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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/recover.c
  37  *
  38  * Author: Mike Shaver <shaver@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42 #include <libcfs/libcfs.h>
  43 #include <obd_support.h>
  44 #include <lustre_ha.h>
  45 #include <lustre_net.h>
  46 #include <lustre_import.h>
  47 #include <lustre_export.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <libcfs/list.h>
  51
  52 #include "ptlrpc_internal.h"
  53
  54 /**
  55  * Start recovery on disconnected import.
  56  * This is done by just attempting a connect
  57  */
  58 void ptlrpc_initiate_recovery(struct obd_import *imp)
  59 {
  60         ENTRY;
  61
  62         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  63         ptlrpc_connect_import(imp);
  64
  65         EXIT;
  66 }
  67
  68 /**
  69  * Identify what request from replay list needs to be replayed next
  70  * (based on what we have already replayed) and send it to server.
  71  */
  72 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  73 {
  74         int rc = 0;
  75         struct list_head *tmp, *pos;
  76         struct ptlrpc_request *req = NULL;
  77         __u64 last_transno;
  78         ENTRY;
  79
  80         *inflight = 0;
  81
  82         /* It might have committed some after we last spoke, so make sure we
  83          * get rid of them now.
  84          */
  85         spin_lock(&imp->imp_lock);
  86         imp->imp_last_transno_checked = 0;
  87         ptlrpc_free_committed(imp);
  88         last_transno = imp->imp_last_replay_transno;
  89         spin_unlock(&imp->imp_lock);
  90
  91         CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
  92                imp, obd2cli_tgt(imp->imp_obd),
  93                imp->imp_peer_committed_transno, last_transno);
  94
  95         /* Do I need to hold a lock across this iteration?  We shouldn't be
  96          * racing with any additions to the list, because we're in recovery
  97          * and are therefore not processing additional requests to add.  Calls
  98          * to ptlrpc_free_committed might commit requests, but nothing "newer"
  99          * than the one we're replaying (it can't be committed until it's
 100          * replayed, and we're doing that here).  l_f_e_safe protects against
 101          * problems with the current request being committed, in the unlikely
 102          * event of that race.  So, in conclusion, I think that it's safe to
 103          * perform this list-walk without the imp_lock held.
 104          *
 105          * But, the {mdc,osc}_replay_open callbacks both iterate
 106          * request lists, and have comments saying they assume the
 107          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 108          * just a little race...
 109          */
 110
 111         /* Replay all the committed open requests on committed_list first */
 112         if (!list_empty(&imp->imp_committed_list)) {
 113                 tmp = imp->imp_committed_list.prev;
 114                 req = list_entry(tmp, struct ptlrpc_request,
 115                                      rq_replay_list);
 116
 117                 /* The last request on committed_list hasn't been replayed */
 118                 if (req->rq_transno > last_transno) {
 119                         /* Since the imp_committed_list is immutable before
 120                          * all of it's requests being replayed, it's safe to
 121                          * use a cursor to accelerate the search */
 122                         imp->imp_replay_cursor = imp->imp_replay_cursor->next;
 123
 124                         while (imp->imp_replay_cursor !=
 125                                &imp->imp_committed_list) {
 126                                 req = list_entry(imp->imp_replay_cursor,
 127                                                      struct ptlrpc_request,
 128                                                      rq_replay_list);
 129                                 if (req->rq_transno > last_transno)
 130                                         break;
 131
 132                                 req = NULL;
 133                                 imp->imp_replay_cursor =
 134                                         imp->imp_replay_cursor->next;
 135                         }
 136                 } else {
 137                         /* All requests on committed_list have been replayed */
 138                         imp->imp_replay_cursor = &imp->imp_committed_list;
 139                         req = NULL;
 140                 }
 141         }
 142
 143         /* All the requests in committed list have been replayed, let's replay
 144          * the imp_replay_list */
 145         if (req == NULL) {
 146                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 147                         req = list_entry(tmp, struct ptlrpc_request,
 148                                              rq_replay_list);
 149
 150                         if (req->rq_transno > last_transno)
 151                                 break;
 152                         req = NULL;
 153                 }
 154         }
 155
 156         /* If need to resend the last sent transno (because a reconnect
 157          * has occurred), then stop on the matching req and send it again.
