drivers/staging/lustre/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, 2015, 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 "../../include/linux/libcfs/libcfs.h"
  43
  44 #include "../include/obd_support.h"
  45 #include "../include/lustre_ha.h"
  46 #include "../include/lustre_net.h"
  47 #include "../include/lustre_import.h"
  48 #include "../include/lustre_export.h"
  49 #include "../include/obd.h"
  50 #include "../include/obd_class.h"
  51 #include <linux/list.h>
  52
  53 #include "ptlrpc_internal.h"
  54
  55 /**
  56  * Start recovery on disconnected import.
  57  * This is done by just attempting a connect
  58  */
  59 void ptlrpc_initiate_recovery(struct obd_import *imp)
  60 {
  61         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  62         ptlrpc_connect_import(imp);
  63 }
  64
  65 /**
  66  * Identify what request from replay list needs to be replayed next
  67  * (based on what we have already replayed) and send it to server.
  68  */
  69 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  70 {
  71         int rc = 0;
  72         struct list_head *tmp, *pos;
  73         struct ptlrpc_request *req = NULL;
  74         __u64 last_transno;
  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, rq_replay_list);
 111
 112                 /* The last request on committed_list hasn't been replayed */
 113                 if (req->rq_transno > last_transno) {
 114                         /* Since the imp_committed_list is immutable before
 115                          * all of it's requests being replayed, it's safe to
 116                          * use a cursor to accelerate the search
 117                          */
 118                         imp->imp_replay_cursor = imp->imp_replay_cursor->next;
 119
 120                         while (imp->imp_replay_cursor !=
 121                                &imp->imp_committed_list) {
 122                                 req = list_entry(imp->imp_replay_cursor,
 123                                                  struct ptlrpc_request,
 124                                                  rq_replay_list);
 125                                 if (req->rq_transno > last_transno)
 126                                         break;
 127
 128                                 req = NULL;
 129                                 imp->imp_replay_cursor =
 130                                         imp->imp_replay_cursor->next;
 131                         }
 132                 } else {
 133                         /* All requests on committed_list have been replayed */
 134                         imp->imp_replay_cursor = &imp->imp_committed_list;
 135                         req = NULL;
 136                 }
 137         }
 138
 139         /* All the requests in committed list have been replayed, let's replay
 140          * the imp_replay_list
 141          */
 142         if (!req) {
 143                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 144                         req = list_entry(tmp, struct ptlrpc_request,
 145                                          rq_replay_list);
 146
 147                         if (req->rq_transno > last_transno)
 148                                 break;
 149                         req = NULL;
 150                 }
 151         }
 152
 153         /* If need to resend the last sent transno (because a reconnect
 154          * has occurred), then stop on the matching req and send it again.
 155          * If, however, the last sent transno has been committed then we
 156          * continue replay from the next request.
 157          */
 158         if (req && imp->imp_resend_replay)
 159                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 160
 161         spin_lock(&imp->imp_lock);
 162         imp->imp_resend_replay = 0;
 163         spin_unlock(&imp->imp_lock);
 164
 165         if (req) {
 166                 rc = ptlrpc_replay_req(req);
 167                 if (rc) {
 168                         CERROR("recovery replay error %d for req %llu\n",
 169                                rc, req->rq_xid);
 170                         return rc;
 171                 }
 172                 *inflight = 1;
 173         }
 174         return rc;
 175 }
 176
 177 /**
 178  * Schedule resending of request on sending_list. This is done after
 179  * we completed replaying of requests and locks.
 180  */
 181 int ptlrpc_resend(struct obd_import *imp)
 182 {
 183         struct ptlrpc_request *req, *next;
 184
 185         /* As long as we're in recovery, nothing should be added to the sending
 186          * list, so we don't need to hold the lock during this iteration and
 187          * resend process.
 188          */
 189         /* Well... what if lctl recover is called twice at the same time?
 190          */
 191         spin_lock(&imp->imp_lock);
 192         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 193                 spin_unlock(&imp->imp_lock);
 194                 return -1;
 195         }
 196
 197         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 198                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 199                          "req %p bad\n", req);
 200                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 201                 if (!ptlrpc_no_resend(req))
 202                         ptlrpc_resend_req(req);
 203         }
 204         spin_unlock(&imp->imp_lock);
 205
 206         return 0;
 207 }
 208 EXPORT_SYMBOL(ptlrpc_resend);
 209
 210 /**
 211  * Go through all requests in delayed list and wake their threads
 212  * for resending
 213  */
 214 void ptlrpc_wake_delayed(struct obd_import *imp)
 215 {
 216         struct list_head *tmp, *pos;
 217         struct ptlrpc_request *req;
 218
 219         spin_lock(&imp->imp_lock);
 220         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 221                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 222
 223                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 224                 ptlrpc_client_wake_req(req);
 225         }
 226         spin_unlock(&imp->imp_lock);
 227 }
 228 EXPORT_SYMBOL(ptlrpc_wake_delayed);
 229
 230 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 231 {
 232         struct obd_import *imp = failed_req->rq_import;
 233
 234         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 235                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 236                imp->imp_connection->c_remote_uuid.uuid);
 237
 238         if (ptlrpc_set_import_discon(imp,
 239                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 240                 if (!imp->imp_replayable) {
 241                         CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n",
 242                                obd2cli_tgt(imp->imp_obd),
 243                                imp->imp_connection->c_remote_uuid.uuid,
 244                                imp->imp_obd->obd_name);
 245                         ptlrpc_deactivate_import(imp);
 246                 }
 247                 /* to control recovery via lctl {disable|enable}_recovery */
 248                 if (imp->imp_deactive == 0)
 249                         ptlrpc_connect_import(imp);
 250         }
 251
 252         /* Wait for recovery to complete and resend. If evicted, then
 253          * this request will be errored out later.
