X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fptlrpc%2Frecover.c;h=4d5f28bc96fdcfc47c0744a275d959c5569231a9;hp=6c3a68b4b34a078535e04d7bf0f5bd4bca3f22ac;hb=f28353b3d810cfbec018a263556ceac84ab9413e;hpb=4d22d9a6a7be61064a76f69ff3b8ca5b84b57f2c diff --git a/lustre/ptlrpc/recover.c b/lustre/ptlrpc/recover.c index 6c3a68b..4d5f28b 100644 --- a/lustre/ptlrpc/recover.c +++ b/lustre/ptlrpc/recover.c @@ -1,186 +1,180 @@ -/* -*- 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. * - * Copyright (c) 2002, 2003 Cluster File Systems, Inc. - * Author: Mike Shaver + * 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. * - * This file is part of Lustre, http://www.lustre.org. + * 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). * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * 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 * - * Lustre 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 for more details. + * GPL HEADER END + */ +/* + * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. + * Use is subject to license terms. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * Copyright (c) 2011, 2017, 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 */ #define DEBUG_SUBSYSTEM S_RPC -#ifdef __KERNEL__ -# include -# include -# include -#else -# include -#endif - -#include -#include -#include -#include -#include -#include -#include -#include -#include /* for IOC_LOV_SET_OSC_ACTIVE */ +#include +#include +#include +#include +#include +#include +#include +#include +#include #include "ptlrpc_internal.h" -static int ptlrpc_recover_import_no_retry(struct obd_import *, char *); - -void ptlrpc_run_recovery_over_upcall(struct obd_device *obd) +/** + * Start recovery on disconnected import. + * This is done by just attempting a connect + */ +void ptlrpc_initiate_recovery(struct obd_import *imp) { - char *argv[4]; - char *envp[3]; - int rc; ENTRY; - argv[0] = obd_lustre_upcall; - argv[1] = "RECOVERY_OVER"; - argv[2] = obd->obd_uuid.uuid; - argv[3] = NULL; - - envp[0] = "HOME=/"; - envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; - envp[2] = NULL; - - rc = USERMODEHELPER(argv[0], argv, envp); - if (rc < 0) { - CERROR("Error invoking recovery upcall %s %s %s: %d; check " - "/proc/sys/lustre/upcall\n", - argv[0], argv[1], argv[2], rc); - - } else { - CERROR("Invoked upcall %s %s %s\n", - argv[0], argv[1], argv[2]); - } -} - -void ptlrpc_run_failed_import_upcall(struct obd_import* imp) -{ -#ifdef __KERNEL__ - unsigned long flags; - char *argv[7]; - char *envp[3]; - int rc; - ENTRY; + CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd)); + ptlrpc_connect_import(imp); - spin_lock_irqsave(&imp->imp_lock, flags); - if (imp->imp_state == LUSTRE_IMP_CLOSED) { - spin_unlock_irqrestore(&imp->imp_lock, flags); - EXIT; - return; - } - spin_unlock_irqrestore(&imp->imp_lock, flags); - - argv[0] = obd_lustre_upcall; - argv[1] = "FAILED_IMPORT"; - argv[2] = imp->imp_target_uuid.uuid; - argv[3] = imp->imp_obd->obd_name; - argv[4] = imp->imp_connection->c_remote_uuid.uuid; - argv[5] = imp->imp_obd->obd_uuid.uuid; - argv[6] = NULL; - - envp[0] = "HOME=/"; - envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; - envp[2] = NULL; - - rc = USERMODEHELPER(argv[0], argv, envp); - if (rc < 0) { - CERROR("Error invoking recovery upcall %s %s %s %s %s: %d; " - "check /proc/sys/lustre/lustre_upcall\n", - argv[0], argv[1], argv[2], argv[3], argv[4],rc); - - } else { - CERROR("Invoked upcall %s %s %s %s %s\n", - argv[0], argv[1], argv[2], argv[3], argv[4]); - } -#else - ptlrpc_recover_import(imp, NULL); -#endif + EXIT; } -int ptlrpc_replay_next(struct obd_import *imp) +/** + * 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; - unsigned long flags; + struct list_head *tmp, *pos; + struct ptlrpc_request *req = NULL; __u64 last_transno; - int sent_req = 0; ENTRY; + *inflight = 0; + /* It might have committed some after we last spoke, so make sure we * get rid of them now. */ - spin_lock_irqsave(&imp->imp_lock, flags); - ptlrpc_free_committed(imp); - last_transno = imp->imp_last_replay_transno; - spin_unlock_irqrestore(&imp->imp_lock, flags); - - CDEBUG(D_HA, "import %p from %s has committed "LPD64"\n", - imp, imp->imp_target_uuid.uuid, imp->imp_peer_committed_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... - */ - 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) { - /* remove from list so ptlrpcd can send the - req, it should be reinserted after it is - sent and replied. Perhaps better solution - would be to add req->rq_replay_list so the - req can be saved for replay and still go - through the normal send thread. */ - rc = ptlrpc_replay_req(req); - if (rc) { - CERROR("recovery replay error %d for req "LPD64"\n", - rc, req->rq_xid); - RETURN(rc); - } - sent_req = 1; - break; - } - - } - - RETURN(sent_req); + spin_lock(&imp->imp_lock); + imp->imp_last_transno_checked = 0; + ptlrpc_free_committed(imp); + last_transno = imp->imp_last_replay_transno; + + 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); + + /* 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) { + 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; + LASSERT(!list_empty(imp->imp_replay_cursor)); + 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); + + /* ptlrpc_prepare_replay() may fail to add the reqeust into unreplied + * list if the request hasn't been added to replay list then. Another + * exception is that resend replay could have been removed from the + * unreplied list. */ + if (req != NULL && list_empty(&req->rq_unreplied_list)) { + DEBUG_REQ(D_HA, req, "resend_replay: %d, last_transno: %llu\n", + imp->imp_resend_replay, last_transno); + 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) { + 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 list_head *tmp, *pos; - struct ptlrpc_request *req; - unsigned long flags; + struct ptlrpc_request *req, *next; ENTRY; @@ -190,224 +184,204 @@ int ptlrpc_resend(struct obd_import *imp) */ /* Well... what if lctl recover is called twice at the same time? */ - spin_lock_irqsave(&imp->imp_lock, flags); - if (imp->imp_state != LUSTRE_IMP_RECOVER) { - spin_unlock_irqrestore(&imp->imp_lock, flags); + spin_lock(&imp->imp_lock); + if (imp->imp_state != LUSTRE_IMP_RECOVER) { + spin_unlock(&imp->imp_lock); RETURN(-1); } - spin_unlock_irqrestore(&imp->imp_lock, flags); - list_for_each_safe(tmp, pos, &imp->imp_sending_list) { - req = list_entry(tmp, struct ptlrpc_request, rq_list); - ptlrpc_resend_req(req); - } - - RETURN(0); + 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); + + OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT, 2); + RETURN(0); } +/** + * Go through all requests in delayed list and wake their threads + * for resending + */ void ptlrpc_wake_delayed(struct obd_import *imp) { - unsigned long flags; - struct list_head *tmp, *pos; - struct ptlrpc_request *req; - - spin_lock_irqsave(&imp->imp_lock, flags); - list_for_each_safe(tmp, pos, &imp->imp_delayed_list) { - req = list_entry(tmp, struct ptlrpc_request, rq_list); - - if (req->rq_set) { - DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set); - wake_up(&req->rq_set->set_waitq); - } else { - DEBUG_REQ(D_HA, req, "waking:"); - wake_up(&req->rq_reply_waitq); - } - } - spin_unlock_irqrestore(&imp->imp_lock, flags); -} + struct list_head *tmp, *pos; + struct ptlrpc_request *req; -inline void ptlrpc_invalidate_import_state(struct obd_import *imp) -{ - struct obd_device *obd = imp->imp_obd; - struct ldlm_namespace *ns = obd->obd_namespace; - - ptlrpc_abort_inflight(imp); - -#if 0 - obd_invalidate_import(obd, imp); -#endif - - ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY); -} - -void ptlrpc_handle_failed_import(struct obd_import *imp) -{ - ENTRY; - - if (!imp->imp_replayable) { - CDEBUG(D_HA, - "import %s@%s for %s not replayable, deactivating\n", - imp->imp_target_uuid.uuid, - imp->imp_connection->c_remote_uuid.uuid, - imp->imp_obd->obd_name); - ptlrpc_set_import_active(imp, 0); - } + spin_lock(&imp->imp_lock); + list_for_each_safe(tmp, pos, &imp->imp_delayed_list) { + req = list_entry(tmp, struct ptlrpc_request, rq_list); - ptlrpc_run_failed_import_upcall(imp); - EXIT; + 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) { - int rc; - struct obd_import *imp= failed_req->rq_import; - unsigned long flags; + struct obd_import *imp = failed_req->rq_import; ENTRY; CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n", - imp->imp_obd->obd_name, - imp->imp_target_uuid.uuid, + imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd), imp->imp_connection->c_remote_uuid.uuid); - - ptlrpc_set_import_discon(imp); - rc = ptlrpc_connect_import(imp, NULL); - + 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); + } + /* Wait for recovery to complete and resend. If evicted, then this request will be errored out later.*/ - spin_lock_irqsave(&failed_req->rq_lock, flags); - if (!failed_req->rq_no_resend) - failed_req->rq_resend = 1; - spin_unlock_irqrestore(&failed_req->rq_lock, flags); - - EXIT; + spin_lock(&failed_req->rq_lock); + if (!failed_req->rq_no_resend) + failed_req->rq_resend = 1; + spin_unlock(&failed_req->rq_lock); + + EXIT; } +/** + * 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; - unsigned long flags; + int rc = 0; + ENTRY; LASSERT(obd); /* When deactivating, mark import invalid, and abort in-flight * requests. */ if (!active) { - spin_lock_irqsave(&imp->imp_lock, flags); - /* This is a bit of a hack, but invalidating replayable - * imports makes a temporary reconnect failure into a much more - * ugly -- and hard to remedy -- situation. */ - if (!imp->imp_replayable) { - CDEBUG(D_HA, "setting import %s INVALID\n", - imp->imp_target_uuid.uuid); - imp->imp_invalid = 1; - } - imp->imp_generation++; - spin_unlock_irqrestore(&imp->imp_lock, flags); - ptlrpc_invalidate_import_state(imp); + LCONSOLE_WARN("setting import %s INACTIVE by administrator " + "request\n", obd2cli_tgt(imp->imp_obd)); + + /* 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); + + obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE); + + ptlrpc_invalidate_import(imp); } - /* When activating, mark import valid */ + /* When activating, mark import valid, and attempt recovery */ if (active) { CDEBUG(D_HA, "setting import %s VALID\n", - imp->imp_target_uuid.uuid); - spin_lock_irqsave(&imp->imp_lock, flags); - imp->imp_invalid = 0; - spin_unlock_irqrestore(&imp->imp_lock, flags); - } - - if (obd->obd_observer) - RETURN(obd_notify(obd->obd_observer, obd, active)); + obd2cli_tgt(imp->imp_obd)); - RETURN(0); -} + spin_lock(&imp->imp_lock); + imp->imp_deactive = 0; + spin_unlock(&imp->imp_lock); + obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE); -int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid) -{ - int rc; - ENTRY; - - /* force import to be disconnected. */ - ptlrpc_set_import_discon(imp); - - rc = ptlrpc_recover_import_no_retry(imp, new_uuid); + rc = ptlrpc_recover_import(imp, NULL, 0); + } RETURN(rc); } +EXPORT_SYMBOL(ptlrpc_set_import_active); -int ptlrpc_import_in_recovery(struct obd_import *imp) +/* Attempt to reconnect an import */ +int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async) { - unsigned long flags; - int in_recovery = 1; - spin_lock_irqsave(&imp->imp_lock, flags); - if (imp->imp_state == LUSTRE_IMP_FULL || - imp->imp_state == LUSTRE_IMP_CLOSED || - imp->imp_state == LUSTRE_IMP_DISCON) - in_recovery = 0; - spin_unlock_irqrestore(&imp->imp_lock, flags); - return in_recovery; -} + 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); -static int ptlrpc_recover_import_no_retry(struct obd_import *imp, - char *new_uuid) -{ - int rc; - unsigned long flags; - int in_recovery = 0; - struct l_wait_info lwi; - ENTRY; + /* force import to be disconnected. */ + ptlrpc_set_import_discon(imp, 0); - spin_lock_irqsave(&imp->imp_lock, flags); - if (imp->imp_state != LUSTRE_IMP_DISCON) { - in_recovery = 1; - } - spin_unlock_irqrestore(&imp->imp_lock, flags); + if (new_uuid) { + struct obd_uuid uuid; - if (in_recovery == 1) - RETURN(-EALREADY); + /* intruct import to use new uuid */ + obd_str2uuid(&uuid, new_uuid); + rc = import_set_conn_priority(imp, &uuid); + if (rc) + GOTO(out, rc); + } - - rc = ptlrpc_connect_import(imp, new_uuid); + /* 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) - RETURN(rc); + GOTO(out, rc); - CDEBUG(D_ERROR, "%s: recovery started, waiting\n", - imp->imp_client->cli_name); + OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE); - lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL); - rc = l_wait_event(imp->imp_recovery_waitq, - !ptlrpc_import_in_recovery(imp), &lwi); - CDEBUG(D_ERROR, "%s: recovery finished\n", - imp->imp_client->cli_name); + rc = ptlrpc_connect_import(imp); + if (rc) + GOTO(out, rc); - RETURN(rc); - + if (!async) { + struct l_wait_info lwi; + long secs = cfs_time_seconds(obd_timeout); + + CDEBUG(D_HA, "%s: recovery started, waiting %lu 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: + return rc; } +EXPORT_SYMBOL(ptlrpc_recover_import); -void ptlrpc_fail_export(struct obd_export *exp) +int ptlrpc_import_in_recovery(struct obd_import *imp) { - int rc, already_failed; - unsigned long flags; - - spin_lock_irqsave(&exp->exp_lock, flags); - already_failed = exp->exp_failed; - exp->exp_failed = 1; - spin_unlock_irqrestore(&exp->exp_lock, flags); - - if (already_failed) { - CDEBUG(D_HA, "disconnecting dead export %p/%s; skipping\n", - exp, exp->exp_client_uuid.uuid); - return; - } + int in_recovery = 1; - CDEBUG(D_HA, "disconnecting export %p/%s\n", - exp, exp->exp_client_uuid.uuid); + spin_lock(&imp->imp_lock); + if (imp->imp_state <= LUSTRE_IMP_DISCON || + imp->imp_state >= LUSTRE_IMP_FULL || + imp->imp_obd->obd_no_recov) + in_recovery = 0; + spin_unlock(&imp->imp_lock); - /* Most callers into obd_disconnect are removing their own reference - * (request, for example) in addition to the one from the hash table. - * We don't have such a reference here, so make one. */ - class_export_get(exp); - rc = obd_disconnect(exp, 0); - if (rc) - CERROR("disconnecting export %p failed: %d\n", exp, rc); + return in_recovery; }