lustre/ptlrpc/recover.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/recover.c
  33  *
  34  * Author: Mike Shaver <shaver@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include <linux/list.h>
  39 #include <libcfs/libcfs.h>
  40 #include <obd_support.h>
  41 #include <lustre_ha.h>
  42 #include <lustre_net.h>
  43 #include <lustre_import.h>
  44 #include <lustre_export.h>
  45 #include <obd.h>
  46 #include <obd_class.h>
  47
  48 #include "ptlrpc_internal.h"
  49
  50 /**
  51  * Start recovery on disconnected import.
  52  * This is done by just attempting a connect
  53  */
  54 void ptlrpc_initiate_recovery(struct obd_import *imp)
  55 {
  56         ENTRY;
  57
  58         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  59         ptlrpc_connect_import(imp);
  60
  61         EXIT;
  62 }
  63
  64 /**
  65  * Identify what request from replay list needs to be replayed next
  66  * (based on what we have already replayed) and send it to server.
  67  */
  68 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  69 {
  70         int rc = 0;
  71         struct list_head *tmp, *pos;
  72         struct ptlrpc_request *req = NULL;
  73         __u64 last_transno;
  74         ENTRY;
  75
  76         *inflight = 0;
  77
  78         /* It might have committed some after we last spoke, so make sure we
  79          * get rid of them now.
  80          */
  81         spin_lock(&imp->imp_lock);
  82         imp->imp_last_transno_checked = 0;
  83         ptlrpc_free_committed(imp);
  84         last_transno = imp->imp_last_replay_transno;
  85         spin_unlock(&imp->imp_lock);
  86
  87         CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
  88                imp, obd2cli_tgt(imp->imp_obd),
  89                imp->imp_peer_committed_transno, last_transno);
  90
  91         /* Do I need to hold a lock across this iteration?  We shouldn't be
  92          * racing with any additions to the list, because we're in recovery
  93          * and are therefore not processing additional requests to add.  Calls
  94          * to ptlrpc_free_committed might commit requests, but nothing "newer"
  95          * than the one we're replaying (it can't be committed until it's
  96          * replayed, and we're doing that here).  l_f_e_safe protects against
  97          * problems with the current request being committed, in the unlikely
  98          * event of that race.  So, in conclusion, I think that it's safe to
  99          * perform this list-walk without the imp_lock held.
 100          *
 101          * But, the {mdc,osc}_replay_open callbacks both iterate
 102          * request lists, and have comments saying they assume the
 103          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 104          * just a little race...
 105          */
 106
 107         /* Replay all the committed open requests on committed_list first */
 108         if (!list_empty(&imp->imp_committed_list)) {
 109                 tmp = imp->imp_committed_list.prev;
 110                 req = list_entry(tmp, struct ptlrpc_request,
 111                                      rq_replay_list);
 112
 113                 /* The last request on committed_list hasn't been replayed */
 114                 if (req->rq_transno > last_transno) {
 115                         /* Since the imp_committed_list is immutable before
 116                          * all of it's requests being replayed, it's safe to
 117                          * use a cursor to accelerate the search */
 118                         if (!imp->imp_resend_replay ||
 119                             imp->imp_replay_cursor == &imp->imp_committed_list)
 120                                 imp->imp_replay_cursor =
 121                                         imp->imp_replay_cursor->next;
 122
 123                         while (imp->imp_replay_cursor !=
 124                                &imp->imp_committed_list) {
 125                                 req = list_entry(imp->imp_replay_cursor,
 126                                                      struct ptlrpc_request,
 127                                                      rq_replay_list);
 128                                 if (req->rq_transno > last_transno)
 129                                         break;
 130
 131                                 req = NULL;
 132                                 imp->imp_replay_cursor =
 133                                         imp->imp_replay_cursor->next;
 134                         }
 135                 } else {
 136                         /* All requests on committed_list have been replayed */
 137                         imp->imp_replay_cursor = &imp->imp_committed_list;
 138                         req = NULL;
 139                 }
 140         }
 141
 142         /* All the requests in committed list have been replayed, let's replay
 143          * the imp_replay_list */
 144         if (req == NULL) {
 145                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 146                         req = list_entry(tmp, struct ptlrpc_request,
 147                                              rq_replay_list);
 148
 149                         if (req->rq_transno > last_transno)
 150                                 break;
 151                         req = NULL;
 152                 }
 153         }
 154
 155         /* If need to resend the last sent transno (because a reconnect
 156          * has occurred), then stop on the matching req and send it again.
