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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/ptlrpc/recover.c
  32  *
  33  * Author: Mike Shaver <shaver@clusterfs.com>
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_RPC
  37 #include <linux/list.h>
  38 #include <libcfs/libcfs.h>
  39 #include <obd_support.h>
  40 #include <lustre_ha.h>
  41 #include <lustre_net.h>
  42 #include <lustre_import.h>
  43 #include <lustre_export.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46
  47 #include "ptlrpc_internal.h"
  48
  49 /**
  50  * Start recovery on disconnected import.
  51  * This is done by just attempting a connect
  52  */
  53 void ptlrpc_initiate_recovery(struct obd_import *imp)
  54 {
  55         ENTRY;
  56
  57         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  58         ptlrpc_connect_import(imp);
  59
  60         EXIT;
  61 }
  62
  63 /**
  64  * Identify what request from replay list needs to be replayed next
  65  * (based on what we have already replayed) and send it to server.
  66  */
  67 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  68 {
  69         int rc = 0;
  70         struct ptlrpc_request *req = NULL;
  71         __u64 last_transno;
  72         ENTRY;
  73
  74         *inflight = 0;
  75
  76         /* It might have committed some after we last spoke, so make sure we
  77          * get rid of them now.
  78          */
  79         spin_lock(&imp->imp_lock);
  80         imp->imp_last_transno_checked = 0;
  81         ptlrpc_free_committed(imp);
  82         last_transno = imp->imp_last_replay_transno;
  83
  84         CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
  85                imp, obd2cli_tgt(imp->imp_obd),
  86                imp->imp_peer_committed_transno, last_transno);
  87
  88         /* Replay all the committed open requests on committed_list first */
  89         if (!list_empty(&imp->imp_committed_list)) {
  90                 req = list_last_entry(&imp->imp_committed_list,
  91                                       struct ptlrpc_request, rq_replay_list);
  92
  93                 /* The last request on committed_list hasn't been replayed */
  94                 if (req->rq_transno > last_transno) {
  95                         if (!imp->imp_resend_replay ||
  96                             imp->imp_replay_cursor == &imp->imp_committed_list)
  97                                 imp->imp_replay_cursor =
  98                                         imp->imp_replay_cursor->next;
  99
 100                         while (imp->imp_replay_cursor !=
 101                                &imp->imp_committed_list) {
 102                                 req = list_entry(imp->imp_replay_cursor,
 103                                                  struct ptlrpc_request,
 104                                                  rq_replay_list);
 105                                 if (req->rq_transno > last_transno)
 106                                         break;
 107
 108                                 req = NULL;
 109                                 LASSERT(!list_empty(imp->imp_replay_cursor));
 110                                 imp->imp_replay_cursor =
 111                                         imp->imp_replay_cursor->next;
 112                         }
 113                 } else {
 114                         /* All requests on committed_list have been replayed */
 115                         imp->imp_replay_cursor = &imp->imp_committed_list;
 116                         req = NULL;
 117                 }
 118         }
 119
 120         /* All the requests in committed list have been replayed, let's replay
 121          * the imp_replay_list */
 122         if (req == NULL) {
 123                 struct ptlrpc_request *tmp;
 124
 125                 list_for_each_entry(tmp, &imp->imp_replay_list,
 126                                     rq_replay_list) {
 127                         if (tmp->rq_transno > last_transno) {
 128                                 req = tmp;
 129                                 break;
 130                         }
 131                 }
 132         }
 133
 134         /* If need to resend the last sent transno (because a reconnect
 135          * has occurred), then stop on the matching req and send it again.
