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.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) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, 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/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42
  43 #include <linux/kthread.h>
  44 #include <obd_support.h>
  45 #include <obd_class.h>
  46 #include "ptlrpc_internal.h"
  47
  48 static int suppress_pings;
  49 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  50
  51 struct mutex pinger_mutex;
  52 static struct list_head pinger_imports =
  53                 LIST_HEAD_INIT(pinger_imports);
  54 static struct list_head timeout_list =
  55                 LIST_HEAD_INIT(timeout_list);
  56
  57 int ptlrpc_pinger_suppress_pings()
  58 {
  59         return suppress_pings;
  60 }
  61 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  62
  63 struct ptlrpc_request *
  64 ptlrpc_prep_ping(struct obd_import *imp)
  65 {
  66         struct ptlrpc_request *req;
  67
  68         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  69                                         LUSTRE_OBD_VERSION, OBD_PING);
  70         if (req) {
  71                 ptlrpc_request_set_replen(req);
  72                 req->rq_no_resend = req->rq_no_delay = 1;
  73         }
  74         return req;
  75 }
  76
  77 int ptlrpc_obd_ping(struct obd_device *obd)
  78 {
  79         int rc;
  80         struct ptlrpc_request *req;
  81         ENTRY;
  82
  83         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  84         if (req == NULL)
  85                 RETURN(-ENOMEM);
  86
  87         req->rq_send_state = LUSTRE_IMP_FULL;
  88
  89         rc = ptlrpc_queue_wait(req);
  90
  91         ptlrpc_req_finished(req);
  92
  93         RETURN(rc);
  94 }
  95 EXPORT_SYMBOL(ptlrpc_obd_ping);
  96
  97 static int ptlrpc_ping(struct obd_import *imp)
  98 {
  99         struct ptlrpc_request   *req;
 100         ENTRY;
 101
 102         req = ptlrpc_prep_ping(imp);
 103         if (req == NULL) {
 104                 CERROR("OOM trying to ping %s->%s\n",
 105                        imp->imp_obd->obd_uuid.uuid,
 106                        obd2cli_tgt(imp->imp_obd));
 107                 RETURN(-ENOMEM);
 108         }
 109
 110         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 111                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 112         ptlrpcd_add_req(req);
 113
 114         RETURN(0);
 115 }
 116
 117 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 118 {
 119 #ifdef ENABLE_PINGER
 120         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 121         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 122                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 123                                   AT_OFF ? 0 :
 124                                   at_get(&imp->imp_at.iat_net_latency));
 125                 time = min(time, dtime);
 126         }
 127         imp->imp_next_ping = cfs_time_shift(time);
 128 #endif /* ENABLE_PINGER */
 129 }
 130
 131 void ptlrpc_ping_import_soon(struct obd_import *imp)
 132 {
 133         imp->imp_next_ping = cfs_time_current();
 134 }
 135
 136 static inline int imp_is_deactive(struct obd_import *imp)
 137 {
 138         return (imp->imp_deactive ||
 139                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 140 }
 141
 142 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 143 {
 144         if (imp->imp_server_timeout)
 145                 return cfs_time_shift(obd_timeout / 2);
 146         else
 147                 return cfs_time_shift(obd_timeout);
 148 }
 149
 150 static cfs_duration_t pinger_check_timeout(cfs_time_t time)
 151 {
 152         struct timeout_item *item;
 153         cfs_time_t timeout = PING_INTERVAL;
 154
 155         /* This list is sorted in increasing timeout order */
 156         mutex_lock(&pinger_mutex);
 157         list_for_each_entry(item, &timeout_list, ti_chain) {
 158                 int ti_timeout = item->ti_timeout;
 159                 if (timeout > ti_timeout)
 160                         timeout = ti_timeout;
 161                 break;
 162         }
 163         mutex_unlock(&pinger_mutex);
 164
 165         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 166                                          cfs_time_current());
 167 }
 168
 169
 170 static bool ir_up;
 171
 172 void ptlrpc_pinger_ir_up(void)
 173 {
 174         CDEBUG(D_HA, "IR up\n");
 175         ir_up = true;
 176 }
 177 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 178
 179 void ptlrpc_pinger_ir_down(void)
 180 {
 181         CDEBUG(D_HA, "IR down\n");
 182         ir_up = false;
 183 }
 184 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 185
 186 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 187                                          unsigned long this_ping)
 188 {
 189         int level;
 190         int force;
 191         int force_next;
 192         int suppress;
 193
 194         spin_lock(&imp->imp_lock);
 195
 196         level = imp->imp_state;
 197         force = imp->imp_force_verify;
 198         force_next = imp->imp_force_next_verify;
 199         /*
 200          * This will be used below only if the import is "FULL".
