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_ost.h>
  55 #include <obd_class.h>
  56 #include <obd_lov.h> /* for IOC_LOV_SET_OSC_ACTIVE */
  57 #include <libcfs/list.h>
  58
  59 #include "ptlrpc_internal.h"
  60
  61 /**
  62  * Start recovery on disconnected import.
  63  * This is done by just attempting a connect
  64  */
  65 void ptlrpc_initiate_recovery(struct obd_import *imp)
  66 {
  67         ENTRY;
  68
  69         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  70         ptlrpc_connect_import(imp);
  71
  72         EXIT;
  73 }
  74
  75 /**
  76  * Identify what request from replay list needs to be replayed next
  77  * (based on what we have already replayed) and send it to server.
  78  */
  79 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  80 {
  81         int rc = 0;
  82         cfs_list_t *tmp, *pos;
  83         struct ptlrpc_request *req = NULL;
  84         __u64 last_transno;
  85         ENTRY;
  86
  87         *inflight = 0;
  88
  89         /* It might have committed some after we last spoke, so make sure we
  90          * get rid of them now.
  91          */
  92         spin_lock(&imp->imp_lock);
  93         imp->imp_last_transno_checked = 0;
  94         ptlrpc_free_committed(imp);
  95         last_transno = imp->imp_last_replay_transno;
  96         spin_unlock(&imp->imp_lock);
  97
  98         CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
  99                imp, obd2cli_tgt(imp->imp_obd),
 100                imp->imp_peer_committed_transno, last_transno);
 101
 102         /* Do I need to hold a lock across this iteration?  We shouldn't be
 103          * racing with any additions to the list, because we're in recovery
 104          * and are therefore not processing additional requests to add.  Calls
 105          * to ptlrpc_free_committed might commit requests, but nothing "newer"
 106          * than the one we're replaying (it can't be committed until it's
 107          * replayed, and we're doing that here).  l_f_e_safe protects against
 108          * problems with the current request being committed, in the unlikely
 109          * event of that race.  So, in conclusion, I think that it's safe to
 110          * perform this list-walk without the imp_lock held.
 111          *
 112          * But, the {mdc,osc}_replay_open callbacks both iterate
 113          * request lists, and have comments saying they assume the
 114          * imp_lock is being held by ptlrpc_replay, but it's not. it's
 115          * just a little race...
 116          */
 117
 118         /* Replay all the committed open requests on committed_list first */
 119         if (!cfs_list_empty(&imp->imp_committed_list)) {
 120                 tmp = imp->imp_committed_list.prev;
 121                 req = cfs_list_entry(tmp, struct ptlrpc_request,
 122                                      rq_replay_list);
 123
 124                 /* The last request on committed_list hasn't been replayed */
 125                 if (req->rq_transno > last_transno) {
 126                         /* Since the imp_committed_list is immutable before
 127                          * all of it's requests being replayed, it's safe to
 128                          * use a cursor to accelerate the search */
 129                         imp->imp_replay_cursor = imp->imp_replay_cursor->next;
 130
 131                         while (imp->imp_replay_cursor !=
 132                                &imp->imp_committed_list) {
 133                                 req = cfs_list_entry(imp->imp_replay_cursor,
 134                                                      struct ptlrpc_request,
 135                                                      rq_replay_list);
 136                                 if (req->rq_transno > last_transno)
 137                                         break;
 138
 139                                 req = NULL;
 140                                 imp->imp_replay_cursor =
 141                                         imp->imp_replay_cursor->next;
 142                         }
 143                 } else {
 144                         /* All requests on committed_list have been replayed */
 145                         imp->imp_replay_cursor = &imp->imp_committed_list;
 146                         req = NULL;
 147                 }
 148         }
 149
 150         /* All the requests in committed list have been replayed, let's replay
 151          * the imp_replay_list */
 152         if (req == NULL) {
 153                 cfs_list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 154                         req = cfs_list_entry(tmp, struct ptlrpc_request,
 155                                              rq_replay_list);
 156
 157                         if (req->rq_transno > last_transno)
 158                                 break;
 159                         req = NULL;
 160                 }
 161         }
 162
 163         /* If need to resend the last sent transno (because a reconnect
 164          * has occurred), then stop on the matching req and send it again.
