X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fldlm%2Fldlm_lib.c;h=2483afb88ce53b843dc66be48aed47d50d0250db;hp=85527949ab1b7d37cc92cb2014709a365619a5f9;hb=389fde827be2ee6fb4ee08e955d773a2a16e70c6;hpb=defdb250fe5020e4aafe9913423d380d76866348 diff --git a/lustre/ldlm/ldlm_lib.c b/lustre/ldlm/ldlm_lib.c index 8552794..2483afb 100644 --- a/lustre/ldlm/ldlm_lib.c +++ b/lustre/ldlm/ldlm_lib.c @@ -26,8 +26,10 @@ * 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/ @@ -80,18 +82,19 @@ 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 +104,19 @@ 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) @@ -140,13 +144,13 @@ 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 +177,7 @@ 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", @@ -182,12 +186,37 @@ 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); } +/** + * 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) { /* drop security policy instance after all rpc finished/aborted @@ -197,6 +226,7 @@ void client_destroy_import(struct obd_import *imp) sptlrpc_import_sec_put(imp); class_import_put(imp); } +EXPORT_SYMBOL(client_destroy_import); /* configure an RPC client OBD device * @@ -212,6 +242,7 @@ 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; @@ -223,12 +254,16 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) connect_op = OST_CONNECT; cli->cl_sp_me = LUSTRE_SP_CLI; cli->cl_sp_to = LUSTRE_SP_OST; + ns_type = LDLM_NS_TYPE_OSC; + } else if (!strcmp(name, LUSTRE_MDC_NAME)) { 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 +271,8 @@ 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 { CERROR("unknown client OBD type \"%s\", can't setup\n", name); @@ -262,8 +299,8 @@ 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), @@ -273,8 +310,8 @@ 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); @@ -284,14 +321,14 @@ int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg) 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); 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; @@ -302,7 +339,7 @@ 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); + cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS); /* This value may be changed at connect time in ptlrpc_connect_interpret. */ @@ -311,11 +348,11 @@ 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; @@ -335,8 +372,6 @@ 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)); @@ -358,12 +393,22 @@ 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); - imp->imp_invalid = 1; - spin_unlock(&imp->imp_lock); + cfs_spin_lock(&imp->imp_lock); + imp->imp_deactive = 1; + cfs_spin_unlock(&imp->imp_lock); } } + obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name, + LDLM_NAMESPACE_CLIENT, + LDLM_NAMESPACE_GREEDY, + ns_type); + if (obddev->obd_namespace == NULL) { + CERROR("Unable to create client namespace - %s\n", + obddev->obd_name); + GOTO(err_import, rc = -ENOMEM); + } + cli->cl_qchk_stat = CL_NOT_QUOTACHECKED; RETURN(rc); @@ -380,6 +425,12 @@ err: int client_obd_cleanup(struct obd_device *obddev) { ENTRY; + + ldlm_namespace_free_post(obddev->obd_namespace); + obddev->obd_namespace = NULL; + + LASSERT(obddev->u.cli.cl_import == NULL); + ldlm_put_ref(); RETURN(0); } @@ -393,36 +444,23 @@ int client_connect_import(const struct lu_env *env, struct client_obd *cli = &obd->u.cli; struct obd_import *imp = cli->cl_import; struct obd_connect_data *ocd; - struct ldlm_namespace *to_be_freed = NULL; struct lustre_handle conn = { 0 }; int rc; 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); rc = class_connect(&conn, obd, cluuid); if (rc) GOTO(out_sem, rc); - + cli->cl_conn_count++; *exp = class_conn2export(&conn); - if (obd->obd_namespace != NULL) - CERROR("already have namespace!\n"); - - /* - * Deadlock case - bug 18380 - */ - up_write(&cli->cl_sem); - obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name, - LDLM_NAMESPACE_CLIENT, - LDLM_NAMESPACE_GREEDY); - down_write(&cli->cl_sem); - if (obd->obd_namespace == NULL) - GOTO(out_disco, rc = -ENOMEM); + LASSERT(obd->obd_namespace); imp->imp_dlm_handle = conn; rc = ptlrpc_init_import(imp); @@ -435,7 +473,7 @@ 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); @@ -455,18 +493,12 @@ int client_connect_import(const struct lu_env *env, if (rc) { out_ldlm: - ldlm_namespace_free_prior(obd->obd_namespace, imp, 0); - to_be_freed = obd->obd_namespace; - obd->obd_namespace = NULL; -out_disco: cli->cl_conn_count--; class_disconnect(*exp); *exp = NULL; } out_sem: - up_write(&cli->cl_sem); - if (to_be_freed) - ldlm_namespace_free_post(to_be_freed); + cfs_up_write(&cli->cl_sem); return rc; } @@ -477,7 +509,6 @@ int client_disconnect_export(struct obd_export *exp) struct client_obd *cli; struct obd_import *imp; int rc = 0, err; - struct ldlm_namespace *to_be_freed = NULL; ENTRY; if (!obd) { @@ -489,7 +520,10 @@ 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); + if (!cli->cl_conn_count) { CERROR("disconnecting disconnected device (%s)\n", obd->obd_name); @@ -503,9 +537,9 @@ 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 @@ -515,32 +549,19 @@ 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); - to_be_freed = obd->obd_namespace; } /* - * there's no necessary to hold sem during diconnecting an import, + * 