lustre/ptlrpc/pinger.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) 2003, 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/pinger.c
  33  *
  34  * Portal-RPC reconnection and replay operations, for use in recovery.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  39 #include <linux/kthread.h>
  40 #include <linux/workqueue.h>
  41 #include <obd_support.h>
  42 #include <obd_class.h>
  43 #include "ptlrpc_internal.h"
  44
  45 static int suppress_pings;
  46 module_param(suppress_pings, int, 0644);
  47 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
  48
  49 struct mutex pinger_mutex;
  50 static struct list_head pinger_imports =
  51                 LIST_HEAD_INIT(pinger_imports);
  52 static struct list_head timeout_list =
  53                 LIST_HEAD_INIT(timeout_list);
  54
  55 int ptlrpc_pinger_suppress_pings()
  56 {
  57         return suppress_pings;
  58 }
  59 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  60
  61 struct ptlrpc_request *
  62 ptlrpc_prep_ping(struct obd_import *imp)
  63 {
  64         struct ptlrpc_request *req;
  65
  66         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  67                                         LUSTRE_OBD_VERSION, OBD_PING);
  68         if (req) {
  69                 ptlrpc_request_set_replen(req);
  70                 req->rq_no_resend = req->rq_no_delay = 1;
  71         }
  72         return req;
  73 }
  74
  75 int ptlrpc_obd_ping(struct obd_device *obd)
  76 {
  77         int rc;
  78         struct ptlrpc_request *req;
  79         ENTRY;
  80
  81         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  82         if (req == NULL)
  83                 RETURN(-ENOMEM);
  84
  85         req->rq_send_state = LUSTRE_IMP_FULL;
  86
  87         rc = ptlrpc_queue_wait(req);
  88
  89         ptlrpc_req_finished(req);
  90
  91         RETURN(rc);
  92 }
  93 EXPORT_SYMBOL(ptlrpc_obd_ping);
  94
  95 static bool ptlrpc_check_import_is_idle(struct obd_import *imp)
  96 {
  97         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
  98         time64_t now;
  99
 100         if (!imp->imp_idle_timeout)
 101                 return false;
 102
 103         /* 4 comes from:
 104          *  - client_obd_setup() - hashed import
 105          *  - ptlrpcd_alloc_work()
 106          *  - ptlrpcd_alloc_work()
 107          *  - ptlrpc_pinger_add_import
 108          */
 109         if (atomic_read(&imp->imp_refcount) > 4)
 110                 return false;
 111
 112         /* any lock increases ns_bref being a resource holder */
 113         if (ns && atomic_read(&ns->ns_bref) > 0)
 114                 return false;
 115
 116         now = ktime_get_real_seconds();
 117         if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
 118                 return false;
 119
 120         return true;
 121 }
 122
 123 static int ptlrpc_ping(struct obd_import *imp)
 124 {
 125         struct ptlrpc_request   *req;
 126         ENTRY;
 127
 128         if (ptlrpc_check_import_is_idle(imp))
 129                 RETURN(ptlrpc_disconnect_and_idle_import(imp));
 130
 131         req = ptlrpc_prep_ping(imp);
 132         if (req == NULL) {
 133                 CERROR("OOM trying to ping %s->%s\n",
 134                        imp->imp_obd->obd_uuid.uuid,
 135                        obd2cli_tgt(imp->imp_obd));
 136                 RETURN(-ENOMEM);
 137         }
 138
 139         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 140                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 141         ptlrpcd_add_req(req);
 142
 143         RETURN(0);
 144 }
 145
 146 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 147 {
 148 #ifdef ENABLE_PINGER
 149         time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 150
 151         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 152                 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
 153                                        AT_OFF ? 0 :
 154                                        at_get(&imp->imp_at.iat_net_latency));
 155                 time = min(time, dtime);
 156         }
 157         imp->imp_next_ping = ktime_get_seconds() + time;
 158 #endif /* ENABLE_PINGER */
 159 }
 160
 161 void ptlrpc_ping_import_soon(struct obd_import *imp)
 162 {
 163         imp->imp_next_ping = ktime_get_seconds();
 164 }
 165
 166 static inline int imp_is_deactive(struct obd_import *imp)
 167 {
 168         return (imp->imp_deactive ||
 169                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 170 }
 171
 172 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
 173 {
 174         if (imp->imp_server_timeout)
 175                 return ktime_get_seconds() + (obd_timeout >> 1);
 176         else
 177                 return ktime_get_seconds() + obd_timeout;
 178 }
 179
 180 static time64_t pinger_check_timeout(time64_t time)
 181 {
 182         struct timeout_item *item;