 201          */
 202         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 203
 204         imp->imp_force_verify = 0;
 205
 206         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 207             !force) {
 208                 spin_unlock(&imp->imp_lock);
 209                 return;
 210         }
 211
 212         imp->imp_force_next_verify = 0;
 213
 214         spin_unlock(&imp->imp_lock);
 215
 216         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 217                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 218                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 219                ptlrpc_import_state_name(level), level, force, force_next,
 220                imp->imp_deactive, imp->imp_pingable, suppress);
 221
 222         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 223                 /* wait for a while before trying recovery again */
 224                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 225                 if (!imp->imp_no_pinger_recover)
 226                         ptlrpc_initiate_recovery(imp);
 227         } else if (level != LUSTRE_IMP_FULL ||
 228                    imp->imp_obd->obd_no_recov ||
 229                    imp_is_deactive(imp)) {
 230                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 231                        "or recovery disabled: %s)\n",
 232                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 233                        ptlrpc_import_state_name(level));
 234                 if (force) {
 235                         spin_lock(&imp->imp_lock);
 236                         imp->imp_force_verify = 1;
 237                         spin_unlock(&imp->imp_lock);
 238                 }
 239         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 240                 ptlrpc_ping(imp);
 241         }
 242 }
 243
 244 static int ptlrpc_pinger_main(void *arg)
 245 {
 246         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 247         ENTRY;
 248
 249         /* Record that the thread is running */
 250         thread_set_flags(thread, SVC_RUNNING);
 251         wake_up(&thread->t_ctl_waitq);
 252
 253         /* And now, loop forever, pinging as needed. */
 254         while (1) {
 255                 cfs_time_t this_ping = cfs_time_current();
 256                 struct l_wait_info lwi;
 257                 cfs_duration_t time_to_next_wake;
 258                 struct timeout_item *item;
 259                 struct list_head *iter;
 260
 261                 mutex_lock(&pinger_mutex);
 262                 list_for_each_entry(item, &timeout_list, ti_chain)
 263                         item->ti_cb(item, item->ti_cb_data);
 264
 265                 list_for_each(iter, &pinger_imports) {
 266                         struct obd_import *imp = list_entry(iter,
 267                                                             struct obd_import,
 268                                                             imp_pinger_chain);
 269
 270                         ptlrpc_pinger_process_import(imp, this_ping);
 271                         /* obd_timeout might have changed */
 272                         if (imp->imp_pingable && imp->imp_next_ping &&
 273                             cfs_time_after(imp->imp_next_ping,
 274                                            cfs_time_add(this_ping,
 275                                                         cfs_time_seconds(PING_INTERVAL))))
 276                                 ptlrpc_update_next_ping(imp, 0);
 277                 }
 278                 mutex_unlock(&pinger_mutex);
 279                 /* update memory usage info */
 280                 obd_update_maxusage();
 281
 282                 /* Wait until the next ping time, or until we're stopped. */
 283                 time_to_next_wake = pinger_check_timeout(this_ping);
 284                 /* The ping sent by ptlrpc_send_rpc may get sent out
 285                    say .01 second after this.