 157          * If, however, the last sent transno has been committed then we
 158          * continue replay from the next request. */
 159         if (req != NULL && imp->imp_resend_replay)
 160                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 161
 162         spin_lock(&imp->imp_lock);
 163         /* The resend replay request may have been removed from the
 164          * unreplied list. */
 165         if (req != NULL && imp->imp_resend_replay &&
 166             list_empty(&req->rq_unreplied_list)) {
 167                 ptlrpc_add_unreplied(req);
 168                 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 169         }
 170
 171         imp->imp_resend_replay = 0;
 172         spin_unlock(&imp->imp_lock);
 173
 174         if (req != NULL) {
 175                 /* The request should have been added back in unreplied list
 176                  * by ptlrpc_prepare_replay(). */
 177                 LASSERT(!list_empty(&req->rq_unreplied_list));
 178
 179                 rc = ptlrpc_replay_req(req);
 180                 if (rc) {
 181                         CERROR("recovery replay error %d for req "
 182                                "%llu\n", rc, req->rq_xid);
 183                         RETURN(rc);
 184                 }
 185                 *inflight = 1;
 186         }
 187         RETURN(rc);
 188 }
 189
 190 /**
 191  * Schedule resending of request on sending_list. This is done after
 192  * we completed replaying of requests and locks.
 193  */
 194 int ptlrpc_resend(struct obd_import *imp)
 195 {
 196         struct ptlrpc_request *req, *next;
 197
 198         ENTRY;
 199
 200         /* As long as we're in recovery, nothing should be added to the sending
 201          * list, so we don't need to hold the lock during this iteration and
 202          * resend process.
 203          */
 204         /* Well... what if lctl recover is called twice at the same time?
 205          */
 206         spin_lock(&imp->imp_lock);
 207         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 208                 spin_unlock(&imp->imp_lock);
 209                 RETURN(-1);
 210         }
 211
 212         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 213                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 214                          "req %p bad\n", req);
 215                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 216
 217                 /* If the request is allowed to be sent during replay and it
 218                  * is not timeout yet, then it does not need to be resent. */
 219                 if (!ptlrpc_no_resend(req) &&
 220                     (req->rq_timedout || !req->rq_allow_replay))
 221                         ptlrpc_resend_req(req);
 222         }
 223         spin_unlock(&imp->imp_lock);
 224
 225         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT, 2);
 226         RETURN(0);
 227 }
 228
 229 /**
 230  * Go through all requests in delayed list and wake their threads
 231  * for resending
 232  */
 233 void ptlrpc_wake_delayed(struct obd_import *imp)
 234 {
 235         struct list_head *tmp, *pos;
 236         struct ptlrpc_request *req;
 237
 238         spin_lock(&imp->imp_lock);
 239         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 240                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 241
 242                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 243                 ptlrpc_client_wake_req(req);
 244         }
 245         spin_unlock(&imp->imp_lock);
 246 }
 247
 248 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 249 {
 250         struct obd_import *imp = failed_req->rq_import;
 251         ENTRY;
 252
 253         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 254                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 255                imp->imp_connection->c_remote_uuid.uuid);
 256
 257         if (ptlrpc_set_import_discon(imp,
 258                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 259                 if (!imp->imp_replayable) {
 260                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 261                                "auto-deactivating\n",
 262                                obd2cli_tgt(imp->imp_obd),
 263                                imp->imp_connection->c_remote_uuid.uuid,
 264                                imp->imp_obd->obd_name);
 265                         ptlrpc_deactivate_import(imp);
 266                 }
 267                 /* to control recovery via lctl {disable|enable}_recovery */
 268                 if (imp->imp_deactive == 0)
 269                         ptlrpc_connect_import(imp);
 270         }
 271
 272         /* Wait for recovery to complete and resend. If evicted, then
 273            this request will be errored out later.*/
 274         spin_lock(&failed_req->rq_lock);
 275         if (!failed_req->rq_no_resend)
 276                 failed_req->rq_resend = 1;
 277         spin_unlock(&failed_req->rq_lock);
 278
 279         EXIT;
 280 }
 281
 282 /**
 283  * Administratively active/deactive a client.