 183         time64_t timeout = PING_INTERVAL;
 184
 185         /* This list is sorted in increasing timeout order */
 186         mutex_lock(&pinger_mutex);
 187         list_for_each_entry(item, &timeout_list, ti_chain) {
 188                 time64_t ti_timeout = item->ti_timeout;
 189
 190                 if (timeout > ti_timeout)
 191                         timeout = ti_timeout;
 192                 break;
 193         }
 194         mutex_unlock(&pinger_mutex);
 195
 196         return time + timeout - ktime_get_seconds();
 197 }
 198
 199 static bool ir_up;
 200
 201 void ptlrpc_pinger_ir_up(void)
 202 {
 203         CDEBUG(D_HA, "IR up\n");
 204         ir_up = true;
 205 }
 206 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 207
 208 void ptlrpc_pinger_ir_down(void)
 209 {
 210         CDEBUG(D_HA, "IR down\n");
 211         ir_up = false;
 212 }
 213 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 214
 215 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 216                                          time64_t this_ping)
 217 {
 218         int level;
 219         int force;
 220         int force_next;
 221         int suppress;
 222
 223         spin_lock(&imp->imp_lock);
 224
 225         level = imp->imp_state;
 226         force = imp->imp_force_verify;
 227         force_next = imp->imp_force_next_verify;
 228         /*
 229          * This will be used below only if the import is "FULL".
 230          */
 231         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 232
 233         imp->imp_force_verify = 0;
 234
 235         if (imp->imp_next_ping - 5 >= this_ping && !force) {
 236                 spin_unlock(&imp->imp_lock);
 237                 return;
 238         }
 239
 240         imp->imp_force_next_verify = 0;
 241
 242         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 243                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 244                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 245                ptlrpc_import_state_name(level), level, force, force_next,
 246                imp->imp_deactive, imp->imp_pingable, suppress);
 247
 248         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 249                 /* wait for a while before trying recovery again */
 250                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 251                 spin_unlock(&imp->imp_lock);
 252                 if (!imp->imp_no_pinger_recover)
 253                         ptlrpc_initiate_recovery(imp);
 254         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
 255                    imp_is_deactive(imp)) {
 256                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 257                        "or recovery disabled: %s)\n",
 258                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 259                        ptlrpc_import_state_name(level));
 260                 if (force)
 261                         imp->imp_force_verify = 1;
 262                 spin_unlock(&imp->imp_lock);
 263         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 264                 spin_unlock(&imp->imp_lock);
 265                 ptlrpc_ping(imp);
 266         }
 267 }
 268
 269 static struct workqueue_struct *pinger_wq;
 270 static void ptlrpc_pinger_main(struct work_struct *ws);
 271 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
 272
 273 static void ptlrpc_pinger_main(struct work_struct *ws)
 274 {
 275         time64_t this_ping = ktime_get_seconds();
 276         time64_t time_to_next_wake;
 277         struct timeout_item *item;
 278         struct obd_import *imp;
 279         struct list_head *iter;
 280
 281         do {
 282                 mutex_lock(&pinger_mutex);
 283                 list_for_each_entry(item, &timeout_list, ti_chain)
 284                         item->ti_cb(item, item->ti_cb_data);
 285
 286                 list_for_each(iter, &pinger_imports) {
 287                         imp = list_entry(iter, struct obd_import,
 288                                          imp_pinger_chain);
 289
 290                         ptlrpc_pinger_process_import(imp, this_ping);
 291                         /* obd_timeout might have changed */
 292                         if (imp->imp_pingable && imp->imp_next_ping &&
 293                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 294                                 ptlrpc_update_next_ping(imp, 0);
 295                 }
 296                 mutex_unlock(&pinger_mutex);
 297                 /* update memory usage info */
 298                 obd_update_maxusage();
 299
 300                 /* Wait until the next ping time, or until we're stopped. */
 301                 time_to_next_wake = pinger_check_timeout(this_ping);
 302                 /* The ping sent by ptlrpc_send_rpc may get sent out
 303                  * say .01 second after this.