 286                    ptlrpc_pinger_sending_on_import will then set the
 287                    next ping time to next_ping + .01 sec, which means
 288                    we will SKIP the next ping at next_ping, and the
 289                    ping will get sent 2 timeouts from now!  Beware. */
 290                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 291                        CFS_TIME_T")\n", time_to_next_wake,
 292                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 293                 if (time_to_next_wake > 0) {
 294                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 295                                                 time_to_next_wake,
 296                                                 cfs_time_seconds(1)),
 297                                           NULL, NULL);
 298                         l_wait_event(thread->t_ctl_waitq,
 299                                      thread_is_stopping(thread) ||
 300                                      thread_is_event(thread),
 301                                      &lwi);
 302                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 303                                 EXIT;
 304                                 break;
 305                         } else {
 306                                 /* woken after adding import to reset timer */
 307                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 308                         }
 309                 }
 310         }
 311
 312         thread_set_flags(thread, SVC_STOPPED);
 313         wake_up(&thread->t_ctl_waitq);
 314
 315         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
 316         return 0;
 317 }
 318
 319 static struct ptlrpc_thread pinger_thread;
 320
 321 int ptlrpc_start_pinger(void)
 322 {
 323         struct l_wait_info lwi = { 0 };
 324         struct task_struct *task;
 325         int rc;
 326 #ifndef ENABLE_PINGER
 327         return 0;
 328 #endif
 329         ENTRY;
 330
 331         if (!thread_is_init(&pinger_thread) &&
 332             !thread_is_stopped(&pinger_thread))
 333                 RETURN(-EALREADY);
 334
 335         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 336
 337         strcpy(pinger_thread.t_name, "ll_ping");
 338
 339         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 340          * just drop the VM and FILES in kthread_run() right away. */
 341         task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
 342                            pinger_thread.t_name);
 343         if (IS_ERR(task)) {
 344                 rc = PTR_ERR(task);
 345                 CERROR("cannot start pinger thread: rc = %d\n", rc);
 346                 RETURN(rc);
 347         }
 348
 349         l_wait_event(pinger_thread.t_ctl_waitq,
 350                      thread_is_running(&pinger_thread), &lwi);
 351
 352         if (suppress_pings)
 353                 CWARN("Pings will be suppressed at the request of the "
 354                       "administrator.  The configuration shall meet the "
 355                       "additional requirements described in the manual.  "
 356                       "(Search for the \"suppress_pings\" kernel module "
 357                       "parameter.)\n");
 358
 359         RETURN(0);
 360 }
 361
 362 int ptlrpc_pinger_remove_timeouts(void);
 363
 364 int ptlrpc_stop_pinger(void)
 365 {
 366         struct l_wait_info lwi = { 0 };
 367 #ifndef ENABLE_PINGER
 368         return 0;
 369 #endif
 370         ENTRY;
 371
 372         if (thread_is_init(&pinger_thread) ||
 373             thread_is_stopped(&pinger_thread))
 374                 RETURN(-EALREADY);
 375
 376         ptlrpc_pinger_remove_timeouts();
 377
 378         thread_set_flags(&pinger_thread, SVC_STOPPING);
 379         wake_up(&pinger_thread.t_ctl_waitq);
 380
 381         l_wait_event(pinger_thread.t_ctl_waitq,
 382                      thread_is_stopped(&pinger_thread), &lwi);
 383         RETURN(0);
 384 }
 385
 386 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 387 {
 388         ptlrpc_update_next_ping(imp, 0);
 389 }
 390
 391 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 392 {
 393         ptlrpc_update_next_ping(imp, 1);
 394         assert_spin_locked(&imp->imp_lock);
 395         /*
 396          * Avoid reading stale imp_connect_data.  When not sure if pings are
 397          * expected or not on next connection, we assume they are not and force
 398          * one anyway to guarantee the chance of updating
 399          * imp_peer_committed_transno.
 400          */
 401         if (imp->imp_state != LUSTRE_IMP_FULL ||
 402             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 403                 imp->imp_force_next_verify = 1;
 404 }
 405
 406 int ptlrpc_pinger_add_import(struct obd_import *imp)
 407 {
 408         ENTRY;
 409         if (!list_empty(&imp->imp_pinger_chain))
 410                 RETURN(-EALREADY);
 411
 412         mutex_lock(&pinger_mutex);
 413         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 414                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 415         /* if we add to pinger we want recovery on this import */
 416         imp->imp_obd->obd_no_recov = 0;
 417         ptlrpc_update_next_ping(imp, 0);
 418         /* XXX sort, blah blah */
 419         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 420         class_import_get(imp);
 421
 422         ptlrpc_pinger_wake_up();
 423         mutex_unlock(&pinger_mutex);
 424
 425         RETURN(0);
 426 }
 427 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 428
 429 int ptlrpc_pinger_del_import(struct obd_import *imp)
 430 {
 431         ENTRY;
 432
 433         if (list_empty(&imp->imp_pinger_chain))
 434                 RETURN(-ENOENT);
 435
 436         mutex_lock(&pinger_mutex);
 437         list_del_init(&imp->imp_pinger_chain);
 438         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 439                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 440         /* if we remove from pinger we don't want recovery on this import */
 441         imp->imp_obd->obd_no_recov = 1;
 442         class_import_put(imp);
 443         mutex_unlock(&pinger_mutex);
 444         RETURN(0);
 445 }
 446 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 447
 448 /**
 449  * Register a timeout callback to the pinger list, and the callback will
 450  * be called when timeout happens.