 284  * This should only be called by the ioctl interface, currently
 285  *  - the lctl deactivate and activate commands
 286  *  - echo 0/1 >> /proc/osc/XXX/active
 287  *  - client umount -f (ll_umount_begin)
 288  */
 289 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 290 {
 291         struct obd_device *obd = imp->imp_obd;
 292         int rc = 0;
 293
 294         ENTRY;
 295         LASSERT(obd);
 296
 297         /* When deactivating, mark import invalid, and abort in-flight
 298          * requests. */
 299         if (!active) {
 300                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 301                               "request\n", obd2cli_tgt(imp->imp_obd));
 302
 303                 /* set before invalidate to avoid messages about imp_inval
 304                  * set without imp_deactive in ptlrpc_import_delay_req */
 305                 spin_lock(&imp->imp_lock);
 306                 imp->imp_deactive = 1;
 307                 spin_unlock(&imp->imp_lock);
 308
 309                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 310
 311                 ptlrpc_invalidate_import(imp);
 312         }
 313
 314         /* When activating, mark import valid, and attempt recovery */
 315         if (active) {
 316                 CDEBUG(D_HA, "setting import %s VALID\n",
 317                        obd2cli_tgt(imp->imp_obd));
 318
 319                 spin_lock(&imp->imp_lock);
 320                 imp->imp_deactive = 0;
 321                 spin_unlock(&imp->imp_lock);
 322                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 323
 324                 rc = ptlrpc_recover_import(imp, NULL, 0);
 325         }
 326
 327         RETURN(rc);
 328 }
 329 EXPORT_SYMBOL(ptlrpc_set_import_active);
 330
 331 /* Attempt to reconnect an import */
 332 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 333 {
 334         int rc = 0;
 335         ENTRY;
 336
 337         spin_lock(&imp->imp_lock);
 338         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 339             atomic_read(&imp->imp_inval_count))
 340                 rc = -EINVAL;
 341         spin_unlock(&imp->imp_lock);
 342         if (rc)
 343                 GOTO(out, rc);
 344
 345         /* force import to be disconnected. */
 346         ptlrpc_set_import_discon(imp, 0);
 347
 348         if (new_uuid) {
 349                 struct obd_uuid uuid;
 350
 351                 /* intruct import to use new uuid */
 352                 obd_str2uuid(&uuid, new_uuid);
 353                 rc = import_set_conn_priority(imp, &uuid);
 354                 if (rc)
 355                         GOTO(out, rc);
 356         }
 357
 358         /* Check if reconnect is already in progress */
 359         spin_lock(&imp->imp_lock);
 360         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 361                 imp->imp_force_verify = 1;
 362                 rc = -EALREADY;
 363         }
 364         spin_unlock(&imp->imp_lock);
 365         if (rc)
 366                 GOTO(out, rc);
 367
 368         rc = ptlrpc_connect_import(imp);
 369         if (rc)
 370                 GOTO(out, rc);
 371
 372         if (!async) {
 373                 struct l_wait_info lwi;
 374                 int secs = cfs_time_seconds(obd_timeout);
 375
 376                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 377                        obd2cli_tgt(imp->imp_obd), secs);
 378
 379                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 380                 rc = l_wait_event(imp->imp_recovery_waitq,
 381                                   !ptlrpc_import_in_recovery(imp), &lwi);
 382                 CDEBUG(D_HA, "%s: recovery finished\n",
 383                        obd2cli_tgt(imp->imp_obd));
 384         }
 385         EXIT;
 386
 387 out:
 388         return rc;
 389 }
 390 EXPORT_SYMBOL(ptlrpc_recover_import);
 391
 392 int ptlrpc_import_in_recovery(struct obd_import *imp)
 393 {
 394         int in_recovery = 1;
 395
 396         spin_lock(&imp->imp_lock);
 397         if (imp->imp_state == LUSTRE_IMP_FULL ||
 398             imp->imp_state == LUSTRE_IMP_CLOSED ||
 399             imp->imp_state == LUSTRE_IMP_DISCON ||
 400             imp->imp_obd->obd_no_recov)
 401                 in_recovery = 0;
 402         spin_unlock(&imp->imp_lock);
 403
 404         return in_recovery;
 405 }