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, 2013, 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 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 static int suppress_pings;
  52 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  53
  54 struct mutex pinger_mutex;
  55 static struct list_head pinger_imports =
  56                 LIST_HEAD_INIT(pinger_imports);
  57 static struct list_head timeout_list =
  58                 LIST_HEAD_INIT(timeout_list);
  59
  60 int ptlrpc_pinger_suppress_pings()
  61 {
  62         return suppress_pings;
  63 }
  64 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  65
  66 struct ptlrpc_request *
  67 ptlrpc_prep_ping(struct obd_import *imp)
  68 {
  69         struct ptlrpc_request *req;
  70
  71         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  72                                         LUSTRE_OBD_VERSION, OBD_PING);
  73         if (req) {
  74                 ptlrpc_request_set_replen(req);
  75                 req->rq_no_resend = req->rq_no_delay = 1;
  76         }
  77         return req;
  78 }
  79
  80 int ptlrpc_obd_ping(struct obd_device *obd)
  81 {
  82         int rc;
  83         struct ptlrpc_request *req;
  84         ENTRY;
  85
  86         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  87         if (req == NULL)
  88                 RETURN(-ENOMEM);
  89
  90         req->rq_send_state = LUSTRE_IMP_FULL;
  91
  92         rc = ptlrpc_queue_wait(req);
  93
  94         ptlrpc_req_finished(req);
  95
  96         RETURN(rc);
  97 }
  98 EXPORT_SYMBOL(ptlrpc_obd_ping);
  99
 100 int ptlrpc_ping(struct obd_import *imp)
 101 {
 102         struct ptlrpc_request *req;
 103         ENTRY;
 104
 105         req = ptlrpc_prep_ping(imp);
 106         if (req == NULL) {
 107                 CERROR("OOM trying to ping %s->%s\n",
 108                        imp->imp_obd->obd_uuid.uuid,
 109                        obd2cli_tgt(imp->imp_obd));
 110                 RETURN(-ENOMEM);
 111         }
 112
 113         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 114                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 115         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 116
 117         RETURN(0);
 118 }
 119
 120 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 121 {
 122 #ifdef ENABLE_PINGER
 123         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 124         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 125                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 126                                   AT_OFF ? 0 :
 127                                   at_get(&imp->imp_at.iat_net_latency));
 128                 time = min(time, dtime);
 129         }
 130         imp->imp_next_ping = cfs_time_shift(time);
 131 #endif /* ENABLE_PINGER */
 132 }
 133
 134 void ptlrpc_ping_import_soon(struct obd_import *imp)
 135 {
 136         imp->imp_next_ping = cfs_time_current();
 137 }
 138
 139 static inline int imp_is_deactive(struct obd_import *imp)
 140 {
 141         return (imp->imp_deactive ||
 142                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 143 }
 144
 145 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 146 {
 147         if (imp->imp_server_timeout)
 148                 return cfs_time_shift(obd_timeout / 2);
 149         else
 150                 return cfs_time_shift(obd_timeout);
 151 }
 152
 153 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 154 {
 155         struct timeout_item *item;
 156         cfs_time_t timeout = PING_INTERVAL;
 157
 158         /* The timeout list is a increase order sorted list */
 159         mutex_lock(&pinger_mutex);
 160         list_for_each_entry(item, &timeout_list, ti_chain) {
 161                 int ti_timeout = item->ti_timeout;
 162                 if (timeout > ti_timeout)
 163                         timeout = ti_timeout;
 164                 break;
 165         }
 166         mutex_unlock(&pinger_mutex);
 167
 168         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 169                                          cfs_time_current());
 170 }
 171
 172 #ifdef __KERNEL__
 173
 174 static bool ir_up;
 175
 176 void ptlrpc_pinger_ir_up(void)
 177 {
 178         CDEBUG(D_HA, "IR up\n");
 179         ir_up = true;
 180 }
 181 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 182
 183 void ptlrpc_pinger_ir_down(void)
 184 {
 185         CDEBUG(D_HA, "IR down\n");
 186         ir_up = false;
 187 }
 188 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 189
 190 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 191                                          unsigned long this_ping)
 192 {
 193         int level;
 194         int force;
 195         int force_next;
 196         int suppress;
 197
 198         spin_lock(&imp->imp_lock);
 199
 200         level = imp->imp_state;
 201         force = imp->imp_force_verify;
 202         force_next = imp->imp_force_next_verify;
 203         /*
 204          * This will be used below only if the import is "FULL".