 304                  * ptlrpc_pinger_sending_on_import will then set the
 305                  * next ping time to next_ping + .01 sec, which means
 306                  * we will SKIP the next ping at next_ping, and the
 307                  * ping will get sent 2 timeouts from now!  Beware. */
 308                 CDEBUG(D_INFO, "next wakeup in %lld (%lld)\n",
 309                        time_to_next_wake, this_ping + PING_INTERVAL);
 310         } while (time_to_next_wake <= 0);
 311
 312         queue_delayed_work(pinger_wq, &ping_work,
 313                            cfs_time_seconds(max(time_to_next_wake, 1LL)));
 314 }
 315
 316 int ptlrpc_start_pinger(void)
 317 {
 318 #ifdef ENABLE_PINGER
 319         if (pinger_wq)
 320                 return -EALREADY;
 321
 322         pinger_wq = alloc_workqueue("ptlrpc_pinger", 0, 1);
 323         if (!pinger_wq) {
 324                 CERROR("cannot start pinger workqueue\n");
 325                 return -ENOMEM;
 326         }
 327
 328         queue_delayed_work(pinger_wq, &ping_work, 0);
 329
 330         if (suppress_pings)
 331                 CWARN("Pings will be suppressed at the request of the "
 332                       "administrator.  The configuration shall meet the "
 333                       "additional requirements described in the manual.  "
 334                       "(Search for the \"suppress_pings\" kernel module "
 335                       "parameter.)\n");
 336 #endif
 337         return 0;
 338 }
 339
 340 int ptlrpc_pinger_remove_timeouts(void);
 341
 342 int ptlrpc_stop_pinger(void)
 343 {
 344 #ifdef ENABLE_PINGER
 345         if (!pinger_wq)
 346                 return -EALREADY;
 347
 348         ptlrpc_pinger_remove_timeouts();
 349
 350         cancel_delayed_work_sync(&ping_work);
 351         destroy_workqueue(pinger_wq);
 352         pinger_wq = NULL;
 353 #endif
 354         return 0;
 355 }
 356
 357 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 358 {
 359         ptlrpc_update_next_ping(imp, 0);
 360 }
 361
 362 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 363 {
 364         ptlrpc_update_next_ping(imp, 1);
 365         assert_spin_locked(&imp->imp_lock);
 366         /*
 367          * Avoid reading stale imp_connect_data.  When not sure if pings are
 368          * expected or not on next connection, we assume they are not and force
 369          * one anyway to guarantee the chance of updating
 370          * imp_peer_committed_transno.
 371          */
 372         if (imp->imp_state != LUSTRE_IMP_FULL ||
 373             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 374                 imp->imp_force_next_verify = 1;
 375 }
 376
 377 int ptlrpc_pinger_add_import(struct obd_import *imp)
 378 {
 379         ENTRY;
 380         if (!list_empty(&imp->imp_pinger_chain))
 381                 RETURN(-EALREADY);
 382
 383         mutex_lock(&pinger_mutex);
 384         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 385                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 386         /* if we add to pinger we want recovery on this import */
 387         imp->imp_obd->obd_no_recov = 0;
 388         ptlrpc_update_next_ping(imp, 0);
 389         /* XXX sort, blah blah */
 390         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 391         class_import_get(imp);
 392
 393         ptlrpc_pinger_wake_up();
 394         mutex_unlock(&pinger_mutex);
 395
 396         RETURN(0);
 397 }
 398 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 399
 400 int ptlrpc_pinger_del_import(struct obd_import *imp)
 401 {
 402         ENTRY;
 403
 404         if (list_empty(&imp->imp_pinger_chain))
 405                 RETURN(-ENOENT);
 406
 407         mutex_lock(&pinger_mutex);
 408         list_del_init(&imp->imp_pinger_chain);
 409         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 410                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 411         /* if we remove from pinger we don't want recovery on this import */
 412         imp->imp_obd->obd_no_recov = 1;
 413         class_import_put(imp);
 414         mutex_unlock(&pinger_mutex);
 415         RETURN(0);
 416 }
 417 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 418
 419 /**
 420  * Register a timeout callback to the pinger list, and the callback will
 421  * be called when timeout happens.