 136          * If, however, the last sent transno has been committed then we
 137          * continue replay from the next request. */
 138         if (req != NULL && imp->imp_resend_replay)
 139                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 140
 141         /* ptlrpc_prepare_replay() may fail to add the reqeust into unreplied
 142          * list if the request hasn't been added to replay list then. Another
 143          * exception is that resend replay could have been removed from the
 144          * unreplied list. */
 145         if (req != NULL && list_empty(&req->rq_unreplied_list)) {
 146                 DEBUG_REQ(D_HA, req, "resend_replay=%d, last_transno=%llu",
 147                           imp->imp_resend_replay, last_transno);
 148                 ptlrpc_add_unreplied(req);
 149                 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 150         }
 151
 152         imp->imp_resend_replay = 0;
 153         spin_unlock(&imp->imp_lock);
 154
 155         if (req != NULL) {
 156                 LASSERT(!list_empty(&req->rq_unreplied_list));
 157
 158                 rc = ptlrpc_replay_req(req);
 159                 if (rc) {
 160                         CERROR("recovery replay error %d for req %llu\n",
 161                                rc, req->rq_xid);
 162                         RETURN(rc);
 163                 }
 164                 *inflight = 1;
 165         }
 166         RETURN(rc);
 167 }
 168
 169 /**
 170  * Schedule resending of request on sending_list. This is done after
 171  * we completed replaying of requests and locks.
 172  */
 173 int ptlrpc_resend(struct obd_import *imp)
 174 {
 175         struct ptlrpc_request *req;
 176
 177         ENTRY;
 178
 179         /* As long as we're in recovery, nothing should be added to the sending
 180          * list, so we don't need to hold the lock during this iteration and
 181          * resend process.
 182          */
 183         /* Well... what if lctl recover is called twice at the same time?
 184          */
 185         spin_lock(&imp->imp_lock);
 186         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 187                 spin_unlock(&imp->imp_lock);
 188                 RETURN(-1);
 189         }
 190
 191         list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
 192                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 193                          "req %p bad\n", req);
 194                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 195
 196                 /* If the request is allowed to be sent during replay and it
 197                  * is not timeout yet, then it does not need to be resent. */
 198                 if (!ptlrpc_no_resend(req) &&
 199                     (req->rq_timedout || !req->rq_allow_replay))
 200                         ptlrpc_resend_req(req);
 201         }
 202         spin_unlock(&imp->imp_lock);
 203
 204         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT, 2);
 205         RETURN(0);
 206 }
 207
 208 /**
 209  * Go through all requests in delayed list and wake their threads
 210  * for resending
 211  */
 212 void ptlrpc_wake_delayed(struct obd_import *imp)
 213 {
 214         struct ptlrpc_request *req;
 215
 216         spin_lock(&imp->imp_lock);
 217         list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
 218                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 219                 ptlrpc_client_wake_req(req);
 220         }
 221         spin_unlock(&imp->imp_lock);
 222 }
 223
 224 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 225 {
 226         struct obd_import *imp = failed_req->rq_import;
 227         int conn = lustre_msg_get_conn_cnt(failed_req->rq_reqmsg);
 228         ENTRY;
 229
 230         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 231                 imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 232                 imp->imp_connection->c_remote_uuid.uuid);
 233
 234         if (ptlrpc_set_import_discon(imp, conn, true)) {
 235                 /* to control recovery via lctl {disable|enable}_recovery */
 236                 if (imp->imp_deactive == 0)
 237                         ptlrpc_connect_import(imp);
 238         }
 239
 240         /* Wait for recovery to complete and resend. If evicted, then
 241            this request will be errored out later.*/
 242         spin_lock(&failed_req->rq_lock);
 243         if (!failed_req->rq_no_resend)
 244                 failed_req->rq_resend = 1;
 245         spin_unlock(&failed_req->rq_lock);
 246
 247         EXIT;
 248 }
 249
 250 /**
 251  * Administratively active/deactive a client.