 205          */
 206         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 207
 208         imp->imp_force_verify = 0;
 209
 210         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 211             !force) {
 212                 spin_unlock(&imp->imp_lock);
 213                 return;
 214         }
 215
 216         imp->imp_force_next_verify = 0;
 217
 218         spin_unlock(&imp->imp_lock);
 219
 220         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 221                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 222                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 223                ptlrpc_import_state_name(level), level, force, force_next,
 224                imp->imp_deactive, imp->imp_pingable, suppress);
 225
 226         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 227                 /* wait for a while before trying recovery again */
 228                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 229                 if (!imp->imp_no_pinger_recover)
 230                         ptlrpc_initiate_recovery(imp);
 231         } else if (level != LUSTRE_IMP_FULL ||
 232                    imp->imp_obd->obd_no_recov ||
 233                    imp_is_deactive(imp)) {
 234                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 235                        "or recovery disabled: %s)\n",
 236                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 237                        ptlrpc_import_state_name(level));
 238                 if (force) {
 239                         spin_lock(&imp->imp_lock);
 240                         imp->imp_force_verify = 1;
 241                         spin_unlock(&imp->imp_lock);
 242                 }
 243         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 244                 ptlrpc_ping(imp);
 245         }
 246 }
 247
 248 static int ptlrpc_pinger_main(void *arg)
 249 {
 250         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 251         ENTRY;
 252
 253         /* Record that the thread is running */
 254         thread_set_flags(thread, SVC_RUNNING);
 255         wake_up(&thread->t_ctl_waitq);
 256
 257         /* And now, loop forever, pinging as needed. */
 258         while (1) {
 259                 cfs_time_t this_ping = cfs_time_current();
 260                 struct l_wait_info lwi;
 261                 cfs_duration_t time_to_next_wake;
 262                 struct timeout_item *item;
 263                 struct list_head *iter;
 264
 265                 mutex_lock(&pinger_mutex);
 266                 list_for_each_entry(item, &timeout_list, ti_chain)
 267                         item->ti_cb(item, item->ti_cb_data);
 268
 269                 list_for_each(iter, &pinger_imports) {
 270                         struct obd_import *imp = list_entry(iter,
 271                                                             struct obd_import,
 272                                                             imp_pinger_chain);
 273
 274                         ptlrpc_pinger_process_import(imp, this_ping);
 275                         /* obd_timeout might have changed */
 276                         if (imp->imp_pingable && imp->imp_next_ping &&
 277                             cfs_time_after(imp->imp_next_ping,
 278                                            cfs_time_add(this_ping,
 279                                                         cfs_time_seconds(PING_INTERVAL))))
 280                                 ptlrpc_update_next_ping(imp, 0);
 281                 }
 282                 mutex_unlock(&pinger_mutex);
 283                 /* update memory usage info */
 284                 obd_update_maxusage();
 285
 286                 /* Wait until the next ping time, or until we're stopped. */
 287                 time_to_next_wake = pinger_check_timeout(this_ping);
 288                 /* The ping sent by ptlrpc_send_rpc may get sent out
 289                    say .01 second after this.
