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
 169         imp->imp_resend_replay = 0;
 170         spin_unlock(&imp->imp_lock);
 171
 172         if (req != NULL) {
 173                 /* The request should have been added back in unreplied list
 174                  * by ptlrpc_prepare_replay(). */
 175                 LASSERT(!list_empty(&req->rq_unreplied_list));
 176
 177                 rc = ptlrpc_replay_req(req);
 178                 if (rc) {
 179                         CERROR("recovery replay error %d for req "
 180                                "%llu\n", rc, req->rq_xid);
 181                         RETURN(rc);
 182                 }
 183                 *inflight = 1;
 184         }
 185         RETURN(rc);
 186 }
 187
 188 /**
 189  * Schedule resending of request on sending_list. This is done after
 190  * we completed replaying of requests and locks.
 191  */
 192 int ptlrpc_resend(struct obd_import *imp)
 193 {
 194         struct ptlrpc_request *req, *next;
 195
 196         ENTRY;
 197
 198         /* As long as we're in recovery, nothing should be added to the sending
 199          * list, so we don't need to hold the lock during this iteration and
 200          * resend process.
 201          */
 202         /* Well... what if lctl recover is called twice at the same time?
 203          */
 204         spin_lock(&imp->imp_lock);
 205         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 206                 spin_unlock(&imp->imp_lock);
 207                 RETURN(-1);
 208         }
 209
 210         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 211                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 212                          "req %p bad\n", req);
 213                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 214
 215                 /* If the request is allowed to be sent during replay and it
 216                  * is not timeout yet, then it does not need to be resent. */
 217                 if (!ptlrpc_no_resend(req) &&
 218                     (req->rq_timedout || !req->rq_allow_replay))
 219                         ptlrpc_resend_req(req);
 220         }
 221         spin_unlock(&imp->imp_lock);
 222
 223         RETURN(0);
 224 }
 225
 226 /**
 227  * Go through all requests in delayed list and wake their threads
 228  * for resending
 229  */
 230 void ptlrpc_wake_delayed(struct obd_import *imp)
 231 {
 232         struct list_head *tmp, *pos;
 233         struct ptlrpc_request *req;
 234
 235         spin_lock(&imp->imp_lock);
 236         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 237                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 238
 239                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 240                 ptlrpc_client_wake_req(req);
 241         }
 242         spin_unlock(&imp->imp_lock);
 243 }
 244
 245 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 246 {
 247         struct obd_import *imp = failed_req->rq_import;
 248         ENTRY;
 249
 250         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 251                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 252                imp->imp_connection->c_remote_uuid.uuid);
 253
 254         if (ptlrpc_set_import_discon(imp,
 255                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 256                 if (!imp->imp_replayable) {
 257                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 258                                "auto-deactivating\n",
 259                                obd2cli_tgt(imp->imp_obd),
 260                                imp->imp_connection->c_remote_uuid.uuid,
 261                                imp->imp_obd->obd_name);
 262                         ptlrpc_deactivate_import(imp);
 263                 }
 264                 /* to control recovery via lctl {disable|enable}_recovery */
 265                 if (imp->imp_deactive == 0)
 266                         ptlrpc_connect_import(imp);
 267         }
 268
 269         /* Wait for recovery to complete and resend. If evicted, then
 270            this request will be errored out later.*/
 271         spin_lock(&failed_req->rq_lock);
 272         if (!failed_req->rq_no_resend)
 273                 failed_req->rq_resend = 1;
 274         spin_unlock(&failed_req->rq_lock);
 275
 276         EXIT;
 277 }
 278
 279 /**
 280  * Administratively active/deactive a client.