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); - /* set obd_namespace to NULL only after invalidate, because we can have - * some connect requests in flight, and his need store a connect flags - * in obd_namespace. bug 14260 */ - obd->obd_namespace = NULL; - - 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; @@ -551,9 +572,41 @@ int client_disconnect_export(struct obd_export *exp) if (!rc && err) rc = err; - up_write(&cli->cl_sem); - if (to_be_freed) - ldlm_namespace_free_post(to_be_freed); + cfs_up_write(&cli->cl_sem); + + RETURN(rc); +} + +int server_disconnect_export(struct obd_export *exp) +{ + int rc; + ENTRY; + + /* Disconnect early so that clients can't keep using export */ + rc = class_disconnect(exp); + /* close import for avoid sending any requests */ + if (exp->exp_imp_reverse) + ptlrpc_cleanup_imp(exp->exp_imp_reverse); + + if (exp->exp_obd->obd_namespace != NULL) + ldlm_cancel_locks_for_export(exp); + + /* complete all outstanding replies */ + cfs_spin_lock(&exp->exp_lock); + while (!cfs_list_empty(&exp->exp_outstanding_replies)) { + struct ptlrpc_reply_state *rs = + cfs_list_entry(exp->exp_outstanding_replies.next, + struct ptlrpc_reply_state, rs_exp_list); + struct ptlrpc_service *svc = rs->rs_service; + + cfs_spin_lock(&svc->srv_rs_lock); + 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); + cfs_spin_unlock(&svc->srv_rs_lock); + } + cfs_spin_unlock(&exp->exp_lock); RETURN(rc); } @@ -567,24 +620,47 @@ 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) + LCONSOLE_WARN("%s: Client %s (at %s) reconnect" + "ing, waiting for %d clients in " + "recovery for %lds\n", target->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), + target->obd_max_recoverable_clients, + cfs_duration_sec(cfs_time_sub( + cfs_timer_deadline( + &target->obd_recovery_timer), + cfs_time_current()))); + 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: The server has already connected " + "client %s (at %s) with handle " LPX64 + ", rejecting a client with the same " + "uuid trying to reconnect with " + "handle " LPX64, 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 @@ -607,15 +683,24 @@ 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) { @@ -629,8 +714,11 @@ int target_handle_connect(struct ptlrpc_request *req) struct obd_uuid remote_uuid; char *str; int rc = 0; + char *target_start; + int target_len; int mds_conn = 0; struct obd_connect_data *data, *tmpdata; + int size, tmpsize; lnet_nid_t *client_nid = NULL; ENTRY; @@ -648,16 +736,17 @@ int target_handle_connect(struct ptlrpc_request *req) 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")); + 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 ? + "no target" : (target->obd_stopping ? + "stopping" : "not set up")); GOTO(out, rc = -ENODEV); } if (target->obd_no_conn) { - LCONSOLE_WARN("%s: temporarily refusing client connection " + LCONSOLE_WARN("%s: Temporarily refusing client connection " "from %s\n", target->obd_name, libcfs_nid2str(req->rq_peer.nid)); GOTO(out, rc = -EAGAIN); @@ -698,6 +787,8 @@ 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); @@ -743,50 +834,53 @@ int target_handle_connect(struct ptlrpc_request *req) if (obd_uuid_equals(&cluuid, &target->obd_uuid)) goto dont_check_exports; - export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid); + export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid); if (!export) goto no_export; /* we've found an export in the hash */ if (export->exp_connecting) { /* bug 9635, et. al. */ - CWARN("%s: exp %p already connecting\n", - export->exp_obd->obd_name, export); + 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) { 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); export = NULL; rc = 0; } else if (export->exp_connection != NULL && - req->rq_peer.nid != export->exp_connection->c_peer.nid) { + req->rq_peer.nid != export->exp_connection->c_peer.nid && + (lustre_msg_get_op_flags(req->rq_reqmsg) & + MSG_CONNECT_INITIAL)) { /* 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); + cfs_spin_lock(&export->exp_lock); export->exp_connecting = 1; - spin_unlock(&export->exp_lock); - class_export_put(export); + cfs_spin_unlock(&export->exp_lock); LASSERT(export->exp_obd == target); rc = target_handle_reconnect(&conn, export, &cluuid); @@ -797,30 +891,35 @@ int target_handle_connect(struct ptlrpc_request *req) no_export: OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout); } else if (req->rq_export == NULL && - 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); @@ -830,21 +929,16 @@ 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 */ + if (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 */ @@ -866,20 +960,27 @@ no_export: 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, - target->obd_recoverable_clients, - cfs_duration_sec(t)); + LCONSOLE_WARN("%s: Denying connection for new client " + "%s (at %s), waiting for %d clients in " + "recovery for "CFS_TIME_T"s\n", + target->obd_name, + libcfs_nid2str(req->rq_peer.nid), + cluuid.uuid, + cfs_atomic_read(&target-> \ + obd_lock_replay_clients), + cfs_duration_sec(t)); rc = -EBUSY; } else { dont_check_exports: rc = obd_connect(req->rq_svc_thread->t_env, &export, target, &cluuid, data, client_nid); - if (rc == 0) + if (rc == 