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                         if (!imp->imp_resend_replay ||
 123                             imp->imp_replay_cursor == &imp->imp_committed_list)
 124                                 imp->imp_replay_cursor =
 125                                         imp->imp_replay_cursor->next;
 126
 127                         while (imp->imp_replay_cursor !=
 128                                &imp->imp_committed_list) {
 129                                 req = list_entry(imp->imp_replay_cursor,
 130                                                      struct ptlrpc_request,
 131                                                      rq_replay_list);
 132                                 if (req->rq_transno > last_transno)
 133                                         break;
 134
 135                                 req = NULL;
 136                                 imp->imp_replay_cursor =
 137                                         imp->imp_replay_cursor->next;
 138                         }
 139                 } else {
 140                         /* All requests on committed_list have been replayed */
 141                         imp->imp_replay_cursor = &imp->imp_committed_list;
 142                         req = NULL;
 143                 }
 144         }
 145
 146         /* All the requests in committed list have been replayed, let's replay
 147          * the imp_replay_list */
 148         if (req == NULL) {
 149                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 150                         req = list_entry(tmp, struct ptlrpc_request,
 151                                              rq_replay_list);
 152
 153                         if (req->rq_transno > last_transno)
 154                                 break;
 155                         req = NULL;
 156                 }
 157         }
 158
 159         /* If need to resend the last sent transno (because a reconnect
 160          * has occurred), then stop on the matching req and send it again.
 161          * If, however, the last sent transno has been committed then we
 162          * continue replay from the next request. */
 163         if (req != NULL && imp->imp_resend_replay)
 164                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 165
 166         spin_lock(&imp->imp_lock);
 167         imp->imp_resend_replay = 0;
 168         spin_unlock(&imp->imp_lock);
 169
 170         if (req != NULL) {
 171                 rc = ptlrpc_replay_req(req);
 172                 if (rc) {
 173                         CERROR("recovery replay error %d for req "
 174                                LPU64"\n", rc, req->rq_xid);
 175                         RETURN(rc);
 176                 }
 177                 *inflight = 1;
 178         }
 179         RETURN(rc);
 180 }
 181
 182 /**
 183  * Schedule resending of request on sending_list. This is done after
 184  * we completed replaying of requests and locks.
 185  */
 186 int ptlrpc_resend(struct obd_import *imp)
 187 {
 188         struct ptlrpc_request *req, *next;
 189
 190         ENTRY;
 191
 192         /* As long as we're in recovery, nothing should be added to the sending
 193          * list, so we don't need to hold the lock during this iteration and
 194          * resend process.
 195          */
 196         /* Well... what if lctl recover is called twice at the same time?
 197          */
 198         spin_lock(&imp->imp_lock);
 199         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 200                 spin_unlock(&imp->imp_lock);
 201                 RETURN(-1);
 202         }
 203
 204         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 205                 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
 206                          "req %p bad\n", req);
 207                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 208
 209                 /* If the request is allowed to be sent during replay and it
 210                  * is not timeout yet, then it does not need to be resent. */
 211                 if (!ptlrpc_no_resend(req) &&
 212                     (req->rq_timedout || !req->rq_allow_replay))
 213                         ptlrpc_resend_req(req);
 214         }
 215         spin_unlock(&imp->imp_lock);
 216
 217         RETURN(0);
 218 }
 219
 220 /**
 221  * Go through all requests in delayed list and wake their threads
 222  * for resending
 223  */
 224 void ptlrpc_wake_delayed(struct obd_import *imp)
 225 {
 226         struct list_head *tmp, *pos;
 227         struct ptlrpc_request *req;
 228
 229         spin_lock(&imp->imp_lock);
 230         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 231                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 232
 233                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 234                 ptlrpc_client_wake_req(req);
 235         }
 236         spin_unlock(&imp->imp_lock);
 237 }
 238
 239 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 240 {
 241         struct obd_import *imp = failed_req->rq_import;
 242         ENTRY;
 243
 244         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 245                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 246                imp->imp_connection->c_remote_uuid.uuid);
 247
 248         if (ptlrpc_set_import_discon(imp,
 249                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 250                 if (!imp->imp_replayable) {
 251                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 252                                "auto-deactivating\n",
 253                                obd2cli_tgt(imp->imp_obd),
 254                                imp->imp_connection->c_remote_uuid.uuid,
 255                                imp->imp_obd->obd_name);
 256                         ptlrpc_deactivate_import(imp);
 257                 }
 258                 /* to control recovery via lctl {disable|enable}_recovery */
 259                 if (imp->imp_deactive == 0)
 260                         ptlrpc_connect_import(imp);
 261         }
 262
 263         /* Wait for recovery to complete and resend. If evicted, then
 264            this request will be errored out later.*/
 265         spin_lock(&failed_req->rq_lock);
 266         if (!failed_req->rq_no_resend)
 267                 failed_req->rq_resend = 1;
 268         spin_unlock(&failed_req->rq_lock);
 269
 270         EXIT;
 271 }
 272
 273 /**
 274  * Administratively active/deactive a client.
