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, 2012, 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 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47
  48 #include <obd_support.h>
  49 #include <lustre_ha.h>
  50 #include <lustre_net.h>
  51 #include <lustre_import.h>
  52 #include <lustre_export.h>
  53 #include <obd.h>
  54 #include <obd_ost.h>
  55 #include <obd_class.h>
  56 #include <libcfs/list.h>
  57
  58 #include "ptlrpc_internal.h"
  59
  60 /**
  61  * Start recovery on disconnected import.
  62  * This is done by just attempting a connect
  63  */
  64 void ptlrpc_initiate_recovery(struct obd_import *imp)
  65 {
  66         ENTRY;
  67
  68         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  69         ptlrpc_connect_import(imp);
  70
  71         EXIT;
  72 }
  73
  74 /**
  75  * Identify what request from replay list needs to be replayed next
  76  * (based on what we have already replayed) and send it to server.
  77  */
  78 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  79 {
  80         int rc = 0;
  81         cfs_list_t *tmp, *pos;
  82         struct ptlrpc_request *req = NULL;
  83         __u64 last_transno;
  84         ENTRY;
  85
  86         *inflight = 0;
  87
  88         /* It might have committed some after we last spoke, so make sure we
  89          * get rid of them now.
  90          */
  91         spin_lock(&imp->imp_lock);
  92         imp->imp_last_transno_checked = 0;
  93         ptlrpc_free_committed(imp);
  94         last_transno = imp->imp_last_replay_transno;
  95         spin_unlock(&imp->imp_lock);
  96
  97         CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
  98                imp, obd2cli_tgt(imp->imp_obd),
  99                imp->imp_peer_committed_transno, last_transno);
 100
 101         /* Do I need to hold a lock across this iteration?  We shouldn't be
 102          * racing with any additions to the list, because we're in recovery
 103          * and are therefore not processing additional requests to add.  Calls
 104          * to ptlrpc_free_committed might commit requests, but nothing "newer"
 105          * than the one we're replaying (it can't be committed until it's
 106          * replayed, and we're doing that here).  l_f_e_safe protects against
 107          * problems with the current request being committed, in the unlikely
 108          * event of that race.  So, in conclusion, I think that it's safe to
 109          * perform this list-walk without the imp_lock held.
 110          *
 111          * But, the {mdc,osc}_replay_open callbacks both iterate
 112          * request lists, and have comments saying they assume the
 113          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 114          * just a little race...
 115          */
 116
 117         /* Replay all the committed open requests on committed_list first */
 118         if (!cfs_list_empty(&imp->imp_committed_list)) {
 119                 tmp = imp->imp_committed_list.prev;
 120                 req = cfs_list_entry(tmp, struct ptlrpc_request,
 121                                      rq_replay_list);
 122
 123                 /* The last request on committed_list hasn't been replayed */
 124                 if (req->rq_transno > last_transno) {
 125                         /* Since the imp_committed_list is immutable before
 126                          * all of it's requests being replayed, it's safe to
 127                          * use a cursor to accelerate the search */
 128                         imp->imp_replay_cursor = imp->imp_replay_cursor->next;
 129
 130                         while (imp->imp_replay_cursor !=
 131                                &imp->imp_committed_list) {
 132                                 req = cfs_list_entry(imp->imp_replay_cursor,
 133                                                      struct ptlrpc_request,
 134                                                      rq_replay_list);
 135                                 if (req->rq_transno > last_transno)
 136                                         break;
 137
 138                                 req = NULL;
 139                                 imp->imp_replay_cursor =
 140                                         imp->imp_replay_cursor->next;
 141                         }
 142                 } else {
 143                         /* All requests on committed_list have been replayed */
 144                         imp->imp_replay_cursor = &imp->imp_committed_list;
 145                         req = NULL;
 146                 }
 147         }
 148
 149         /* All the requests in committed list have been replayed, let's replay
 150          * the imp_replay_list */
 151         if (req == NULL) {
 152                 cfs_list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 153                         req = cfs_list_entry(tmp, struct ptlrpc_request,
 154                                              rq_replay_list);
 155
 156                         if (req->rq_transno > last_transno)
 157                                 break;
 158                         req = NULL;
 159                 }
 160         }
 161
 162         /* If need to resend the last sent transno (because a reconnect
 163          * has occurred), then stop on the matching req and send it again.
