X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fldlm%2Fldlm_lib.c;h=a7ada1b34434006c44823e85af2289923c887a74;hb=f9d1ac5ab499b0b9a31a957da35455515e567c49;hp=2de9fea187c38b8084066bedf368fbe7ec1d2f3b;hpb=c95d64fe1f2e415327e62c605e7adddc8284cd13;p=fs%2Flustre-release.git diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 2de9fea..a7ada1b 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,17 +24,16 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2012, Whamcloud, Inc. */ /* * This file is part of Lustre, http://www.lustre.org/ * Lustre is a trademark of Sun Microsystems, Inc. */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_LDLM #ifdef __KERNEL__ @@ -45,7 +42,7 @@ # include #endif #include -#include +#include #include #include #include @@ -80,18 +77,19 @@ static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid, } } - spin_lock(&imp->imp_lock); - list_for_each_entry(item, &imp->imp_conn_list, oic_item) { + cfs_spin_lock(&imp->imp_lock); + cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) { if (obd_uuid_equals(uuid, &item->oic_uuid)) { if (priority) { - list_del(&item->oic_item); - list_add(&item->oic_item, &imp->imp_conn_list); + cfs_list_del(&item->oic_item); + cfs_list_add(&item->oic_item, + &imp->imp_conn_list); item->oic_last_attempt = 0; } CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n", imp, imp->imp_obd->obd_name, uuid->uuid, (priority ? ", moved to head" : "")); - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); GOTO(out_free, rc = 0); } } @@ -101,18 +99,19 @@ 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) - list_add(&imp_conn->oic_item, &imp->imp_conn_list); + cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list); else - list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list); + cfs_list_add_tail(&imp_conn->oic_item, + &imp->imp_conn_list); CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n", imp, imp->imp_obd->obd_name, uuid->uuid, (priority ? "head" : "tail")); } else { - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); GOTO(out_free, rc = -ENOENT); } - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); RETURN(0); out_free: if (imp_conn) @@ -132,6 +131,7 @@ int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid, { return import_set_conn(imp, uuid, priority, 1); } +EXPORT_SYMBOL(client_import_add_conn); int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) { @@ -140,13 +140,13 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) int rc = -ENOENT; ENTRY; - spin_lock(&imp->imp_lock); - if (list_empty(&imp->imp_conn_list)) { + cfs_spin_lock(&imp->imp_lock); + if (cfs_list_empty(&imp->imp_conn_list)) { LASSERT(!imp->imp_connection); GOTO(out, rc); } - list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) { + cfs_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); @@ -173,7 +173,7 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) } } - list_del(&imp_conn->oic_item); + cfs_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", @@ -182,11 +182,37 @@ int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid) break; } out: - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); if (rc == -ENOENT) CERROR("connection %s not found\n", uuid->uuid); RETURN(rc); } +EXPORT_SYMBOL(client_import_del_conn); + +/** + * Find conn uuid by peer nid. @peer is a server nid. This function is used + * to find a conn uuid of @imp which can reach @peer. + */ +int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer, + struct obd_uuid *uuid) +{ + struct obd_import_conn *conn; + int rc = -ENOENT; + ENTRY; + + cfs_spin_lock(&imp->imp_lock); + cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) { + /* check if conn uuid does have this peer nid */ + if (class_check_uuid(&conn->oic_uuid, peer)) { + *uuid = conn->oic_uuid; + rc = 0; + break; + } + } + cfs_spin_unlock(&imp->imp_lock); + RETURN(rc); +} +EXPORT_SYMBOL(client_import_find_conn); void client_destroy_import(struct obd_import *imp) { @@ -197,6 +223,30 @@ void client_destroy_import(struct obd_import *imp) sptlrpc_import_sec_put(imp); class_import_put(imp); } +EXPORT_SYMBOL(client_destroy_import); + +/** + * check whether the osc is on MDT or not + * In the config log, + * osc on MDT + * setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID + * osc on client + * setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID + * + **/ +static int osc_on_mdt(char *obdname) +{ + char *ptr; + + ptr = strrchr(obdname, '-'); + if (ptr == NULL) + return 0; + + if (strncmp(ptr + 1, "MDT", 3) == 0) + return 1; + + return 0; +} /* configure an RPC client OBD device * @@ -212,23 +262,32 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) 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; /* 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)) { + if (!strcmp(name, LUSTRE_OSC_NAME) || + (!strcmp(name, LUSTRE_OSP_NAME) && + !is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) { 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; - } else if (!strcmp(name, LUSTRE_MDC_NAME)) { + ns_type = LDLM_NS_TYPE_OSC; + + } else if (!strcmp(name, LUSTRE_MDC_NAME) || + (!strcmp(name, LUSTRE_OSP_NAME) && + is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) { 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; + } else if (!strcmp(name, LUSTRE_MGC_NAME)) { rq_portal = MGS_REQUEST_PORTAL; rp_portal = MGC_REPLY_PORTAL; @@ -236,6 +295,16 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) cli->cl_sp_me = LUSTRE_SP_MGC; cli->cl_sp_to = LUSTRE_SP_MGS; cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID; + ns_type = LDLM_NS_TYPE_MGC; + + } else if (!strcmp(name, LUSTRE_OSP_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 { CERROR("unknown client OBD type \"%s\", can't setup\n", name); @@ -262,8 +331,8 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) RETURN(-EINVAL); } - init_rwsem(&cli->cl_sem); - sema_init(&cli->cl_mgc_sem, 1); + cfs_init_rwsem(&cli->cl_sem); + cfs_sema_init(&cli->cl_mgc_sem, 1); cli->cl_conn_count = 0; memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2), min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2), @@ -273,25 +342,36 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) 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 >> CFS_PAGE_SHIFT > num_physpages / 8) - cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3); + if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8) + cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_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_r_in_flight = 0; cli->cl_w_in_flight = 0; - spin_lock_init(&cli->cl_read_rpc_hist.oh_lock); - spin_lock_init(&cli->cl_write_rpc_hist.oh_lock); - spin_lock_init(&cli->cl_read_page_hist.oh_lock); - spin_lock_init(&cli->cl_write_page_hist.oh_lock); - spin_lock_init(&cli->cl_read_offset_hist.oh_lock); - spin_lock_init(&cli->cl_write_offset_hist.oh_lock); + cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock); + cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock); + cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock); + cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock); + cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock); + cfs_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); - atomic_set(&cli->cl_destroy_in_flight, 0); + cfs_atomic_set(&cli->cl_destroy_in_flight, 0); #ifdef ENABLE_CHECKSUM /* Turn on checksumming by default. */ cli->cl_checksum = 1; @@ -302,8 +382,7 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) */ cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32; #endif - atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); - atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS); + cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); /* This value may be changed at connect time in ptlrpc_connect_interpret. */ @@ -312,16 +391,18 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) if (!strcmp(name, LUSTRE_MDC_NAME)) { cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT; - } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) { + } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) { cli->cl_max_rpcs_in_flight = 2; - } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) { + } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) { cli->cl_max_rpcs_in_flight = 3; - } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) { + } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) { cli->cl_max_rpcs_in_flight = 4; } else { - cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT; + if (osc_on_mdt(obddev->obd_name)) + cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT; + else + cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT; } - rc = ldlm_get_ref(); if (rc) { CERROR("ldlm_get_ref failed: %d\n", rc); @@ -336,8 +417,6 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) GOTO(err_ldlm, rc = -ENOENT); imp->imp_client = &obddev->obd_ldlm_client; imp->imp_connect_op = connect_op; - imp->imp_initial_recov = 1; - imp->imp_initial_recov_bk = 0; CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain); memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1), LUSTRE_CFG_BUFLEN(lcfg, 1)); @@ -359,15 +438,16 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) CDEBUG(D_HA, "marking %s %s->%s as inactive\n", name, obddev->obd_name, cli->cl_target_uuid.uuid); - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); imp->imp_deactive = 1; - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); } } obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name, LDLM_NAMESPACE_CLIENT, - LDLM_NAMESPACE_GREEDY); + LDLM_NAMESPACE_GREEDY, + ns_type); if (obddev->obd_namespace == NULL) { CERROR("Unable to create client namespace - %s\n", obddev->obd_name); @@ -386,6 +466,7 @@ err: RETURN(rc); } +EXPORT_SYMBOL(client_obd_setup); int client_obd_cleanup(struct obd_device *obddev) { @@ -394,9 +475,12 @@ int client_obd_cleanup(struct obd_device *obddev) ldlm_namespace_free_post(obddev->obd_namespace); obddev->obd_namespace = NULL; + LASSERT(obddev->u.cli.cl_import == NULL); + ldlm_put_ref(); RETURN(0); } +EXPORT_SYMBOL(client_obd_cleanup); /* ->o_connect() method for client side (OSC and MDC and MGC) */ int client_connect_import(const struct lu_env *env, @@ -412,7 +496,7 @@ int client_connect_import(const struct lu_env *env, ENTRY; *exp = NULL; - down_write(&cli->cl_sem); + cfs_down_write(&cli->cl_sem); if (cli->cl_conn_count > 0 ) GOTO(out_sem, rc = -EALREADY); @@ -436,7 +520,7 @@ int client_connect_import(const struct lu_env *env, imp->imp_connect_flags_orig = data->ocd_connect_flags; } - rc = ptlrpc_connect_import(imp, NULL); + rc = ptlrpc_connect_import(imp); if (rc != 0) { LASSERT (imp->imp_state == LUSTRE_IMP_DISCON); GOTO(out_ldlm, rc); @@ -461,10 +545,11 @@ out_ldlm: *exp = NULL; } out_sem: - up_write(&cli->cl_sem); + cfs_up_write(&cli->cl_sem); return rc; } +EXPORT_SYMBOL(client_connect_import); int client_disconnect_export(struct obd_export *exp) { @@ -483,7 +568,7 @@ int client_disconnect_export(struct obd_export *exp) cli = &obd->u.cli; imp = cli->cl_import; - down_write(&cli->cl_sem); + cfs_down_write(&cli->cl_sem); CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name, cli->cl_conn_count); @@ -500,9 +585,9 @@ int client_disconnect_export(struct obd_export *exp) /* Mark import deactivated now, so we don't try to reconnect if any * of the cleanup RPCs fails (e.g. ldlm cancel, etc). We don't * fully deactivate the import, or that would drop all requests. */ - spin_lock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); imp->imp_deactive = 1; - spin_unlock(&imp->imp_lock); + cfs_spin_unlock(&imp->imp_lock); /* Some non-replayable imports (MDS's OSCs) are pinged, so just * delete it regardless. (It's safe to delete an import that was @@ -512,8 +597,7 @@ int client_disconnect_export(struct obd_export *exp) if (obd->obd_namespace != NULL) { /* obd_force == local only */ ldlm_cli_cancel_unused(obd->obd_namespace, NULL, - obd->obd_force ? LDLM_FL_LOCAL_ONLY:0, - NULL); + obd->obd_force ? LCF_LOCAL : 0, NULL); ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force); } @@ -521,19 +605,12 @@ int client_disconnect_export(struct obd_export *exp) * there's no need to hold sem during disconnecting an import, * and actually it may cause deadlock in gss. */ - up_write(&cli->cl_sem); + cfs_up_write(&cli->cl_sem); rc = ptlrpc_disconnect_import(imp, 0); - down_write(&cli->cl_sem); + cfs_down_write(&cli->cl_sem); ptlrpc_invalidate_import(imp); - if (imp->imp_rq_pool) { - ptlrpc_free_rq_pool(imp->imp_rq_pool); - imp->imp_rq_pool = NULL; - } - client_destroy_import(imp); - cli->cl_import = NULL; - EXIT; out_disconnect: @@ -543,11 +620,13 @@ int client_disconnect_export(struct obd_export *exp) if (!rc && err) rc = err; - up_write(&cli->cl_sem); + cfs_up_write(&cli->cl_sem); RETURN(rc); } +EXPORT_SYMBOL(client_disconnect_export); +#ifdef HAVE_SERVER_SUPPORT int server_disconnect_export(struct obd_export *exp) { int rc; @@ -563,27 +642,27 @@ int server_disconnect_export(struct obd_export *exp) ldlm_cancel_locks_for_export(exp); /* complete all outstanding replies */ - spin_lock(&exp->exp_lock); - while (!list_empty(&exp->exp_outstanding_replies)) { + cfs_spin_lock(&exp->exp_lock); + while (!cfs_list_empty(&exp->exp_outstanding_replies)) { struct ptlrpc_reply_state *rs = - list_entry(exp->exp_outstanding_replies.next, - struct ptlrpc_reply_state, rs_exp_list); - struct ptlrpc_service *svc = rs->rs_service; + cfs_list_entry(exp->exp_outstanding_replies.next, + struct ptlrpc_reply_state, rs_exp_list); + struct ptlrpc_service_part *svcpt = rs->rs_svcpt; - spin_lock(&svc->srv_lock); - list_del_init(&rs->rs_exp_list); - spin_lock(&rs->rs_lock); - ptlrpc_schedule_difficult_reply(rs); - spin_unlock(&rs->rs_lock); - spin_unlock(&svc->srv_lock); - } - spin_unlock(&exp->exp_lock); + cfs_spin_lock(&svcpt->scp_rep_lock); - /* release nid stat refererence */ - lprocfs_exp_cleanup(exp); + cfs_list_del_init(&rs->rs_exp_list); + cfs_spin_lock(&rs->rs_lock); + ptlrpc_schedule_difficult_reply(rs); + cfs_spin_unlock(&rs->rs_lock); - RETURN(rc); + cfs_spin_unlock(&svcpt->scp_rep_lock); + } + cfs_spin_unlock(&exp->exp_lock); + + RETURN(rc); } +EXPORT_SYMBOL(server_disconnect_export); /* -------------------------------------------------------------------------- * from old lib/target.c @@ -594,24 +673,49 @@ static int target_handle_reconnect(struct lustre_handle *conn, struct obd_uuid *cluuid) { ENTRY; + if (exp->exp_connection && exp->exp_imp_reverse) { struct lustre_handle *hdl; + struct obd_device *target; + hdl = &exp->exp_imp_reverse->imp_remote_handle; + target = exp->exp_obd; + /* Might be a re-connect after a partition. */ if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) { - CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name, - cluuid->uuid); + if (target->obd_recovering) { + int timeout = cfs_duration_sec(cfs_time_sub( + cfs_timer_deadline( + &target->obd_recovery_timer), + cfs_time_current())); + + LCONSOLE_WARN("%s: Client %s (at %s) reconnect" + "ing, waiting for %d clients in recov" + "ery for %d:%.02d\n", target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), + target->obd_max_recoverable_clients, + timeout / 60, timeout % 60); + } else { + LCONSOLE_WARN("%s: Client %s (at %s) " + "reconnecting\n", target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + } + conn->cookie = exp->exp_handle.h_cookie; /* target_handle_connect() treats EALREADY and * -EALREADY differently. EALREADY means we are * doing a valid reconnect from the same client. */ RETURN(EALREADY); } else { - CERROR("%s reconnecting from %s, " - "handle mismatch (ours "LPX64", theirs " - LPX64")\n", cluuid->uuid, - exp->exp_connection->c_remote_uuid.uuid, - hdl->cookie, conn->cookie); + 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 @@ -634,23 +738,31 @@ void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data, CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n", obd->obd_name, exp->exp_client_uuid.uuid); - spin_lock(&exp->exp_lock); + cfs_spin_lock(&exp->exp_lock); exp->exp_need_sync = 0; - spin_unlock(&exp->exp_lock); + cfs_spin_unlock(&exp->exp_lock); class_export_cb_put(exp); } EXPORT_SYMBOL(target_client_add_cb); +#ifdef __KERNEL__ static void -target_start_and_reset_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, - int new_client); +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) +{ +} +#endif int target_handle_connect(struct ptlrpc_request *req) { - struct obd_device *target, *targref = NULL; + struct obd_device *target = NULL, *targref = NULL; struct obd_export *export = NULL; struct obd_import *revimp; + struct obd_import *tmp_imp = NULL; struct lustre_handle conn; struct lustre_handle *tmp; struct obd_uuid tgtuuid; @@ -658,10 +770,13 @@ int target_handle_connect(struct ptlrpc_request *req) struct obd_uuid remote_uuid; char *str; int rc = 0; - int mds_conn = 0; + char *target_start; + int target_len; + bool mds_conn = false, lw_client = false; struct obd_connect_data *data, *tmpdata; + int size, tmpsize; lnet_nid_t *client_nid = NULL; - ENTRY; + ENTRY; OBD_RACE(OBD_FAIL_TGT_CONN_RACE); @@ -676,17 +791,30 @@ int target_handle_connect(struct ptlrpc_request *req) if (!target) target = class_name2obd(str); - if (!target || target->obd_stopping || !target->obd_set_up) { - LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available " - " for connect (%s)\n", str, - !target ? "no target" : - (target->obd_stopping ? "stopping" : - "not set up")); - GOTO(out, rc = -ENODEV); - } + if (!target) { + deuuidify(str, NULL, &target_start, &target_len); + LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for " + "connect (no target)\n", str); + GOTO(out, rc = -ENODEV); + } + + cfs_spin_lock(&target->obd_dev_lock); + if (target->obd_stopping || !target->obd_set_up) { + cfs_spin_unlock(&target->obd_dev_lock); + + deuuidify(str, NULL, &target_start, &target_len); + LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect " + "from %s (%s)\n", target_len, target_start, + libcfs_nid2str(req->rq_peer.nid), + (target->obd_stopping ? + "stopping" : "not set up")); + GOTO(out, rc = -ENODEV); + } if (target->obd_no_conn) { - LCONSOLE_WARN("%s: temporarily refusing client connection " + cfs_spin_unlock(&target->obd_dev_lock); + + LCONSOLE_WARN("%s: Temporarily refusing client connection " "from %s\n", target->obd_name, libcfs_nid2str(req->rq_peer.nid)); GOTO(out, rc = -EAGAIN); @@ -697,6 +825,8 @@ int target_handle_connect(struct ptlrpc_request *req) Really, class_uuid2obd should take the ref. */ targref = class_incref(target, __FUNCTION__, cfs_current()); + target->obd_conn_inprogress++; + cfs_spin_unlock(&target->obd_dev_lock); str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID); if (str == NULL) { @@ -727,6 +857,8 @@ int target_handle_connect(struct ptlrpc_request *req) conn = *tmp; + size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_CLIENT); data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA); if (!data) GOTO(out, rc = -EPROTO); @@ -766,7 +898,10 @@ int target_handle_connect(struct ptlrpc_request *req) if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) && (data->ocd_connect_flags & OBD_CONNECT_MDS)) - mds_conn = 1; + mds_conn = true; + + if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0) + lw_client = true; /* lctl gets a backstage, all-access pass. */ if (obd_uuid_equals(&cluuid, &target->obd_uuid)) @@ -777,23 +912,29 @@ int target_handle_connect(struct ptlrpc_request *req) goto no_export; /* we've found an export in the hash */ + + cfs_spin_lock(&export->exp_lock); + if (export->exp_connecting) { /* bug 9635, et. al. */ - CWARN("%s: exp %p already