 290                    ptlrpc_pinger_sending_on_import will then set the
 291                    next ping time to next_ping + .01 sec, which means
 292                    we will SKIP the next ping at next_ping, and the
 293                    ping will get sent 2 timeouts from now!  Beware. */
 294                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 295                        CFS_TIME_T")\n", time_to_next_wake,
 296                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 297                 if (time_to_next_wake > 0) {
 298                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 299                                                 time_to_next_wake,
 300                                                 cfs_time_seconds(1)),
 301                                           NULL, NULL);
 302                         l_wait_event(thread->t_ctl_waitq,
 303                                      thread_is_stopping(thread) ||
 304                                      thread_is_event(thread),
 305                                      &lwi);
 306                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 307                                 EXIT;
 308                                 break;
 309                         } else {
 310                                 /* woken after adding import to reset timer */
 311                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 312                         }
 313                 }
 314         }
 315
 316         thread_set_flags(thread, SVC_STOPPED);
 317         wake_up(&thread->t_ctl_waitq);
 318
 319         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
 320         return 0;
 321 }
 322
 323 static struct ptlrpc_thread pinger_thread;
 324
 325 int ptlrpc_start_pinger(void)
 326 {
 327         struct l_wait_info lwi = { 0 };
 328         struct task_struct *task;
 329         int rc;
 330 #ifndef ENABLE_PINGER
 331         return 0;
 332 #endif
 333         ENTRY;
 334
 335         if (!thread_is_init(&pinger_thread) &&
 336             !thread_is_stopped(&pinger_thread))
 337                 RETURN(-EALREADY);
 338
 339         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 340
 341         strcpy(pinger_thread.t_name, "ll_ping");
 342
 343         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 344          * just drop the VM and FILES in kthread_run() right away. */
 345         task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
 346                            pinger_thread.t_name);
 347         if (IS_ERR(task)) {
 348                 rc = PTR_ERR(task);
 349                 CERROR("cannot start pinger thread: rc = %d\n", rc);
 350                 RETURN(rc);
 351         }
 352
 353         l_wait_event(pinger_thread.t_ctl_waitq,
 354                      thread_is_running(&pinger_thread), &lwi);
 355
 356         if (suppress_pings)
 357                 CWARN("Pings will be suppressed at the request of the "
 358                       "administrator.  The configuration shall meet the "
 359                       "additional requirements described in the manual.  "
 360                       "(Search for the \"suppress_pings\" kernel module "
 361                       "parameter.)\n");
 362
 363         RETURN(0);
 364 }
 365
 366 int ptlrpc_pinger_remove_timeouts(void);
 367
 368 int ptlrpc_stop_pinger(void)
 369 {
 370         struct l_wait_info lwi = { 0 };
 371 #ifndef ENABLE_PINGER
 372         return 0;
 373 #endif
 374         ENTRY;
 375
 376         if (thread_is_init(&pinger_thread) ||
 377             thread_is_stopped(&pinger_thread))
 378                 RETURN(-EALREADY);
 379
 380         ptlrpc_pinger_remove_timeouts();
 381
 382         thread_set_flags(&pinger_thread, SVC_STOPPING);
 383         wake_up(&pinger_thread.t_ctl_waitq);
 384
 385         l_wait_event(pinger_thread.t_ctl_waitq,
 386                      thread_is_stopped(&pinger_thread), &lwi);
 387         RETURN(0);
 388 }
 389
 390 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 391 {
 392         ptlrpc_update_next_ping(imp, 0);
 393 }
 394 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 395
 396 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 397 {
 398         ptlrpc_update_next_ping(imp, 1);
 399         assert_spin_locked(&imp->imp_lock);
 400         /*
 401          * Avoid reading stale imp_connect_data.  When not sure if pings are
 402          * expected or not on next connection, we assume they are not and force
 403          * one anyway to guarantee the chance of updating
 404          * imp_peer_committed_transno.