 422  */
 423 static struct timeout_item *ptlrpc_new_timeout(time64_t time,
 424                                                enum timeout_event event,
 425                                                timeout_cb_t cb, void *data)
 426 {
 427         struct timeout_item *ti;
 428
 429         OBD_ALLOC_PTR(ti);
 430         if (!ti)
 431                 return(NULL);
 432
 433         INIT_LIST_HEAD(&ti->ti_obd_list);
 434         INIT_LIST_HEAD(&ti->ti_chain);
 435         ti->ti_timeout = time;
 436         ti->ti_event = event;
 437         ti->ti_cb = cb;
 438         ti->ti_cb_data = data;
 439
 440         return ti;
 441 }
 442
 443 /**
 444  * Register timeout event on the the pinger thread.
 445  * Note: the timeout list is an sorted list with increased timeout value.
 446  */
 447 static struct timeout_item*
 448 ptlrpc_pinger_register_timeout(time64_t time, enum timeout_event event,
 449                                timeout_cb_t cb, void *data)
 450 {
 451         struct timeout_item *item, *tmp;
 452
 453         LASSERT(mutex_is_locked(&pinger_mutex));
 454
 455         list_for_each_entry(item, &timeout_list, ti_chain)
 456                 if (item->ti_event == event)
 457                         goto out;
 458
 459         item = ptlrpc_new_timeout(time, event, cb, data);
 460         if (item) {
 461                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 462                         if (tmp->ti_timeout < time) {
 463                                 list_add(&item->ti_chain, &tmp->ti_chain);
 464                                 goto out;
 465                         }
 466                 }
 467                 list_add(&item->ti_chain, &timeout_list);
 468         }
 469 out:
 470         return item;
 471 }
 472
 473 /* Add a client_obd to the timeout event list, when timeout(@time)
 474  * happens, the callback(@cb) will be called.
 475  */
 476 int ptlrpc_add_timeout_client(time64_t time, enum timeout_event event,
 477                               timeout_cb_t cb, void *data,
 478                               struct list_head *obd_list)
 479 {
 480         struct timeout_item *ti;
 481
 482         mutex_lock(&pinger_mutex);
 483         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 484         if (!ti) {
 485                 mutex_unlock(&pinger_mutex);
 486                 return (-EINVAL);
 487         }
 488         list_add(obd_list, &ti->ti_obd_list);
 489         mutex_unlock(&pinger_mutex);
 490         return 0;
 491 }
 492 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 493
 494 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 495                               enum timeout_event event)
 496 {
 497         struct timeout_item *ti = NULL, *item;
 498
 499         if (list_empty(obd_list))
 500                 return 0;
 501         mutex_lock(&pinger_mutex);
 502         list_del_init(obd_list);
 503         /**
 504          * If there are no obd attached to the timeout event
 505          * list, remove this timeout event from the pinger
 506          */
 507         list_for_each_entry(item, &timeout_list, ti_chain) {
 508                 if (item->ti_event == event) {
 509                         ti = item;
 510                         break;
 511                 }
 512         }
 513         LASSERTF(ti != NULL, "ti is NULL !\n");
 514         if (list_empty(&ti->ti_obd_list)) {
 515                 list_del(&ti->ti_chain);
 516                 OBD_FREE_PTR(ti);
 517         }
 518         mutex_unlock(&pinger_mutex);
 519         return 0;
 520 }
 521 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 522
 523 int ptlrpc_pinger_remove_timeouts(void)
 524 {
 525         struct timeout_item *item, *tmp;
 526
 527         mutex_lock(&pinger_mutex);
 528         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 529                 LASSERT(list_empty(&item->ti_obd_list));
 530                 list_del(&item->ti_chain);
 531                 OBD_FREE_PTR(item);
 532         }
 533         mutex_unlock(&pinger_mutex);
 534         return 0;
 535 }
 536
 537 void ptlrpc_pinger_wake_up()
 538 {
 539 #ifdef ENABLE_PINGER
 540         mod_delayed_work(pinger_wq, &ping_work, 0);
 541 #endif
 542 }
 543
 544 /* Ping evictor thread */
 545 #define PET_READY     1
 546 #define PET_TERMINATE 2
 547
 548 static int               pet_refcount = 0;
 549 static int               pet_state;
 550 static wait_queue_head_t pet_waitq;
 551 static struct list_head  pet_list;
 552 static DEFINE_SPINLOCK(pet_lock);
 553
 554 int ping_evictor_wake(struct obd_export *exp)
 555 {
 556         struct obd_device *obd;
 557
 558         spin_lock(&pet_lock);
 559         if (pet_state != PET_READY) {
 560                 /* eventually the new obd will call here again. */
 561                 spin_unlock(&pet_lock);
 562                 return 1;
 563         }
 564
 565         obd = class_exp2obd(exp);
 566         if (list_empty(&obd->obd_evict_list)) {
 567                 class_incref(obd, "evictor", obd);
 568                 list_add(&obd->obd_evict_list, &pet_list);
