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.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, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/recover.c
  33  *
  34  * Author: Mike Shaver <shaver@clusterfs.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include "../../include/linux/libcfs/libcfs.h"
  39
  40 #include "../include/obd_support.h"
  41 #include "../include/lustre_ha.h"
  42 #include "../include/lustre_net.h"
  43 #include "../include/lustre_import.h"
  44 #include "../include/lustre_export.h"
  45 #include "../include/obd.h"
  46 #include "../include/obd_class.h"
  47 #include <linux/list.h>
  48
  49 #include "ptlrpc_internal.h"
  50
  51 /**
  52  * Start recovery on disconnected import.
  53  * This is done by just attempting a connect
  54  */
  55 void ptlrpc_initiate_recovery(struct obd_import *imp)
  56 {
  57         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  58         ptlrpc_connect_import(imp);
  59 }
  60
  61 /**
  62  * Identify what request from replay list needs to be replayed next
  63  * (based on what we have already replayed) and send it to server.
  64  */
  65 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  66 {
  67         int rc = 0;
  68         struct list_head *tmp, *pos;
  69         struct ptlrpc_request *req = NULL;
  70         __u64 last_transno;
  71
  72         *inflight = 0;
  73
  74         /* It might have committed some after we last spoke, so make sure we
  75          * get rid of them now.
  76          */
  77         spin_lock(&imp->imp_lock);
  78         imp->imp_last_transno_checked = 0;
  79         ptlrpc_free_committed(imp);
  80         last_transno = imp->imp_last_replay_transno;
  81         spin_unlock(&imp->imp_lock);
  82
  83         CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
  84                imp, obd2cli_tgt(imp->imp_obd),
  85                imp->imp_peer_committed_transno, last_transno);
  86
  87         /* Do I need to hold a lock across this iteration?  We shouldn't be
  88          * racing with any additions to the list, because we're in recovery
  89          * and are therefore not processing additional requests to add.  Calls
  90          * to ptlrpc_free_committed might commit requests, but nothing "newer"
  91          * than the one we're replaying (it can't be committed until it's
  92          * replayed, and we're doing that here).  l_f_e_safe protects against
  93          * problems with the current request being committed, in the unlikely
  94          * event of that race.  So, in conclusion, I think that it's safe to
  95          * perform this list-walk without the imp_lock held.
  96          *
  97          * But, the {mdc,osc}_replay_open callbacks both iterate
  98          * request lists, and have comments saying they assume the
  99          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 100          * just a little race...
 101          */
 102
 103         /* Replay all the committed open requests on committed_list first */
 104         if (!list_empty(&imp->imp_committed_list)) {
 105                 tmp = imp->imp_committed_list.prev;
 106                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 107
 108                 /* The last request on committed_list hasn't been replayed */
 109                 if (req->rq_transno > last_transno) {
 110                         /* Since the imp_committed_list is immutable before
 111                          * all of it's requests being replayed, it's safe to
 112                          * use a cursor to accelerate the search
 113                          */
 114                         imp->imp_replay_cursor = imp->imp_replay_cursor->next;
 115
 116                         while (imp->imp_replay_cursor !=
 117                                &imp->imp_committed_list) {
 118                                 req = list_entry(imp->imp_replay_cursor,
 119                                                  struct ptlrpc_request,
 120                                                  rq_replay_list);
 121                                 if (req->rq_transno > last_transno)
 122                                         break;
 123
 124                                 req = NULL;
 125                                 imp->imp_replay_cursor =
 126                                         imp->imp_replay_cursor->next;
 127                         }
 128                 } else {
 129                         /* All requests on committed_list have been replayed */
 130                         imp->imp_replay_cursor = &imp->imp_committed_list;
 131                         req = NULL;
 132                 }
 133         }
 134
 135         /* All the requests in committed list have been replayed, let's replay
 136          * the imp_replay_list
 137          */
 138         if (!req) {
 139                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 140                         req = list_entry(tmp, struct ptlrpc_request,
 141                                          rq_replay_list);
 142
 143                         if (req->rq_transno > last_transno)
 144                                 break;
 145                         req = NULL;
 146                 }
 147         }
 148
 149         /* If need to resend the last sent transno (because a reconnect
 150          * has occurred), then stop on the matching req and send it again.