 405          */
 406         if (imp->imp_state != LUSTRE_IMP_FULL ||
 407             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 408                 imp->imp_force_next_verify = 1;
 409 }
 410
 411 int ptlrpc_pinger_add_import(struct obd_import *imp)
 412 {
 413         ENTRY;
 414         if (!list_empty(&imp->imp_pinger_chain))
 415                 RETURN(-EALREADY);
 416
 417         mutex_lock(&pinger_mutex);
 418         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 419                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 420         /* if we add to pinger we want recovery on this import */
 421         imp->imp_obd->obd_no_recov = 0;
 422         ptlrpc_update_next_ping(imp, 0);
 423         /* XXX sort, blah blah */
 424         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 425         class_import_get(imp);
 426
 427         ptlrpc_pinger_wake_up();
 428         mutex_unlock(&pinger_mutex);
 429
 430         RETURN(0);
 431 }
 432 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 433
 434 int ptlrpc_pinger_del_import(struct obd_import *imp)
 435 {
 436         ENTRY;
 437
 438         if (list_empty(&imp->imp_pinger_chain))
 439                 RETURN(-ENOENT);
 440
 441         mutex_lock(&pinger_mutex);
 442         list_del_init(&imp->imp_pinger_chain);
 443         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 444                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 445         /* if we remove from pinger we don't want recovery on this import */
 446         imp->imp_obd->obd_no_recov = 1;
 447         class_import_put(imp);
 448         mutex_unlock(&pinger_mutex);
 449         RETURN(0);
 450 }
 451 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 452
 453 /**
 454  * Register a timeout callback to the pinger list, and the callback will
 455  * be called when timeout happens.
 456  */
 457 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 458                                         timeout_cb_t cb, void *data)
 459 {
 460         struct timeout_item *ti;
 461
 462         OBD_ALLOC_PTR(ti);
 463         if (!ti)
 464                 return(NULL);
 465
 466         INIT_LIST_HEAD(&ti->ti_obd_list);
 467         INIT_LIST_HEAD(&ti->ti_chain);
 468         ti->ti_timeout = time;
 469         ti->ti_event = event;
 470         ti->ti_cb = cb;
 471         ti->ti_cb_data = data;
 472
 473         return ti;
 474 }
 475
 476 /**
 477  * Register timeout event on the the pinger thread.
 478  * Note: the timeout list is an sorted list with increased timeout value.
 479  */
 480 static struct timeout_item*
 481 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 482                                timeout_cb_t cb, void *data)
 483 {
 484         struct timeout_item *item, *tmp;
 485
 486         LASSERT(mutex_is_locked(&pinger_mutex));
 487
 488         list_for_each_entry(item, &timeout_list, ti_chain)
 489                 if (item->ti_event == event)
 490                         goto out;
 491
 492         item = ptlrpc_new_timeout(time, event, cb, data);
 493         if (item) {
 494                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 495                         if (tmp->ti_timeout < time) {
 496                                 list_add(&item->ti_chain, &tmp->ti_chain);
 497                                 goto out;
 498                         }
 499                 }
 500                 list_add(&item->ti_chain, &timeout_list);
 501         }
 502 out:
 503         return item;
 504 }
 505
 506 /* Add a client_obd to the timeout event list, when timeout(@time)
 507  * happens, the callback(@cb) will be called.