 569         }
 570         spin_unlock(&pet_lock);
 571
 572         wake_up(&pet_waitq);
 573         return 0;
 574 }
 575
 576 static int ping_evictor_main(void *arg)
 577 {
 578         struct obd_device *obd;
 579         struct obd_export *exp;
 580         struct l_wait_info lwi = { 0 };
 581         time64_t expire_time;
 582
 583         ENTRY;
 584         unshare_fs_struct();
 585
 586         CDEBUG(D_HA, "Starting Ping Evictor\n");
 587         pet_state = PET_READY;
 588         while (1) {
 589                 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
 590                             (pet_state == PET_TERMINATE), &lwi);
 591
 592                 /* loop until all obd's will be removed */
 593                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 594                         break;
 595
 596                 /* we only get here if pet_exp != NULL, and the end of this
 597                  * loop is the only place which sets it NULL again, so lock
 598                  * is not strictly necessary. */
 599                 spin_lock(&pet_lock);
 600                 obd = list_entry(pet_list.next, struct obd_device,
 601                                  obd_evict_list);
 602                 spin_unlock(&pet_lock);
 603
 604                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 605
 606                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 607                        obd->obd_name, expire_time);
 608
 609                 /* Exports can't be deleted out of the list while we hold
 610                  * the obd lock (class_unlink_export), which means we can't
 611                  * lose the last ref on the export.  If they've already been
 612                  * removed from the list, we won't find them here. */
 613                 spin_lock(&obd->obd_dev_lock);
 614                 while (!list_empty(&obd->obd_exports_timed)) {
 615                         exp = list_entry(obd->obd_exports_timed.next,
 616                                          struct obd_export,
 617                                          exp_obd_chain_timed);
 618                         if (expire_time > exp->exp_last_request_time) {
 619                                 class_export_get(exp);
 620                                 spin_unlock(&obd->obd_dev_lock);
 621                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 622                                               " (at %s) in %lld seconds. I think"
 623                                               " it's dead, and I am evicting"
 624                                               " it. exp %p, cur %lld expire %lld"
 625                                               " last %lld\n",
 626                                               obd->obd_name,
 627                                               obd_uuid2str(&exp->exp_client_uuid),
 628                                               obd_export_nid2str(exp),
 629                                               ktime_get_real_seconds() -
 630                                               exp->exp_last_request_time,
 631                                               exp, ktime_get_real_seconds(),
 632                                               expire_time,
 633                                               exp->exp_last_request_time);
 634                                 CDEBUG(D_HA, "Last request was at %lld\n",
 635                                        exp->exp_last_request_time);
 636                                 class_fail_export(exp);
 637                                 class_export_put(exp);
 638                                 spin_lock(&obd->obd_dev_lock);
 639                         } else {
 640                                 /* List is sorted, so everyone below is ok */
 641                                 break;
 642                         }
 643                 }
 644                 spin_unlock(&obd->obd_dev_lock);
 645
 646                 spin_lock(&pet_lock);
 647                 list_del_init(&obd->obd_evict_list);
 648                 spin_unlock(&pet_lock);
 649
 650                 class_decref(obd, "evictor", obd);
 651         }
 652         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 653
 654         RETURN(0);
 655 }
 656
 657 void ping_evictor_start(void)
 658 {
 659         struct task_struct *task;
 660
 661         if (++pet_refcount > 1)
 662                 return;
 663
 664         INIT_LIST_HEAD(&pet_list);
 665         init_waitqueue_head(&pet_waitq);
 666
 667         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 668         if (IS_ERR(task)) {
 669                 pet_refcount--;
 670                 CERROR("Cannot start ping evictor thread: %ld\n",
 671                         PTR_ERR(task));
 672         }
 673 }
 674 EXPORT_SYMBOL(ping_evictor_start);
 675
 676 void ping_evictor_stop(void)
 677 {
 678         if (--pet_refcount > 0)
 679                 return;
 680
 681         pet_state = PET_TERMINATE;
 682         wake_up(&pet_waitq);
 683 }
 684 EXPORT_SYMBOL(ping_evictor_stop);