 164          * If, however, the last sent transno has been committed then we
 165          * continue replay from the next request. */
 166         if (req != NULL && imp->imp_resend_replay)
 167                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 168
 169         spin_lock(&imp->imp_lock);
 170         imp->imp_resend_replay = 0;
 171         spin_unlock(&imp->imp_lock);
 172
 173         if (req != NULL) {
 174                 rc = ptlrpc_replay_req(req);
 175                 if (rc) {
 176                         CERROR("recovery replay error %d for req "
 177                                LPU64"\n", rc, req->rq_xid);
 178                         RETURN(rc);
 179                 }
 180                 *inflight = 1;
 181         }
 182         RETURN(rc);
 183 }
 184
 185 /**
 186  * Schedule resending of request on sending_list. This is done after
 187  * we completed replaying of requests and locks.
 188  */
 189 int ptlrpc_resend(struct obd_import *imp)
 190 {
 191         struct ptlrpc_request *req, *next;
 192
 193         ENTRY;
 194
 195         /* As long as we're in recovery, nothing should be added to the sending
 196          * list, so we don't need to hold the lock during this iteration and
 197          * resend process.
 198          */
 199         /* Well... what if lctl recover is called twice at the same time?
 200          */
 201         spin_lock(&imp->imp_lock);
 202         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 203                 spin_unlock(&imp->imp_lock);
 204                 RETURN(-1);
 205         }
 206
 207         cfs_list_for_each_entry_safe(req, next, &imp->imp_sending_list,
 208                                      rq_list) {
 209                 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
 210                          "req %p bad\n", req);
 211                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 212                 if (!ptlrpc_no_resend(req))
 213                         ptlrpc_resend_req(req);
 214         }
 215         spin_unlock(&imp->imp_lock);
 216
 217         RETURN(0);
 218 }
 219 EXPORT_SYMBOL(ptlrpc_resend);
 220
 221 /**
 222  * Go through all requests in delayed list and wake their threads
 223  * for resending
 224  */
 225 void ptlrpc_wake_delayed(struct obd_import *imp)
 226 {
 227         cfs_list_t *tmp, *pos;
 228         struct ptlrpc_request *req;
 229
 230         spin_lock(&imp->imp_lock);
 231         cfs_list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 232                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 233
 234                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 235                 ptlrpc_client_wake_req(req);
 236         }
 237         spin_unlock(&imp->imp_lock);
 238 }
 239 EXPORT_SYMBOL(ptlrpc_wake_delayed);
 240
 241 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 242 {
 243         struct obd_import *imp = failed_req->rq_import;
 244         ENTRY;
 245
 246         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 247                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 248                imp->imp_connection->c_remote_uuid.uuid);
 249
 250         if (ptlrpc_set_import_discon(imp,
 251                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 252                 if (!imp->imp_replayable) {
 253                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 254                                "auto-deactivating\n",
 255                                obd2cli_tgt(imp->imp_obd),
 256                                imp->imp_connection->c_remote_uuid.uuid,
 257                                imp->imp_obd->obd_name);
 258                         ptlrpc_deactivate_import(imp);
 259                 }
 260                 /* to control recovery via lctl {disable|enable}_recovery */
 261                 if (imp->imp_deactive == 0)
 262                         ptlrpc_connect_import(imp);
 263         }
 264
 265         /* Wait for recovery to complete and resend. If evicted, then
 266            this request will be errored out later.*/
 267         spin_lock(&failed_req->rq_lock);
 268         if (!failed_req->rq_no_resend)
 269                 failed_req->rq_resend = 1;
 270         spin_unlock(&failed_req->rq_lock);
 271
 272         EXIT;
 273 }
 274
 275 /**
 276  * Administratively active/deactive a client.