 508  */
 509 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 510                               timeout_cb_t cb, void *data,
 511                               struct list_head *obd_list)
 512 {
 513         struct timeout_item *ti;
 514
 515         mutex_lock(&pinger_mutex);
 516         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 517         if (!ti) {
 518                 mutex_unlock(&pinger_mutex);
 519                 return (-EINVAL);
 520         }
 521         list_add(obd_list, &ti->ti_obd_list);
 522         mutex_unlock(&pinger_mutex);
 523         return 0;
 524 }
 525 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 526
 527 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 528                               enum timeout_event event)
 529 {
 530         struct timeout_item *ti = NULL, *item;
 531
 532         if (list_empty(obd_list))
 533                 return 0;
 534         mutex_lock(&pinger_mutex);
 535         list_del_init(obd_list);
 536         /**
 537          * If there are no obd attached to the timeout event
 538          * list, remove this timeout event from the pinger
 539          */
 540         list_for_each_entry(item, &timeout_list, ti_chain) {
 541                 if (item->ti_event == event) {
 542                         ti = item;
 543                         break;
 544                 }
 545         }
 546         LASSERTF(ti != NULL, "ti is NULL !\n");
 547         if (list_empty(&ti->ti_obd_list)) {
 548                 list_del(&ti->ti_chain);
 549                 OBD_FREE_PTR(ti);
 550         }
 551         mutex_unlock(&pinger_mutex);
 552         return 0;
 553 }
 554 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 555
 556 int ptlrpc_pinger_remove_timeouts(void)
 557 {
 558         struct timeout_item *item, *tmp;
 559
 560         mutex_lock(&pinger_mutex);
 561         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 562                 LASSERT(list_empty(&item->ti_obd_list));
 563                 list_del(&item->ti_chain);
 564                 OBD_FREE_PTR(item);
 565         }
 566         mutex_unlock(&pinger_mutex);
 567         return 0;
 568 }
 569
 570 void ptlrpc_pinger_wake_up()
 571 {
 572 #ifdef ENABLE_PINGER
 573         thread_add_flags(&pinger_thread, SVC_EVENT);
 574         wake_up(&pinger_thread.t_ctl_waitq);
 575 #endif
 576 }
 577
 578 /* Ping evictor thread */
 579 #define PET_READY     1
 580 #define PET_TERMINATE 2
 581
 582 static int               pet_refcount = 0;
 583 static int               pet_state;
 584 static wait_queue_head_t pet_waitq;
 585 struct list_head         pet_list;
 586 static DEFINE_SPINLOCK(pet_lock);
 587
 588 int ping_evictor_wake(struct obd_export *exp)
 589 {
 590         struct obd_device *obd;
 591
 592         spin_lock(&pet_lock);
 593         if (pet_state != PET_READY) {
 594                 /* eventually the new obd will call here again. */
 595                 spin_unlock(&pet_lock);
 596                 return 1;
 597         }
 598
 599         obd = class_exp2obd(exp);
 600         if (list_empty(&obd->obd_evict_list)) {
 601                 class_incref(obd, "evictor", obd);
 602                 list_add(&obd->obd_evict_list, &pet_list);
 603         }
 604         spin_unlock(&pet_lock);
 605
 606         wake_up(&pet_waitq);
 607         return 0;
 608 }
 609
 610 static int ping_evictor_main(void *arg)
 611 {
 612         struct obd_device *obd;
 613         struct obd_export *exp;
 614         struct l_wait_info lwi = { 0 };
 615         time_t expire_time;
 616         ENTRY;
 617
 618         unshare_fs_struct();
 619
 620         CDEBUG(D_HA, "Starting Ping Evictor\n");
 621         pet_state = PET_READY;
 622         while (1) {
 623                 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
 624                             (pet_state == PET_TERMINATE), &lwi);
 625
 626                 /* loop until all obd's will be removed */
 627                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 628                         break;
 629
 630                 /* we only get here if pet_exp != NULL, and the end of this
 631                  * loop is the only place which sets it NULL again, so lock
 632                  * is not strictly necessary. */
 633                 spin_lock(&pet_lock);
 634                 obd = list_entry(pet_list.next, struct obd_device,
 635                                  obd_evict_list);
 636                 spin_unlock(&pet_lock);
 637
 638                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 639
 640                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 641                        obd->obd_name, expire_time);
 642
 643                 /* Exports can't be deleted out of the list while we hold