 275  * This should only be called by the ioctl interface, currently
 276  *  - the lctl deactivate and activate commands
 277  *  - echo 0/1 >> /proc/osc/XXX/active
 278  *  - client umount -f (ll_umount_begin)
 279  */
 280 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 281 {
 282         struct obd_device *obd = imp->imp_obd;
 283         int rc = 0;
 284
 285         ENTRY;
 286         LASSERT(obd);
 287
 288         /* When deactivating, mark import invalid, and abort in-flight
 289          * requests. */
 290         if (!active) {
 291                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 292                               "request\n", obd2cli_tgt(imp->imp_obd));
 293
 294                 /* set before invalidate to avoid messages about imp_inval
 295                  * set without imp_deactive in ptlrpc_import_delay_req */
 296                 spin_lock(&imp->imp_lock);
 297                 imp->imp_deactive = 1;
 298                 spin_unlock(&imp->imp_lock);
 299
 300                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 301
 302                 ptlrpc_invalidate_import(imp);
 303         }
 304
 305         /* When activating, mark import valid, and attempt recovery */
 306         if (active) {
 307                 CDEBUG(D_HA, "setting import %s VALID\n",
 308                        obd2cli_tgt(imp->imp_obd));
 309
 310                 spin_lock(&imp->imp_lock);
 311                 imp->imp_deactive = 0;
 312                 spin_unlock(&imp->imp_lock);
 313                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 314
 315                 rc = ptlrpc_recover_import(imp, NULL, 0);
 316         }
 317
 318         RETURN(rc);
 319 }
 320 EXPORT_SYMBOL(ptlrpc_set_import_active);
 321
 322 /* Attempt to reconnect an import */
 323 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 324 {
 325         int rc = 0;
 326         ENTRY;
 327
 328         spin_lock(&imp->imp_lock);
 329         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 330             atomic_read(&imp->imp_inval_count))
 331                 rc = -EINVAL;
 332         spin_unlock(&imp->imp_lock);
 333         if (rc)
 334                 GOTO(out, rc);
 335
 336         /* force import to be disconnected. */
 337         ptlrpc_set_import_discon(imp, 0);
 338
 339         if (new_uuid) {
 340                 struct obd_uuid uuid;
 341
 342                 /* intruct import to use new uuid */
 343                 obd_str2uuid(&uuid, new_uuid);
 344                 rc = import_set_conn_priority(imp, &uuid);
 345                 if (rc)
 346                         GOTO(out, rc);
 347         }
 348
 349         /* Check if reconnect is already in progress */
 350         spin_lock(&imp->imp_lock);
 351         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 352                 imp->imp_force_verify = 1;
 353                 rc = -EALREADY;
 354         }
 355         spin_unlock(&imp->imp_lock);
 356         if (rc)
 357                 GOTO(out, rc);
 358
 359         rc = ptlrpc_connect_import(imp);
 360         if (rc)
 361                 GOTO(out, rc);
 362
 363         if (!async) {
 364                 struct l_wait_info lwi;
 365                 int secs = cfs_time_seconds(obd_timeout);
 366
 367                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 368                        obd2cli_tgt(imp->imp_obd), secs);
 369
 370                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 371                 rc = l_wait_event(imp->imp_recovery_waitq,
 372                                   !ptlrpc_import_in_recovery(imp), &lwi);
 373                 CDEBUG(D_HA, "%s: recovery finished\n",
 374                        obd2cli_tgt(imp->imp_obd));
 375         }
 376         EXIT;
 377
 378 out:
 379         return rc;
 380 }
 381 EXPORT_SYMBOL(ptlrpc_recover_import);
 382
 383 int ptlrpc_import_in_recovery(struct obd_import *imp)
 384 {
 385         int in_recovery = 1;
 386
 387         spin_lock(&imp->imp_lock);
 388         if (imp->imp_state == LUSTRE_IMP_FULL ||
 389             imp->imp_state == LUSTRE_IMP_CLOSED ||
 390             imp->imp_state == LUSTRE_IMP_DISCON ||
 391             imp->imp_obd->obd_no_recov)
 392                 in_recovery = 0;
 393         spin_unlock(&imp->imp_lock);
 394
 395         return in_recovery;
 396 }