 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; - 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. */ + /* 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; @@ -286,7 +287,12 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) rq_portal = MDS_REQUEST_PORTAL; rp_portal = MDC_REPLY_PORTAL; connect_op = MDS_CONNECT; - cli->cl_sp_me = LUSTRE_SP_CLI; + if (is_lwp_on_ost(cli_name)) + cli->cl_sp_me = LUSTRE_SP_OST; + else if (is_lwp_on_mdt(cli_name)) + cli->cl_sp_me = LUSTRE_SP_MDT; + else + cli->cl_sp_me = LUSTRE_SP_CLI; cli->cl_sp_to = LUSTRE_SP_MDT; ns_type = LDLM_NS_TYPE_MDC; } else if (!strcmp(name, LUSTRE_OSP_NAME)) { @@ -304,7 +310,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) rq_portal = OST_REQUEST_PORTAL; } rp_portal = OSC_REPLY_PORTAL; - cli->cl_sp_me = LUSTRE_SP_CLI; + cli->cl_sp_me = LUSTRE_SP_MDT; } else if (!strcmp(name, LUSTRE_MGC_NAME)) { rq_portal = MGS_REQUEST_PORTAL; rp_portal = MGC_REPLY_PORTAL; @@ -357,7 +363,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list); INIT_LIST_HEAD(&cli->cl_loi_write_list); INIT_LIST_HEAD(&cli->cl_loi_read_list); - client_obd_list_lock_init(&cli->cl_loi_list_lock); + spin_lock_init(&cli->cl_loi_list_lock); atomic_set(&cli->cl_pending_w_pages, 0); atomic_set(&cli->cl_pending_r_pages, 0); cli->cl_r_in_flight = 0; @@ -373,11 +379,12 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) /* lru for osc. */ INIT_LIST_HEAD(&cli->cl_lru_osc); atomic_set(&cli->cl_lru_shrinkers, 0); - atomic_set(&cli->cl_lru_busy, 0); - atomic_set(&cli->cl_lru_in_list, 0); + atomic_long_set(&cli->cl_lru_busy, 0); + atomic_long_set(&cli->cl_lru_in_list, 0); INIT_LIST_HEAD(&cli->cl_lru_list); - client_obd_list_lock_init(&cli->cl_lru_list_lock); - atomic_set(&cli->cl_unstable_count, 0); + spin_lock_init(&cli->cl_lru_list_lock); + atomic_long_set(&cli->cl_unstable_count, 0); + INIT_LIST_HEAD(&cli->cl_shrink_list); init_waitqueue_head(&cli->cl_destroy_waitq); atomic_set(&cli->cl_destroy_in_flight, 0); @@ -393,24 +400,21 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) #endif atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); - /* This value may be reduced at connect time in - * ptlrpc_connect_interpret() . We initialize it to only - * 1MB until we know what the performance looks like. - * In the future this should likely be increased. LU-1431 */ - cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES, - LNET_MTU >> PAGE_CACHE_SHIFT); + /* Set it to possible maximum size. It may be reduced by ocd_brw_size + * from OFD after connecting. */ + cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES; - /* set cl_chunkbits default value to PAGE_CACHE_SHIFT, + /* set cl_chunkbits default value to PAGE_SHIFT, * it will be updated at OSC connection time. */ - cli->cl_chunkbits = PAGE_CACHE_SHIFT; + cli->cl_chunkbits = PAGE_SHIFT; if (!strcmp(name, LUSTRE_MDC_NAME)) { cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) { cli->cl_max_rpcs_in_flight = 2; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) { cli->cl_max_rpcs_in_flight = 3; - } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) { + } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) { cli->cl_max_rpcs_in_flight = 4; } else { if (osc_on_mdt(obddev->obd_name)) @@ -418,6 +422,23 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) else cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT; } + + spin_lock_init(&cli->cl_mod_rpcs_lock); + spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock); + cli->cl_max_mod_rpcs_in_flight = 0; + cli->cl_mod_rpcs_in_flight = 0; + cli->cl_close_rpcs_in_flight = 0; + init_waitqueue_head(&cli->cl_mod_rpcs_waitq); + cli->cl_mod_tag_bitmap = NULL; + + 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); @@ -442,10 +463,9 @@ 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")) { @@ -468,8 +488,6 @@ 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: @@ -477,6 +495,10 @@ 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); } @@ -484,6 +506,7 @@ EXPORT_SYMBOL(client_obd_setup); int client_obd_cleanup(struct obd_device *obddev) { + struct client_obd *cli = &obddev->u.cli; ENTRY; ldlm_namespace_free_post(obddev->obd_namespace); @@ -493,6 +516,12 @@ int client_obd_cleanup(struct obd_device *obddev) LASSERT(obddev->u.cli.cl_import == NULL); ldlm_put_ref(); + + if (cli->cl_mod_tag_bitmap != NULL) + OBD_FREE(cli->cl_mod_tag_bitmap, + BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long)); + cli->cl_mod_tag_bitmap = NULL; + RETURN(0); } EXPORT_SYMBOL(client_obd_cleanup); @@ -508,57 +537,69 @@ 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; + *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); - if (rc) - GOTO(out_sem, rc); - - cli->cl_conn_count++; - *exp = class_conn2export(&conn); - - LASSERT(obd->obd_namespace); - - imp->imp_dlm_handle = conn; - rc = ptlrpc_init_import(imp); - if (rc != 0) - GOTO(out_ldlm, rc); - - ocd = &imp->imp_connect_data; - if (data) { - *ocd = *data; - imp->imp_connect_flags_orig = data->ocd_connect_flags; - } + rc = class_connect(&conn, obd, cluuid); + if (rc) + GOTO(out_sem, rc); + + cli->cl_conn_count++; + *exp = class_conn2export(&conn); + + LASSERT(obd->obd_namespace); + + imp->imp_dlm_handle = conn; + rc = ptlrpc_init_import(imp); + if (rc != 0) + GOTO(out_ldlm, rc); + + ocd = &imp->imp_connect_data; + if (data) { + *ocd = *data; + is_mdc = strncmp(imp->imp_obd->obd_type->typ_name, + LUSTRE_MDC_NAME, 3) == 0; + if (is_mdc) + data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS; + imp->imp_connect_flags_orig = data->ocd_connect_flags; + imp->imp_connect_flags2_orig = data->ocd_connect_flags2; + } - rc = ptlrpc_connect_import(imp); - if (rc != 0) { - LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); - GOTO(out_ldlm, rc); - } + rc = ptlrpc_connect_import(imp); + if (rc != 0) { + if (data && is_mdc) + data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS; + LASSERT(imp->imp_state == LUSTRE_IMP_DISCON); + GOTO(out_ldlm, rc); + } LASSERT(*exp != NULL && (*exp)->exp_connection); - if (data) { - LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) == - ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n", - data->ocd_connect_flags, ocd->ocd_connect_flags); - data->ocd_connect_flags = ocd->ocd_connect_flags; - } + if (data) { + LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) == + ocd->ocd_connect_flags, "old %#llx, new %#llx\n", + data->ocd_connect_flags, ocd->ocd_connect_flags); + data->ocd_connect_flags = ocd->ocd_connect_flags; + /* 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); + ptlrpc_pinger_add_import(imp); - EXIT; + EXIT; - if (rc) { + if (rc) { out_ldlm: - cli->cl_conn_count--; - class_disconnect(*exp); - *exp = NULL; - } + cli->cl_conn_count--; + class_disconnect(*exp); + *exp = NULL; + } out_sem: up_write(&cli->cl_sem); @@ -575,7 +616,7 @@ int client_disconnect_export(struct obd_export *exp) ENTRY; if (!obd) { - CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n", + CERROR("invalid export for disconnect: exp %p cookie %#llx\n", exp, exp ? exp->exp_handle.h_cookie : -1); RETURN(-EINVAL); } @@ -584,17 +625,17 @@ 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 @@ -651,8 +692,7 @@ int server_disconnect_export(struct obd_export *exp) if (exp->exp_imp_reverse) ptlrpc_cleanup_imp(exp->exp_imp_reverse); - if (exp->exp_obd->obd_namespace != NULL) - ldlm_cancel_locks_for_export(exp); + ldlm_bl_thread_wakeup(); /* complete all outstanding replies */ spin_lock(&exp->exp_lock); @@ -685,62 +725,95 @@ static int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp, struct obd_uuid *cluuid) { - ENTRY; + struct obd_device *target; + struct lustre_handle *hdl; + cfs_time_t now; + cfs_time_t deadline; + int timeout; + int rc = 0; + ENTRY; - if (exp->exp_connection && exp->exp_imp_reverse) { - struct lustre_handle *hdl; - struct obd_device *target; - - hdl = &exp->exp_imp_reverse->imp_remote_handle; - target = exp->exp_obd; - - /* Might be a re-connect after a partition. */ - if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) { - if (target->obd_recovering) { - int timeout = cfs_duration_sec(cfs_time_sub( - cfs_timer_deadline( - &target->obd_recovery_timer), - cfs_time_current())); - - LCONSOLE_WARN("%s: Client %s (at %s) reconnect" - "ing, waiting for %d clients in recov" - "ery for %d:%.02d\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - target->obd_max_recoverable_clients, - timeout / 60, timeout % 60); - } else { - LCONSOLE_WARN("%s: Client %s (at %s) " - "reconnecting\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp)); - } + hdl = &exp->exp_imp_reverse->imp_remote_handle; + if (!exp->exp_connection || !lustre_handle_is_used(hdl)) { + conn->cookie = exp->exp_handle.h_cookie; + CDEBUG(D_HA, "connect export for UUID '%s' at %p," + " cookie %#llx\n", cluuid->uuid, exp, conn->cookie); + RETURN(0); + } - conn->cookie = exp->exp_handle.h_cookie; - /* target_handle_connect() treats EALREADY and - * -EALREADY differently. EALREADY means we are - * doing a valid reconnect from the same client. */ - RETURN(EALREADY); - } else { - LCONSOLE_WARN("%s: already connected client %s (at %s) " - "with handle "LPX64". Rejecting client " - "with the same UUID trying to reconnect " - "with handle "LPX64"\n", target->obd_name, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp), - hdl->cookie, conn->cookie); - memset(conn, 0, sizeof *conn); - /* target_handle_connect() treats EALREADY and - * -EALREADY differently. -EALREADY is an error - * (same UUID, different handle). */ - RETURN(-EALREADY); - } - } + target = exp->exp_obd; + + /* Might be a re-connect after a partition. */ + if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) { + LCONSOLE_WARN("%s: already connected client %s (at %s) " + "with handle %#llx. Rejecting client " + "with the same UUID trying to reconnect " + "with handle %#llx\n", target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), + hdl->cookie, conn->cookie); + memset(conn, 0, sizeof *conn); + /* target_handle_connect() treats EALREADY and + * -EALREADY differently. -EALREADY is an error + * (same UUID, different handle). */ + RETURN(-EALREADY); + } - conn->cookie = exp->exp_handle.h_cookie; - CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n", - cluuid->uuid, exp, conn->cookie); - RETURN(0); + if (!target->obd_recovering) { + LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n", + target->obd_name, obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + GOTO(out_already, rc); + } + + now = cfs_time_current(); + deadline = cfs_timer_deadline(&target->obd_recovery_timer); + if (cfs_time_before(now, deadline)) { + struct target_distribute_txn_data *tdtd = + class_exp2tgt(exp)->lut_tdtd; + int size = 0; + int count = 0; + char *buf = NULL; + + timeout = cfs_duration_sec(cfs_time_sub(deadline, now)); + if (tdtd && tdtd->tdtd_show_update_logs_retrievers) + buf = tdtd->tdtd_show_update_logs_retrievers( + tdtd->tdtd_show_retrievers_cbdata, + &size, &count); + + if (count > 0) + LCONSOLE_WARN("%s: Recovery already passed deadline " + "%d:%.02d. It is due to DNE recovery " + "failed/stuck on the %d MDT(s):%s. " + "Please wait until all MDTs recovered " + "or abort the recovery by force.\n", + target->obd_name, timeout / 60, + timeout % 60, count, + buf ? buf : "unknown (not enough RAM)"); + else + LCONSOLE_WARN("%s: Recovery already passed deadline " + "%d:%.02d. If you do not want to wait " + "more, please abort the recovery by " + "force.\n", target->obd_name, + timeout / 60, timeout % 60); + + if (buf != NULL) + OBD_FREE(buf, size); + } else { + timeout = cfs_duration_sec(cfs_time_sub(now, deadline)); + LCONSOLE_WARN("%s: Recovery already passed deadline" + " %d:%.02d, It is most likely due to DNE" + " recovery is failed or stuck, please wait a" + " few more minutes or abort the recovery.\n", + target->obd_name, timeout / 60, timeout % 60); + } + +out_already: + conn->cookie = exp->exp_handle.h_cookie; + /* target_handle_connect() treats EALREADY and + * -EALREADY differently. EALREADY means we are + * doing a valid reconnect from the same client. */ + RETURN(EALREADY); } void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, @@ -756,36 +829,137 @@ 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; + bool target_referenced = false; struct obd_connect_data *data, *tmpdata; int size, tmpsize; lnet_nid_t *client_nid = NULL; @@ -839,7 +1013,8 @@ int target_handle_connect(struct ptlrpc_request *req) /* Make sure the target isn't cleaned up while we're here. Yes, * there's still a race between the above check and our incref here. * Really, class_uuid2obd should take the ref. */ - targref = class_incref(target, __FUNCTION__, current); + class_incref(target, __func__, current); + target_referenced = true; target->obd_conn_inprogress++; spin_unlock(&target->obd_dev_lock); @@ -852,21 +1027,6 @@ int target_handle_connect(struct ptlrpc_request *req) obd_str2uuid(&cluuid, str); - /* XXX Extract a nettype and format accordingly. */ - switch (sizeof(lnet_nid_t)) { - /* NB the casts only avoid compiler warnings. */ - case 8: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid); - break; - case 4: - snprintf(remote_uuid.uuid, sizeof remote_uuid, - "NET_%x_UUID", (__u32)req->rq_peer.nid); - break; - default: - LBUG(); - } - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); if (tmp == NULL) GOTO(out, rc = -EPROTO); @@ -883,6 +1043,20 @@ int target_handle_connect(struct ptlrpc_request *req) if (rc) GOTO(out, rc); +#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0) + /* Don't allow clients to connect that are using old 1.8 format + * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18, + * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc). The + * FULL20 flag should be set on all connections since 2.0, but no + * longer affects behaviour. + * + * Later this check will be disabled and the flag can be retired + * completely once interop with 3.0 is no longer needed. + */ + if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20)) + GOTO(out, rc = -EPROTO); +#endif + if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { if (data->ocd_version < LUSTRE_VERSION_CODE - LUSTRE_VERSION_ALLOWED_OFFSET || @@ -908,10 +1082,47 @@ int target_handle_connect(struct ptlrpc_request *req) } } + /* Note: lw_client is needed in MDS-MDS failover during update log + * processing, so we needs to allow lw_client to be connected at + * anytime, instead of only the initial connection */ + lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0; + if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) { mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0; - lw_client = (data->ocd_connect_flags & - OBD_CONNECT_LIGHTWEIGHT) != 0; + mds_mds_conn = (data->ocd_connect_flags & + OBD_CONNECT_MDS_MDS) != 0; + + /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS + * for Imperative Recovery connection from MGC to MGS. + * + * Via check OBD_CONNECT_FID, we can distinguish whether + * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from + * MGC or MDT. */ + if (!lw_client && + (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && + (data->ocd_connect_flags & OBD_CONNECT_FID) && + (data->ocd_connect_flags & OBD_CONNECT_VERSION)) { + __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version); + __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version); + __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version); + + /* We do not support the MDT-MDT interoperations with + * different version MDT because of protocol changes. */ + if (unlikely(major != LUSTRE_MAJOR || + minor != LUSTRE_MINOR || + abs(patch - LUSTRE_PATCH) > 3)) { + LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the " + "connection from different version MDT " + "(%d.%d.%d.%d) %s %s\n", + target->obd_name, LUSTRE_MAJOR, + LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX, + major, minor, patch, + OBD_OCD_VERSION_FIX(data->ocd_version), + libcfs_nid2str(req->rq_peer.nid), str); + + GOTO(out, rc = -EPROTO); + } + } } /* lctl gets a backstage, all-access pass. */ @@ -934,25 +1145,43 @@ int target_handle_connect(struct ptlrpc_request *req) class_export_put(export); export = NULL; rc = -EALREADY; - } else if ((mds_conn || lw_client) && export->exp_connection != NULL) { + } else if ((mds_conn || lw_client || + data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) && + export->exp_connection != NULL) { spin_unlock(&export->exp_lock); - if (req->rq_peer.nid != export->exp_connection->c_peer.nid) + if (req->rq_peer.nid != export->exp_connection->c_peer.nid) { /* MDS or LWP reconnected after failover. */ LCONSOLE_WARN("%s: Received %s connection from " "%s, removing former export from %s\n", target->obd_name, mds_conn ? "MDS" : "LWP", libcfs_nid2str(req->rq_peer.nid), libcfs_nid2str(export->exp_connection->c_peer.nid)); - else + } else { /* New MDS connection from the same NID. */ LCONSOLE_WARN("%s: Received new %s connection from " "%s, removing former export from same NID\n", target->obd_name, mds_conn ? "MDS" : "LWP", libcfs_nid2str(req->rq_peer.nid)); - class_fail_export(export); - class_export_put(export); - export = NULL; - rc = 0; + } + + if (req->rq_peer.nid == export->exp_connection->c_peer.nid && + data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) { + /* Because exports between MDTs will always be + * kept, let's do not fail such export if they + * come from the same NID, otherwise it might + * cause eviction between MDTs, which might + * cause namespace inconsistency */ + spin_lock(&export->exp_lock); + export->exp_connecting = 1; + spin_unlock(&export->exp_lock); + conn.cookie = export->exp_handle.h_cookie; + rc = EALREADY; + } else { + class_fail_export(export); + class_export_put(export); + export = NULL; + rc = 0; + } } else if (export->exp_connection != NULL && req->rq_peer.nid != export->exp_connection->c_peer.nid && (lustre_msg_get_op_flags(req->rq_reqmsg) & @@ -1004,7 +1233,7 @@ no_export: GOTO(out, rc); } - CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n", + CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), target->obd_recovering ? "recovering/" : "", data->ocd_transno, export, (long)cfs_time_current_sec(), @@ -1031,7 +1260,9 @@ no_export: if (export == NULL) { /* allow lightweight connections during recovery */ - if (target->obd_recovering && !lw_client) { + /* allow "new" MDT to be connected during recovery, since we + * need retrieve recovery update records from it */ + if (target->obd_recovering && !lw_client && !mds_mds_conn) { cfs_time_t t; int c; /* connected */ int i; /* in progress */ @@ -1045,49 +1276,40 @@ no_export: t = cfs_timer_deadline(&target->obd_recovery_timer); t = cfs_time_sub(t, cfs_time_current()); t = cfs_duration_sec(t); - LCONSOLE_WARN("%s: Denying connection for new client " - "%s (at %s), waiting for all %d known " - "clients (%d recovered, %d in progress, " - "and %d evicted) to recover in %d:%.02d\n", + LCONSOLE_WARN("%s: Denying connection for new client %s" + "(at %s), waiting for %d known clients " + "(%d recovered, %d in progress, and %d " + "evicted) to recover in %d:%.02d\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), k, c - i, i, s, (int)t / 60, (int)t % 60); - rc = -EBUSY; - } else { + rc = -EBUSY; + } else { dont_check_exports: - rc = obd_connect(req->rq_svc_thread->t_env, - &export, target, &cluuid, data, - client_nid); + rc = obd_connect(req->rq_svc_thread->t_env, + &export, target, &cluuid, data, + client_nid); if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG)) lustre_msg_add_op_flags(req->rq_repmsg, - MSG_CONNECT_RECOVERING); - if (rc == 0) - conn.cookie = export->exp_handle.h_cookie; - } - } else { - rc = obd_reconnect(req->rq_svc_thread->t_env, - export, target, &cluuid, data, client_nid); - } - if (rc) - GOTO(out, rc); + MSG_CONNECT_RECOVERING); + if (rc == 0) { + conn.cookie = export->exp_handle.h_cookie; + rc = rev_import_init(export); + } -#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 52, 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 + 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 +1324,6 @@ 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. @@ -1153,14 +1370,19 @@ dont_check_exports: ptlrpc_connection_put(export->exp_connection); } - export->exp_connection = ptlrpc_connection_get(req->rq_peer, - req->rq_self, - &remote_uuid); - if (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; @@ -1195,6 +1417,14 @@ dont_check_exports: atomic_inc(&target->obd_req_replay_clients); atomic_inc(&target->obd_lock_replay_clients); + /* Note: MDS-MDS connection is allowed to be connected during + * recovery, no matter if the exports needs to be recoveried. + * Because we need retrieve updates logs from all other MDTs. + * So if the MDS-MDS export is new, obd_max_recoverable_clients + * also needs to be increased to match other recovery checking + * condition. */ + if (new_mds_mds_conn) + target->obd_max_recoverable_clients++; if (atomic_inc_return(&target->obd_connected_clients) == target->obd_max_recoverable_clients) wake_up(&target->obd_next_transno_waitq); @@ -1204,59 +1434,7 @@ dont_check_exports: if (target->obd_recovering && !lw_client) lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); - conn = *tmp; - - /* Return -ENOTCONN in case of errors to let client reconnect. */ - revimp = class_new_import(target); - if (revimp == NULL) { - CERROR("fail to alloc new reverse import.\n"); - GOTO(out, rc = -ENOTCONN); - } - - spin_lock(&export->exp_lock); - if (export->exp_imp_reverse != NULL) - /* destroyed import can be still referenced in ctxt */ - tmp_imp = export->exp_imp_reverse; - export->exp_imp_reverse = revimp; - spin_unlock(&export->exp_lock); - - revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection); - revimp->imp_client = &export->exp_obd->obd_ldlm_client; - revimp->imp_remote_handle = conn; - revimp->imp_dlm_fake = 1; - revimp->imp_state = LUSTRE_IMP_FULL; - - /* Unknown versions will be caught in - * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */ - revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; - - if ((data->ocd_connect_flags & OBD_CONNECT_AT) && - (revimp->imp_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; @@ -1264,18 +1442,16 @@ out: class_export_put(export); } - if (targref) { + if (target_referenced == true && target != NULL) { spin_lock(&target->obd_dev_lock); target->obd_conn_inprogress--; spin_unlock(&target->obd_dev_lock); - class_decref(targref, __func__, current); + class_decref(target, __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) { @@ -1291,7 +1467,6 @@ int target_handle_disconnect(struct ptlrpc_request *req) RETURN(0); } -EXPORT_SYMBOL(target_handle_disconnect); void target_destroy_export(struct obd_export *exp) { @@ -1344,6 +1519,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); struct obd_export *exp = req->rq_export; struct ptlrpc_request *reqiter; + struct ptlrpc_request *dup_req = NULL; int dup = 0; LASSERT(exp); @@ -1352,6 +1528,7 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) list_for_each_entry(reqiter, &exp->exp_req_replay_queue, rq_replay_list) { if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { + dup_req = reqiter; dup = 1; break; } @@ -1363,6 +1540,16 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY)) CERROR("invalid flags %x of resent replay\n", lustre_msg_get_flags(req->rq_reqmsg)); + + if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) { + __u32 new_conn; + + new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg); + if (new_conn > + lustre_msg_get_conn_cnt(dup_req->rq_reqmsg)) + lustre_msg_set_conn_cnt(dup_req->rq_reqmsg, + new_conn); + } } else { list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue); @@ -1382,9 +1569,9 @@ static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) 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. */ @@ -1483,7 +1670,6 @@ void target_cleanup_recovery(struct obd_device *obd) { struct ptlrpc_request *req, *n; struct list_head clean_list; - ENTRY; INIT_LIST_HEAD(&clean_list); spin_lock(&obd->obd_dev_lock); @@ -1501,7 +1687,7 @@ void target_cleanup_recovery(struct obd_device *obd) spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); + LASSERT(req->rq_reply_state == NULL); target_exp_dequeue_req_replay(req); target_request_copy_put(req); } @@ -1512,7 +1698,7 @@ void target_cleanup_recovery(struct obd_device *obd) spin_unlock(&obd->obd_recovery_task_lock); list_for_each_entry_safe(req, n, &clean_list, rq_list) { - LASSERT(req->rq_reply_state == 0); + LASSERT(req->rq_reply_state == NULL); target_request_copy_put(req); } @@ -1526,7 +1712,6 @@ 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) { @@ -1590,22 +1775,20 @@ 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 */ @@ -1662,6 +1845,14 @@ static inline int exp_req_replay_healthy(struct obd_export *exp) return (!exp->exp_req_replay_needed || atomic_read(&exp->exp_replay_count) > 0); } + + +static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp) +{ + return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) || + exp_req_replay_healthy(exp); +} + /** if export done lock_replay or has replay in queue */ static inline int exp_lock_replay_healthy(struct obd_export *exp) { @@ -1679,23 +1870,21 @@ static inline int exp_finished(struct obd_export *exp) return (exp->exp_in_recovery && !exp->exp_lock_replay_needed); } -/** Checking routines for recovery */ -static int check_for_clients(struct obd_device *obd) +static inline int exp_finished_or_from_mdt(struct obd_export *exp) { - unsigned int clnts = atomic_read(&obd->obd_connected_clients); - - if (obd->obd_abort_recovery || obd->obd_recovery_expired) - return 1; - LASSERT(clnts <= obd->obd_max_recoverable_clients); - return (clnts + obd->obd_stale_clients == - obd->obd_max_recoverable_clients); + return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) || + exp_finished(exp); } -static int check_for_next_transno(struct obd_device *obd) +static int check_for_next_transno(struct lu_target *lut) { struct ptlrpc_request *req = NULL; + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; int wake_up = 0, connected, completed, queue_len; - __u64 next_transno, req_transno; + __u64 req_transno = 0; + __u64 update_transno = 0; + __u64 next_transno = 0; ENTRY; spin_lock(&obd->obd_recovery_task_lock); @@ -1703,17 +1892,18 @@ static int check_for_next_transno(struct obd_device *obd) req = list_entry(obd->obd_req_replay_queue.next, struct ptlrpc_request, rq_list); req_transno = lustre_msg_get_transno(req->rq_reqmsg); - } else { - req_transno = 0; } + if (tdtd != NULL) + update_transno = distribute_txn_get_next_transno(tdtd); + connected = atomic_read(&obd->obd_connected_clients); completed = connected - atomic_read(&obd->obd_req_replay_clients); queue_len = obd->obd_requests_queued_for_recovery; next_transno = obd->obd_next_recovery_transno; CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, " - "req_transno: "LPU64", next_transno: "LPU64"\n", + "req_transno: %llu, next_transno: %llu\n", obd->obd_max_recoverable_clients, connected, completed, queue_len, req_transno, next_transno); @@ -1723,26 +1913,30 @@ static int check_for_next_transno(struct obd_device *obd) } else if (obd->obd_recovery_expired) { CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; - } else if (req_transno == next_transno) { - CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno); + } else if (tdtd != NULL && req != NULL && + is_req_replayed_by_update(req)) { + LASSERTF(req_transno < next_transno, "req_transno %llu" + "next_transno%llu\n", req_transno, next_transno); + CDEBUG(D_HA, "waking for duplicate req (%llu)\n", + req_transno); + wake_up = 1; + } else if (req_transno == next_transno || + (update_transno != 0 && update_transno <= next_transno)) { + CDEBUG(D_HA, "waking for next (%lld)\n", next_transno); wake_up = 1; } else if (queue_len > 0 && queue_len == atomic_read(&obd->obd_req_replay_clients)) { - int d_lvl = D_HA; /** handle gaps occured due to lost reply or VBR */ LASSERTF(req_transno >= next_transno, - "req_transno: "LPU64", next_transno: "LPU64"\n", + "req_transno: %llu, next_transno: %llu\n", req_transno, next_transno); - if (req_transno > obd->obd_last_committed && - !obd->obd_version_recov) - d_lvl = D_ERROR; - CDEBUG(d_lvl, + CDEBUG(D_HA, "%s: waking for gap in transno, VBR is %s (skip: " - LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64 - ", last_committed: "LPD64")\n", + "%lld, ql: %d, comp: %d, conn: %d, next: %lld" + ", next_update %lld last_committed: %lld)\n", obd->obd_name, obd->obd_version_recov ? "ON" : "OFF", next_transno, queue_len, completed, connected, - req_transno, obd->obd_last_committed); + req_transno, update_transno, obd->obd_last_committed); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } else if (atomic_read(&obd->obd_req_replay_clients) == 0) { @@ -1750,7 +1944,7 @@ static int check_for_next_transno(struct obd_device *obd) wake_up = 1; } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) { CDEBUG(D_HA, "accepting transno gaps is explicitly allowed" - " by fail_lock, waking up ("LPD64")\n", next_transno); + " by fail_lock, waking up (%lld)\n", next_transno); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } @@ -1758,8 +1952,9 @@ static int check_for_next_transno(struct obd_device *obd) return wake_up; } -static int check_for_next_lock(struct obd_device *obd) +static int check_for_next_lock(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; int wake_up = 0; spin_lock(&obd->obd_recovery_task_lock); @@ -1786,14 +1981,87 @@ static int check_for_next_lock(struct obd_device *obd) * check its status with help of check_routine * evict dead clients via health_check */ -static int target_recovery_overseer(struct obd_device *obd, - int (*check_routine)(struct obd_device *), +static int target_recovery_overseer(struct lu_target *lut, + int (*check_routine)(struct lu_target *), int (*health_check)(struct obd_export *)) { + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd; repeat: - wait_event(obd->obd_next_transno_waitq, check_routine(obd)); + if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >= + (obd->obd_recovery_start + obd->obd_recovery_time_hard))) { + __u64 next_update_transno = 0; + + /* Only abort the recovery if there are no update recovery + * left in the queue */ + spin_lock(&obd->obd_recovery_task_lock); + if (lut->lut_tdtd != NULL) { + next_update_transno = + distribute_txn_get_next_transno(lut->lut_tdtd); + + tdtd = lut->lut_tdtd; + /* If next_update_transno == 0, it probably because + * updatelog retrieve threads did not get any records + * yet, let's wait those threads stopped */ + if (next_update_transno == 0) { + struct l_wait_info lwi = { 0 }; + + l_wait_event(tdtd->tdtd_recovery_threads_waitq, + atomic_read( + &tdtd->tdtd_recovery_threads_count) == 0, + &lwi); + + next_update_transno = + distribute_txn_get_next_transno( + lut->lut_tdtd); + } + } + + if (next_update_transno != 0 && !obd->obd_abort_recovery) { + obd->obd_next_recovery_transno = next_update_transno; + spin_unlock(&obd->obd_recovery_task_lock); + /* Disconnect unfinished exports from clients, and + * keep connection from MDT to make sure the update + * recovery will still keep trying until some one + * manually abort the recovery */ + class_disconnect_stale_exports(obd, + exp_finished_or_from_mdt); + /* Abort all of replay and replay lock req from + * clients */ + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + CDEBUG(D_HA, "%s: there are still update replay (%#llx" + ")in the queue.\n", obd->obd_name, + next_update_transno); + } else { + obd->obd_abort_recovery = 1; + spin_unlock(&obd->obd_recovery_task_lock); + CWARN("%s recovery is aborted by hard timeout\n", + obd->obd_name); + } + } + + while (wait_event_timeout(obd->obd_next_transno_waitq, + check_routine(lut), + msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0) + /* wait indefinitely for event, but don't trigger watchdog */; + if (obd->obd_abort_recovery) { CWARN("recovery is aborted, evict exports in recovery\n"); + if (lut->lut_tdtd != NULL) { + struct l_wait_info lwi = { 0 }; + + tdtd = lut->lut_tdtd; + /* Let's wait all of the update log recovery thread + * finished */ + l_wait_event(tdtd->tdtd_recovery_threads_waitq, + atomic_read(&tdtd->tdtd_recovery_threads_count) == 0, + &lwi); + /* Then abort the update recovery list */ + dtrq_list_dump(lut->lut_tdtd, D_ERROR); + dtrq_list_destroy(lut->lut_tdtd); + } + /** evict exports which didn't finish recovery yet */ class_disconnect_stale_exports(obd, exp_finished); return 1; @@ -1804,6 +2072,7 @@ repeat: "evict stale exports\n", obd->obd_name); /** evict cexports with no replay in queue, they are stalled */ class_disconnect_stale_exports(obd, health_check); + /** continue with VBR */ spin_lock(&obd->obd_dev_lock); obd->obd_version_recov = 1; @@ -1819,43 +2088,13 @@ repeat: return 0; } -static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) -{ - struct ptlrpc_request *req = NULL; - ENTRY; - - CDEBUG(D_HA, "Waiting for transno "LPD64"\n", - obd->obd_next_recovery_transno); - - if (target_recovery_overseer(obd, check_for_next_transno, - exp_req_replay_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } - - spin_lock(&obd->obd_recovery_task_lock); - if (!list_empty(&obd->obd_req_replay_queue)) { - req = list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); - list_del_init(&req->rq_list); - obd->obd_requests_queued_for_recovery--; - spin_unlock(&obd->obd_recovery_task_lock); - } else { - spin_unlock(&obd->obd_recovery_task_lock); - LASSERT(list_empty(&obd->obd_req_replay_queue)); - LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); - /** evict exports failed VBR */ - class_disconnect_stale_exports(obd, exp_vbr_healthy); - } - RETURN(req); -} - -static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) +static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut) { + struct obd_device *obd = lut->lut_obd; struct ptlrpc_request *req = NULL; CDEBUG(D_HA, "Waiting