 151          * If, however, the last sent transno has been committed then we
 152          * continue replay from the next request.
 153          */
 154         if (req && imp->imp_resend_replay)
 155                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 156
 157         spin_lock(&imp->imp_lock);
 158         imp->imp_resend_replay = 0;
 159         spin_unlock(&imp->imp_lock);
 160
 161         if (req) {
 162                 rc = ptlrpc_replay_req(req);
 163                 if (rc) {
 164                         CERROR("recovery replay error %d for req %llu\n",
 165                                rc, req->rq_xid);
 166                         return rc;
 167                 }
 168                 *inflight = 1;
 169         }
 170         return rc;
 171 }
 172
 173 /**
 174  * Schedule resending of request on sending_list. This is done after
 175  * we completed replaying of requests and locks.
 176  */
 177 int ptlrpc_resend(struct obd_import *imp)
 178 {
 179         struct ptlrpc_request *req, *next;
 180
 181         /* As long as we're in recovery, nothing should be added to the sending
 182          * list, so we don't need to hold the lock during this iteration and
 183          * resend process.
 184          */
 185         /* Well... what if lctl recover is called twice at the same time?
 186          */
 187         spin_lock(&imp->imp_lock);
 188         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 189                 spin_unlock(&imp->imp_lock);
 190                 return -1;
 191         }
 192
 193         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 194                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 195                          "req %p bad\n", req);
 196                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 197                 if (!ptlrpc_no_resend(req))
 198                         ptlrpc_resend_req(req);
 199         }
 200         spin_unlock(&imp->imp_lock);
 201
 202         return 0;
 203 }
 204 EXPORT_SYMBOL(ptlrpc_resend);
 205
 206 /**
 207  * Go through all requests in delayed list and wake their threads
 208  * for resending
 209  */
 210 void ptlrpc_wake_delayed(struct obd_import *imp)
 211 {
 212         struct list_head *tmp, *pos;
 213         struct ptlrpc_request *req;
 214
 215         spin_lock(&imp->imp_lock);
 216         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 217                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 218
 219                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 220                 ptlrpc_client_wake_req(req);
 221         }
 222         spin_unlock(&imp->imp_lock);
 223 }
 224 EXPORT_SYMBOL(ptlrpc_wake_delayed);
 225
 226 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 227 {
 228         struct obd_import *imp = failed_req->rq_import;
 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,
 235                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 236                 if (!imp->imp_replayable) {
 237                         CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n",
 238                                obd2cli_tgt(imp->imp_obd),
 239                                imp->imp_connection->c_remote_uuid.uuid,
 240                                imp->imp_obd->obd_name);
 241                         ptlrpc_deactivate_import(imp);
 242                 }
 243                 /* to control recovery via lctl {disable|enable}_recovery */
 244                 if (imp->imp_deactive == 0)
 245                         ptlrpc_connect_import(imp);
 246         }
 247
 248         /* Wait for recovery to complete and resend. If evicted, then
 249          * this request will be errored out later.
 250          */
 251         spin_lock(&failed_req->rq_lock);
 252         if (!failed_req->rq_no_resend)
 253                 failed_req->rq_resend = 1;
 254         spin_unlock(&failed_req->rq_lock);
 255 }
 256
 257 /**
 258  * Administratively active/deactive a client.
 259  * This should only be called by the ioctl interface, currently
 260  *  - the lctl deactivate and activate commands
 261  *  - echo 0/1 >> /sys/fs/lustre/osc/XXX/active
 262  *  - client umount -f (ll_umount_begin)
 263  */
 264 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 265 {
 266         struct obd_device *obd = imp->imp_obd;
 267         int rc = 0;
 268
 269         LASSERT(obd);
 270
 271         /* When deactivating, mark import invalid, and abort in-flight
 272          * requests.