 644                  * the obd lock (class_unlink_export), which means we can't
 645                  * lose the last ref on the export.  If they've already been
 646                  * removed from the list, we won't find them here. */
 647                 spin_lock(&obd->obd_dev_lock);
 648                 while (!list_empty(&obd->obd_exports_timed)) {
 649                         exp = list_entry(obd->obd_exports_timed.next,
 650                                          struct obd_export,
 651                                          exp_obd_chain_timed);
 652                         if (expire_time > exp->exp_last_request_time) {
 653                                 class_export_get(exp);
 654                                 spin_unlock(&obd->obd_dev_lock);
 655                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 656                                               " (at %s) in %ld seconds. I think"
 657                                               " it's dead, and I am evicting"
 658                                               " it. exp %p, cur %ld expire %ld"
 659                                               " last %ld\n",
 660                                               obd->obd_name,
 661                                               obd_uuid2str(&exp->exp_client_uuid),
 662                                               obd_export_nid2str(exp),
 663                                               (long)(cfs_time_current_sec() -
 664                                                      exp->exp_last_request_time),
 665                                               exp, (long)cfs_time_current_sec(),
 666                                               (long)expire_time,
 667                                               (long)exp->exp_last_request_time);
 668                                 CDEBUG(D_HA, "Last request was at %ld\n",
 669                                        exp->exp_last_request_time);
 670                                 class_fail_export(exp);
 671                                 class_export_put(exp);
 672                                 spin_lock(&obd->obd_dev_lock);
 673                         } else {
 674                                 /* List is sorted, so everyone below is ok */
 675                                 break;
 676                         }
 677                 }
 678                 spin_unlock(&obd->obd_dev_lock);
 679
 680                 spin_lock(&pet_lock);
 681                 list_del_init(&obd->obd_evict_list);
 682                 spin_unlock(&pet_lock);
 683
 684                 class_decref(obd, "evictor", obd);
 685         }
 686         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 687
 688         RETURN(0);
 689 }
 690
 691 void ping_evictor_start(void)
 692 {
 693         struct task_struct *task;
 694
 695         if (++pet_refcount > 1)
 696                 return;
 697
 698         INIT_LIST_HEAD(&pet_list);
 699         init_waitqueue_head(&pet_waitq);
 700
 701         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 702         if (IS_ERR(task)) {
 703                 pet_refcount--;
 704                 CERROR("Cannot start ping evictor thread: %ld\n",
 705                         PTR_ERR(task));
 706         }
 707 }
 708 EXPORT_SYMBOL(ping_evictor_start);
 709
 710 void ping_evictor_stop(void)
 711 {
 712         if (--pet_refcount > 0)
 713                 return;
 714
 715         pet_state = PET_TERMINATE;
 716         wake_up(&pet_waitq);
 717 }
 718 EXPORT_SYMBOL(ping_evictor_stop);
 719 #else /* !__KERNEL__ */
 720
 721 /* XXX
 722  * the current implementation of pinger in liblustre is not optimized
 723  */
 724
 725 #ifdef ENABLE_PINGER
 726 static struct pinger_data {
 727         int             pd_recursion;
 728         cfs_time_t      pd_this_ping;   /* jiffies */
 729         cfs_time_t      pd_next_ping;   /* jiffies */
 730         struct ptlrpc_request_set *pd_set;
 731 } pinger_args;
 732
 733 static int pinger_check_rpcs(void *arg)
 734 {
 735         cfs_time_t curtime = cfs_time_current();
 736         struct ptlrpc_request *req;
 737         struct ptlrpc_request_set *set;
 738         struct list_head *iter;
 739         struct obd_import *imp;
 740         struct pinger_data *pd = &pinger_args;
 741         int rc;
 742
 743         /* prevent recursion */
 744         if (pd->pd_recursion++) {
 745                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 746                 LASSERT(pd->pd_set);
 747                 pd->pd_recursion--;
 748                 return 0;
 749         }
 750
 751         /* have we reached ping point? */
 752         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 753                 pd->pd_recursion--;
 754                 return 0;
 755         }
 756
 757         /* if we have rpc_set already, continue processing it */
