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, 2016, 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 <linux/list.h>
  39 #include <libcfs/libcfs.h>
  40 #include <obd_support.h>
  41 #include <lustre_ha.h>
  42 #include <lustre_net.h>
  43 #include <lustre_import.h>
  44 #include <lustre_export.h>
  45 #include <obd.h>
  46 #include <obd_class.h>
  47
  48 #include "ptlrpc_internal.h"
  49
  50 /**
  51  * Start recovery on disconnected import.
  52  * This is done by just attempting a connect
  53  */
  54 void ptlrpc_initiate_recovery(struct obd_import *imp)
  55 {
  56         ENTRY;
  57
  58         CDEBUG(D_HA, "%s: starting recovery\n", obd2cli_tgt(imp->imp_obd));
  59         ptlrpc_connect_import(imp);
  60
  61         EXIT;
  62 }
  63
  64 /**
  65  * Identify what request from replay list needs to be replayed next
  66  * (based on what we have already replayed) and send it to server.
  67  */
  68 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
  69 {
  70         int rc = 0;
  71         struct list_head *tmp, *pos;
  72         struct ptlrpc_request *req = NULL;
  73         __u64 last_transno;
  74         ENTRY;
  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
  86         CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
  87                imp, obd2cli_tgt(imp->imp_obd),
  88                imp->imp_peer_committed_transno, last_transno);
  89
  90         /* Replay all the committed open requests on committed_list first */
  91         if (!list_empty(&imp->imp_committed_list)) {
  92                 tmp = imp->imp_committed_list.prev;
  93                 req = list_entry(tmp, struct ptlrpc_request,
  94                                      rq_replay_list);
  95
  96                 /* The last request on committed_list hasn't been replayed */
  97                 if (req->rq_transno > last_transno) {
  98                         if (!imp->imp_resend_replay ||
  99                             imp->imp_replay_cursor == &imp->imp_committed_list)
 100                                 imp->imp_replay_cursor =
 101                                         imp->imp_replay_cursor->next;
 102
 103                         while (imp->imp_replay_cursor !=
 104                                &imp->imp_committed_list) {
 105                                 req = list_entry(imp->imp_replay_cursor,
 106                                                      struct ptlrpc_request,
 107                                                      rq_replay_list);
 108                                 if (req->rq_transno > last_transno)
 109                                         break;
 110
 111                                 req = NULL;
 112                                 LASSERT(!list_empty(imp->imp_replay_cursor));
 113                                 imp->imp_replay_cursor =
 114                                         imp->imp_replay_cursor->next;
 115                         }
 116                 } else {
 117                         /* All requests on committed_list have been replayed */
 118                         imp->imp_replay_cursor = &imp->imp_committed_list;
 119                         req = NULL;
 120                 }
 121         }
 122
 123         /* All the requests in committed list have been replayed, let's replay
 124          * the imp_replay_list */
 125         if (req == NULL) {
 126                 list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
 127                         req = list_entry(tmp, struct ptlrpc_request,
 128                                              rq_replay_list);
 129
 130                         if (req->rq_transno > last_transno)
 131                                 break;
 132                         req = NULL;
 133                 }
 134         }
 135
 136         /* If need to resend the last sent transno (because a reconnect
 137          * has occurred), then stop on the matching req and send it again.
 138          * If, however, the last sent transno has been committed then we
 139          * continue replay from the next request. */
 140         if (req != NULL && imp->imp_resend_replay)
 141                 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 142
 143         /* The resend replay request may have been removed from the
 144          * unreplied list. */
 145         if (req != NULL && imp->imp_resend_replay &&
 146             list_empty(&req->rq_unreplied_list)) {
 147                 ptlrpc_add_unreplied(req);
 148                 imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
 149         }
 150
 151         imp->imp_resend_replay = 0;
 152         spin_unlock(&imp->imp_lock);
 153
 154         if (req != NULL) {
 155                 /* The request should have been added back in unreplied list
 156                  * by ptlrpc_prepare_replay(). */
 157                 LASSERT(!list_empty(&req->rq_unreplied_list));
 158
 159                 rc = ptlrpc_replay_req(req);
 160                 if (rc) {
 161                         CERROR("recovery replay error %d for req "
 162                                "%llu\n", rc, req->rq_xid);
 163                         RETURN(rc);
 164                 }
 165                 *inflight = 1;
 166         }
 167         RETURN(rc);
 168 }
 169
 170 /**
 171  * Schedule resending of request on sending_list. This is done after
 172  * we completed replaying of requests and locks.
 173  */
 174 int ptlrpc_resend(struct obd_import *imp)
 175 {
 176         struct ptlrpc_request *req, *next;
 177
 178         ENTRY;
 179
 180         /* As long as we're in recovery, nothing should be added to the sending
 181          * list, so we don't need to hold the lock during this iteration and
 182          * resend process.