 158          * If, however, the last sent transno has been committed then we
 159          * continue replay from the next request. */
 160         if (req != NULL && imp->imp_resend_replay)
 161                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 162
 163         spin_lock(&imp->imp_lock);
 164         imp->imp_resend_replay = 0;
 165         spin_unlock(&imp->imp_lock);
 166
 167         if (req != NULL) {
 168                 rc = ptlrpc_replay_req(req);
 169                 if (rc) {
 170                         CERROR("recovery replay error %d for req "
 171                                LPU64"\n", rc, req->rq_xid);
 172                         RETURN(rc);
 173                 }
 174                 *inflight = 1;
 175         }
 176         RETURN(rc);
 177 }
 178
 179 /**
 180  * Schedule resending of request on sending_list. This is done after
 181  * we completed replaying of requests and locks.
 182  */
 183 int ptlrpc_resend(struct obd_import *imp)
 184 {
 185         struct ptlrpc_request *req, *next;
 186
 187         ENTRY;
 188
 189         /* As long as we're in recovery, nothing should be added to the sending
 190          * list, so we don't need to hold the lock during this iteration and
 191          * resend process.
 192          */
 193         /* Well... what if lctl recover is called twice at the same time?
 194          */
 195         spin_lock(&imp->imp_lock);
 196         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 197                 spin_unlock(&imp->imp_lock);
 198                 RETURN(-1);
 199         }
 200
 201         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 202                 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
 203                          "req %p bad\n", req);
 204                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 205
 206                 /* If the request is allowed to be sent during replay and it
 207                  * is not timeout yet, then it does not need to be resent. */
 208                 if (!ptlrpc_no_resend(req) &&
 209                     (req->rq_timedout || !req->rq_allow_replay))
 210                         ptlrpc_resend_req(req);
 211         }
 212         spin_unlock(&imp->imp_lock);
 213
 214         RETURN(0);
 215 }
 216
 217 /**
 218  * Go through all requests in delayed list and wake their threads
 219  * for resending
 220  */
 221 void ptlrpc_wake_delayed(struct obd_import *imp)
 222 {
 223         struct list_head *tmp, *pos;
 224         struct ptlrpc_request *req;
 225
 226         spin_lock(&imp->imp_lock);
 227         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 228                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 229
 230                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 231                 ptlrpc_client_wake_req(req);
 232         }
 233         spin_unlock(&imp->imp_lock);
 234 }
 235
 236 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 237 {
 238         struct obd_import *imp = failed_req->rq_import;
 239         ENTRY;
 240
 241         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 242                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 243                imp->imp_connection->c_remote_uuid.uuid);
 244
 245         if (ptlrpc_set_import_discon(imp,
 246                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 247                 if (!imp->imp_replayable) {
 248                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 249                                "auto-deactivating\n",
 250                                obd2cli_tgt(imp->imp_obd),
 251                                imp->imp_connection->c_remote_uuid.uuid,
 252                                imp->imp_obd->obd_name);
 253                         ptlrpc_deactivate_import(imp);
 254                 }
 255                 /* to control recovery via lctl {disable|enable}_recovery */
 256                 if (imp->imp_deactive == 0)
 257                         ptlrpc_connect_import(imp);
 258         }
 259
 260         /* Wait for recovery to complete and resend. If evicted, then
 261            this request will be errored out later.*/
 262         spin_lock(&failed_req->rq_lock);
 263         if (!failed_req->rq_no_resend)
 264                 failed_req->rq_resend = 1;
 265         spin_unlock(&failed_req->rq_lock);
 266
 267         EXIT;
 268 }
 269
 270 /**
 271  * Administratively active/deactive a client.