 273          */
 274         if (!active) {
 275                 LCONSOLE_WARN("setting import %s INACTIVE by administrator request\n",
 276                               obd2cli_tgt(imp->imp_obd));
 277
 278                 /* set before invalidate to avoid messages about imp_inval
 279                  * set without imp_deactive in ptlrpc_import_delay_req
 280                  */
 281                 spin_lock(&imp->imp_lock);
 282                 imp->imp_deactive = 1;
 283                 spin_unlock(&imp->imp_lock);
 284
 285                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 286
 287                 ptlrpc_invalidate_import(imp);
 288         }
 289
 290         /* When activating, mark import valid, and attempt recovery */
 291         if (active) {
 292                 CDEBUG(D_HA, "setting import %s VALID\n",
 293                        obd2cli_tgt(imp->imp_obd));
 294
 295                 spin_lock(&imp->imp_lock);
 296                 imp->imp_deactive = 0;
 297                 spin_unlock(&imp->imp_lock);
 298                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 299
 300                 rc = ptlrpc_recover_import(imp, NULL, 0);
 301         }
 302
 303         return rc;
 304 }
 305 EXPORT_SYMBOL(ptlrpc_set_import_active);
 306
 307 /* Attempt to reconnect an import */
 308 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 309 {
 310         int rc = 0;
 311
 312         spin_lock(&imp->imp_lock);
 313         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 314             atomic_read(&imp->imp_inval_count))
 315                 rc = -EINVAL;
 316         spin_unlock(&imp->imp_lock);
 317         if (rc)
 318                 goto out;
 319
 320         /* force import to be disconnected. */
 321         ptlrpc_set_import_discon(imp, 0);
 322
 323         if (new_uuid) {
 324                 struct obd_uuid uuid;
 325
 326                 /* intruct import to use new uuid */
 327                 obd_str2uuid(&uuid, new_uuid);
 328                 rc = import_set_conn_priority(imp, &uuid);
 329                 if (rc)
 330                         goto out;
 331         }
 332
 333         /* Check if reconnect is already in progress */
 334         spin_lock(&imp->imp_lock);
 335         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 336                 imp->imp_force_verify = 1;
 337                 rc = -EALREADY;
 338         }
 339         spin_unlock(&imp->imp_lock);
 340         if (rc)
 341                 goto out;
 342
 343         rc = ptlrpc_connect_import(imp);
 344         if (rc)
 345                 goto out;
 346
 347         if (!async) {
 348                 struct l_wait_info lwi;
 349                 int secs = cfs_time_seconds(obd_timeout);
 350
 351                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 352                        obd2cli_tgt(imp->imp_obd), secs);
 353
 354                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 355                 rc = l_wait_event(imp->imp_recovery_waitq,
 356                                   !ptlrpc_import_in_recovery(imp), &lwi);
 357                 CDEBUG(D_HA, "%s: recovery finished\n",
 358                        obd2cli_tgt(imp->imp_obd));
 359         }
 360
 361 out:
 362         return rc;
 363 }
 364 EXPORT_SYMBOL(ptlrpc_recover_import);
 365
 366 int ptlrpc_import_in_recovery(struct obd_import *imp)
 367 {
 368         int in_recovery = 1;
 369
 370         spin_lock(&imp->imp_lock);
 371         if (imp->imp_state == LUSTRE_IMP_FULL ||
 372             imp->imp_state == LUSTRE_IMP_CLOSED ||
 373             imp->imp_state == LUSTRE_IMP_DISCON ||
 374             imp->imp_obd->obd_no_recov)
 375                 in_recovery = 0;
 376         spin_unlock(&imp->imp_lock);
 377
 378         return in_recovery;
 379 }