 183          */
 184         /* Well... what if lctl recover is called twice at the same time?
 185          */
 186         spin_lock(&imp->imp_lock);
 187         if (imp->imp_state != LUSTRE_IMP_RECOVER) {
 188                 spin_unlock(&imp->imp_lock);
 189                 RETURN(-1);
 190         }
 191
 192         list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
 193                 LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
 194                          "req %p bad\n", req);
 195                 LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
 196
 197                 /* If the request is allowed to be sent during replay and it
 198                  * is not timeout yet, then it does not need to be resent. */
 199                 if (!ptlrpc_no_resend(req) &&
 200                     (req->rq_timedout || !req->rq_allow_replay))
 201                         ptlrpc_resend_req(req);
 202         }
 203         spin_unlock(&imp->imp_lock);
 204
 205         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT, 2);
 206         RETURN(0);
 207 }
 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
 228 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 229 {
 230         struct obd_import *imp = failed_req->rq_import;
 231         ENTRY;
 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, "
 241                                "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         spin_lock(&failed_req->rq_lock);
 255         if (!failed_req->rq_no_resend)
 256                 failed_req->rq_resend = 1;
 257         spin_unlock(&failed_req->rq_lock);
 258
 259         EXIT;
 260 }
 261
 262 /**
 263  * Administratively active/deactive a client.
 264  * This should only be called by the ioctl interface, currently
 265  *  - the lctl deactivate and activate commands
 266  *  - echo 0/1 >> /proc/osc/XXX/active
 267  *  - client umount -f (ll_umount_begin)
 268  */
 269 int ptlrpc_set_import_active(struct obd_import *imp, int active)
 270 {
 271         struct obd_device *obd = imp->imp_obd;
 272         int rc = 0;
 273
 274         ENTRY;
 275         LASSERT(obd);
 276
 277         /* When deactivating, mark import invalid, and abort in-flight
 278          * requests. */
 279         if (!active) {
 280                 LCONSOLE_WARN("setting import %s INACTIVE by administrator "
 281                               "request\n", obd2cli_tgt(imp->imp_obd));
 282
 283                 /* set before invalidate to avoid messages about imp_inval
 284                  * set without imp_deactive in ptlrpc_import_delay_req */
 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         ENTRY;
 316
 317         spin_lock(&imp->imp_lock);
 318         if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
 319             atomic_read(&imp->imp_inval_count))
 320                 rc = -EINVAL;
 321         spin_unlock(&imp->imp_lock);
 322         if (rc)
 323                 GOTO(out, rc);
 324
 325         /* force import to be disconnected. */
 326         ptlrpc_set_import_discon(imp, 0);
 327
 328         if (new_uuid) {
 329                 struct obd_uuid uuid;
 330
 331                 /* intruct import to use new uuid */
 332                 obd_str2uuid(&uuid, new_uuid);
 333                 rc = import_set_conn_priority(imp, &uuid);
 334                 if (rc)
 335                         GOTO(out, rc);
 336         }
 337
 338         /* Check if reconnect is already in progress */
 339         spin_lock(&imp->imp_lock);
 340         if (imp->imp_state != LUSTRE_IMP_DISCON) {
 341                 imp->imp_force_verify = 1;
 342                 rc = -EALREADY;
 343         }
 344         spin_unlock(&imp->imp_lock);
 345         if (rc)
 346                 GOTO(out, rc);
 347
 348         rc = ptlrpc_connect_import(imp);
 349         if (rc)
 350                 GOTO(out, rc);
 351
 352         if (!async) {
 353                 struct l_wait_info lwi;
 354                 int secs = cfs_time_seconds(obd_timeout);
 355
 356                 CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 357                        obd2cli_tgt(imp->imp_obd), secs);
 358
 359                 lwi = LWI_TIMEOUT(secs, NULL, NULL);
 360                 rc = l_wait_event(imp->imp_recovery_waitq,
 361                                   !ptlrpc_import_in_recovery(imp), &lwi);
 362                 CDEBUG(D_HA, "%s: recovery finished\n",
 363                        obd2cli_tgt(imp->imp_obd));
 364         }
 365         EXIT;
 366
 367 out:
 368         return rc;
 369 }
 370 EXPORT_SYMBOL(ptlrpc_recover_import);
 371
 372 int ptlrpc_import_in_recovery(struct obd_import *imp)
 373 {
 374         int in_recovery = 1;
 375
 376         spin_lock(&imp->imp_lock);
 377         if (imp->imp_state == LUSTRE_IMP_FULL ||
 378             imp->imp_state == LUSTRE_IMP_CLOSED ||
 379             imp->imp_state == LUSTRE_IMP_DISCON ||
 380             imp->imp_obd->obd_no_recov)
 381                 in_recovery = 0;
 382         spin_unlock(&imp->imp_lock);
 383
 384         return in_recovery;
 385 }