 451  */
 452 static struct timeout_item *ptlrpc_new_timeout(int time,
 453                                                enum timeout_event event,
 454                                                timeout_cb_t cb, void *data)
 455 {
 456         struct timeout_item *ti;
 457
 458         OBD_ALLOC_PTR(ti);
 459         if (!ti)
 460                 return(NULL);
 461
 462         INIT_LIST_HEAD(&ti->ti_obd_list);
 463         INIT_LIST_HEAD(&ti->ti_chain);
 464         ti->ti_timeout = time;
 465         ti->ti_event = event;
 466         ti->ti_cb = cb;
 467         ti->ti_cb_data = data;
 468
 469         return ti;
 470 }
 471
 472 /**
 473  * Register timeout event on the the pinger thread.
 474  * Note: the timeout list is an sorted list with increased timeout value.
 475  */
 476 static struct timeout_item*
 477 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 478                                timeout_cb_t cb, void *data)
 479 {
 480         struct timeout_item *item, *tmp;
 481
 482         LASSERT(mutex_is_locked(&pinger_mutex));
 483
 484         list_for_each_entry(item, &timeout_list, ti_chain)
 485                 if (item->ti_event == event)
 486                         goto out;
 487
 488         item = ptlrpc_new_timeout(time, event, cb, data);
 489         if (item) {
 490                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 491                         if (tmp->ti_timeout < time) {
 492                                 list_add(&item->ti_chain, &tmp->ti_chain);
 493                                 goto out;
 494                         }
 495                 }
 496                 list_add(&item->ti_chain, &timeout_list);
 497         }
 498 out:
 499         return item;
 500 }
 501
 502 /* Add a client_obd to the timeout event list, when timeout(@time)
 503  * happens, the callback(@cb) will be called.
 504  */
 505 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 506                               timeout_cb_t cb, void *data,
 507                               struct list_head *obd_list)
 508 {
 509         struct timeout_item *ti;
 510
 511         mutex_lock(&pinger_mutex);
 512         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 513         if (!ti) {
 514                 mutex_unlock(&pinger_mutex);
 515                 return (-EINVAL);
 516         }
 517         list_add(obd_list, &ti->ti_obd_list);
 518         mutex_unlock(&pinger_mutex);
 519         return 0;
 520 }
 521 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 522
 523 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 524                               enum timeout_event event)
 525 {
 526         struct timeout_item *ti = NULL, *item;
 527
 528         if (list_empty(obd_list))
 529                 return 0;
 530         mutex_lock(&pinger_mutex);
 531         list_del_init(obd_list);
 532         /**
 533          * If there are no obd attached to the timeout event
 534          * list, remove this timeout event from the pinger
 535          */
 536         list_for_each_entry(item, &timeout_list, ti_chain) {
 537                 if (item->ti_event == event) {
 538                         ti = item;
 539                         break;
 540                 }
 541         }
 542         LASSERTF(ti != NULL, "ti is NULL !\n");
 543         if (list_empty(&ti->ti_obd_list)) {
 544                 list_del(&ti->ti_chain);
 545                 OBD_FREE_PTR(ti);
 546         }
 547         mutex_unlock(&pinger_mutex);
 548         return 0;
 549 }
 550 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 551
 552 int ptlrpc_pinger_remove_timeouts(void)
 553 {
 554         struct timeout_item *item, *tmp;
 555
 556         mutex_lock(&pinger_mutex);
 557         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 558                 LASSERT(list_empty(&item->ti_obd_list));
 559                 list_del(&item->ti_chain);
 560                 OBD_FREE_PTR(item);
 561         }
 562         mutex_unlock(&pinger_mutex);
 563         return 0;
 564 }
 565
 566 void ptlrpc_pinger_wake_up()
 567 {
 568 #ifdef ENABLE_PINGER
 569         thread_add_flags(&pinger_thread, SVC_EVENT);
 570         wake_up(&pinger_thread.t_ctl_waitq);
 571 #endif
 572 }
 573
 574 /* Ping evictor thread */
 575 #define PET_READY     1
 576 #define PET_TERMINATE 2
 577
 578 static int               pet_refcount = 0;
 579 static int               pet_state;
 580 static wait_queue_head_t pet_waitq;
 581 static struct list_head  pet_list;
 582 static DEFINE_SPINLOCK(pet_lock);
 583
 584 int ping_evictor_wake(struct obd_export *exp)
 585 {
 586         struct obd_device *obd;
 587
 588         spin_lock(&pet_lock);
 589         if (pet_state != PET_READY) {
 590                 /* eventually the new obd will call here again. */
 591                 spin_unlock(&pet_lock);
 592                 return 1;
 593         }
 594
 595         obd = class_exp2obd(exp);
 596         if (list_empty(&obd->obd_evict_list)) {
 597                 class_incref(obd, "evictor", obd);
 598                 list_add(&obd->obd_evict_list, &pet_list);
