X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fptlrpc%2Frecover.c;h=02f79d08e306d6f3d1a32e400527c63e76840942;hb=51f1b46dab6c2cc16f71d3f3be483f530b5e4001;hp=0b3a1b8e4b3bd4e31dff6033c812b73aa61f57d7;hpb=762e5b7708660eb86c614b6ac1c5aaa6bf7acc4c;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/recover.c b/lustre/ptlrpc/recover.c index 0b3a1b8..02f79d0 100644 --- a/lustre/ptlrpc/recover.c +++ b/lustre/ptlrpc/recover.c @@ -1,225 +1,404 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: +/* + * GPL HEADER START * - * Portal-RPC reconnection and replay operations, for use in recovery. + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * - * This code is issued under the GNU General Public License. - * See the file COPYING in this distribution + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 only, + * as published by the Free Software Foundation. * - * Copryright (C) 1996 Peter J. Braam - * Copryright (C) 1999 Stelias Computing Inc. - * Copryright (C) 1999 Seagate Technology Inc. - * Copryright (C) 2001 Mountain View Data, Inc. - * Copryright (C) 2002 Cluster File Systems, Inc. + * This program is distributed in the hope that it will be useful, but + * WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * General Public License version 2 for more details (a copy is included + * in the LICENSE file that accompanied this code). * + * You should have received a copy of the GNU General Public License + * version 2 along with this program; If not, see + * http://www.gnu.org/licenses/gpl-2.0.html + * + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. + * + * Copyright (c) 2011, 2015, Intel Corporation. + */ +/* + * This file is part of Lustre, http://www.lustre.org/ + * Lustre is a trademark of Sun Microsystems, Inc. + * + * lustre/ptlrpc/recover.c + * + * Author: Mike Shaver */ - -#include -#include -#include #define DEBUG_SUBSYSTEM S_RPC +#include +#include +#include +#include +#include +#include +#include +#include +#include -#include -#include -#include +#include "ptlrpc_internal.h" -static int ptlrpc_reconnect(struct ptlrpc_connection *conn) +/** + * Start recovery on disconnected import. + * This is done by just attempting a connect + */ +void ptlrpc_initiate_recovery(struct obd_import *imp) { - struct list_head *tmp; - int rc = -EINVAL; - - /* XXX c_lock semantics! */ - conn->c_level = LUSTRE_CONN_CON; - - /* XXX this code MUST be shared with class_obd_connect! */ - list_for_each(tmp, &conn->c_imports) { - struct obd_import *imp = list_entry(tmp, struct obd_import, - imp_chain); - struct obd_device *obd = imp->imp_obd; - struct client_obd *cli = &obd->u.cli; - int rq_opc = (obd->obd_type->typ_ops->o_brw) - ? OST_CONNECT : MDS_CONNECT; - int size[] = { sizeof(cli->cl_target_uuid), - sizeof(obd->obd_uuid) }; - char *tmp[] = {cli->cl_target_uuid, obd->obd_uuid }; - struct lustre_handle old_hdl; - struct ptlrpc_request *request; - struct obd_export *ldlmexp; - - LASSERT(imp->imp_connection == conn); - request = ptlrpc_prep_req(imp, rq_opc, 2, size, tmp); - request->rq_level = LUSTRE_CONN_NEW; - request->rq_replen = lustre_msg_size(0, NULL); - /* - * This address is the export that represents our client-side - * LDLM service (for ASTs). We should only have one on this - * list, so we just grab the first one. - * - * XXX tear down export, call class_obd_connect! - */ - ldlmexp = list_entry(obd->obd_exports.next, struct obd_export, - exp_obd_chain); - request->rq_reqmsg->addr = (__u64)(unsigned long)ldlmexp; - request->rq_reqmsg->cookie = ldlmexp->exp_cookie; - rc = ptlrpc_queue_wait(request); - rc = ptlrpc_check_status(request, rc); - if (rc) { - CERROR("cannot connect to %s@%s: rc = %d\n", - cli->cl_target_uuid, conn->c_remote_uuid, rc); - ptlrpc_free_req(request); - GOTO(out_disc, rc = -ENOTCONN); - } + ENTRY; + + CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd)); + ptlrpc_connect_import(imp); + + EXIT; +} + +/** + * Identify what request from replay list needs to be replayed next + * (based on what we have already replayed) and send it to server. + */ +int ptlrpc_replay_next(struct obd_import *imp, int *inflight) +{ + int rc = 0; + struct list_head *tmp, *pos; + struct ptlrpc_request *req = NULL; + __u64 last_transno; + ENTRY; + + *inflight = 0; + + /* It might have committed some after we last spoke, so make sure we + * get rid of them now. + */ + spin_lock(&imp->imp_lock); + imp->imp_last_transno_checked = 0; + ptlrpc_free_committed(imp); + last_transno = imp->imp_last_replay_transno; + spin_unlock(&imp->imp_lock); - old_hdl = imp->imp_handle; - imp->imp_handle.addr = request->rq_repmsg->addr; - imp->imp_handle.cookie = request->rq_repmsg->cookie; - CERROR("reconnected to %s@%s (%Lx/%Lx, was %Lx/%Lx)!\n", - cli->cl_target_uuid, conn->c_remote_uuid, - imp->imp_handle.addr, imp->imp_handle.cookie, - old_hdl.addr, old_hdl.cookie); - ptlrpc_req_finished(request); + CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n", + imp, obd2cli_tgt(imp->imp_obd), + imp->imp_peer_committed_transno, last_transno); + + /* Do I need to hold a lock across this iteration? We shouldn't be + * racing with any additions to the list, because we're in recovery + * and are therefore not processing additional requests to add. Calls + * to ptlrpc_free_committed might commit requests, but nothing "newer" + * than the one we're replaying (it can't be committed until it's + * replayed, and we're doing that here). l_f_e_safe protects against + * problems with the current request being committed, in the unlikely + * event of that race. So, in conclusion, I think that it's safe to + * perform this list-walk without the imp_lock held. + * + * But, the {mdc,osc}_replay_open callbacks both iterate + * request lists, and have comments saying they assume the + * imp_lock is being held by ptlrpc_replay, but it's not. it's + * just a little race... + */ + + /* Replay all the committed open requests on committed_list first */ + if (!list_empty(&imp->imp_committed_list)) { + tmp = imp->imp_committed_list.prev; + req = list_entry(tmp, struct ptlrpc_request, + rq_replay_list); + + /* The last request on committed_list hasn't been replayed */ + if (req->rq_transno > last_transno) { + /* Since the imp_committed_list is immutable before + * all of it's requests being replayed, it's safe to + * use a cursor to accelerate the search */ + if (!imp->imp_resend_replay || + imp->imp_replay_cursor == &imp->imp_committed_list) + imp->imp_replay_cursor = + imp->imp_replay_cursor->next; + + while (imp->imp_replay_cursor != + &imp->imp_committed_list) { + req = list_entry(imp->imp_replay_cursor, + struct ptlrpc_request, + rq_replay_list); + if (req->rq_transno > last_transno) + break; + + req = NULL; + imp->imp_replay_cursor = + imp->imp_replay_cursor->next; + } + } else { + /* All requests on committed_list have been replayed */ + imp->imp_replay_cursor = &imp->imp_committed_list; + req = NULL; + } + } + + /* All the requests in committed list have been replayed, let's replay + * the imp_replay_list */ + if (req == NULL) { + list_for_each_safe(tmp, pos, &imp->imp_replay_list) { + req = list_entry(tmp, struct ptlrpc_request, + rq_replay_list); + + if (req->rq_transno > last_transno) + break; + req = NULL; + } + } + + /* If need to resend the last sent transno (because a reconnect + * has occurred), then stop on the matching req and send it again. + * If, however, the last sent transno has been committed then we + * continue replay from the next request. */ + if (req != NULL && imp->imp_resend_replay) + lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT); + + spin_lock(&imp->imp_lock); + /* The resend replay request may have been removed from the + * unreplied list. */ + if (req != NULL && imp->imp_resend_replay && + list_empty(&req->rq_unreplied_list)) { + ptlrpc_add_unreplied(req); + imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp); + } + + imp->imp_resend_replay = 0; + spin_unlock(&imp->imp_lock); + + if (req != NULL) { + /* The request should have been added back in unreplied list + * by ptlrpc_prepare_replay(). */ + LASSERT(!list_empty(&req->rq_unreplied_list)); + + rc = ptlrpc_replay_req(req); + if (rc) { + CERROR("recovery replay error %d for req " + "%llu\n", rc, req->rq_xid); + RETURN(rc); + } + *inflight = 1; + } + RETURN(rc); +} + +/** + * Schedule resending of request on sending_list. This is done after + * we completed replaying of requests and locks. + */ +int ptlrpc_resend(struct obd_import *imp) +{ + struct ptlrpc_request *req, *next; + + ENTRY; + + /* As long as we're in recovery, nothing should be added to the sending + * list, so we don't need to hold the lock during this iteration and + * resend process. + */ + /* Well... what if lctl recover is called twice at the same time? + */ + spin_lock(&imp->imp_lock); + if (imp->imp_state != LUSTRE_IMP_RECOVER) { + spin_unlock(&imp->imp_lock); + RETURN(-1); } - conn->c_level = LUSTRE_CONN_RECOVD; - out_disc: - return rc; + list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) { + LASSERTF((long)req > PAGE_SIZE && req != LP_POISON, + "req %p bad\n", req); + LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req); + + /* If the request is allowed to be sent during replay and it + * is not timeout yet, then it does not need to be resent. */ + if (!ptlrpc_no_resend(req) && + (req->rq_timedout || !req->rq_allow_replay)) + ptlrpc_resend_req(req); + } + spin_unlock(&imp->imp_lock); + + RETURN(0); } -int ptlrpc_run_recovery_upcall(struct ptlrpc_connection *conn) +/** + * Go through all requests in delayed list and wake their threads + * for resending + */ +void ptlrpc_wake_delayed(struct obd_import *imp) { - char *argv[3]; - char *envp[3]; - int rc; + struct list_head *tmp, *pos; + struct ptlrpc_request *req; + + spin_lock(&imp->imp_lock); + list_for_each_safe(tmp, pos, &imp->imp_delayed_list) { + req = list_entry(tmp, struct ptlrpc_request, rq_list); + DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set); + ptlrpc_client_wake_req(req); + } + spin_unlock(&imp->imp_lock); +} + +void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req) +{ + struct obd_import *imp = failed_req->rq_import; ENTRY; - conn->c_level = LUSTRE_CONN_RECOVD; - - argv[0] = obd_recovery_upcall; - argv[1] = conn->c_remote_uuid; - argv[2] = NULL; - - envp[0] = "HOME=/"; - envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; - envp[2] = NULL; - - rc = call_usermodehelper(argv[0], argv, envp); - if (rc < 0) { - CERROR("Error invoking recovery upcall %s for %s: %d\n", - argv[0], argv[1], rc); - CERROR("Check /proc/sys/lustre/recovery_upcall?