 281  * This should only be called by the ioctl interface, currently
 282  *  - the lctl deactivate and activate commands
 283  *  - echo 0/1 >> /proc/osc/XXX/active
 284  *  - client umount -f (ll_umount_begin)
 285  */
 286 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 287 {
 288         struct obd_device *obd = imp->imp_obd;
 289         int rc = 0;
 290
 291         ENTRY;
 292         LASSERT(obd);
 293
 294         /* When deactivating, mark import invalid, and abort in-flight
 295          * requests. */
 296         if (!active) {
 297                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 298                               "request\n", obd2cli_tgt(imp->imp_obd));
 299
 300                 /* set before invalidate to avoid messages about imp_inval
 301                  * set without imp_deactive in ptlrpc_import_delay_req */
 302                 spin_lock(&imp->imp_lock);
 303                 imp->imp_deactive = 1;
 304                 spin_unlock(&imp->imp_lock);
 305
 306                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 307
 308                 ptlrpc_invalidate_import(imp);
 309         }
 310
 311         /* When activating, mark import valid, and attempt recovery */
 312         if (active) {
 313                 CDEBUG(D_HA, "setting import %s VALID\n",
 314                        obd2cli_tgt(imp->imp_obd));
 315
 316                 spin_lock(&imp->imp_lock);
 317                 imp->imp_deactive = 0;
 318                 spin_unlock(&imp->imp_lock);
 319                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 320
 321                 rc = ptlrpc_recover_import(imp, NULL, 0);
 322         }
 323
 324         RETURN(rc);
 325 }
 326 EXPORT_SYMBOL(ptlrpc_set_import_active);
 327
 328 /* Attempt to reconnect an import */
 329 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 330 {
 331         int rc = 0;
 332         ENTRY;
 333
 334         spin_lock(&imp->imp_lock);
 335         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 336             atomic_read(&imp->imp_inval_count))
 337                 rc = -EINVAL;
 338         spin_unlock(&imp->imp_lock);
 339         if (rc)
 340                 GOTO(out, rc);
 341
 342         /* force import to be disconnected. */
 343         ptlrpc_set_import_discon(imp, 0);
 344
 345         if (new_uuid) {
 346                 struct obd_uuid uuid;
 347
 348                 /* intruct import to use new uuid */
 349                 obd_str2uuid(&uuid, new_uuid);
 350                 rc = import_set_conn_priority(imp, &uuid);
 351                 if (rc)
 352                         GOTO(out, rc);
 353         }
 354
 355         /* Check if reconnect is already in progress */
 356         spin_lock(&imp->imp_lock);
 357         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 358                 imp->imp_force_verify = 1;
 359                 rc = -EALREADY;
 360         }
 361         spin_unlock(&imp->imp_lock);
 362         if (rc)
 363                 GOTO(out, rc);
 364
 365         rc = ptlrpc_connect_import(imp);
 366         if (rc)
 367                 GOTO(out, rc);
 368
 369         if (!async) {
 370                 struct l_wait_info lwi;
 371                 int secs = cfs_time_seconds(obd_timeout);
 372
 373                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 374                        obd2cli_tgt(imp->imp_obd), secs);
 375
 376                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 377                 rc = l_wait_event(imp->imp_recovery_waitq,
 378                                   !ptlrpc_import_in_recovery(imp), &lwi);
 379                 CDEBUG(D_HA, "%s: recovery finished\n",
 380                        obd2cli_tgt(imp->imp_obd));
 381         }
 382         EXIT;
 383
 384 out:
 385         return rc;
 386 }
 387 EXPORT_SYMBOL(ptlrpc_recover_import);
 388
 389 int ptlrpc_import_in_recovery(struct obd_import *imp)
 390 {
 391         int in_recovery = 1;
 392
 393         spin_lock(&imp->imp_lock);
 394         if (imp->imp_state == LUSTRE_IMP_FULL ||
 395             imp->imp_state == LUSTRE_IMP_CLOSED ||
 396             imp->imp_state == LUSTRE_IMP_DISCON ||
 397             imp->imp_obd->obd_no_recov)
 398                 in_recovery = 0;
 399         spin_unlock(&imp->imp_lock);
 400
 401         return in_recovery;
 402 }