connecting\n", - export->exp_obd->obd_name, export); + cfs_spin_unlock(&export->exp_lock); + LCONSOLE_WARN("%s: Export %p already connecting from %s\n", + export->exp_obd->obd_name, export, + libcfs_nid2str(req->rq_peer.nid)); class_export_put(export); export = NULL; rc = -EALREADY; } else if (mds_conn && export->exp_connection) { + cfs_spin_unlock(&export->exp_lock); if (req->rq_peer.nid != export->exp_connection->c_peer.nid) /* mds reconnected after failover */ - CWARN("%s: received MDS connection from NID %s," - " removing former export from NID %s\n", + LCONSOLE_WARN("%s: Received MDS connection from " + "%s, removing former export from %s\n", target->obd_name, libcfs_nid2str(req->rq_peer.nid), libcfs_nid2str(export->exp_connection->c_peer.nid)); else /* new mds connection from the same nid */ - CWARN("%s: received new MDS connection from NID %s," - " removing former export from same NID\n", + 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)); class_fail_export(export); class_export_put(export); @@ -803,55 +944,60 @@ int target_handle_connect(struct ptlrpc_request *req) req->rq_peer.nid != export->exp_connection->c_peer.nid && (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)) { + cfs_spin_unlock(&export->exp_lock); /* in mds failover we have static uuid but nid can be * changed*/ - CWARN("%s: cookie %s seen on new NID %s when " - "existing NID %s is already connected\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - libcfs_nid2str(export->exp_connection->c_peer.nid)); + LCONSOLE_WARN("%s: Client %s seen on new nid %s when " + "existing nid %s is already connected\n", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid), + libcfs_nid2str( + export->exp_connection->c_peer.nid)); rc = -EALREADY; class_export_put(export); export = NULL; } else { - spin_lock(&export->exp_lock); - export->exp_connecting = 1; - spin_unlock(&export->exp_lock); - class_export_put(export); - LASSERT(export->exp_obd == target); + export->exp_connecting = 1; + cfs_spin_unlock(&export->exp_lock); + LASSERT(export->exp_obd == target); - rc = target_handle_reconnect(&conn, export, &cluuid); - } + rc = target_handle_reconnect(&conn, export, &cluuid); + } /* If we found an export, we already unlocked. */ if (!export) { no_export: OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout); } else if (req->rq_export == NULL && - atomic_read(&export->exp_rpc_count) > 0) { - CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - export, atomic_read(&export->exp_refcount)); + cfs_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)); GOTO(out, rc = -EBUSY); } else if (req->rq_export != NULL && - (atomic_read(&export->exp_rpc_count) > 1)) { + (cfs_atomic_read(&export->exp_rpc_count) > 1)) { /* the current connect rpc has increased exp_rpc_count */ - CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n", - target->obd_name, cluuid.uuid, - libcfs_nid2str(req->rq_peer.nid), - export, atomic_read(&export->exp_rpc_count) - 1); - spin_lock(&export->exp_lock); + 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); + cfs_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); + cfs_spin_unlock(&export->exp_lock); GOTO(out, rc = -EBUSY); } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) { - CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; " - "cookies not random?\n", target->obd_name, - libcfs_nid2str(req->rq_peer.nid), cluuid.uuid); + if (!strstr(cluuid.uuid, "mdt")) + LCONSOLE_WARN("%s: Rejecting reconnect from the " + "known client %s (at %s) because it " + "is indicating it is a new client", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid)); GOTO(out, rc = -EALREADY); } else { OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout); @@ -861,21 +1007,16 @@ no_export: GOTO(out, rc); } - CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n", + CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n", target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), target->obd_recovering ? "recovering/" : "", data->ocd_transno, export, (long)cfs_time_current_sec(), export ? (long)export->exp_last_request_time : 0); - /* Tell the client if we're in recovery. */ - if (target->obd_recovering) { - lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); - /* If this is the first time a client connects, - reset the recovery timer */ - if (rc == 0) - target_start_and_reset_recovery_timer(target, req, - !export); - } + /* If this is the first time a client connects, reset the recovery + * timer. Discard lightweight connections which might be local */ + if (!lw_client && rc == 0 && target->obd_recovering) + check_and_start_recovery_timer(target, req, export == NULL); /* We want to handle EALREADY but *not* -EALREADY from * target_handle_reconnect(), return reconnection state in a flag */ @@ -892,17 +1033,27 @@ no_export: client_nid = &req->rq_peer.nid; if (export == NULL) { - if (target->obd_recovering) { + /* allow lightweight connections during recovery */ + if (target->obd_recovering && !lw_client) { cfs_time_t t; - - t = cfs_timer_deadline(&target->obd_recovery_timer); - t = cfs_time_sub(t, cfs_time_current()); - CERROR("%s: denying connection for new client %s (%s): " - "%d clients in recovery for "CFS_TIME_T"s\n", - target->obd_name, - libcfs_nid2str(req->rq_peer.nid), cluuid.uuid, - atomic_read(&target->obd_lock_replay_clients), - cfs_duration_sec(t)); + int c; /* connected */ + int i; /* in progress */ + int k; /* known */ + + c = cfs_atomic_read(&target->obd_connected_clients); + i = cfs_atomic_read(&target->obd_lock_replay_clients); + k = target->obd_max_recoverable_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 unseen) to recover in %d:%.02d\n", + target->obd_name, cluuid.uuid, + libcfs_nid2str(req->rq_peer.nid), k, + c - i, i, k - c, (int)t / 60, + (int)t % 60); rc = -EBUSY; } else { dont_check_exports: @@ -915,19 +1066,38 @@ dont_check_exports: } else { rc = obd_reconnect(req->rq_svc_thread->t_env, export, target, &cluuid, data, client_nid); - if (rc == 0) - /* prevous done via class_conn2export */ - class_export_get(export); } 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 + + LASSERT(target->u.obt.obt_magic == OBT_MAGIC); + data->ocd_instance = target->u.obt.obt_instance; + /* Return only the parts of obd_connect_data that we understand, so the * client knows that we don't understand the rest. */ if (data) { - tmpdata = req_capsule_server_get(&req->rq_pill, - &RMF_CONNECT_DATA); - //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED; - *tmpdata = *data; + tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA, + RCL_SERVER); + tmpdata = req_capsule_server_get(&req->rq_pill, + &RMF_CONNECT_DATA); + /* Don't use struct assignment here, because the client reply + * buffer may be smaller/larger than the local struct + * obd_connect_data. */ + memcpy(tmpdata, data, min(tmpsize, size)); } /* If all else goes well, this is our RPC return code. */ @@ -945,12 +1115,14 @@ dont_check_exports: if (req->rq_export != NULL) class_export_put(req->rq_export); - req->rq_export = export; + /* request takes one export refcount */ + req->rq_export = class_export_get(export); - spin_lock(&export->exp_lock); + cfs_spin_lock(&export->exp_lock); if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) { - spin_unlock(&export->exp_lock); - CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n", + cfs_spin_unlock(&export->exp_lock); + CDEBUG(D_RPCTRACE, "%s: %s already connected at greater " + "or equal conn_cnt: %d >= %d\n", cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), export->exp_conn_cnt, lustre_msg_get_conn_cnt(req->rq_reqmsg)); @@ -964,19 +1136,19 @@ dont_check_exports: /* request from liblustre? Don't evict it for not pinging. */ if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) { export->exp_libclient = 1; - spin_unlock(&export->exp_lock); + cfs_spin_unlock(&export->exp_lock); - spin_lock(&target->obd_dev_lock); - list_del_init(&export->exp_obd_chain_timed); - spin_unlock(&target->obd_dev_lock); + cfs_spin_lock(&target->obd_dev_lock); + cfs_list_del_init(&export->exp_obd_chain_timed); + cfs_spin_unlock(&target->obd_dev_lock); } else { - spin_unlock(&export->exp_lock); + cfs_spin_unlock(&export->exp_lock); } 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) && - !hlist_unhashed(&export->exp_nid_hash)) + !cfs_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); @@ -987,53 +1159,73 @@ dont_check_exports: export->exp_connection = ptlrpc_connection_get(req->rq_peer, req->rq_self, &remote_uuid); - if (hlist_unhashed(&export->exp_nid_hash)) { - cfs_hash_add_unique(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); - } - - spin_lock_bh(&target->obd_processing_task_lock); - if (target->obd_recovering && !export->exp_in_recovery) { - spin_lock(&export->exp_lock); + 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); + } + + if (target->obd_recovering && !export->exp_in_recovery && !lw_client) { + int has_transno; + __u64 transno = data->ocd_transno; + + cfs_spin_lock(&export->exp_lock); + /* possible race with class_disconnect_stale_exports, + * export may be already in the eviction process */ + if (export->exp_failed) { + cfs_spin_unlock(&export->exp_lock); + GOTO(out, rc = -ENODEV); + } export->exp_in_recovery = 1; export->exp_req_replay_needed = 1; export->exp_lock_replay_needed = 1; - spin_unlock(&export->exp_lock); - if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO) - && (data->ocd_transno == 0)) + cfs_spin_unlock(&export->exp_lock); + + has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) & + MSG_CONNECT_TRANSNO); + if (has_transno && transno == 0) CWARN("Connect with zero transno!