 599         }
 600         spin_unlock(&pet_lock);
 601
 602         wake_up(&pet_waitq);
 603         return 0;
 604 }
 605
 606 static int ping_evictor_main(void *arg)
 607 {
 608         struct obd_device *obd;
 609         struct obd_export *exp;
 610         struct l_wait_info lwi = { 0 };
 611         time_t expire_time;
 612         ENTRY;
 613
 614         unshare_fs_struct();
 615
 616         CDEBUG(D_HA, "Starting Ping Evictor\n");
 617         pet_state = PET_READY;
 618         while (1) {
 619                 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
 620                             (pet_state == PET_TERMINATE), &lwi);
 621
 622                 /* loop until all obd's will be removed */
 623                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 624                         break;
 625
 626                 /* we only get here if pet_exp != NULL, and the end of this
 627                  * loop is the only place which sets it NULL again, so lock
 628                  * is not strictly necessary. */
 629                 spin_lock(&pet_lock);
 630                 obd = list_entry(pet_list.next, struct obd_device,
 631                                  obd_evict_list);
 632                 spin_unlock(&pet_lock);
 633
 634                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 635
 636                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 637                        obd->obd_name, expire_time);
 638
 639                 /* Exports can't be deleted out of the list while we hold
 640                  * the obd lock (class_unlink_export), which means we can't
 641                  * lose the last ref on the export.  If they've already been
 642                  * removed from the list, we won't find them here. */
 643                 spin_lock(&obd->obd_dev_lock);
 644                 while (!list_empty(&obd->obd_exports_timed)) {
 645                         exp = list_entry(obd->obd_exports_timed.next,
 646                                          struct obd_export,
 647                                          exp_obd_chain_timed);
 648                         if (expire_time > exp->exp_last_request_time) {
 649                                 class_export_get(exp);
 650                                 spin_unlock(&obd->obd_dev_lock);
 651                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 652                                               " (at %s) in %ld seconds. I think"
 653                                               " it's dead, and I am evicting"
 654                                               " it. exp %p, cur %ld expire %ld"
 655                                               " last %ld\n",
 656                                               obd->obd_name,
 657                                               obd_uuid2str(&exp->exp_client_uuid),
 658                                               obd_export_nid2str(exp),
 659                                               (long)(cfs_time_current_sec() -
 660                                                      exp->exp_last_request_time),
 661                                               exp, (long)cfs_time_current_sec(),
 662                                               (long)expire_time,
 663                                               (long)exp->exp_last_request_time);
 664                                 CDEBUG(D_HA, "Last request was at %ld\n",
 665                                        exp->exp_last_request_time);
 666                                 class_fail_export(exp);
 667                                 class_export_put(exp);
 668                                 spin_lock(&obd->obd_dev_lock);
 669                         } else {
 670                                 /* List is sorted, so everyone below is ok */
 671                                 break;
 672                         }
 673                 }
 674                 spin_unlock(&obd->obd_dev_lock);
 675
 676                 spin_lock(&pet_lock);
 677                 list_del_init(&obd->obd_evict_list);
 678                 spin_unlock(&pet_lock);
 679
 680                 class_decref(obd, "evictor", obd);
 681         }
 682         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 683
 684         RETURN(0);
 685 }
 686
 687 void ping_evictor_start(void)
 688 {
 689         struct task_struct *task;
 690
 691         if (++pet_refcount > 1)
 692                 return;
 693
 694         INIT_LIST_HEAD(&pet_list);
 695         init_waitqueue_head(&pet_waitq);
 696
 697         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 698         if (IS_ERR(task)) {
 699                 pet_refcount--;
 700                 CERROR("Cannot start ping evictor thread: %ld\n",
 701                         PTR_ERR(task));
 702         }
 703 }
 704 EXPORT_SYMBOL(ping_evictor_start);
 705
 706 void ping_evictor_stop(void)
 707 {
 708         if (--pet_refcount > 0)
 709                 return;
 710
 711         pet_state = PET_TERMINATE;
 712         wake_up(&pet_waitq);
 713 }
 714 EXPORT_SYMBOL(ping_evictor_stop);