\n"); - } else { - CERROR("Invoked upcall %s for connection %s\n", - argv[0], argv[1]); + + CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n", + imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid); + + if (ptlrpc_set_import_discon(imp, + lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) { + if (!imp->imp_replayable) { + CDEBUG(D_HA, "import %s@%s for %s not replayable, " + "auto-deactivating\n", + obd2cli_tgt(imp->imp_obd), + imp->imp_connection->c_remote_uuid.uuid, + imp->imp_obd->obd_name); + ptlrpc_deactivate_import(imp); + } + /* to control recovery via lctl {disable|enable}_recovery */ + if (imp->imp_deactive == 0) + ptlrpc_connect_import(imp); } - /* - * We don't want to make this a "failed" recovery, because the system - * administrator -- or, perhaps, tester -- may well be able to rescue - * things by running the correct upcall. - */ - RETURN(0); + /* Wait for recovery to complete and resend. If evicted, then + this request will be errored out later.*/ + spin_lock(&failed_req->rq_lock); + if (!failed_req->rq_no_resend) + failed_req->rq_resend = 1; + spin_unlock(&failed_req->rq_lock); + + EXIT; } -int ptlrpc_reconnect_and_replay(struct ptlrpc_connection *conn) +/** + * Administratively active/deactive a client. + * This should only be called by the ioctl interface, currently + * - the lctl deactivate and activate commands + * - echo 0/1 >> /proc/osc/XXX/active + * - client umount -f (ll_umount_begin) + */ +int ptlrpc_set_import_active(struct obd_import *imp, int active) { + struct obd_device *obd = imp->imp_obd; int rc = 0; - struct list_head *tmp, *pos; - struct ptlrpc_request *req; + ENTRY; + LASSERT(obd); - /* 1. reconnect */ - rc = ptlrpc_reconnect(conn); - if (rc) - RETURN(rc); - - /* 2. walk the request list */ - spin_lock(&conn->c_lock); - - CDEBUG(D_HA, "connection %p to %s has last_xid "LPD64"\n", - conn, conn->c_remote_uuid, conn->c_last_xid); - - list_for_each_safe(tmp, pos, &conn->c_sending_head) { - req = list_entry(tmp, struct ptlrpc_request, rq_list); - - /* replay what needs to be replayed */ - if (req->rq_flags & PTL_RPC_FL_REPLAY) { - CDEBUG(D_HA, "FL_REPLAY: xid "LPD64" transno "LPD64" op %d @ %d\n", - req->rq_xid, req->rq_repmsg->transno, req->rq_reqmsg->opc, - req->rq_import->imp_client->cli_request_portal); - rc = ptlrpc_replay_req(req); -#if 0 -#error We should not hold a spinlock over such a lengthy operation. -#error If necessary, drop spinlock, do operation, re-get spinlock, restart loop. -#error If we need to avoid re-processint items, then delete them from the list -#error as they are replayed and re-add at the tail of this list, so the next -#error item to process will always be at the head of the list. -#endif - if (rc) { - CERROR("recovery replay error %d for req %Ld\n", - rc, req->rq_xid); - GOTO(out, rc); - } - } + /* When deactivating, mark import invalid, and abort in-flight + * requests. */ + if (!active) { + LCONSOLE_WARN("setting import %s INACTIVE by administrator " + "request\n", obd2cli_tgt(imp->imp_obd)); - /* server has seen req, we have reply: skip */ - if ((req->rq_flags & PTL_RPC_FL_REPLIED) && - req->rq_xid <= conn->c_last_xid) { - CDEBUG(D_HA, "REPLIED SKIP: xid "LPD64" transno "LPD64" op %d @ %d\n", - req->rq_xid, req->rq_repmsg->transno, req->rq_reqmsg->opc, - req->rq_import->imp_client->cli_request_portal); - continue; - } + /* set before invalidate to avoid messages about imp_inval + * set without imp_deactive in ptlrpc_import_delay_req */ + spin_lock(&imp->imp_lock); + imp->imp_deactive = 1; + spin_unlock(&imp->imp_lock); - /* server has lost req, we have reply: resend, ign reply */ - if ((req->rq_flags & PTL_RPC_FL_REPLIED) && - req->rq_xid > conn->c_last_xid) { - CDEBUG(D_HA, "REPLIED RESEND: xid "LPD64" transno "LPD64" op %d @ %d\n", - req->rq_xid, req->rq_repmsg->transno, req->rq_reqmsg->opc, - req->rq_import->imp_client->cli_request_portal); - rc = ptlrpc_replay_req(req); - if (rc) { - CERROR("request resend error %d for req %Ld\n", - rc, req->rq_xid); - GOTO(out, rc); - } - } + obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE); - /* server has seen req, we have lost reply: -ERESTARTSYS */ - if ( !