\n"); - if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO) - && data->ocd_transno < target->obd_next_recovery_transno) - target->obd_next_recovery_transno = data->ocd_transno; - target->obd_connected_clients++; - atomic_inc(&target->obd_req_replay_clients); - atomic_inc(&target->obd_lock_replay_clients); - if (target->obd_connected_clients == + if (has_transno && transno > 0 && + transno < target->obd_next_recovery_transno && + transno > target->obd_last_committed) { + /* another way is to use cmpxchg() so it will be + * lock free */ + cfs_spin_lock(&target->obd_recovery_task_lock); + if (transno < target->obd_next_recovery_transno) + target->obd_next_recovery_transno = transno; + cfs_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); } - spin_unlock_bh(&target->obd_processing_task_lock); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); - conn = *tmp; - if (export->exp_imp_reverse != NULL) { - /* destroyed import can be still referenced in ctxt */ - obd_set_info_async(export, sizeof(KEY_REVIMP_UPD), - KEY_REVIMP_UPD, 0, NULL, NULL); + /* 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); - client_destroy_import(export->exp_imp_reverse); - } + tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); + conn = *tmp; - /* for the rest part, we return -ENOTCONN in case of errors - * in order to let client initialize connection again. - */ - revimp = export->exp_imp_reverse = class_new_import(target); - if (!revimp) { - CERROR("fail to alloc new reverse import.\n"); - GOTO(out, rc = -ENOTCONN); - } + /* for the rest part, we return -ENOTCONN in case of errors + * in order to let client initialize connection again. + */ + revimp = class_new_import(target); + if (revimp == NULL) { + CERROR("fail to alloc new reverse import.\n"); + GOTO(out, rc = -ENOTCONN); + } + + cfs_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; + cfs_spin_unlock(&export->exp_lock); revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection); revimp->imp_client = &export->exp_obd->obd_ldlm_client; @@ -1045,32 +1237,51 @@ dont_check_exports: * ptlrpc_handle_server_req_in->lustre_unpack_msg() */ revimp->imp_msg_magic = req->rq_reqmsg->lm_magic; - if ((export->exp_connect_flags & OBD_CONNECT_AT) && + 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_AT_SUPPORT; + revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; else - revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; - rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); - if (rc) { - CERROR("Failed to get sec for reverse import: %d\n", rc); - export->exp_imp_reverse = NULL; - class_destroy_import(revimp); - } + 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); + cfs_spin_lock(&export->exp_lock); + export->exp_imp_reverse = NULL; + cfs_spin_unlock(&export->exp_lock); + class_destroy_import(revimp); + } + + class_import_put(revimp); - class_import_put(revimp); out: + if (tmp_imp != NULL) + client_destroy_import(tmp_imp); if (export) { - spin_lock(&export->exp_lock); + cfs_spin_lock(&export->exp_lock); export->exp_connecting = 0; - spin_unlock(&export->exp_lock); + cfs_spin_unlock(&export->exp_lock); + + class_export_put(export); } - if (targref) + if (targref) { + cfs_spin_lock(&target->obd_dev_lock); + target->obd_conn_inprogress--; + cfs_spin_unlock(&target->obd_dev_lock); + class_decref(targref, __FUNCTION__, cfs_current()); + } if (rc) req->rq_status = rc; RETURN(rc); } +EXPORT_SYMBOL(target_handle_connect); int target_handle_disconnect(struct ptlrpc_request *req) { @@ -1086,19 +1297,28 @@ int target_handle_disconnect(struct ptlrpc_request *req) RETURN(0); } +EXPORT_SYMBOL(target_handle_disconnect); void target_destroy_export(struct obd_export *exp) { - /* exports created from last_rcvd data, and "fake" - exports created by lctl don't have an import */ - if (exp->exp_imp_reverse != NULL) - client_destroy_import(exp->exp_imp_reverse); - - LASSERT(atomic_read(&exp->exp_locks_count) == 0); - LASSERT(atomic_read(&exp->exp_rpc_count) == 0); - LASSERT(atomic_read(&exp->exp_cb_count) == 0); - LASSERT(atomic_read(&exp->exp_replay_count) == 0); + struct obd_import *imp = NULL; + /* exports created from last_rcvd data, and "fake" + exports created by lctl don't have an import */ + cfs_spin_lock(&exp->exp_lock); + if (exp->exp_imp_reverse != NULL) { + imp = exp->exp_imp_reverse; + exp->exp_imp_reverse = NULL; + } + cfs_spin_unlock(&exp->exp_lock); + if (imp != NULL) + client_destroy_import(imp); + + LASSERT_ATOMIC_ZERO(&exp->exp_locks_count); + LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count); + LASSERT_ATOMIC_ZERO(&exp->exp_cb_count); + LASSERT_ATOMIC_ZERO(&exp->exp_replay_count); } +EXPORT_SYMBOL(target_destroy_export); /* * Recovery functions @@ -1106,20 +1326,21 @@ void target_destroy_export(struct obd_export *exp) static void target_request_copy_get(struct ptlrpc_request *req) { class_export_rpc_get(req->rq_export); - LASSERT(list_empty(&req->rq_list)); + LASSERT(cfs_list_empty(&req->rq_list)); CFS_INIT_LIST_HEAD(&req->rq_replay_list); + /* increase refcount to keep request in queue */ - LASSERT(atomic_read(&req->rq_refcount)); - atomic_inc(&req->rq_refcount); + cfs_atomic_inc(&req->rq_refcount); /** let export know it has replays to be handled */ - atomic_inc(&req->rq_export->exp_replay_count); + cfs_atomic_inc(&req->rq_export->exp_replay_count); } static void target_request_copy_put(struct ptlrpc_request *req) { - LASSERT(list_empty(&req->rq_replay_list)); - LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0); - atomic_dec(&req->rq_export->exp_replay_count); + LASSERT(cfs_list_empty(&req->rq_replay_list)); + LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count); + + cfs_atomic_dec(&req->rq_export->exp_replay_count); class_export_rpc_put(req->rq_export); ptlrpc_server_drop_request(req); } @@ -1133,9 +1354,9 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) LASSERT(exp); - spin_lock(&exp->exp_lock); - list_for_each_entry(reqiter, &exp->exp_req_replay_queue, - rq_replay_list) { + cfs_spin_lock(&exp->exp_lock); + cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue, + rq_replay_list) { if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) { dup = 1; break; @@ -1149,44 +1370,58 @@ static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) CERROR("invalid flags %x of resent replay\n", lustre_msg_get_flags(req->rq_reqmsg)); } else { - list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue); + cfs_list_add_tail(&req->rq_replay_list, + &exp->exp_req_replay_queue); } - spin_unlock(&exp->exp_lock); + cfs_spin_unlock(&exp->exp_lock); return dup; } static void target_exp_dequeue_req_replay(struct ptlrpc_request *req) { - LASSERT(!list_empty(&req->rq_replay_list)); + LASSERT(!cfs_list_empty(&req->rq_replay_list)); LASSERT(req->rq_export); - spin_lock(&req->rq_export->exp_lock); - list_del_init(&req->rq_replay_list); - spin_unlock(&req->rq_export->exp_lock); + cfs_spin_lock(&req->rq_export->exp_lock); + cfs_list_del_init(&req->rq_replay_list); + cfs_spin_unlock(&req->rq_export->exp_lock); } #ifdef __KERNEL__ static void target_finish_recovery(struct obd_device *obd) { ENTRY; - LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n", - obd->obd_name); + + /* only log a recovery message when recovery has occurred */ + if (obd->obd_recovery_start) { + time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() - + obd->obd_recovery_start); + LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients " + "%d recovered and %d %s evicted.\n", obd->obd_name, + (int)elapsed_time / 60, (int)elapsed_time % 60, + obd->obd_max_recoverable_clients, + cfs_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_bh(&obd->obd_processing_task_lock); - if (!list_empty(&obd->obd_req_replay_queue) || - !list_empty(&obd->obd_lock_replay_queue) || - !list_empty(&obd->obd_final_req_queue)) { + cfs_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)) { CERROR("%s: Recovery queues ( %s%s%s) are not empty\n", obd->obd_name, - list_empty(&obd->obd_req_replay_queue) ? "" : "req ", - list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ", - list_empty(&obd->obd_final_req_queue) ? "" : "final "); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ", + cfs_list_empty(&obd->obd_lock_replay_queue) ? \ + "" : "lock ", + cfs_list_empty(&obd->obd_final_req_queue) ? \ + "" : "final "); + cfs_spin_unlock(&obd->obd_recovery_task_lock); LBUG(); } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); obd->obd_recovery_end = cfs_time_current_sec(); @@ -1203,13 +1438,13 @@ static void target_finish_recovery(struct obd_device *obd) static void abort_req_replay_queue(struct obd_device *obd) { struct ptlrpc_request *req, *n; - struct list_head abort_list; + cfs_list_t abort_list; CFS_INIT_LIST_HEAD(&abort_list); - spin_lock_bh(&obd->obd_processing_task_lock); - list_splice_init(&obd->obd_req_replay_queue, &abort_list); - spin_unlock_bh(&obd->obd_processing_task_lock); - list_for_each_entry_safe(req, n, &abort_list, rq_list) { + cfs_spin_lock(&obd->obd_recovery_task_lock); + cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list); + cfs_spin_unlock(&obd->obd_recovery_task_lock); + cfs_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)) { @@ -1224,13 +1459,13 @@ 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; - struct list_head abort_list; + cfs_list_t abort_list; CFS_INIT_LIST_HEAD(&abort_list); - spin_lock_bh(&obd->obd_processing_task_lock); - list_splice_init(&obd->obd_lock_replay_queue, &abort_list); - spin_unlock_bh(&obd->obd_processing_task_lock); - list_for_each_entry_safe(req, n, &abort_list, rq_list){ + cfs_spin_lock(&obd->obd_recovery_task_lock); + cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list); + cfs_spin_unlock(&obd->obd_recovery_task_lock); + cfs_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)) { @@ -1240,7 +1475,6 @@ static void abort_lock_replay_queue(struct obd_device *obd) target_request_copy_put(req); } } -#endif /* Called from a cleanup function if the device is being cleaned up forcefully. The exports should all have been disconnected already, @@ -1254,104 +1488,129 @@ static void abort_lock_replay_queue(struct obd_device *obd) void target_cleanup_recovery(struct obd_device *obd) { struct ptlrpc_request *req, *n; - struct list_head clean_list; + cfs_list_t clean_list; ENTRY; CFS_INIT_LIST_HEAD(&clean_list); - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering) { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); EXIT; return; } obd->obd_recovering = obd->obd_abort_recovery = 0; - target_cancel_recovery_timer(obd); + cfs_spin_unlock(&obd->obd_dev_lock); - list_splice_init(&obd->obd_req_replay_queue, &clean_list); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); + target_cancel_recovery_timer(obd); + cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list); + cfs_spin_unlock(&obd->obd_recovery_task_lock); - list_for_each_entry_safe(req, n, &clean_list, rq_list) { + cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) { LASSERT(req->rq_reply_state == 0); target_exp_dequeue_req_replay(req); target_request_copy_put(req); } - spin_lock_bh(&obd->obd_processing_task_lock); - list_splice_init(&obd->obd_lock_replay_queue, &clean_list); - list_splice_init(&obd->obd_final_req_queue, &clean_list); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_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); + cfs_spin_unlock(&obd->obd_recovery_task_lock); - list_for_each_entry_safe(req, n, &clean_list, rq_list){ + cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){ LASSERT(req->rq_reply_state == 0); target_request_copy_put(req); } EXIT; } +EXPORT_SYMBOL(target_cleanup_recovery); -/* obd_processing_task_lock should be held */ +/* obd_recovery_task_lock should be held */ void target_cancel_recovery_timer(struct obd_device *obd) { CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name); cfs_timer_disarm(&obd->obd_recovery_timer); } +EXPORT_SYMBOL(target_cancel_recovery_timer); -/* extend = 1 means require at least "duration" seconds left in the timer, - extend = 0 means set the total duration (start_recovery_timer) */ -static void reset_recovery_timer(struct obd_device *obd, int duration, - int extend) +static void target_start_recovery_timer(struct obd_device *obd) { - cfs_time_t now = cfs_time_current_sec(); - cfs_duration_t left; + if (obd->obd_recovery_start != 0) + return; - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); if (!obd->obd_recovering || obd->obd_abort_recovery) { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); return; } - left = cfs_time_sub(obd->obd_recovery_end, now); + LASSERT(obd->obd_recovery_timeout != 0); - if (extend && (duration > left)) - obd->obd_recovery_timeout += duration - left; - else if (!extend && (duration > obd->obd_recovery_timeout)) - /* Track the client's largest expected replay time */ - obd->obd_recovery_timeout = duration; -#ifdef CRAY_XT3 - /* - * If total recovery time already exceed the - * obd_recovery_max_time, then CRAY XT3 will - * abort the recovery - */ - if(obd->obd_recovery_timeout > obd->obd_recovery_max_time) - obd->obd_recovery_timeout = obd->obd_recovery_max_time; -#endif - obd->obd_recovery_end = obd->obd_recovery_start + - obd->obd_recovery_timeout; - if (!cfs_timer_is_armed(&obd->obd_recovery_timer) || - cfs_time_before(now, obd->obd_recovery_end)) { - left = cfs_time_sub(obd->obd_recovery_end, now); - cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left)); - } - spin_unlock_bh(&obd->obd_processing_task_lock); - CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", - obd->obd_name, (unsigned)left); + if (obd->obd_recovery_start != 0) { + cfs_spin_unlock(&obd->obd_dev_lock); + return; + } + + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(obd->obd_recovery_timeout)); + obd->obd_recovery_start = cfs_time_current_sec(); + cfs_spin_unlock(&obd->obd_dev_lock); + + LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, " + "or until %d client%s reconnect%s\n", + obd->obd_name, + obd->obd_recovery_timeout / 60, + obd->obd_recovery_timeout % 60, + obd->obd_max_recoverable_clients, + (obd->obd_max_recoverable_clients == 1) ? "" : "s", + (obd->obd_max_recoverable_clients == 1) ? "s": ""); } -static void check_and_start_recovery_timer(struct obd_device *obd) +/** + * extend recovery window. + * + * if @extend is true, extend recovery window to have @drt remaining at least; + * otherwise, make sure the recovery timeout value is not less than @drt. + */ +static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend) { - spin_lock_bh(&obd->obd_processing_task_lock); - if (cfs_timer_is_armed(&obd->obd_recovery_timer)) { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_time_t now; + cfs_time_t end; + cfs_duration_t left; + int to; + + cfs_spin_lock(&obd->obd_dev_lock); + if (!obd->obd_recovering || obd->obd_abort_recovery) { + cfs_spin_unlock(&obd->obd_dev_lock); return; } - CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name); - obd->obd_recovery_start = cfs_time_current_sec(); - /* minimum */ - obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout; - spin_unlock_bh(&obd->obd_processing_task_lock); + LASSERT(obd->obd_recovery_start != 0); - reset_recovery_timer(obd, obd->obd_recovery_timeout, 0); + now = cfs_time_current_sec(); + to = obd->obd_recovery_timeout; + end = obd->obd_recovery_start + to; + left = cfs_time_sub(end, now); + + if (extend && (drt > left)) { + to += drt - left; + } else if (!extend && (drt > to)) { + to = drt; + /* 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 = to; + cfs_timer_arm(&obd->obd_recovery_timer, + cfs_time_shift(drt)); + } + cfs_spin_unlock(&obd->obd_dev_lock); + + CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n", + obd->obd_name, (unsigned)drt); } /* Reset the timer with each new client connection */ @@ -1366,19 +1625,21 @@ static void check_and_start_recovery_timer(struct obd_device *obd) */ static void -target_start_and_reset_recovery_timer(struct obd_device *obd, - struct ptlrpc_request *req, - int new_client) +check_and_start_recovery_timer(struct obd_device *obd, + struct ptlrpc_request *req, + int new_client) { int service_time = lustre_msg_get_service_time(req->rq_reqmsg); + 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 * at how long new requests will take. */ - at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate, - service_time); + at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate, + service_time); - check_and_start_recovery_timer(obd); + target_start_recovery_timer(obd); /* convert the service time to rpc timeout, * reuse service_time to limit stack usage */ @@ -1387,14 +1648,17 @@ target_start_and_reset_recovery_timer(struct obd_device *obd, /* We expect other clients to timeout within service_time, then try * to reconnect, then try the failover server. The max delay between * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */ - service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC + - INITIAL_CONNECT_TIMEOUT); + 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) - reset_recovery_timer(obd, service_time, 0); + extend_recovery_timer(obd, service_time, false); } -#ifdef __KERNEL__ - /** Health checking routines */ static inline int exp_connect_healthy(struct obd_export *exp) { @@ -1405,13 +1669,13 @@ static inline int exp_connect_healthy(struct obd_export *exp) static inline int exp_req_replay_healthy(struct obd_export *exp) { return (!exp->exp_req_replay_needed || - atomic_read(&exp->exp_replay_count) > 0); + cfs_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 || - atomic_read(&exp->exp_replay_count) > 0); + cfs_atomic_read(&exp->exp_replay_count) > 0); } static inline int exp_vbr_healthy(struct obd_export *exp) @@ -1427,12 +1691,13 @@ static inline int exp_finished(struct obd_export *exp) /** 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(obd->obd_connected_clients <= obd->obd_max_recoverable_clients); + LASSERT(clnts <= obd->obd_max_recoverable_clients); if (obd->obd_no_conn == 0 && - obd->obd_connected_clients + obd->obd_stale_clients == - obd->obd_max_recoverable_clients) + clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients) return 1; return 0; } @@ -1443,18 +1708,18 @@ static int check_for_next_transno(struct obd_device *obd) int wake_up = 0, connected, completed, queue_len; __u64 next_transno, req_transno; ENTRY; - spin_lock_bh(&obd->obd_processing_task_lock); - if (!list_empty(&obd->obd_req_replay_queue)) { - req = list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); + cfs_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; } - connected = obd->obd_connected_clients; - completed = connected - atomic_read(&obd->obd_req_replay_clients); + 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; @@ -1469,13 +1734,13 @@ 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 (atomic_read(&obd->obd_req_replay_clients) == 0) { + } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) { CDEBUG(D_HA, "waking for completed recovery\n"); wake_up = 1; } else if (req_transno == next_transno) { CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno); wake_up = 1; - } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) { + } else if (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, @@ -1499,7 +1764,7 @@ static int check_for_next_transno(struct obd_device *obd) obd->obd_next_recovery_transno = req_transno; wake_up = 1; } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } @@ -1507,11 +1772,11 @@ static int check_for_next_lock(struct obd_device *obd) { int wake_up = 0; - spin_lock_bh(&obd->obd_processing_task_lock); - if (!list_empty(&obd->obd_lock_replay_queue)) { + cfs_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 (atomic_read(&obd->obd_lock_replay_clients) == 0) { + } 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) { @@ -1521,7 +1786,7 @@ static int check_for_next_lock(struct obd_device *obd) CDEBUG(D_HA, "waking for expired