 165          * If, however, the last sent transno has been committed then we
 166          * continue replay from the next request. */
 167         if (req != NULL && imp->imp_resend_replay)
 168                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 169
 170         spin_lock(&imp->imp_lock);
 171         imp->imp_resend_replay = 0;
 172         spin_unlock(&imp->imp_lock);
 173
 174         if (req != NULL) {
 175                 rc = ptlrpc_replay_req(req);
 176                 if (rc) {
 177                         CERROR("recovery replay error %d for req "
 178                                LPU64"\n", rc, req->rq_xid);
 179                         RETURN(rc);
 180                 }
 181                 *inflight = 1;
 182         }
 183         RETURN(rc);
 184 }
 185
 186 /**
 187  * Schedule resending of request on sending_list. This is done after
 188  * we completed replaying of requests and locks.
 189  */
 190 int ptlrpc_resend(struct obd_import *imp)
 191 {
 192         struct ptlrpc_request *req, *next;
 193
 194         ENTRY;
 195
 196         /* As long as we're in recovery, nothing should be added to the sending
 197          * list, so we don't need to hold the lock during this iteration and
 198          * resend process.
 199          */
 200         /* Well... what if lctl recover is called twice at the same time?
 201          */
 202         spin_lock(&imp->imp_lock);
 203         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 204                 spin_unlock(&imp->imp_lock);
 205                 RETURN(-1);
 206         }
 207
 208         cfs_list_for_each_entry_safe(req, next, &imp->imp_sending_list,
 209                                      rq_list) {
 210                 LASSERTF((long)req > PAGE_CACHE_SIZE && req != LP_POISON,
 211                          "req %p bad\n", req);
 212                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 213                 if (!ptlrpc_no_resend(req))
 214                         ptlrpc_resend_req(req);
 215         }
 216         spin_unlock(&imp->imp_lock);
 217
 218         RETURN(0);
 219 }
 220 EXPORT_SYMBOL(ptlrpc_resend);
 221
 222 /**
 223  * Go through all requests in delayed list and wake their threads
 224  * for resending
 225  */
 226 void ptlrpc_wake_delayed(struct obd_import *imp)
 227 {
 228         cfs_list_t *tmp, *pos;
 229         struct ptlrpc_request *req;
 230
 231         spin_lock(&imp->imp_lock);
 232         cfs_list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
 233                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 234
 235                 DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
 236                 ptlrpc_client_wake_req(req);
 237         }
 238         spin_unlock(&imp->imp_lock);
 239 }
 240 EXPORT_SYMBOL(ptlrpc_wake_delayed);
 241
 242 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 243 {
 244         struct obd_import *imp = failed_req->rq_import;
 245         ENTRY;
 246
 247         CDEBUG(D_HA, "import %s of %s@%s abruptly disconnected: reconnecting\n",
 248                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 249                imp->imp_connection->c_remote_uuid.uuid);
 250
 251         if (ptlrpc_set_import_discon(imp,
 252                               lustre_msg_get_conn_cnt(failed_req->rq_reqmsg))) {
 253                 if (!imp->imp_replayable) {
 254                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 255                                "auto-deactivating\n",
 256                                obd2cli_tgt(imp->imp_obd),
 257                                imp->imp_connection->c_remote_uuid.uuid,
 258                                imp->imp_obd->obd_name);
 259                         ptlrpc_deactivate_import(imp);
 260                 }
 261                 /* to control recovery via lctl {disable|enable}_recovery */
 262                 if (imp->imp_deactive == 0)
 263                         ptlrpc_connect_import(imp);
 264         }
 265
 266         /* Wait for recovery to complete and resend. If evicted, then
 267            this request will be errored out later.*/
 268         spin_lock(&failed_req->rq_lock);
 269         if (!failed_req->rq_no_resend)
 270                 failed_req->rq_resend = 1;
 271         spin_unlock(&failed_req->rq_lock);
 272
 273         EXIT;
 274 }
 275
 276 /**
 277  * Administratively active/deactive a client.
