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