0) { conn.cookie = export->exp_handle.h_cookie; + /* LU-1092 reconnect put export refcount in the + * end, connect needs take one here too. */ + class_export_get(export); + } } } else { rc = obd_reconnect(req->rq_svc_thread->t_env, @@ -890,13 +991,21 @@ dont_check_exports: } if (rc) GOTO(out, rc); + + 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. */ @@ -916,66 +1025,92 @@ dont_check_exports: req->rq_export = 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 higher " + "conn_cnt: %d > %d\n", cluuid.uuid, libcfs_nid2str(req->rq_peer.nid), export->exp_conn_cnt, lustre_msg_get_conn_cnt(req->rq_reqmsg)); GOTO(out, rc = -EALREADY); } + LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0); export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg); export->exp_abort_active_req = 0; /* 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) + if (export->exp_connection != NULL) { + /* Check to see if connection came from another NID */ + if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) && + !cfs_hlist_unhashed(&export->exp_nid_hash)) + cfs_hash_del(export->exp_obd->obd_nid_hash, + &export->exp_connection->c_peer.nid, + &export->exp_nid_hash); + ptlrpc_connection_put(export->exp_connection); + } + export->exp_connection = ptlrpc_connection_get(req->rq_peer, req->rq_self, &remote_uuid); - if (hlist_unhashed(&export->exp_nid_hash)) { - lustre_hash_add_unique(export->exp_obd->obd_nid_hash, - &export->exp_connection->c_peer.nid, - &export->exp_nid_hash); + 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); } - - spin_lock_bh(&target->obd_processing_task_lock); + /** + class_disconnect->class_export_recovery_cleanup() race + */ if (target->obd_recovering && !export->exp_in_recovery) { - spin_lock(&export->exp_lock); + int has_transno; + __u64 transno = data->ocd_transno; + + cfs_spin_lock(&export->exp_lock); 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++; - /* each connected client is counted as recoverable */ - target->obd_recoverable_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); + + /* Tell the client we're in recovery, when client is involved in it. */ + if (target->obd_recovering) + lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING); + tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN); conn = *tmp; @@ -984,19 +1119,6 @@ dont_check_exports: obd_set_info_async(export, sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD, 0, NULL, NULL); - /* in some recovery senarios, previous ctx init rpc handled - * in sptlrpc_target_export_check() might be used to install - * a reverse ctx in this reverse import, and later OBD_CONNECT - * using the same gss ctx could reach here and following new - * reverse import. note all reverse ctx in new/old import are - * actually based on the same gss ctx. so we invalidate ctx - * here before destroy import, otherwise flush old import will - * lead to remote reverse ctx be destroied, thus the reverse - * ctx of new import will lost its peer. - * there might be a better way to deal with this??? - */ - sptlrpc_import_inval_all_ctx(export->exp_imp_reverse); - client_destroy_import(export->exp_imp_reverse); } @@ -1025,6 +1147,12 @@ dont_check_exports: else revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT; + if ((export->exp_connect_flags & OBD_CONNECT_FULL20) && + (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1)) + revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18; + else + revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18; + rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr); if (rc) { CERROR("Failed to get sec for reverse import: %d\n", rc); @@ -1035,9 +1163,11 @@ dont_check_exports: class_import_put(revimp); out: 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) class_decref(targref, __FUNCTION__, cfs_current()); @@ -1068,80 +1198,35 @@ void target_destroy_export(struct obd_export *exp) if (exp->exp_imp_reverse != NULL) client_destroy_import(exp->exp_imp_reverse); - /* We cancel locks at disconnect time, but this will catch any locks - * granted in a race with recovery-induced disconnect. */ - if (exp->exp_obd->obd_namespace != NULL) - ldlm_cancel_locks_for_export(exp); + 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); } /* * Recovery functions */ - -struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req) +static void target_request_copy_get(struct ptlrpc_request *req) { - struct ptlrpc_request *copy_req; - struct lustre_msg *copy_reqmsg; - struct ptlrpc_user_desc *udesc = NULL; - - OBD_ALLOC_PTR(copy_req); - if (!copy_req) - return NULL; - OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen); - if (!copy_reqmsg){ - OBD_FREE_PTR(copy_req); - return NULL; - } - - if (orig_req->rq_user_desc) { - int ngroups = orig_req->rq_user_desc->pud_ngroups; - - OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups)); - if (!udesc) { - OBD_FREE(copy_reqmsg, orig_req->rq_reqlen); - OBD_FREE_PTR(copy_req); - return NULL; - } - memcpy(udesc, orig_req->rq_user_desc, - sptlrpc_user_desc_size(ngroups)); - } - - *copy_req = *orig_req; - memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen); - copy_req->rq_reqmsg = copy_reqmsg; - copy_req->rq_user_desc = udesc; - - class_export_rpc_get(copy_req->rq_export); - CFS_INIT_LIST_HEAD(©_req->rq_list); - CFS_INIT_LIST_HEAD(©_req->rq_replay_list); - sptlrpc_svc_ctx_addref(copy_req); - - if (copy_req->rq_reply_state) { - /* the copied req takes over the reply state */ - orig_req->rq_reply_state = NULL; - /* to catch further access */ - orig_req->rq_repmsg = NULL; - orig_req->rq_replen = 0; - } - - return copy_req; + class_export_rpc_get(req->rq_export); + LASSERT(cfs_list_empty(&req->rq_list)); + CFS_INIT_LIST_HEAD(&req->rq_replay_list); + + /* increase refcount to keep request in queue */ + cfs_atomic_inc(&req->rq_refcount); + /** let export know it has replays to be handled */ + cfs_atomic_inc(&req->rq_export->exp_replay_count); } -void ptlrpc_free_clone(struct ptlrpc_request *req) +static void target_request_copy_put(struct ptlrpc_request *req) { - LASSERT(list_empty(&req->rq_replay_list)); + LASSERT(cfs_list_empty(&req->rq_replay_list)); + LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count); - ptlrpc_req_drop_rs(req); - sptlrpc_svc_ctx_decref(req); + cfs_atomic_dec(&req->rq_export->exp_replay_count); class_export_rpc_put(req->rq_export); - list_del_init(&req->rq_list); - - if (req->rq_user_desc) { - int ngroups = req->rq_user_desc->pud_ngroups; - OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups)); - } - OBD_FREE(req->rq_reqmsg, req->rq_reqlen); - OBD_FREE_PTR(req); + ptlrpc_server_drop_request(req); } static int target_exp_enqueue_req_replay(struct ptlrpc_request *req) @@ -1153,9 +1238,9 @@ 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; @@ -1169,46 +1254,57 @@ 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) { + time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() - + obd->obd_recovery_start); ENTRY; - LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n", - obd->obd_name); + + 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)) { - obd->obd_processing_task = 0; - } else { + 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(); /* when recovery finished, cleanup orphans on mds and ost */ if (OBT(obd) && OBP(obd, postrecov)) { @@ -1217,21 +1313,19 @@ static void target_finish_recovery(struct obd_device *obd) LCONSOLE_WARN("%s: Post recovery failed, rc %d\n", obd->obd_name, rc); } - - obd->obd_recovery_end = cfs_time_current_sec(); EXIT; } 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)) { @@ -1239,30 +1333,29 @@ static void abort_req_replay_queue(struct obd_device *obd) "failed abort_req_reply; skipping"); } target_exp_dequeue_req_replay(req); - ptlrpc_free_clone(req); + target_request_copy_put(req); } } 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)) { DEBUG_REQ(D_ERROR, req, "failed abort_lock_reply; skipping"); } - ptlrpc_free_clone(req); + 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, @@ -1276,106 +1369,127 @@ 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; - LASSERT(obd->obd_stopping); - 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); - ptlrpc_free_clone(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); - ptlrpc_free_clone(req); + target_request_copy_put(req); } EXIT; } -/* 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); } -/* 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); + + now = cfs_time_current_sec(); + to = obd->obd_recovery_timeout; + end = obd->obd_recovery_start + to; + left = cfs_time_sub(end, now); - reset_recovery_timer(obd, obd->obd_recovery_timeout, 0); + 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 */ @@ -1385,24 +1499,26 @@ static void check_and_start_recovery_timer(struct obd_device *obd) * attempt timing. So the initial recovery time_out will be set to * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming * from client is bigger than this, then the recovery time_out will - * be extend to make sure the client could be reconnected, in the + * be extended to make sure the client could be reconnected, in the * process, the timeout from the new client should be ignored. */ 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_service->srv_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 */ @@ -1411,60 +1527,114 @@ 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 | LSI_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); +} + +/** Health checking routines */ +static inline int exp_connect_healthy(struct obd_export *exp) +{ + return (exp->exp_in_recovery); +} + +/** if export done req_replay or has replay in queue */ +static inline int exp_req_replay_healthy(struct obd_export *exp) +{ + return (!exp->exp_req_replay_needed || + cfs_atomic_read(&exp->exp_replay_count) > 0); +} +/** if export done lock_replay or has replay in queue */ +static inline int exp_lock_replay_healthy(struct obd_export *exp) +{ + return (!exp->exp_lock_replay_needed || + cfs_atomic_read(&exp->exp_replay_count) > 0); +} + +static inline int exp_vbr_healthy(struct obd_export *exp) +{ + return (!exp->exp_vbr_failed); +} + +static inline int exp_finished(struct obd_export *exp) +{ + return (exp->exp_in_recovery && !exp->exp_lock_replay_needed); +} + +/** Checking routines for recovery */ +static int check_for_clients(struct obd_device *obd) +{ + unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients); + + if (obd->obd_abort_recovery || obd->obd_recovery_expired) + return 