 252  * This should only be called by the ioctl interface, currently
 253  *  - the lctl deactivate and activate commands
 254  *  - echo 0/1 >> /proc/osc/XXX/active
 255  *  - client umount -f (ll_umount_begin)
 256  */
 257 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 258 {
 259         struct obd_device *obd = imp->imp_obd;
 260         int rc = 0;
 261
 262         ENTRY;
 263         LASSERT(obd);
 264
 265         /* When deactivating, mark import invalid, and abort in-flight
 266          * requests. */
 267         if (!active) {
 268                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 269                               "request\n", obd2cli_tgt(imp->imp_obd));
 270
 271                 /* set before invalidate to avoid messages about imp_inval
 272                  * set without imp_deactive in ptlrpc_import_delay_req */
 273                 spin_lock(&imp->imp_lock);
 274                 imp->imp_deactive = 1;
 275                 spin_unlock(&imp->imp_lock);
 276
 277                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 278
 279                 ptlrpc_invalidate_import(imp);
 280         }
 281
 282         /* When activating, mark import valid, and attempt recovery */
 283         if (active) {
 284                 CDEBUG(D_HA, "setting import %s VALID\n",
 285                        obd2cli_tgt(imp->imp_obd));
 286
 287                 spin_lock(&imp->imp_lock);
 288                 imp->imp_deactive = 0;
 289                 spin_unlock(&imp->imp_lock);
 290                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 291
 292                 rc = ptlrpc_recover_import(imp, NULL, 0);
 293         }
 294
 295         RETURN(rc);
 296 }
 297 EXPORT_SYMBOL(ptlrpc_set_import_active);
 298
 299 /* Attempt to reconnect an import */
 300 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 301 {
 302         int rc = 0;
 303         ENTRY;
 304
 305         spin_lock(&imp->imp_lock);
 306         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 307             atomic_read(&imp->imp_inval_count))
 308                 rc = -EINVAL;
 309         spin_unlock(&imp->imp_lock);
 310         if (rc)
 311                 GOTO(out, rc);
 312
 313         /* force import to be disconnected. */
 314         ptlrpc_set_import_discon(imp, 0, false);
 315
 316         if (new_uuid) {
 317                 struct obd_uuid uuid;
 318
 319                 /* intruct import to use new uuid */
 320                 obd_str2uuid(&uuid, new_uuid);
 321                 rc = import_set_conn_priority(imp, &uuid);
 322                 if (rc)
 323                         GOTO(out, rc);
 324         }
 325
 326         /* Check if reconnect is already in progress */
 327         spin_lock(&imp->imp_lock);
 328         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 329                 imp->imp_force_verify = 1;
 330                 rc = -EALREADY;
 331         }
 332         spin_unlock(&imp->imp_lock);
 333         if (rc)
 334                 GOTO(out, rc);
 335
 336         OBD_RACE(OBD_FAIL_PTLRPC_CONNECT_RACE);
 337
 338         rc = ptlrpc_connect_import(imp);
 339         if (rc)
 340                 GOTO(out, rc);
 341
 342         if (!async) {
 343                 long timeout = cfs_time_seconds(obd_timeout);
 344
 345                 CDEBUG(D_HA, "%s: recovery started, waiting %u jiffies\n",
 346                        obd2cli_tgt(imp->imp_obd), obd_timeout);
 347
 348                 rc = wait_event_idle_timeout(imp->imp_recovery_waitq,
 349                                              !ptlrpc_import_in_recovery(imp),
 350                                              timeout);
 351                 if (rc == 0)
 352                         rc = -ETIMEDOUT;
 353                 else
 354                         rc = 0;
 355                 CDEBUG(D_HA, "%s: recovery finished\n",
 356                        obd2cli_tgt(imp->imp_obd));
 357         }
 358         EXIT;
 359
 360 out:
 361         return rc;
 362 }
 363 EXPORT_SYMBOL(ptlrpc_recover_import);
 364
 365 int ptlrpc_import_in_recovery(struct obd_import *imp)
 366 {
 367         int in_recovery = 1;
 368
 369         spin_lock(&imp->imp_lock);
 370         if (imp->imp_state <= LUSTRE_IMP_DISCON ||
 371             imp->imp_state >= LUSTRE_IMP_FULL ||
 372             imp->imp_obd->obd_no_recov)
 373                 in_recovery = 0;
 374         spin_unlock(&imp->imp_lock);
 375
 376         return in_recovery;
 377 }