 758         if (pd->pd_set) {
 759                 LASSERT(pd->pd_this_ping);
 760                 set = pd->pd_set;
 761                 goto do_check_set;
 762         }
 763
 764         pd->pd_this_ping = curtime;
 765         pd->pd_set = ptlrpc_prep_set();
 766         if (pd->pd_set == NULL)
 767                 goto out;
 768         set = pd->pd_set;
 769
 770         /* add rpcs into set */
 771         mutex_lock(&pinger_mutex);
 772         list_for_each(iter, &pinger_imports) {
 773                 struct obd_import *imp = list_entry(iter, struct obd_import,
 774                                                         imp_pinger_chain);
 775                 int generation, level;
 776
 777                 if (cfs_time_aftereq(pd->pd_this_ping,
 778                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 779                         /* Add a ping. */
 780                         spin_lock(&imp->imp_lock);
 781                         generation = imp->imp_generation;
 782                         level = imp->imp_state;
 783                         spin_unlock(&imp->imp_lock);
 784
 785                         if (level != LUSTRE_IMP_FULL) {
 786                                 CDEBUG(D_HA,
 787                                        "not pinging %s (in recovery)\n",
 788                                        obd2cli_tgt(imp->imp_obd));
 789                                 continue;
 790                         }
 791
 792                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 793                                                         LUSTRE_OBD_VERSION,
 794                                                         OBD_PING);
 795                         if (req == NULL) {
 796                                 CERROR("OOM trying to ping %s->%s\n",
 797                                        imp->imp_obd->obd_uuid.uuid,
 798                                        obd2cli_tgt(imp->imp_obd));
 799                                 break;
 800                         }
 801
 802                         req->rq_no_resend = 1;
 803                         ptlrpc_request_set_replen(req);
 804                         req->rq_send_state = LUSTRE_IMP_FULL;
 805                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 806                         req->rq_import_generation = generation;
 807                         ptlrpc_set_add_req(set, req);
 808                 } else {
 809                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 810                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 811                                imp->imp_next_ping, pd->pd_this_ping);
 812                 }
 813         }
 814         pd->pd_this_ping = curtime;
 815         mutex_unlock(&pinger_mutex);
 816
 817         /* Might be empty, that's OK. */
 818         if (atomic_read(&set->set_remaining) == 0)
 819                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 820
 821         list_for_each(iter, &set->set_requests) {
 822                 struct ptlrpc_request *req =
 823                         list_entry(iter, struct ptlrpc_request,
 824                                    rq_set_chain);
 825                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 826                           req->rq_import->imp_obd->obd_uuid.uuid,
 827                           obd2cli_tgt(req->rq_import->imp_obd));
 828                 (void)ptl_send_rpc(req, 0);
 829         }
 830
 831 do_check_set:
 832         rc = ptlrpc_check_set(NULL, set);
 833
 834         /* not finished, and we are not expired, simply return */
 835         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 836                                             cfs_time_seconds(PING_INTERVAL)))) {
 837                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 838                 pd->pd_recursion--;
 839                 return 0;
 840         }
 841
 842         /* Expire all the requests that didn't come back. */
 843         mutex_lock(&pinger_mutex);
 844         list_for_each(iter, &set->set_requests) {
 845                 req = list_entry(iter, struct ptlrpc_request,
 846                                  rq_set_chain);
 847
 848                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 849                         continue;
 850
 851                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 852                        req);
 853
 854                 /* This will also unregister reply. */
 855                 ptlrpc_expire_one_request(req, 0);
 856
 857                 /* We're done with this req, let's finally move it to complete
 858                  * phase and take care of inflights. */
 859                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 860                 imp = req->rq_import;