1; + LASSERT(clnts <= obd->obd_max_recoverable_clients); + if (obd->obd_no_conn == 0 && + clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients) + return 1; + return 0; } -#ifdef __KERNEL__ static int check_for_next_transno(struct obd_device *obd) { struct ptlrpc_request *req = NULL; - int wake_up = 0, connected, completed, queue_len, max; + 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; } - max = obd->obd_max_recoverable_clients; - connected = obd->obd_connected_clients; - completed = connected - obd->obd_recoverable_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; CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, " "req_transno: "LPU64", next_transno: "LPU64"\n", - max, connected, completed, queue_len, req_transno, next_transno); + obd->obd_max_recoverable_clients, connected, completed, + queue_len, req_transno, next_transno); if (obd->obd_abort_recovery) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; - } else if (atomic_read(&obd->obd_req_replay_clients) == 0) { + } else if (obd->obd_recovery_expired) { + CDEBUG(D_HA, "waking for expired recovery\n"); + wake_up = 1; + } 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 + completed == max) { - /* handle gaps occured due to lost reply. It is allowed gaps - * because all clients are connected and there will be resend - * for missed transaction */ + } 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, "req_transno: "LPU64", next_transno: "LPU64"\n", req_transno, next_transno); - - CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA, - "waking for skipped transno (skip: "LPD64 - ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n", - next_transno, queue_len, completed, connected, req_transno); + if (req_transno > obd->obd_last_committed && + !obd->obd_version_recov) + d_lvl = D_ERROR; + CDEBUG(d_lvl, + "%s: waking for gap in transno, VBR is %s (skip: " + LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64 + ", last_committed: "LPD64")\n", + obd->obd_name, obd->obd_version_recov ? "ON" : "OFF", + next_transno, queue_len, completed, connected, + req_transno, obd->obd_last_committed); obd->obd_next_recovery_transno = req_transno; wake_up = 1; } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) { @@ -1472,154 +1642,149 @@ static int check_for_next_transno(struct obd_device *obd) " by fail_lock, waking up ("LPD64")\n", next_transno); obd->obd_next_recovery_transno = req_transno; wake_up = 1; - } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) { - /* some clients haven't connected in time, but we can try - * to replay requests that demand on already committed ones - * also, we can replay first non-committed transation */ - LASSERT(req_transno != 0); - if (obd->obd_version_recov || - req_transno == obd->obd_last_committed + 1) { - obd->obd_next_recovery_transno = req_transno; - } else if (req_transno > obd->obd_last_committed) { - /* can't continue recovery: have no needed transno */ - obd->obd_abort_recovery = 1; - CDEBUG(D_ERROR, "abort due to missed clients. max: %d, " - "connected: %d, completed: %d, queue_len: %d, " - "req_transno: "LPU64", next_transno: "LPU64"\n", - max, connected, completed, queue_len, - req_transno, next_transno); - } - wake_up = 1; } - - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } -static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd) -{ - struct l_wait_info lwi = { 0 }; - struct ptlrpc_request *req; - - CDEBUG(D_HA, "Waiting for transno "LPD64"\n", - obd->obd_next_recovery_transno); - l_wait_event(obd->obd_next_transno_waitq, - check_for_next_transno(obd), &lwi); - - spin_lock_bh(&obd->obd_processing_task_lock); - if (obd->obd_abort_recovery) { - req = NULL; - } else if (!list_empty(&obd->obd_req_replay_queue)) { - req = list_entry(obd->obd_req_replay_queue.next, - struct ptlrpc_request, rq_list); - target_exp_dequeue_req_replay(req); - list_del_init(&req->rq_list); - obd->obd_requests_queued_for_recovery--; - } else { - req = NULL; - } - spin_unlock_bh(&obd->obd_processing_task_lock); - RETURN(req); -} - static int check_for_next_lock(struct obd_device *obd) { - struct ptlrpc_request *req = NULL; int wake_up = 0; - 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); + 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) { CDEBUG(D_HA, "waking for aborted recovery\n"); wake_up = 1; + } else if (obd->obd_recovery_expired) { + CDEBUG(D_HA, "waking for expired recovery\n"); + wake_up = 1; } - spin_unlock_bh(&obd->obd_processing_task_lock); + cfs_spin_unlock(&obd->obd_recovery_task_lock); return wake_up; } +/** + * wait for recovery events, + * check its status with help of check_routine + * evict dead clients via health_check + */ +static int target_recovery_overseer(struct obd_device *obd, + int (*check_routine)(struct obd_device *), + int (*health_check)(struct obd_export *)) +{ +repeat: + cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd)); + if (obd->obd_abort_recovery) { + CWARN("recovery is aborted, evict exports in recovery\n"); + /** evict exports which didn't finish recovery yet */ + class_disconnect_stale_exports(obd, exp_finished); + return 1; + } else if (obd->obd_recovery_expired) { + obd->obd_recovery_expired = 0; + /** If some clients died being recovered, evict them */ + 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 */ + 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) +{ + struct ptlrpc_request *req = NULL; + ENTRY; + + CDEBUG(D_HA, "Waiting for transno "LPD64"\n", + obd->obd_next_recovery_transno); + + if (target_recovery_overseer(obd, check_for_next_transno, + exp_req_replay_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); + } + + 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--; + cfs_spin_unlock(&obd->obd_recovery_task_lock); + } else { + 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); + } + RETURN(req); +} + static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd) { - struct l_wait_info lwi = { 0 }; - struct ptlrpc_request *req; + struct ptlrpc_request *req = NULL; CDEBUG(D_HA, "Waiting for lock\n"); - l_wait_event(obd->obd_next_transno_waitq, - check_for_next_lock(obd), &lwi); + if (target_recovery_overseer(obd, check_for_next_lock, + exp_lock_replay_healthy)) + abort_lock_replay_queue(obd); - spin_lock_bh(&obd->obd_processing_task_lock); - if (obd->obd_abort_recovery) { - req = NULL; - } else 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); + 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 { - req = NULL; + 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); } - spin_unlock_bh(&obd->obd_processing_task_lock); return req; } static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd) { - struct ptlrpc_request *req; + 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 { - req = NULL; + cfs_spin_unlock(&obd->obd_recovery_task_lock); } - spin_unlock_bh(&obd->obd_processing_task_lock); return req; } -static inline int req_vbr_done(struct obd_export *exp) -{ - return (exp->exp_vbr_failed == 0); -} - -static inline int req_replay_done(struct obd_export *exp) -{ - return (exp->exp_req_replay_needed == 0); -} - -static inline int lock_replay_done(struct obd_export *exp) -{ - return (exp->exp_lock_replay_needed == 0); -} - -static inline int connect_done(struct obd_export *exp) -{ - return (exp->exp_in_recovery != 0); -} - -static int check_for_clients(struct obd_device *obd) -{ - if (obd->obd_abort_recovery || obd->obd_version_recov) - return 1; - LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients); - if (obd->obd_no_conn == 0 && - obd->obd_connected_clients == obd->obd_max_recoverable_clients) - return 1; - return 0; -} - static int handle_recovery_req(struct ptlrpc_thread *thread, struct ptlrpc_request *req, svc_handler_t handler) @@ -1627,31 +1792,56 @@ static int handle_recovery_req(struct ptlrpc_thread *thread, int rc; ENTRY; - rc = lu_context_init(&req->rq_session, LCT_SESSION); + /** + * 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); - return rc; + GOTO(reqcopy_put, rc); } - req->rq_session.lc_thread = thread; - lu_context_enter(&req->rq_session); + + req->rq_recov_session.lc_thread = thread; + lu_context_enter(&req->rq_recov_session); req->rq_svc_thread = thread; - req->rq_svc_thread->t_env->le_ses = &req->rq_session; + req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session; /* thread context */ lu_context_enter(&thread->t_env->le_ctx); (void)handler(req); lu_context_exit(&thread->t_env->le_ctx); - lu_context_exit(&req->rq_session); - lu_context_fini(&req->rq_session); + lu_context_exit(&req->rq_recov_session); + lu_context_fini(&req->rq_recov_session); /* don't reset timer for final stage */ - if (!req_replay_done(req->rq_export) || - !lock_replay_done(req->rq_export)) - reset_recovery_timer(class_exp2obd(req->rq_export), - AT_OFF ? obd_timeout : - at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1); - ptlrpc_free_clone(req); - RETURN(0); + 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 + */ + if (!AT_OFF) { + struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service; + /* 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(&svc->srv_at_estimate)), + (int)lustre_msg_get_timeout(req->rq_reqmsg)); + /* Add net_latency (see ptlrpc_replay_req) */ + to += lustre_msg_get_service_time(req->rq_reqmsg); + } + extend_recovery_timer(class_exp2obd(req->rq_export), to, true); + } +reqcopy_put: + RETURN(rc); } static int target_recovery_thread(void *arg) @@ -1660,62 +1850,63 @@ static int target_recovery_thread(void *arg) struct obd_device *obd = lut->lut_obd; struct ptlrpc_request *req; struct target_recovery_data *trd = &obd->obd_recovery_data; - struct l_wait_info lwi = { 0 }; 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); + 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 */ - obd->obd_abort_recovery = 0; - l_wait_event(obd->obd_next_transno_waitq, - check_for_clients(obd), &lwi); - - /* If some clients haven't connected in time, evict them */ - if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) { - CWARN("Some clients haven't connect in time (%d/%d)," - "evict them\n", obd->obd_connected_clients, - obd->obd_max_recoverable_clients); - class_disconnect_stale_exports(obd, connect_done, - exp_flags_from_obd(obd) | - OBD_OPT_ABORT_RECOV); - /** - * if recovery proceeds with versions then some clients may be - * timed out waiting for others and trying to reconnect. - * Extend timer for such reconnect cases. - */ - if (obd->obd_version_recov) - reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1); + if (target_recovery_overseer(obd, check_for_clients, + exp_connect_healthy)) { + abort_req_replay_queue(obd); + abort_lock_replay_queue(obd); } /* next stage: replay requests */ delta = jiffies; - 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()); @@ -1724,81 +1915,73 @@ static int target_recovery_thread(void *arg) libcfs_nid2str(req->rq_peer.nid)); handle_recovery_req(thread, req, trd->trd_recovery_handler); - obd->obd_replayed_requests++; - spin_lock_bh(&obd->obd_processing_task_lock); + /** + * bz18031: increase next_recovery_transno before + * target_request_copy_put() will drop exp_rpc reference + */ + cfs_spin_lock(&obd->obd_recovery_task_lock); obd->obd_next_recovery_transno++; - spin_unlock_bh(&obd->obd_processing_task_lock); - } - - /* If some clients haven't replayed requests in time, evict them */ - if (obd->obd_abort_recovery) { - CDEBUG(D_WARNING, "req replay is aborted\n"); - class_disconnect_stale_exports(obd, req_replay_done, - exp_flags_from_obd(obd) | - OBD_OPT_ABORT_RECOV); - abort_req_replay_queue(obd); + cfs_spin_unlock(&obd->obd_recovery_task_lock); + target_exp_dequeue_req_replay(req); + target_request_copy_put(req); + obd->obd_replayed_requests++; } - LASSERT(list_empty(&obd->obd_req_replay_queue)); - /* The second stage: replay locks */ + /** + * 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: ", libcfs_nid2str(req->rq_peer.nid)); handle_recovery_req(thread, req, trd->trd_recovery_handler); + target_request_copy_put(req); obd->obd_replayed_locks++; } - /* If some clients haven't replayed requests in time, evict them */ - if (obd->obd_abort_recovery) { - CERROR("lock replay is aborted\n"); - class_disconnect_stale_exports(obd, lock_replay_done, - exp_flags_from_obd(obd) | - OBD_OPT_ABORT_RECOV); - abort_lock_replay_queue(obd); - } - LASSERT(list_empty(&obd->obd_lock_replay_queue)); - - /* The third stage: reply on final pings */ + /** + * The third stage: reply on final pings, at this moment all clients + * must have request in final queue + */ CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n"); - /** evict exports failed VBR */ - class_disconnect_stale_exports(obd, req_vbr_done, - exp_flags_from_obd(obd) | - OBD_OPT_ABORT_RECOV); /** Update server last boot epoch */ lut_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: ", libcfs_nid2str(req->rq_peer.nid)); handle_recovery_req(thread, req, trd->trd_recovery_handler); + 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); - LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); - LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0); - 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); } @@ -1810,12 +1993,12 @@ 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; @@ -1825,16 +2008,17 @@ 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; - 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); + /** 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); + } + cfs_spin_unlock(&obd->obd_dev_lock); + cfs_wait_for_completion(&trd->trd_finishing); } } @@ -1849,26 +2033,24 @@ EXPORT_SYMBOL(target_recovery_fini); static void target_recovery_expired(unsigned long castmeharder) { struct obd_device *obd = (struct obd_device *)castmeharder; - CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected " - "after %lds (%d clients did)\n", - obd->obd_name, obd->obd_recoverable_clients, + CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery" + " after %lds (%d clients connected)\n", + obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients), cfs_time_current_sec()- obd->obd_recovery_start, - obd->obd_connected_clients); + cfs_atomic_read(&obd->obd_connected_clients)); - spin_lock_bh(&obd->obd_processing_task_lock); - obd->obd_version_recov = 1; - CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n", - atomic_read(&obd->obd_req_replay_clients), - obd->obd_next_recovery_transno); + 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) { struct obd_device *obd = lut->lut_obd; - if (obd->obd_max_recoverable_clients == 0) + if (obd->obd_max_recoverable_clients == 0) { + /** Update server last boot epoch */ + lut_boot_epoch_update(lut); return; + } CWARN("RECOVERY: service %s, %d recoverable clients, " "last_transno "LPU64"\n", obd->obd_name, @@ -1877,9 +2059,7 @@ 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); } @@ -1894,43 +2074,38 @@ 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); - LASSERT(obd->obd_recoverable_clients > 0); - obd->obd_recoverable_clients--; - if (atomic_read(&obd->obd_req_replay_clients) == 0) - CDEBUG(D_HA, "all clients have replayed reqs\n"); + 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); - if (atomic_read(&obd->obd_lock_replay_clients) == 0) - CDEBUG(D_HA, "all clients have replayed locks\n"); + 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; @@ -1945,50 +2120,43 @@ int target_queue_recovery_request(struct ptlrpc_request *req, if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) { /* client declares he's ready to complete recovery * so, we put the request on th final queue */ - req = ptlrpc_clone_req(req); - if (req == NULL) - RETURN(-ENOMEM); + 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); - if (obd->obd_recovering) - list_add_tail(&req->rq_list, &obd->obd_final_req_queue); - else { - spin_unlock_bh(&obd->obd_processing_task_lock); - ptlrpc_free_clone(req); - if (obd->obd_stopping) { - RETURN(-ENOTCONN); - } else { - RETURN(1); - } + cfs_spin_lock(&obd->obd_recovery_task_lock); + if (obd->obd_recovering) { + cfs_list_add_tail(&req->rq_list, + &obd->obd_final_req_queue); + } else { + 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 */ - req = ptlrpc_clone_req(req); - if (req == NULL) - RETURN(-ENOMEM); + 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); - ptlrpc_free_clone(req); + 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); @@ -1996,8 +2164,6 @@ 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. * @@ -2008,46 +2174,42 @@ 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) { + CWARN("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); - req = ptlrpc_clone_req(req); - if (req == NULL) - RETURN(-ENOMEM); - - spin_lock_bh(&obd->obd_processing_task_lock); - LASSERT(obd->obd_recovering); + target_request_copy_get(req); if (!req->rq_export->exp_in_recovery) { - spin_unlock_bh(&obd->obd_processing_task_lock); - ptlrpc_free_clone(req); + 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"); - ptlrpc_free_clone(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; } @@ -2056,33 +2218,22 @@ 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); - ptlrpc_free_clone(req); + target_request_copy_put(req); RETURN(0); } } 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); } -struct obd_device * target_req2obd(struct ptlrpc_request *req) -{ - return req->rq_export->exp_obd; -} - -static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp) -{ - LASSERT(exp != NULL); - return &exp->exp_obd->obd_namespace->ns_pool; -} - /** * Packs current SLV and Limit into \a req. */ @@ -2107,10 +2258,10 @@ 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); } @@ -2123,7 +2274,7 @@ 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 { @@ -2137,7 +2288,6 @@ void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id) { int netrc; struct ptlrpc_reply_state *rs; - struct obd_device *obd; struct obd_export *exp; struct ptlrpc_service *svc; ENTRY; @@ -2167,63 +2317,62 @@ 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 onto srv_reply_queue while I'm setting up */ + /* disable reply scheduling while I'm setting up */ rs->rs_scheduled = 1; rs->rs_on_net = 1; rs->rs_xid = req->rq_xid; rs->rs_transno = req->rq_transno; rs->rs_export = exp; + rs->rs_opc = lustre_msg_get_opc(rs->rs_msg); - 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(&svc->srv_rs_lock); - atomic_inc(&svc->srv_n_difficult_replies); + cfs_atomic_inc(&svc->srv_n_difficult_replies); if (netrc != 0) { /* error sending: reply is off the net. Also we need +1 - * reply ref until ptlrpc_server_handle_reply() is done + * reply ref until ptlrpc_handle_rs() is done * with the reply state (if the send was successful, there * would have been +1 ref for the net, which * reply_out_callback leaves alone) */ rs->rs_on_net = 0; ptlrpc_rs_addref(rs); - 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); + cfs_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); + cfs_spin_unlock(&rs->rs_lock); + cfs_spin_unlock(&svc->srv_rs_lock); EXIT; } @@ -2271,7 +2420,7 @@ 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; + struct obd_device *master_obd = NULL, *lov_obd = NULL; struct obd_device_target *obt; struct lustre_quota_ctxt *qctxt; struct qunit_data *qdata = NULL; @@ -2298,13 +2447,13 @@ int target_handle_dqacq_callback(struct ptlrpc_request *req) } /* we use the observer */ - if (!obd->obd_observer || !obd->obd_observer->obd_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); } - master_obd = obd->obd_observer->obd_observer; obt = &master_obd->u.obt; qctxt = &obt->obt_qctxt; @@ -2318,9 +2467,9 @@ int target_handle_dqacq_callback(struct ptlrpc_request *req) GOTO(out, rc); } - down_read(&obt->obt_rwsem); + cfs_down_read(&obt->obt_rwsem); if (qctxt->lqc_lqs_hash == NULL) { - up_read(&obt->obt_rwsem); + cfs_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 */ @@ -2333,7 +2482,7 @@ int target_handle_dqacq_callback(struct ptlrpc_request *req) LASSERT(qctxt->lqc_handler); rc = qctxt->lqc_handler(master_obd, qdata, lustre_msg_get_opc(req->rq_reqmsg)); - up_read(&obt->obt_rwsem); + cfs_up_read(&obt->obt_rwsem); if (rc && rc != -EDQUOT) CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR, "dqacq/dqrel failed! (rc:%d)\n", rc); @@ -2350,6 +2499,11 @@ int target_handle_dqacq_callback(struct ptlrpc_request *req) EXIT; out: + if (master_obd) + obd_unpin_observer(lov_obd); + if (lov_obd) + obd_unpin_observer(obd); + rc = ptlrpc_reply(req); return rc; #else @@ -2403,7 +2557,7 @@ 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; } } @@ -2447,3 +2601,122 @@ ldlm_error_t ldlm_errno2error(int err_no) } EXPORT_SYMBOL(ldlm_errno2error); +#if LUSTRE_TRACKS_LOCK_EXP_REFS +void ldlm_dump_export_locks(struct obd_export *exp) +{ + 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); + cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link) + ldlm_lock_dump(D_ERROR, lock, 0); + } + cfs_spin_unlock(&exp->exp_locks_list_guard); +} +#endif + +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);