(req->rq_flags & PTL_RPC_FL_REPLIED) && - req->rq_xid <= conn->c_last_xid) { - CDEBUG(D_HA, "RESTARTSYS: xid "LPD64" op %d @ %d\n", - req->rq_xid, req->rq_reqmsg->opc, - req->rq_import->imp_client->cli_request_portal); - ptlrpc_restart_req(req); - } + ptlrpc_invalidate_import(imp); + } - /* service has not seen req, no reply: resend */ - if ( !(req->rq_flags & PTL_RPC_FL_REPLIED) && - req->rq_xid > conn->c_last_xid) { - CDEBUG(D_HA, "RESEND: xid "LPD64" transno "LPD64 - " op %d @ %d\n", req->rq_xid, - req->rq_repmsg ? req->rq_repmsg->transno : 0, - req->rq_reqmsg->opc, - req->rq_import->imp_client->cli_request_portal); - ptlrpc_resend_req(req); - } + /* When activating, mark import valid, and attempt recovery */ + if (active) { + CDEBUG(D_HA, "setting import %s VALID\n", + obd2cli_tgt(imp->imp_obd)); + spin_lock(&imp->imp_lock); + imp->imp_deactive = 0; + spin_unlock(&imp->imp_lock); + obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE); + + rc = ptlrpc_recover_import(imp, NULL, 0); } - conn->c_level = LUSTRE_CONN_FULL; - recovd_conn_fixed(conn); + RETURN(rc); +} +EXPORT_SYMBOL(ptlrpc_set_import_active); + +/* Attempt to reconnect an import */ +int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async) +{ + int rc = 0; + ENTRY; + + spin_lock(&imp->imp_lock); + if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive || + atomic_read(&imp->imp_inval_count)) + rc = -EINVAL; + spin_unlock(&imp->imp_lock); + if (rc) + GOTO(out, rc); - CERROR("recovery complete on conn %p(%s), waking delayed reqs\n", - conn, conn->c_remote_uuid); - /* Finally, continue what we delayed since recovery started */ - list_for_each_safe(tmp, pos, &conn->c_delayed_head) { - req = list_entry(tmp, struct ptlrpc_request, rq_list); - ptlrpc_continue_req(req); + /* force import to be disconnected. */ + ptlrpc_set_import_discon(imp, 0); + + if (new_uuid) { + struct obd_uuid uuid; + + /* intruct import to use new uuid */ + obd_str2uuid(&uuid, new_uuid); + rc = import_set_conn_priority(imp, &uuid); + if (rc) + GOTO(out, rc); } + /* Check if reconnect is already in progress */ + spin_lock(&imp->imp_lock); + if (imp->imp_state != LUSTRE_IMP_DISCON) { + imp->imp_force_verify = 1; + rc = -EALREADY; + } + spin_unlock(&imp->imp_lock); + if (rc) + GOTO(out, rc); + + rc = ptlrpc_connect_import(imp); + if (rc) + GOTO(out, rc); + + if (!async) { + struct l_wait_info lwi; + int secs = cfs_time_seconds(obd_timeout); + + CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n", + obd2cli_tgt(imp->imp_obd), secs); + + lwi = LWI_TIMEOUT(secs, NULL, NULL); + rc = l_wait_event(imp->imp_recovery_waitq, + !ptlrpc_import_in_recovery(imp), &lwi); + CDEBUG(D_HA, "%s: recovery finished\n", + obd2cli_tgt(imp->imp_obd)); + } EXIT; - out: - spin_unlock(&conn->c_lock); + +out: return rc; } +EXPORT_SYMBOL(ptlrpc_recover_import); + +int ptlrpc_import_in_recovery(struct obd_import *imp) +{ + int in_recovery = 1; + + spin_lock(&imp->imp_lock); + if (imp->imp_state == LUSTRE_IMP_FULL || + imp->imp_state == LUSTRE_IMP_CLOSED || + imp->imp_state == LUSTRE_IMP_DISCON || + imp->imp_obd->obd_no_recov) + in_recovery = 0; + spin_unlock(&imp->imp_lock); + + return in_recovery; +}