 278  * This should only be called by the ioctl interface, currently
 279  *  - the lctl deactivate and activate commands
 280  *  - echo 0/1 >> /proc/osc/XXX/active
 281  *  - client umount -f (ll_umount_begin)
 282  */
 283 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 284 {
 285         struct obd_device *obd = imp->imp_obd;
 286         int rc = 0;
 287
 288         ENTRY;
 289         LASSERT(obd);
 290
 291         /* When deactivating, mark import invalid, and abort in-flight
 292          * requests. */
 293         if (!active) {
 294                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 295                               "request\n", obd2cli_tgt(imp->imp_obd));
 296
 297                 /* set before invalidate to avoid messages about imp_inval
 298                  * set without imp_deactive in ptlrpc_import_delay_req */
 299                 spin_lock(&imp->imp_lock);
 300                 imp->imp_deactive = 1;
 301                 spin_unlock(&imp->imp_lock);
 302
 303                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DEACTIVATE);
 304
 305                 ptlrpc_invalidate_import(imp);
 306         }
 307
 308         /* When activating, mark import valid, and attempt recovery */
 309         if (active) {
 310                 CDEBUG(D_HA, "setting import %s VALID\n",
 311                        obd2cli_tgt(imp->imp_obd));
 312
 313                 spin_lock(&imp->imp_lock);
 314                 imp->imp_deactive = 0;
 315                 spin_unlock(&imp->imp_lock);
 316                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_ACTIVATE);
 317
 318                 rc = ptlrpc_recover_import(imp, NULL, 0);
 319         }
 320
 321         RETURN(rc);
 322 }
 323 EXPORT_SYMBOL(ptlrpc_set_import_active);
 324
 325 /* Attempt to reconnect an import */
 326 int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 327 {
 328         int rc = 0;
 329         ENTRY;
 330
 331         spin_lock(&imp->imp_lock);
 332         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 333             cfs_atomic_read(&imp->imp_inval_count))
 334                 rc = -EINVAL;
 335         spin_unlock(&imp->imp_lock);
 336         if (rc)
 337                 GOTO(out, rc);
 338
 339         /* force import to be disconnected. */
 340         ptlrpc_set_import_discon(imp, 0);
 341
 342         if (new_uuid) {
 343                 struct obd_uuid uuid;
 344
 345                 /* intruct import to use new uuid */
 346                 obd_str2uuid(&uuid, new_uuid);
 347                 rc = import_set_conn_priority(imp, &uuid);
 348                 if (rc)
 349                         GOTO(out, rc);
 350         }
 351
 352         /* Check if reconnect is already in progress */
 353         spin_lock(&imp->imp_lock);
 354         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 355                 imp->imp_force_verify = 1;
 356                 rc = -EALREADY;
 357         }
 358         spin_unlock(&imp->imp_lock);
 359         if (rc)
 360                 GOTO(out, rc);
 361
 362         rc = ptlrpc_connect_import(imp);
 363         if (rc)
 364                 GOTO(out, rc);
 365
 366         if (!async) {
 367                 struct l_wait_info lwi;
 368                 int secs = cfs_time_seconds(obd_timeout);
 369
 370                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 371                        obd2cli_tgt(imp->imp_obd), secs);
 372
 373                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 374                 rc = l_wait_event(imp->imp_recovery_waitq,
 375                                   !ptlrpc_import_in_recovery(imp), &lwi);
 376                 CDEBUG(D_HA, "%s: recovery finished\n",
 377                        obd2cli_tgt(imp->imp_obd));
 378         }
 379         EXIT;
 380
 381 out:
 382         return rc;
 383 }
 384 EXPORT_SYMBOL(ptlrpc_recover_import);
 385
 386 int ptlrpc_import_in_recovery(struct obd_import *imp)
 387 {
 388         int in_recovery = 1;
 389         spin_lock(&imp->imp_lock);
 390         if (imp->imp_state == LUSTRE_IMP_FULL ||
 391             imp->imp_state == LUSTRE_IMP_CLOSED ||
 392             imp->imp_state == LUSTRE_IMP_DISCON)
 393                 in_recovery = 0;
 394         spin_unlock(&imp->imp_lock);
 395         return in_recovery;
 396 }