 254          */
 255         spin_lock(&failed_req->rq_lock);
 256         if (!failed_req->rq_no_resend)
 257                 failed_req->rq_resend = 1;
 258         spin_unlock(&failed_req->rq_lock);
 259 }
 260
 261 /**
 262  * Administratively active/deactive a client.
 263  * This should only be called by the ioctl interface, currently
 264  *  - the lctl deactivate and activate commands
 265  *  - echo 0/1 >> /sys/fs/lustre/osc/XXX/active
 266  *  - client umount -f (ll_umount_begin)
 267  */
 268 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 269 {
 270         struct obd_device *obd = imp->imp_obd;
 271         int rc = 0;
 272
 273         LASSERT(obd);
 274
 275         /* When deactivating, mark import invalid, and abort in-flight
 276          * requests.
 277          */
 278         if (!active) {
 279                 LCONSOLE_WARN("setting import %s INACTIVE by administrator request\n",
 280                               obd2cli_tgt(imp->imp_obd));
 281
 282                 /* set before invalidate to avoid messages about imp_inval
 283                  * set without imp_deactive in ptlrpc_import_delay_req
 284                  */
 285                 spin_lock(&imp->imp_lock);
 286                 imp->imp_deactive = 1;
 287                 spin_unlock(&imp->imp_lock);
 288
 289                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 290
 291                 ptlrpc_invalidate_import(imp);
 292         }
 293
 294         /* When activating, mark import valid, and attempt recovery */
 295         if (active) {
 296                 CDEBUG(D_HA, "setting import %s VALID\n",
 297                        obd2cli_tgt(imp->imp_obd));
 298
 299                 spin_lock(&imp->imp_lock);
 300                 imp->imp_deactive = 0;
 301                 spin_unlock(&imp->imp_lock);
 302                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 303
 304                 rc = ptlrpc_recover_import(imp, NULL, 0);
 305         }
 306
 307         return rc;
 308 }
 309 EXPORT_SYMBOL(ptlrpc_set_import_active);
 310
 311 /* Attempt to reconnect an import */
 312 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 313 {
 314         int rc = 0;
 315
 316         spin_lock(&imp->imp_lock);
 317         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 318             atomic_read(&imp->imp_inval_count))
 319                 rc = -EINVAL;
 320         spin_unlock(&imp->imp_lock);
 321         if (rc)
 322                 goto out;
 323
 324         /* force import to be disconnected. */
 325         ptlrpc_set_import_discon(imp, 0);
 326
 327         if (new_uuid) {
 328                 struct obd_uuid uuid;
 329
 330                 /* intruct import to use new uuid */
 331                 obd_str2uuid(&uuid, new_uuid);
 332                 rc = import_set_conn_priority(imp, &uuid);
 333                 if (rc)
 334                         goto out;
 335         }
 336
 337         /* Check if reconnect is already in progress */
 338         spin_lock(&imp->imp_lock);
 339         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 340                 imp->imp_force_verify = 1;
 341                 rc = -EALREADY;
 342         }
 343         spin_unlock(&imp->imp_lock);
 344         if (rc)
 345                 goto out;
 346
 347         rc = ptlrpc_connect_import(imp);
 348         if (rc)
 349                 goto out;
 350
 351         if (!async) {
 352                 struct l_wait_info lwi;
 353                 int secs = cfs_time_seconds(obd_timeout);
 354
 355                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 356                        obd2cli_tgt(imp->imp_obd), secs);
 357
 358                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 359                 rc = l_wait_event(imp->imp_recovery_waitq,
 360                                   !ptlrpc_import_in_recovery(imp), &lwi);
 361                 CDEBUG(D_HA, "%s: recovery finished\n",
 362                        obd2cli_tgt(imp->imp_obd));
 363         }
 364
 365 out:
 366         return rc;
 367 }
 368 EXPORT_SYMBOL(ptlrpc_recover_import);
 369
 370 int ptlrpc_import_in_recovery(struct obd_import *imp)
 371 {
 372         int in_recovery = 1;
 373
 374         spin_lock(&imp->imp_lock);
 375         if (imp->imp_state == LUSTRE_IMP_FULL ||
 376             imp->imp_state == LUSTRE_IMP_CLOSED ||
 377             imp->imp_state == LUSTRE_IMP_DISCON ||
 378             imp->imp_obd->obd_no_recov)
 379                 in_recovery = 0;
 380         spin_unlock(&imp->imp_lock);
 381
 382         return in_recovery;
 383 }