recovery\n"); wake_up = 1; } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } @@ -1535,38 +1800,33 @@ static int target_recovery_overseer(struct obd_device *obd, int (*check_routine)(struct obd_device *), int (*health_check)(struct obd_export *)) { - int abort = 0, expired = 1; - - do { - cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); - spin_lock_bh(&obd->obd_processing_task_lock); - abort = obd->obd_abort_recovery; - expired = obd->obd_recovery_expired; +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; - spin_unlock_bh(&obd->obd_processing_task_lock); - if (abort) { - CWARN("recovery is aborted, evict exports in recovery\n"); - /** evict exports which didn't finish recovery yet */ - class_disconnect_stale_exports(obd, exp_finished); - } else if (expired) { - /** If some clients died being recovered, evict them */ - CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n"); - /** evict cexports with no replay in queue, they are stalled */ - class_disconnect_stale_exports(obd, health_check); - /** continue with VBR */ - spin_lock_bh(&obd->obd_processing_task_lock); - obd->obd_version_recov = 1; - spin_unlock_bh(&obd->obd_processing_task_lock); - /** - * reset timer, recovery will proceed with versions now, - * timeout is set just to handle reconnection delays - */ - reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1); - /** Wait for recovery events again, after evicting bad clients */ - } - } while (!abort && expired); - - return abort; + /** 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 */ + cfs_spin_lock(&obd->obd_dev_lock); + obd->obd_version_recov = 1; + cfs_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; } static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) @@ -1583,17 +1843,17 @@ static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) abort_lock_replay_queue(obd); } - spin_lock_bh(&obd->obd_processing_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); + cfs_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); + cfs_list_del_init(&req->rq_list); obd->obd_requests_queued_for_recovery--; - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); } else { - spin_unlock_bh(&obd->obd_processing_task_lock); - LASSERT(list_empty(&obd->obd_req_replay_queue)); - LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); + cfs_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); /** evict exports failed VBR */ class_disconnect_stale_exports(obd, exp_vbr_healthy); } @@ -1609,16 +1869,16 @@ static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) exp_lock_replay_healthy)) abort_lock_replay_queue(obd); - spin_lock_bh(&obd->obd_processing_task_lock); - if (!list_empty(&obd->obd_lock_replay_queue)) { - req = list_entry(obd->obd_lock_replay_queue.next, - struct ptlrpc_request, rq_list); - list_del_init(&req->rq_list); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_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); + cfs_spin_unlock(&obd->obd_recovery_task_lock); } else { - spin_unlock_bh(&obd->obd_processing_task_lock); - LASSERT(list_empty(&obd->obd_lock_replay_queue)); - LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0); + cfs_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); } @@ -1629,18 +1889,20 @@ static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) { struct ptlrpc_request *req = NULL; - spin_lock_bh(&obd->obd_processing_task_lock); - if (!list_empty(&obd->obd_final_req_queue)) { - req = list_entry(obd->obd_final_req_queue.next, - struct ptlrpc_request, rq_list); - list_del_init(&req->rq_list); + cfs_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, + struct ptlrpc_request, rq_list); + cfs_list_del_init(&req->rq_list); + cfs_spin_unlock(&obd->obd_recovery_task_lock); if (req->rq_export->exp_in_recovery) { - spin_lock(&req->rq_export->exp_lock); + cfs_spin_lock(&req->rq_export->exp_lock); req->rq_export->exp_in_recovery = 0; - spin_unlock(&req->rq_export->exp_lock); + cfs_spin_unlock(&req->rq_export->exp_lock); } + } else { + cfs_spin_unlock(&obd->obd_recovery_task_lock); } - spin_unlock_bh(&obd->obd_processing_task_lock); return req; } @@ -1651,17 +1913,18 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, int rc; 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) + GOTO(reqcopy_put, rc = 0); + + rc = lu_context_init(&req->rq_recov_session, LCT_SESSION); + if (rc) { + CERROR("Failure to initialize session: %d\n", rc); GOTO(reqcopy_put, rc); + } req->rq_recov_session.lc_thread = thread; lu_context_enter(&req->rq_recov_session); @@ -1677,13 +1940,31 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, lu_context_fini(&req->rq_recov_session); /* don't reset timer for final stage */ if (!exp_finished(req->rq_export)) { + int to = obd_timeout; + /** * Add request timeout to the recovery time so next request from * this client may come in recovery time */ - reset_recovery_timer(class_exp2obd(req->rq_export), - AT_OFF ? obd_timeout : - lustre_msg_get_timeout(req->rq_reqmsg), 1); + if (!AT_OFF) { + struct ptlrpc_service_part *svcpt; + + svcpt = req->rq_rqbd->rqbd_svcpt; + /* If the server sent early reply for this request, + * the client will recalculate the timeout according to + * current server estimate service time, so we will + * use the maxium timeout here for waiting the client + * sending the next req */ + to = max((int)at_est2timeout( + at_get(&svcpt->scp_at_estimate)), + (int)lustre_msg_get_timeout(req->rq_reqmsg)); + /* Add 2 net_latency, one for balance rq_deadline + * (see ptl_send_rpc), one for resend the req to server, + * Note: client will pack net_latency in replay req + * (see ptlrpc_replay_req) */ + to += 2 * lustre_msg_get_service_time(req->rq_reqmsg); + } + extend_recovery_timer(class_exp2obd(req->rq_export), to, true); } reqcopy_put: RETURN(rc); @@ -1697,31 +1978,49 @@ static int target_recovery_thread(void *arg) struct target_recovery_data *trd = &obd->obd_recovery_data; unsigned long delta; unsigned long flags; - struct lu_env env; - struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread; + struct lu_env *env; + struct ptlrpc_thread *thread = NULL; int rc = 0; ENTRY; - cfs_daemonize("tgt_recov"); + cfs_daemonize_ctxt("tgt_recov"); SIGNAL_MASK_LOCK(current, flags); sigfillset(¤t->blocked); RECALC_SIGPENDING; SIGNAL_MASK_UNLOCK(current, flags); - rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD); - if (rc) + OBD_ALLOC_PTR(thread); + if (thread == NULL) + RETURN(-ENOMEM); + + OBD_ALLOC_PTR(env); + if (env == NULL) { + OBD_FREE_PTR(thread); + RETURN(-ENOMEM); + } + + rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD); + if (rc) { + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); RETURN(rc); + } - thread->t_env = &env; - env.le_ctx.lc_thread = thread; + 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; - CERROR("%s: started recovery thread pid %d\n", obd->obd_name, + CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name, cfs_curproc_pid()); trd->trd_processing_task = cfs_curproc_pid(); + cfs_spin_lock(&obd->obd_dev_lock); obd->obd_recovering = 1; - complete(&trd->trd_starting); + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_complete(&trd->trd_starting); /* first of all, we have to know the first transno to replay */ if (target_recovery_overseer(obd, check_for_clients, @@ -1732,9 +2031,8 @@ static int target_recovery_thread(void *arg) /* next stage: replay requests */ delta = jiffies; - obd->obd_req_replaying = 1; CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n", - atomic_read(&obd->obd_req_replay_clients), + 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()); @@ -1747,9 +2045,9 @@ static int target_recovery_thread(void *arg) * bz18031: increase next_recovery_transno before * target_request_copy_put() will drop exp_rpc reference */ - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); obd->obd_next_recovery_transno++; - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); obd->obd_replayed_requests++; @@ -1759,7 +2057,7 @@ static int target_recovery_thread(void *arg) * The second stage: replay locks */ CDEBUG(D_INFO, "2: lock replay stage - %d clients\n", - atomic_read(&obd->obd_lock_replay_clients)); + 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: ", @@ -1776,13 +2074,15 @@ static int target_recovery_thread(void *arg) */ CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); /** Update server last boot epoch */ - lut_boot_epoch_update(lut); + tgt_boot_epoch_update(lut); /* We drop recoverying flag to forward all new requests * to regular mds_handle() since now */ - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&obd->obd_dev_lock); obd->obd_recovering = obd->obd_abort_recovery = 0; + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_spin_lock(&obd->obd_recovery_task_lock); target_cancel_recovery_timer(obd); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_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: ", @@ -1792,19 +2092,22 @@ static int target_recovery_thread(void *arg) target_request_copy_put(req); } - delta = (jiffies - delta) / HZ; + 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_timeout * OBD_RECOVERY_FACTOR) { + if (delta > OBD_RECOVERY_TIME_SOFT) { CWARN("too long recovery - read logs\n"); libcfs_debug_dumplog(); } target_finish_recovery(obd); - lu_context_fini(&env.le_ctx); + lu_context_fini(&env->le_ctx); trd->trd_processing_task = 0; - complete(&trd->trd_finishing); + cfs_complete(&trd->trd_finishing); + + OBD_FREE_PTR(thread); + OBD_FREE_PTR(env); RETURN(rc); } @@ -1816,12 +2119,12 @@ static int target_start_recovery_thread(struct lu_target *lut, struct target_recovery_data *trd = &obd->obd_recovery_data; memset(trd, 0, sizeof(*trd)); - init_completion(&trd->trd_starting); - init_completion(&trd->trd_finishing); + cfs_init_completion(&trd->trd_starting); + cfs_init_completion(&trd->trd_finishing); trd->trd_recovery_handler = handler; - if (kernel_thread(target_recovery_thread, lut, 0) > 0) { - wait_for_completion(&trd->trd_starting); + if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) { + cfs_wait_for_completion(&trd->trd_starting); LASSERT(obd->obd_recovering != 0); } else rc = -ECHILD; @@ -1831,21 +2134,20 @@ static int target_start_recovery_thread(struct lu_target *lut, void target_stop_recovery_thread(struct obd_device *obd) { - spin_lock_bh(&obd->obd_processing_task_lock); if (obd->obd_recovery_data.trd_processing_task > 0) { struct target_recovery_data *trd = &obd->obd_recovery_data; /** recovery can be done but postrecovery is not yet */ + cfs_spin_lock(&obd->obd_dev_lock); 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); } - spin_unlock_bh(&obd->obd_processing_task_lock); - wait_for_completion(&trd->trd_finishing); - } else { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_wait_for_completion(&trd->trd_finishing); } } +EXPORT_SYMBOL(target_stop_recovery_thread); void target_recovery_fini(struct obd_device *obd) { @@ -1860,14 +2162,12 @@ 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, atomic_read(&obd->obd_lock_replay_clients), + obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients), cfs_time_current_sec()- obd->obd_recovery_start, - obd->obd_connected_clients); + cfs_atomic_read(&obd->obd_connected_clients)); - spin_lock_bh(&obd->obd_processing_task_lock); obd->obd_recovery_expired = 1; cfs_waitq_signal(&obd->obd_next_transno_waitq); - spin_unlock_bh(&obd->obd_processing_task_lock); } void target_recovery_init(struct lu_target *lut, svc_handler_t handler) @@ -1875,7 +2175,7 @@ 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 */ - lut_boot_epoch_update(lut); + tgt_boot_epoch_update(lut); return; } @@ -1886,15 +2186,13 @@ void target_recovery_init(struct lu_target *lut, svc_handler_t handler) obd->obd_next_recovery_transno = obd->obd_last_committed + 1; obd->obd_recovery_start = 0; obd->obd_recovery_end = 0; - obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout; - /* bz13079: this should be set to desired value for ost but not for mds */ - obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME; + cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd); target_start_recovery_thread(lut, handler); } EXPORT_SYMBOL(target_recovery_init); -#endif +#endif /* __KERNEL__ */ static int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req) @@ -1903,37 +2201,38 @@ static int target_process_req_flags(struct obd_device *obd, LASSERT(exp != NULL); if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) { /* client declares he's ready to replay locks */ - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&exp->exp_lock); if (exp->exp_req_replay_needed) { - LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0); - spin_lock(&exp->exp_lock); exp->exp_req_replay_needed = 0; - spin_unlock(&exp->exp_lock); - atomic_dec(&obd->obd_req_replay_clients); + cfs_spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients); + cfs_atomic_dec(&obd->obd_req_replay_clients); + } else { + cfs_spin_unlock(&exp->exp_lock); } - spin_unlock_bh(&obd->obd_processing_task_lock); } 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 */ - spin_lock_bh(&obd->obd_processing_task_lock); + cfs_spin_lock(&exp->exp_lock); if (exp->exp_lock_replay_needed) { - LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0); - spin_lock(&exp->exp_lock); exp->exp_lock_replay_needed = 0; - spin_unlock(&exp->exp_lock); - atomic_dec(&obd->obd_lock_replay_clients); + cfs_spin_unlock(&exp->exp_lock); + + LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients); + cfs_atomic_dec(&obd->obd_lock_replay_clients); + } else { + cfs_spin_unlock(&exp->exp_lock); } - spin_unlock_bh(&obd->obd_processing_task_lock); } - return 0; } int target_queue_recovery_request(struct ptlrpc_request *req, struct obd_device *obd) { - struct list_head *tmp; + cfs_list_t *tmp; int inserted = 0; __u64 transno = lustre_msg_get_transno(req->rq_reqmsg); ENTRY; @@ -1950,40 +2249,41 @@ int target_queue_recovery_request(struct ptlrpc_request *req, * so, we put the request on th final queue */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue final req"); - spin_lock_bh(&obd->obd_processing_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); + cfs_spin_lock(&obd->obd_recovery_task_lock); if (obd->obd_recovering) { - list_add_tail(&req->rq_list, &obd->obd_final_req_queue); + cfs_list_add_tail(&req->rq_list, + &obd->obd_final_req_queue); } else { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(obd->obd_stopping ? -ENOTCONN : 1); } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_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 */ target_request_copy_get(req); DEBUG_REQ(D_HA, req, "queue lock replay req"); - spin_lock_bh(&obd->obd_processing_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); + cfs_spin_lock(&obd->obd_recovery_task_lock); LASSERT(obd->obd_recovering); /* usually due to recovery abort */ if (!req->rq_export->exp_in_recovery) { - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_request_copy_put(req); RETURN(-ENOTCONN); } LASSERT(req->rq_export->exp_lock_replay_needed); - list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue); + cfs_spin_unlock(&obd->obd_recovery_task_lock); RETURN(0); } /* CAVEAT EMPTOR: The incoming request message has been swabbed * (i.e. buflens etc are in my own byte order), but type-dependent - * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */ + * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */ if (!transno) { CFS_INIT_LIST_HEAD(&req->rq_list); @@ -1991,8 +2291,6 @@ int target_queue_recovery_request(struct ptlrpc_request *req, RETURN(1); } - spin_lock_bh(&obd->obd_processing_task_lock); - /* If we're processing the queue, we want don't want to queue this * message. * @@ -2003,43 +2301,43 @@ 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. */ - CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n", - obd->obd_next_recovery_transno, transno, obd->obd_req_replaying); - if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) { + CDEBUG(D_HA, "Next recovery transno: "LPU64 + ", current: "LPU64", replaying\n", + obd->obd_next_recovery_transno, transno); + cfs_spin_lock(&obd->obd_recovery_task_lock); + if (transno < obd->obd_next_recovery_transno) { /* Processing the queue right now, don't re-add. */ - LASSERT(list_empty(&req->rq_list)); - spin_unlock_bh(&obd->obd_processing_task_lock); + LASSERT(cfs_list_empty(&req->rq_list)); + cfs_spin_unlock(&obd->obd_recovery_task_lock); RETURN(1); } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP)) RETURN(0); target_request_copy_get(req); - spin_lock_bh(&obd->obd_processing_task_lock); - LASSERT(obd->obd_recovering); if (!req->rq_export->exp_in_recovery) { - spin_unlock_bh(&obd->obd_processing_task_lock); target_request_copy_put(req); RETURN(-ENOTCONN); } LASSERT(req->rq_export->exp_req_replay_needed); if (target_exp_enqueue_req_replay(req)) { - spin_unlock_bh(&obd->obd_processing_task_lock); DEBUG_REQ(D_ERROR, req, "dropping resent queued req"); target_request_copy_put(req); RETURN(0); } /* XXX O(n^2) */ - list_for_each(tmp, &obd->obd_req_replay_queue) { + cfs_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 = - list_entry(tmp, struct ptlrpc_request, rq_list); + cfs_list_entry(tmp, struct ptlrpc_request, rq_list); if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) { - list_add_tail(&req->rq_list, &reqiter->rq_list); + cfs_list_add_tail(&req->rq_list, &reqiter->rq_list); inserted = 1; break; } @@ -2048,7 +2346,7 @@ int target_queue_recovery_request(struct ptlrpc_request *req, transno)) { DEBUG_REQ(D_ERROR, req, "dropping replay: transno " "has been claimed by another client"); - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); target_exp_dequeue_req_replay(req); target_request_copy_put(req); RETURN(0); @@ -2056,13 +2354,40 @@ int target_queue_recovery_request(struct ptlrpc_request *req, } if (!inserted) - list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); + cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue); obd->obd_requests_queued_for_recovery++; + cfs_spin_unlock(&obd->obd_recovery_task_lock); cfs_waitq_signal(&obd->obd_next_transno_waitq); - spin_unlock_bh(&obd->obd_processing_task_lock); 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) +{ + struct obd_export *exp = req->rq_export; + + if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) + lustre_msg_set_last_committed(req->rq_repmsg, + exp->exp_last_committed); + else + DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/" + "%d)", exp->exp_obd->obd_no_transno, + req->rq_repmsg == NULL); + + 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 */ /** * Packs current SLV and Limit into \a req. @@ -2088,13 +2413,14 @@ int target_pack_pool_reply(struct ptlrpc_request *req) */ obd = req->rq_export->exp_obd; - read_lock(&obd->obd_pool_lock); + cfs_read_lock(&obd->obd_pool_lock); lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv); lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit); - read_unlock(&obd->obd_pool_lock); + cfs_read_unlock(&obd->obd_pool_lock); RETURN(0); } +EXPORT_SYMBOL(target_pack_pool_reply); int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) { @@ -2104,7 +2430,7 @@ int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) } if (unlikely(rc)) { - DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc); + DEBUG_REQ(D_NET, req, "processing error (%d)", rc); req->rq_status = rc; return (ptlrpc_send_error(req, 1)); } else { @@ -2116,11 +2442,10 @@ int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id) void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) { + struct ptlrpc_service_part *svcpt; int netrc; struct ptlrpc_reply_state *rs; - struct obd_device *obd; struct obd_export *exp; - struct ptlrpc_service *svc; ENTRY; if (req->rq_no_reply) { @@ -2128,7 +2453,7 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) return; } - svc = req->rq_rqbd->rqbd_service; + svcpt = req->rq_rqbd->rqbd_svcpt; rs = req->rq_reply_state; if (rs == NULL || !rs->rs_difficult) { /* no notifiers */ @@ -2140,7 +2465,7 @@ 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); /* req/reply consistent */ - LASSERT (rs->rs_service == svc); + LASSERT(rs->rs_svcpt == svcpt); /* "fresh" reply */ LASSERT (!rs->rs_scheduled); @@ -2148,11 +2473,10 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) 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)); + LASSERT (cfs_list_empty(&rs->rs_obd_list)); + LASSERT (cfs_list_empty(&rs->rs_exp_list)); exp = class_export_get (req->rq_export); - obd = exp->exp_obd; /* disable reply scheduling while I'm setting up */ rs->rs_scheduled = 1; @@ -2160,27 +2484,27 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) rs->rs_xid = req->rq_xid; rs->rs_transno = req->rq_transno; rs->rs_export = exp; - rs->rs_opc = lustre_msg_get_opc(rs->rs_msg); + rs->rs_opc = lustre_msg_get_opc(req->rq_reqmsg); - spin_lock(&exp->exp_uncommitted_replies_lock); + cfs_spin_lock(&exp->exp_uncommitted_replies_lock); CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n", rs->rs_transno, exp->exp_last_committed); if (rs->rs_transno > exp->exp_last_committed) { /* not committed already */ - list_add_tail(&rs->rs_obd_list, - &exp->exp_uncommitted_replies); + cfs_list_add_tail(&rs->rs_obd_list, + &exp->exp_uncommitted_replies); } - spin_unlock (&exp->exp_uncommitted_replies_lock); + cfs_spin_unlock (&exp->exp_uncommitted_replies_lock); - spin_lock(&exp->exp_lock); - list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies); - spin_unlock(&exp->exp_lock); + cfs_spin_lock(&exp->exp_lock); + cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies); + cfs_spin_unlock(&exp->exp_lock); netrc = target_send_reply_msg (req, rc, fail_id); - spin_lock(&svc->srv_lock); + cfs_spin_lock(&svcpt->scp_rep_lock); - atomic_inc(&svc->srv_n_difficult_replies); + cfs_atomic_inc(&svcpt->scp_nreps_difficult); if (netrc != 0) { /* error sending: reply is off the net. Also we need +1 @@ -2190,207 +2514,24 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) * reply_out_callback leaves alone) */ rs->rs_on_net = 0; ptlrpc_rs_addref(rs); - atomic_inc (&svc->srv_outstanding_replies); } - spin_lock(&rs->rs_lock); + cfs_spin_lock(&rs->rs_lock); if (rs->rs_transno <= exp->exp_last_committed || (!rs->rs_on_net && !rs->rs_no_ack) || - list_empty(&rs->rs_exp_list) || /* completed already */ - list_empty(&rs->rs_obd_list)) { + cfs_list_empty(&rs->rs_exp_list) || /* completed already */ + cfs_list_empty(&rs->rs_obd_list)) { CDEBUG(D_HA, "Schedule reply immediately\n"); ptlrpc_dispatch_difficult_reply(rs); } else { - list_add (&rs->rs_list, &svc->srv_active_replies); - rs->rs_scheduled = 0; /* allow notifier to schedule */ - } - spin_unlock(&rs->rs_lock); - spin_unlock(&svc->srv_lock); - EXIT; -} - -int target_handle_ping(struct ptlrpc_request *req) -{ - obd_ping(req->rq_export); - return req_capsule_server_pack(&req->rq_pill); -} - -void target_committed_to_req(struct ptlrpc_request *req) -{ - struct obd_export *exp = req->rq_export; - - if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) - lustre_msg_set_last_committed(req->rq_repmsg, - exp->exp_last_committed); - else - DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/" - "%d)", exp->exp_obd->obd_no_transno, - req->rq_repmsg == NULL); - - CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n", - exp->exp_last_committed, req->rq_transno, req->rq_xid); -} -EXPORT_SYMBOL(target_committed_to_req); - -int target_handle_qc_callback(struct ptlrpc_request *req) -{ - struct obd_quotactl *oqctl; - struct client_obd *cli = &req->rq_export->exp_obd->u.cli; - - oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL); - if (oqctl == NULL) { - CERROR("Can't unpack obd_quotactl\n"); - RETURN(-EPROTO); - } - - cli->cl_qchk_stat = oqctl->qc_stat; - - return 0; -} - -#ifdef HAVE_QUOTA_SUPPORT -int target_handle_dqacq_callback(struct ptlrpc_request *req) -{ -#ifdef __KERNEL__ - struct obd_device *obd = req->rq_export->exp_obd; - struct obd_device *master_obd = NULL, *lov_obd = NULL; - struct obd_device_target *obt; - struct lustre_quota_ctxt *qctxt; - struct qunit_data *qdata = NULL; - int rc = 0; - ENTRY; - - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ)) - RETURN(rc); - - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) { - CERROR("packing reply failed!: rc = %d\n", rc); - RETURN(rc); - } - - LASSERT(req->rq_export); - - qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT); - if (IS_ERR(qdata)) { - rc = PTR_ERR(qdata); - CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc); - req->rq_status = rc; - GOTO(out, rc); - } - - /* we use the observer */ - if (obd_pin_observer(obd, &lov_obd) || - obd_pin_observer(lov_obd, &master_obd)) { - CERROR("Can't find the observer, it is recovering\n"); - req->rq_status = -EAGAIN; - GOTO(out, rc); - } - - obt = &master_obd->u.obt; - qctxt = &obt->obt_qctxt; - - if (!qctxt->lqc_setup || !qctxt->lqc_valid) { - /* quota_type has not been processed yet, return EAGAIN - * until we know whether or not quotas are supposed to - * be enabled */ - CDEBUG(D_QUOTA, "quota_type not processed yet, return " - "-EAGAIN\n"); - req->rq_status = -EAGAIN; - GOTO(out, rc); - } - - down_read(&obt->obt_rwsem); - if (qctxt->lqc_lqs_hash == NULL) { - up_read(&obt->obt_rwsem); - /* quota_type has not been processed yet, return EAGAIN - * until we know whether or not quotas are supposed to - * be enabled */ - CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return " - "-EAGAIN\n"); - req->rq_status = -EAGAIN; - GOTO(out, rc); - } - - LASSERT(qctxt->lqc_handler); - rc = qctxt->lqc_handler(master_obd, qdata, - lustre_msg_get_opc(req->rq_reqmsg)); - up_read(&obt->obt_rwsem); - if (rc && rc != -EDQUOT) - CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR, - "dqacq/dqrel failed! (rc:%d)\n", rc); - req->rq_status = rc; - - rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT); - if (rc < 0) { - CERROR("Can't pack qunit_data(rc: %d)\n", rc); - GOTO(out, rc); - } - - /* Block the quota req. b=14840 */ - OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout); - EXIT; - -out: - if (master_obd) - obd_unpin_observer(lov_obd); - if (lov_obd) - obd_unpin_observer(obd); - - rc = ptlrpc_reply(req); - return rc; -#else - return 0; -#endif /* !__KERNEL__ */ -} -#endif /* HAVE_QUOTA_SUPPORT */ - -/* Send a remote set_info_async. - * This may go from client to server or server to client - */ -int target_set_info_rpc(struct obd_import *imp, int opcode, - obd_count keylen, void *key, - obd_count vallen, void *val, - struct ptlrpc_request_set *set) -{ - struct ptlrpc_request *req; - char *tmp; - int rc; - ENTRY; - - req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO); - if (req == NULL) - RETURN(-ENOMEM); - - req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY, - RCL_CLIENT, keylen); - req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL, - RCL_CLIENT, vallen); - rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode); - if (rc) { - ptlrpc_request_free(req); - RETURN(rc); - } - - tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY); - memcpy(tmp, key, keylen); - tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL); - memcpy(tmp, val, vallen); - - ptlrpc_request_set_replen(req); - - if (set) { - ptlrpc_set_add_req(set, req); - ptlrpc_check_set(NULL, set); - } else { - rc = ptlrpc_queue_wait(req); - ptlrpc_req_finished(req); - } - - RETURN(rc); + cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active); + rs->rs_scheduled = 0; /* allow notifier to schedule */ + } + cfs_spin_unlock(&rs->rs_lock); + cfs_spin_unlock(&svcpt->scp_rep_lock); + EXIT; } -EXPORT_SYMBOL(target_set_info_rpc); - +EXPORT_SYMBOL(target_send_reply); ldlm_mode_t lck_compat_array[] = { [LCK_EX] LCK_COMPAT_EX, @@ -2437,7 +2578,7 @@ int ldlm_error2errno(ldlm_error_t error) if (((int)error) < 0) /* cast to signed type */ result = error; /* as ldlm_error_t can be unsigned */ else { - CERROR("Invalid DLM result code: %i\n", error); + CERROR("Invalid DLM result code: %d\n", error); result = -EPROTO; } } @@ -2484,15 +2625,122 @@ 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 (!list_empty(&exp->exp_locks_list)) { + cfs_spin_lock(&exp->exp_locks_list_guard); + if (!cfs_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); - list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link) - ldlm_lock_dump(D_ERROR, lock, 0); + cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link) + LDLM_ERROR(lock, "lock:"); } - spin_unlock(&exp->exp_locks_list_guard); + cfs_spin_unlock(&exp->exp_locks_list_guard); } #endif + +#ifdef HAVE_SERVER_SUPPORT +static int target_bulk_timeout(void *data) +{ + ENTRY; + /* We don't fail the connection here, because having the export + * killed makes the (vital) call to commitrw very sad. + */ + RETURN(1); +} + +static inline char *bulk2type(struct ptlrpc_bulk_desc *desc) +{ + return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT"; +} + +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; + + /* Check if there is eviction in progress, and if so, 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); + } + + /* 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); + } + + 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_success || + desc->bd_nob_transferred != desc->bd_nob) { + DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)", + desc->bd_success ? + "truncated" : "network error on", + bulk2type(desc), + desc->bd_nob_transferred, + desc->bd_nob); + /* 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); +} +EXPORT_SYMBOL(target_bulk_io); + +#endif /* HAVE_SERVER_SUPPORT */