 277  * This should only be called by the ioctl interface, currently
 278  *  - the lctl deactivate and activate commands
 279  *  - echo 0/1 >> /proc/osc/XXX/active
 280  *  - client umount -f (ll_umount_begin)
 281  */
 282 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 283 {
 284         struct obd_device *obd = imp->imp_obd;
 285         int rc = 0;
 286
 287         ENTRY;
 288         LASSERT(obd);
 289
 290         /* When deactivating, mark import invalid, and abort in-flight
 291          * requests. */
 292         if (!active) {
 293                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 294                               "request\n", obd2cli_tgt(imp->imp_obd));
 295
 296                 /* set before invalidate to avoid messages about imp_inval
 297                  * set without imp_deactive in ptlrpc_import_delay_req */
 298                 spin_lock(&imp->imp_lock);
 299                 imp->imp_deactive = 1;
 300                 spin_unlock(&imp->imp_lock);
 301
 302                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 303
 304                 ptlrpc_invalidate_import(imp);
 305         }
 306
 307         /* When activating, mark import valid, and attempt recovery */
 308         if (active) {
 309                 CDEBUG(D_HA, "setting import %s VALID\n",
 310                        obd2cli_tgt(imp->imp_obd));
 311
 312                 spin_lock(&imp->imp_lock);
 313                 imp->imp_deactive = 0;
 314                 spin_unlock(&imp->imp_lock);
 315                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 316
 317                 rc = ptlrpc_recover_import(imp, NULL, 0);
 318         }
 319
 320         RETURN(rc);
 321 }
 322 EXPORT_SYMBOL(ptlrpc_set_import_active);
 323
 324 /* Attempt to reconnect an import */
 325 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 326 {
 327         int rc = 0;
 328         ENTRY;
 329
 330         spin_lock(&imp->imp_lock);
 331         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 332             cfs_atomic_read(&imp->imp_inval_count))
 333                 rc = -EINVAL;
 334         spin_unlock(&imp->imp_lock);
 335         if (rc)
 336                 GOTO(out, rc);
 337
 338         /* force import to be disconnected. */
 339         ptlrpc_set_import_discon(imp, 0);
 340
 341         if (new_uuid) {
 342                 struct obd_uuid uuid;
 343
 344                 /* intruct import to use new uuid */
 345                 obd_str2uuid(&uuid, new_uuid);
 346                 rc = import_set_conn_priority(imp, &uuid);
 347                 if (rc)
 348                         GOTO(out, rc);
 349         }
 350
 351         /* Check if reconnect is already in progress */
 352         spin_lock(&imp->imp_lock);
 353         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 354                 imp->imp_force_verify = 1;
 355                 rc = -EALREADY;
 356         }
 357         spin_unlock(&imp->imp_lock);
 358         if (rc)
 359                 GOTO(out, rc);
 360
 361         rc = ptlrpc_connect_import(imp);
 362         if (rc)
 363                 GOTO(out, rc);
 364
 365         if (!async) {
 366                 struct l_wait_info lwi;
 367                 int secs = cfs_time_seconds(obd_timeout);
 368
 369                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 370                        obd2cli_tgt(imp->imp_obd), secs);
 371
 372                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 373                 rc = l_wait_event(imp->imp_recovery_waitq,
 374                                   !ptlrpc_import_in_recovery(imp), &lwi);
 375                 CDEBUG(D_HA, "%s: recovery finished\n",
 376                        obd2cli_tgt(imp->imp_obd));
 377         }
 378         EXIT;
 379
 380 out:
 381         return rc;
 382 }
 383 EXPORT_SYMBOL(ptlrpc_recover_import);
 384
 385 int ptlrpc_import_in_recovery(struct obd_import *imp)
 386 {
 387         int in_recovery = 1;
 388
 389         spin_lock(&imp->imp_lock);
 390         if (imp->imp_state == LUSTRE_IMP_FULL ||
 391             imp->imp_state == LUSTRE_IMP_CLOSED ||
 392             imp->imp_state == LUSTRE_IMP_DISCON ||
 393             imp->imp_obd->obd_no_recov)
 394                 in_recovery = 0;
 395         spin_unlock(&imp->imp_lock);
 396
 397         return in_recovery;
 398 }