for lock\n"); - if (target_recovery_overseer(obd, check_for_next_lock, + if (target_recovery_overseer(lut, check_for_next_lock, exp_lock_replay_healthy)) abort_lock_replay_queue(obd); @@ -1949,6 +2188,256 @@ static void handle_recovery_req(struct ptlrpc_thread *thread, EXIT; } +/** Checking routines for recovery */ +static int check_for_recovery_ready(struct lu_target *lut) +{ + struct obd_device *obd = lut->lut_obd; + unsigned int clnts = atomic_read(&obd->obd_connected_clients); + + CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d" + " abort %d expired %d\n", clnts, obd->obd_stale_clients, + obd->obd_max_recoverable_clients, obd->obd_abort_recovery, + obd->obd_recovery_expired); + + if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) { + LASSERT(clnts <= obd->obd_max_recoverable_clients); + if (clnts + obd->obd_stale_clients < + obd->obd_max_recoverable_clients) + return 0; + } + + if (lut->lut_tdtd != NULL) { + if (!lut->lut_tdtd->tdtd_replay_ready && + !obd->obd_abort_recovery) { + /* Let's extend recovery timer, in case the recovery + * timer expired, and some clients got evicted */ + extend_recovery_timer(obd, obd->obd_recovery_timeout, + true); + CDEBUG(D_HA, "%s update recovery is not ready," + " extend recovery %d\n", obd->obd_name, + obd->obd_recovery_timeout); + 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; +} + +static __u64 get_next_transno(struct lu_target *lut, int *type) +{ + struct obd_device *obd = lut->lut_obd; + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + __u64 transno = 0; + __u64 update_transno; + ENTRY; + + transno = get_next_replay_req_transno(obd); + if (type != NULL) + *type = REQUEST_RECOVERY; + + if (tdtd == NULL) + RETURN(transno); + + update_transno = distribute_txn_get_next_transno(tdtd); + if (transno == 0 || (transno >= update_transno && + update_transno != 0)) { + transno = update_transno; + if (type != NULL) + *type = UPDATE_RECOVERY; + } + + RETURN(transno); +} + +/** + * drop duplicate replay request + * + * Because the operation has been replayed by update recovery, the request + * with the same transno will be dropped and also notify the client to send + * next replay request. + * + * \param[in] env execution environment + * \param[in] obd failover obd device + * \param[in] req request to be dropped + */ +static void drop_duplicate_replay_req(struct lu_env *env, + struct obd_device *obd, + struct ptlrpc_request *req) +{ + DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate" + " update records are found.\n", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + /* Right now, only for MDS reint operation update replay and + * normal request replay can have the same transno */ + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) { + req_capsule_set(&req->rq_pill, &RQF_MDS_REINT); + req->rq_status = req_capsule_server_pack(&req->rq_pill); + if (likely(req->rq_export)) + target_committed_to_req(req); + lustre_msg_set_transno(req->rq_repmsg, req->rq_transno); + target_send_reply(req, req->rq_status, 0); + } else { + DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n", + libcfs_nid2str(req->rq_peer.nid)); + } + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; +} + +static void replay_request_or_update(struct lu_env *env, + struct lu_target *lut, + struct target_recovery_data *trd, + struct ptlrpc_thread *thread) +{ + struct obd_device *obd = lut->lut_obd; + struct ptlrpc_request *req = NULL; + int type; + __u64 transno; + ENTRY; + + CDEBUG(D_HA, "Waiting for transno %lld\n", + obd->obd_next_recovery_transno); + + /* Replay all of request and update by transno */ + do { + struct target_distribute_txn_data *tdtd = lut->lut_tdtd; + + CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val); + + /** It is needed to extend recovery window above + * recovery_time_soft. Extending is possible only in the + * end of recovery window (see more details in + * handle_recovery_req()). + */ + CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300); + + if (target_recovery_overseer(lut, check_for_next_transno, + exp_req_replay_healthy_or_from_mdt)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + goto abort; + } + + spin_lock(&obd->obd_recovery_task_lock); + transno = get_next_transno(lut, &type); + if (type == REQUEST_RECOVERY && transno != 0) { + /* Drop replay request from client side, if the + * replay has been executed by update with the + * same transno */ + req = list_entry(obd->obd_req_replay_queue.next, + struct ptlrpc_request, rq_list); + + list_del_init(&req->rq_list); + obd->obd_requests_queued_for_recovery--; + spin_unlock(&obd->obd_recovery_task_lock); + + /* Let's check if the request has been redone by + * update replay */ + if (is_req_replayed_by_update(req)) { + struct distribute_txn_replay_req *dtrq; + + dtrq = distribute_txn_lookup_finish_list(tdtd, + req->rq_xid); + LASSERT(dtrq != NULL); + spin_lock(&tdtd->tdtd_replay_list_lock); + list_del_init(&dtrq->dtrq_list); + spin_unlock(&tdtd->tdtd_replay_list_lock); + dtrq_destroy(dtrq); + + drop_duplicate_replay_req(env, obd, req); + + continue; + } + + LASSERT(trd->trd_processing_task == current_pid()); + DEBUG_REQ(D_HA, req, "processing t%lld from %s", + lustre_msg_get_transno(req->rq_reqmsg), + libcfs_nid2str(req->rq_peer.nid)); + + handle_recovery_req(thread, req, + trd->trd_recovery_handler); + /** + * bz18031: increase next_recovery_transno before + * target_request_copy_put() will drop exp_rpc reference + */ + spin_lock(&obd->obd_recovery_task_lock); + obd->obd_next_recovery_transno++; + spin_unlock(&obd->obd_recovery_task_lock); + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; + } else if (type == UPDATE_RECOVERY && transno != 0) { + struct distribute_txn_replay_req *dtrq; + int rc; + + spin_unlock(&obd->obd_recovery_task_lock); + + LASSERT(tdtd != NULL); + dtrq = distribute_txn_get_next_req(tdtd); + lu_context_enter(&thread->t_env->le_ctx); + rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq); + lu_context_exit(&thread->t_env->le_ctx); + extend_recovery_timer(obd, obd_timeout, true); + + if (rc == 0 && dtrq->dtrq_xid != 0) { + CDEBUG(D_HA, "Move x%llu t%llu" + " to finish list\n", dtrq->dtrq_xid, + dtrq->dtrq_master_transno); + + /* Add it to the replay finish list */ + spin_lock(&tdtd->tdtd_replay_list_lock); + list_add(&dtrq->dtrq_list, + &tdtd->tdtd_replay_finish_list); + spin_unlock(&tdtd->tdtd_replay_list_lock); + + spin_lock(&obd->obd_recovery_task_lock); + if (transno == obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno++; + else if (transno > + obd->obd_next_recovery_transno) + obd->obd_next_recovery_transno = + transno + 1; + spin_unlock(&obd->obd_recovery_task_lock); + } else { + dtrq_destroy(dtrq); + } + } else { + spin_unlock(&obd->obd_recovery_task_lock); +abort: + LASSERT(list_empty(&obd->obd_req_replay_queue)); + LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); + /** evict exports failed VBR */ + class_disconnect_stale_exports(obd, exp_vbr_healthy); + break; + } + } while (1); +} + static int target_recovery_thread(void *arg) { struct lu_target *lut = arg; @@ -1994,43 +2483,28 @@ static int target_recovery_thread(void *arg) spin_unlock(&obd->obd_dev_lock); complete(&trd->trd_starting); - /* first of all, we have to know the first transno to replay */ - if (target_recovery_overseer(obd, check_for_clients, - exp_connect_healthy)) { - abort_req_replay_queue(obd); - abort_lock_replay_queue(obd); - } + /* first of all, we have to know the first transno to replay */ + if (target_recovery_overseer(lut, check_for_recovery_ready, + exp_connect_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + if (lut->lut_tdtd != NULL) + dtrq_list_destroy(lut->lut_tdtd); + } - /* next stage: replay requests */ + /* next stage: replay requests or update */ delta = jiffies; - CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", + CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n", atomic_read(&obd->obd_req_replay_clients), obd->obd_next_recovery_transno); - while ((req = target_next_replay_req(obd))) { - LASSERT(trd->trd_processing_task == current_pid()); - DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s", - lustre_msg_get_transno(req->rq_reqmsg), - libcfs_nid2str(req->rq_peer.nid)); - handle_recovery_req(thread, req, - trd->trd_recovery_handler); - /** - * bz18031: increase next_recovery_transno before - * target_request_copy_put() will drop exp_rpc reference - */ - spin_lock(&obd->obd_recovery_task_lock); - obd->obd_next_recovery_transno++; - spin_unlock(&obd->obd_recovery_task_lock); - target_exp_dequeue_req_replay(req); - target_request_copy_put(req); - obd->obd_replayed_requests++; - } + replay_request_or_update(env, lut, trd, thread); /** * The second stage: replay locks */ CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", atomic_read(&obd->obd_lock_replay_clients)); - while ((req = target_next_replay_lock(obd))) { + while ((req = target_next_replay_lock(lut))) { LASSERT(trd->trd_processing_task == current_pid()); DEBUG_REQ(D_HA, req, "processing lock from %s: ", libcfs_nid2str(req->rq_peer.nid)); @@ -2044,6 +2518,7 @@ static int target_recovery_thread(void *arg) * The third stage: reply on final pings, at this moment all clients * must have request in final queue */ + CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val); CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); /** Update server last boot epoch */ tgt_boot_epoch_update(lut); @@ -2066,9 +2541,9 @@ static int target_recovery_thread(void *arg) * export is being evicted */ ptlrpc_update_export_timer(req->rq_export, 0); target_request_copy_put(req); - } + } - delta = (jiffies - delta) / HZ; + delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC; CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n", delta, obd->obd_replayed_requests, obd->obd_replayed_locks); if (delta > OBD_RECOVERY_TIME_SOFT) { @@ -2076,7 +2551,7 @@ static int target_recovery_thread(void *arg) libcfs_debug_dumplog(); } - target_finish_recovery(obd); + target_finish_recovery(lut); lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; @@ -2094,14 +2569,19 @@ static int target_start_recovery_thread(struct lu_target *lut, struct obd_device *obd = lut->lut_obd; int rc = 0; struct target_recovery_data *trd = &obd->obd_recovery_data; + int index; memset(trd, 0, sizeof(*trd)); init_completion(&trd->trd_starting); init_completion(&trd->trd_finishing); trd->trd_recovery_handler = handler; + rc = server_name2index(obd->obd_name, &index, NULL); + if (rc < 0) + return rc; + if (!IS_ERR(kthread_run(target_recovery_thread, - lut, "tgt_recov"))) { + lut, "tgt_recover_%d", index))) { wait_for_completion(&trd->trd_starting); LASSERT(obd->obd_recovering != 0); } else { @@ -2152,6 +2632,7 @@ static void target_recovery_expired(unsigned long castmeharder) void target_recovery_init(struct lu_target *lut, svc_handler_t handler) { struct obd_device *obd = lut->lut_obd; + if (obd->obd_max_recoverable_clients == 0) { /** Update server last boot epoch */ tgt_boot_epoch_update(lut); @@ -2159,7 +2640,7 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) } CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, " - "last_transno "LPU64"\n", obd->obd_name, + "last_transno %llu\n", obd->obd_name, obd->obd_max_recoverable_clients, obd->obd_last_committed); LASSERT(obd->obd_stopping == 0); obd->obd_next_recovery_transno = obd->obd_last_committed + 1; @@ -2171,8 +2652,6 @@ 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) { @@ -2231,15 +2710,46 @@ int target_queue_recovery_request(struct ptlrpc_request *req, wake_up(&obd->obd_next_transno_waitq); spin_lock(&obd->obd_recovery_task_lock); if (obd->obd_recovering) { + struct ptlrpc_request *tmp; + struct ptlrpc_request *duplicate = NULL; + + if (likely(!req->rq_export->exp_replay_done)) { + req->rq_export->exp_replay_done = 1; + list_add_tail(&req->rq_list, + &obd->obd_final_req_queue); + spin_unlock(&obd->obd_recovery_task_lock); + RETURN(0); + } + + /* XXX O(n), but only happens if final ping is + * timed out, probably reorganize the list as + * a hash list later */ + list_for_each_entry_safe(reqiter, tmp, + &obd->obd_final_req_queue, + rq_list) { + if (reqiter->rq_export == req->rq_export) { + list_del_init(&reqiter->rq_list); + duplicate = reqiter; + break; + } + } + list_add_tail(&req->rq_list, - &obd->obd_final_req_queue); + &obd->obd_final_req_queue); + req->rq_export->exp_replay_done = 1; + spin_unlock(&obd->obd_recovery_task_lock); + + if (duplicate != NULL) { + DEBUG_REQ(D_HA, duplicate, + "put prev final req\n"); + target_request_copy_put(duplicate); + } + RETURN(0); } else { spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(obd->obd_stopping ? -ENOTCONN : 1); } - spin_unlock(&obd->obd_recovery_task_lock); - RETURN(0); } if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { /* client declares he's ready to replay locks */ @@ -2280,11 +2790,16 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * Also, a resent, replayed request that has already been * handled will pass through here and be processed immediately. */ - CDEBUG(D_HA, "Next recovery transno: "LPU64 - ", current: "LPU64", replaying\n", + CDEBUG(D_HA, "Next recovery transno: %llu" + ", current: %llu, replaying\n", obd->obd_next_recovery_transno, transno); + + /* If the request has been replayed by update replay, then sends this + * request to the recovery thread (replay_request_or_update()), where + * it will be handled */ spin_lock(&obd->obd_recovery_task_lock); - if (transno < obd->obd_next_recovery_transno) { + if (transno < obd->obd_next_recovery_transno && + !is_req_replayed_by_update(req)) { /* Processing the queue right now, don't re-add. */ LASSERT(list_empty(&req->rq_list)); spin_unlock(&obd->obd_recovery_task_lock); @@ -2337,14 +2852,12 @@ added: wake_up(&obd->obd_next_transno_waitq); RETURN(0); } -EXPORT_SYMBOL(target_queue_recovery_request); int target_handle_ping(struct ptlrpc_request *req) { obd_ping(req->rq_svc_thread->t_env, req->rq_export); return req_capsule_server_pack(&req->rq_pill); } -EXPORT_SYMBOL(target_handle_ping); void target_committed_to_req(struct ptlrpc_request *req) { @@ -2358,10 +2871,9 @@ void target_committed_to_req(struct ptlrpc_request *req) "%d)", exp->exp_obd->obd_no_transno, req->rq_repmsg == NULL); - CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", + CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n", exp->exp_last_committed, req->rq_transno, req->rq_xid); } -EXPORT_SYMBOL(target_committed_to_req); #endif /* HAVE_SERVER_SUPPORT */ @@ -2392,24 +2904,30 @@ 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 (unlikely(rc)) { - DEBUG_REQ(D_NET, req, "processing error (%d)", rc); - req->rq_status = rc; - return (ptlrpc_send_error(req, 1)); - } else { - DEBUG_REQ(D_NET, req, "sending reply"); - } + if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) { + DEBUG_REQ(D_ERROR, req, "dropping reply"); + return -ECOMM; + } + /* We can have a null rq_reqmsg in the event of bad signature or + * no context when unwrapping */ + if (req->rq_reqmsg && + unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT && + OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP))) + return -ECOMM; + + if (unlikely(rc)) { + DEBUG_REQ(D_NET, req, "processing error (%d)", rc); + req->rq_status = rc; + return ptlrpc_send_error(req, 1); + } else { + DEBUG_REQ(D_NET, req, "sending reply"); + } - return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT)); + return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT); } void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) @@ -2459,7 +2977,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg); spin_lock(&exp->exp_uncommitted_replies_lock); - CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n", + CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n", rs->rs_transno, exp->exp_last_committed); if (rs->rs_transno > exp->exp_last_committed) { /* not committed already */ @@ -2503,9 +3021,8 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) spin_unlock(&svcpt->scp_rep_lock); EXIT; } -EXPORT_SYMBOL(target_send_reply); -ldlm_mode_t lck_compat_array[] = { +enum ldlm_mode lck_compat_array[] = { [LCK_EX] = LCK_COMPAT_EX, [LCK_PW] = LCK_COMPAT_PW, [LCK_PR] = LCK_COMPAT_PR, @@ -2520,12 +3037,13 @@ ldlm_mode_t lck_compat_array[] = { * Rather arbitrary mapping from LDLM error codes to errno values. This should * not escape to the user level. */ -int ldlm_error2errno(ldlm_error_t error) +int ldlm_error2errno(enum ldlm_error error) { - int result; + int result; - switch (error) { - case ELDLM_OK: + switch (error) { + case ELDLM_OK: + case ELDLM_LOCK_MATCHED: result = 0; break; case ELDLM_LOCK_CHANGED: @@ -2546,22 +3064,22 @@ 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; @@ -2592,7 +3110,6 @@ ldlm_error_t ldlm_errno2error(int err_no) } return error; } -EXPORT_SYMBOL(ldlm_errno2error); #if LUSTRE_TRACKS_LOCK_EXP_REFS void ldlm_dump_export_locks(struct obd_export *exp) @@ -2621,9 +3138,13 @@ static int target_bulk_timeout(void *data) RETURN(1); } -static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) +static inline const char *bulk2type(struct ptlrpc_request *req) { - return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT"; + if (req->rq_bulk_read) + return "READ"; + if (req->rq_bulk_write) + return "WRITE"; + return "UNKNOWN"; } int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, @@ -2631,6 +3152,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, { struct ptlrpc_request *req = desc->bd_req; time_t start = cfs_time_current_sec(); + time_t deadline; int rc = 0; ENTRY; @@ -2640,7 +3162,7 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, *lwi = LWI_INTR(NULL, NULL); rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, !atomic_read(&exp->exp_obd-> - obd_evict_inprogress), + obd_evict_inprogress), lwi); } @@ -2649,15 +3171,21 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { rc = -ENOTCONN; } else { - if (desc->bd_type == BULK_PUT_SINK) + if (req->rq_bulk_read) rc = sptlrpc_svc_wrap_bulk(req, desc); + + if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS)) + req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg); + else /* old version, bulk matchbits is rq_xid */ + req->rq_mbits = req->rq_xid; + if (rc == 0) rc = ptlrpc_start_bulk_transfer(desc); } if (rc < 0) { DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d", - bulk2type(desc), rc); + bulk2type(req), rc); RETURN(rc); } @@ -2666,10 +3194,16 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *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 = req->rq_deadline - cfs_time_current_sec(); + 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); @@ -2680,37 +3214,46 @@ int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc, lustre_msg_get_conn_cnt(req->rq_reqmsg), lwi); LASSERT(rc == 0 || rc == -ETIMEDOUT); - /* Wait again if we changed deadline. */ + /* 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) && - (req->rq_deadline > cfs_time_current_sec())); + (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); + bulk2type(req), deadline - start, + cfs_time_current_sec() - deadline); ptlrpc_abort_bulk(desc); } else if (exp->exp_failed) { DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s", - bulk2type(desc)); + bulk2type(req)); rc = -ENOTCONN; ptlrpc_abort_bulk(desc); } else if (exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) { DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s", - bulk2type(desc)); + bulk2type(req)); /* We don't reply anyway. */ rc = -ETIMEDOUT; ptlrpc_abort_bulk(desc); - } else if (desc->bd_failure || - desc->bd_nob_transferred != desc->bd_nob) { - DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", - desc->bd_failure ? "network error on" : "truncated", - bulk2type(desc), desc->bd_nob_transferred, - desc->bd_nob); - /* XXX Should this be a different errno? */ + } else if (desc->bd_failure) { + DEBUG_REQ(D_ERROR, req, "network error on bulk %s", + bulk2type(req)); + /* XXX should this be a different errno? */ rc = -ETIMEDOUT; - } else if (desc->bd_type == BULK_GET_SINK) { - rc = sptlrpc_svc_unwrap_bulk(req, desc); + } else { + if (req->rq_bulk_write) + rc = sptlrpc_svc_unwrap_bulk(req, desc); + if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)", + bulk2type(req), desc->bd_nob_transferred, + desc->bd_nob); + /* XXX should this be a different errno? */ + rc = -ETIMEDOUT; + } } RETURN(rc);