 272  * This should only be called by the ioctl interface, currently
 273  *  - the lctl deactivate and activate commands
 274  *  - echo 0/1 >> /proc/osc/XXX/active
 275  *  - client umount -f (ll_umount_begin)
 276  */
 277 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 278 {
 279         struct obd_device *obd = imp->imp_obd;
 280         int rc = 0;
 281
 282         ENTRY;
 283         LASSERT(obd);
 284
 285         /* When deactivating, mark import invalid, and abort in-flight
 286          * requests. */
 287         if (!active) {
 288                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 289                               "request\n", obd2cli_tgt(imp->imp_obd));
 290
 291                 /* set before invalidate to avoid messages about imp_inval
 292                  * set without imp_deactive in ptlrpc_import_delay_req */
 293                 spin_lock(&imp->imp_lock);
 294                 imp->imp_deactive = 1;
 295                 spin_unlock(&imp->imp_lock);
 296
 297                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 298
 299                 ptlrpc_invalidate_import(imp);
 300         }
 301
 302         /* When activating, mark import valid, and attempt recovery */
 303         if (active) {
 304                 CDEBUG(D_HA, "setting import %s VALID\n",
 305                        obd2cli_tgt(imp->imp_obd));
 306
 307                 spin_lock(&imp->imp_lock);
 308                 imp->imp_deactive = 0;
 309                 spin_unlock(&imp->imp_lock);
 310                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 311
 312                 rc = ptlrpc_recover_import(imp, NULL, 0);
 313         }
 314
 315         RETURN(rc);
 316 }
 317 EXPORT_SYMBOL(ptlrpc_set_import_active);
 318
 319 /* Attempt to reconnect an import */
 320 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 321 {
 322         int rc = 0;
 323         ENTRY;
 324
 325         spin_lock(&imp->imp_lock);
 326         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 327             atomic_read(&imp->imp_inval_count))
 328                 rc = -EINVAL;
 329         spin_unlock(&imp->imp_lock);
 330         if (rc)
 331                 GOTO(out, rc);
 332
 333         /* force import to be disconnected. */
 334         ptlrpc_set_import_discon(imp, 0);
 335
 336         if (new_uuid) {
 337                 struct obd_uuid uuid;
 338
 339                 /* intruct import to use new uuid */
 340                 obd_str2uuid(&uuid, new_uuid);
 341                 rc = import_set_conn_priority(imp, &uuid);
 342                 if (rc)
 343                         GOTO(out, rc);
 344         }
 345
 346         /* Check if reconnect is already in progress */
 347         spin_lock(&imp->imp_lock);
 348         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 349                 imp->imp_force_verify = 1;
 350                 rc = -EALREADY;
 351         }
 352         spin_unlock(&imp->imp_lock);
 353         if (rc)
 354                 GOTO(out, rc);
 355
 356         rc = ptlrpc_connect_import(imp);
 357         if (rc)
 358                 GOTO(out, rc);
 359
 360         if (!async) {
 361                 struct l_wait_info lwi;
 362                 int secs = cfs_time_seconds(obd_timeout);
 363
 364                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 365                        obd2cli_tgt(imp->imp_obd), secs);
 366
 367                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 368                 rc = l_wait_event(imp->imp_recovery_waitq,
 369                                   !ptlrpc_import_in_recovery(imp), &lwi);
 370                 CDEBUG(D_HA, "%s: recovery finished\n",
 371                        obd2cli_tgt(imp->imp_obd));
 372         }
 373         EXIT;
 374
 375 out:
 376         return rc;
 377 }
 378 EXPORT_SYMBOL(ptlrpc_recover_import);
 379
 380 int ptlrpc_import_in_recovery(struct obd_import *imp)
 381 {
 382         int in_recovery = 1;
 383
 384         spin_lock(&imp->imp_lock);
 385         if (imp->imp_state == LUSTRE_IMP_FULL ||
 386             imp->imp_state == LUSTRE_IMP_CLOSED ||
 387             imp->imp_state == LUSTRE_IMP_DISCON ||
 388             imp->imp_obd->obd_no_recov)
 389                 in_recovery = 0;
 390         spin_unlock(&imp->imp_lock);
 391
 392         return in_recovery;
 393 }