kernel/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, 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 #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,
 111                                      rq_replay_list);
 112
 113                 /* The last request on committed_list hasn't been replayed */
 114                 if (req->rq_transno > last_transno) {
 115                         /* Since the imp_committed_list is immutable before
 116                          * all of it's requests being replayed, it's safe to
 117                          * use a cursor to accelerate the search */
 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         if (req == NULL) {
 142                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 143                         req = list_entry(tmp, struct ptlrpc_request,
 144                                          rq_replay_list);
 145
 146                         if (req->rq_transno > last_transno)
 147                                 break;
 148                         req = NULL;
 149                 }
 150         }
 151
 152         /* If need to resend the last sent transno (because a reconnect
 153          * has occurred), then stop on the matching req and send it again.
 154          * If, however, the last sent transno has been committed then we
 155          * continue replay from the next request. */
 156         if (req != NULL && imp->imp_resend_replay)
 157                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 158
 159         spin_lock(&imp->imp_lock);
 160         imp->imp_resend_replay = 0;
 161         spin_unlock(&imp->imp_lock);
 162
 163         if (req != NULL) {
 164                 rc = ptlrpc_replay_req(req);
 165                 if (rc) {
 166                         CERROR("recovery replay error %d for req %llu\n",
 167                                rc, req->rq_xid);
 168                         return rc;
 169                 }
 170                 *inflight = 1;
 171         }
 172         return rc;
 173 }
 174
 175 /**
 176  * Schedule resending of request on sending_list. This is done after
 177  * we completed replaying of requests and locks.
 178  */
 179 int ptlrpc_resend(struct obd_import *imp)
 180 {
 181         struct ptlrpc_request *req, *next;
 182
 183         /* As long as we're in recovery, nothing should be added to the sending
 184          * list, so we don't need to hold the lock during this iteration and
 185          * resend process.
 186          */
 187         /* Well... what if lctl recover is called twice at the same time?
 188          */
 189         spin_lock(&imp->imp_lock);
 190         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 191                 spin_unlock(&imp->imp_lock);
 192                 return -1;
 193         }
 194
 195         list_for_each_entry_safe(req, next, &imp->imp_sending_list,
 196                                      rq_list) {
 197                 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
 198                          "req %p bad\n", req);
 199                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 200                 if (!ptlrpc_no_resend(req))
 201                         ptlrpc_resend_req(req);
 202         }
 203         spin_unlock(&imp->imp_lock);
 204
 205         return 0;
 206 }
 207 EXPORT_SYMBOL(ptlrpc_resend);
 208
 209 /**
 210  * Go through all requests in delayed list and wake their threads
 211  * for resending
 212  */
 213 void ptlrpc_wake_delayed(struct obd_import *imp)
 214 {
 215         struct list_head *tmp, *pos;
 216         struct ptlrpc_request *req;
 217
 218         spin_lock(&imp->imp_lock);
 219         list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 220                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 221
 222                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 223                 ptlrpc_client_wake_req(req);
 224         }
 225         spin_unlock(&imp->imp_lock);
 226 }
 227 EXPORT_SYMBOL(ptlrpc_wake_delayed);
 228
 229 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 230 {
 231         struct obd_import *imp = failed_req->rq_import;
 232
 233         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 234                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 235                imp->imp_connection->c_remote_uuid.uuid);
 236
 237         if (ptlrpc_set_import_discon(imp,
 238                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 239                 if (!imp->imp_replayable) {
 240                         CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n",
 241                                obd2cli_tgt(imp->imp_obd),
 242                                imp->imp_connection->c_remote_uuid.uuid,
 243                                imp->imp_obd->obd_name);
 244                         ptlrpc_deactivate_import(imp);
 245                 }
 246                 /* to control recovery via lctl {disable|enable}_recovery */
 247                 if (imp->imp_deactive == 0)
 248                         ptlrpc_connect_import(imp);
 249         }
 250
 251         /* Wait for recovery to complete and resend. If evicted, then
 252            this request will be errored out later.*/
 253         spin_lock(&failed_req->rq_lock);
 254         if (!failed_req->rq_no_resend)
 255                 failed_req->rq_resend = 1;
 256         spin_unlock(&failed_req->rq_lock);
 257 }
 258
 259 /**
 260  * Administratively active/deactive a client.
 261  * This should only be called by the ioctl interface, currently
 262  *  - the lctl deactivate and activate commands
 263  *  - echo 0/1 >> /proc/osc/XXX/active
 264  *  - client umount -f (ll_umount_begin)
 265  */
 266 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 267 {
 268         struct obd_device *obd = imp->imp_obd;
 269         int rc = 0;
 270
 271         LASSERT(obd);
 272
 273         /* When deactivating, mark import invalid, and abort in-flight
 274          * requests. */
 275         if (!active) {
 276                 LCONSOLE_WARN("setting import %s INACTIVE by administrator request\n",
 277                               obd2cli_tgt(imp->imp_obd));
 278
 279                 /* set before invalidate to avoid messages about imp_inval
 280                  * set without imp_deactive in ptlrpc_import_delay_req */
 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 }