 861                 spin_lock(&imp->imp_lock);
 862                 if (!list_empty(&req->rq_list)) {
 863                         list_del_init(&req->rq_list);
 864                         atomic_dec(&imp->imp_inflight);
 865                 }
 866                 spin_unlock(&imp->imp_lock);
 867                 atomic_dec(&set->set_remaining);
 868         }
 869         mutex_unlock(&pinger_mutex);
 870
 871         ptlrpc_set_destroy(set);
 872         pd->pd_set = NULL;
 873
 874 out:
 875         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 876                                         cfs_time_seconds(PING_INTERVAL));
 877         pd->pd_this_ping = 0; /* XXX for debug */
 878
 879         CDEBUG(D_INFO, "finished a round ping\n");
 880         pd->pd_recursion--;
 881         return 0;
 882 }
 883
 884 static void *pinger_callback = NULL;
 885 #endif /* ENABLE_PINGER */
 886
 887 int ptlrpc_start_pinger(void)
 888 {
 889 #ifdef ENABLE_PINGER
 890         memset(&pinger_args, 0, sizeof(pinger_args));
 891         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 892                                                            &pinger_check_rpcs,
 893                                                            &pinger_args);
 894 #endif
 895         return 0;
 896 }
 897
 898 int ptlrpc_stop_pinger(void)
 899 {
 900 #ifdef ENABLE_PINGER
 901         if (pinger_callback)
 902                 liblustre_deregister_wait_callback(pinger_callback);
 903 #endif
 904         return 0;
 905 }
 906
 907 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 908 {
 909 #ifdef ENABLE_PINGER
 910         mutex_lock(&pinger_mutex);
 911         ptlrpc_update_next_ping(imp, 0);
 912         if (pinger_args.pd_set == NULL &&
 913             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 914                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 915                         imp->imp_next_ping, cfs_time_current());
 916                 pinger_args.pd_next_ping = imp->imp_next_ping;
 917         }
 918         mutex_unlock(&pinger_mutex);
 919 #endif
 920 }
 921
 922 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 923 {
 924 #ifdef ENABLE_PINGER
 925         mutex_lock(&pinger_mutex);
 926         ptlrpc_update_next_ping(imp, 1);
 927         if (pinger_args.pd_set == NULL &&
 928             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 929                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 930                         imp->imp_next_ping, cfs_time_current());
 931                 pinger_args.pd_next_ping = imp->imp_next_ping;
 932         }
 933         mutex_unlock(&pinger_mutex);
 934 #endif
 935 }
 936
 937 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 938                               timeout_cb_t cb, void *data,
 939                               struct list_head *obd_list)
 940 {
 941         return 0;
 942 }
 943
 944 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 945                               enum timeout_event event)
 946 {
 947         return 0;
 948 }
 949
 950 int ptlrpc_pinger_add_import(struct obd_import *imp)
 951 {
 952         ENTRY;
 953
 954         if (!list_empty(&imp->imp_pinger_chain))
 955                 RETURN(-EALREADY);
 956
 957         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 958                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 959         ptlrpc_pinger_sending_on_import(imp);
 960
 961         mutex_lock(&pinger_mutex);
 962         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 963         class_import_get(imp);
 964         mutex_unlock(&pinger_mutex);
 965
 966         RETURN(0);
 967 }
 968
 969 int ptlrpc_pinger_del_import(struct obd_import *imp)
 970 {
 971         ENTRY;
 972
 973         if (list_empty(&imp->imp_pinger_chain))
 974                 RETURN(-ENOENT);
 975
 976         mutex_lock(&pinger_mutex);
 977         list_del_init(&imp->imp_pinger_chain);
 978         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 979                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 980         class_import_put(imp);
 981         mutex_unlock(&pinger_mutex);
 982         RETURN(0);
 983 }
 984
 985 void ptlrpc_pinger_wake_up()
 986 {
 987 #ifdef ENABLE_PINGER
 988         /* XXX force pinger to run, if needed */
 989         struct obd_import *imp;
 990         ENTRY;
 991         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 992                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 993                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 994 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 995                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 996                     !imp_is_deactive(imp))
 997 #else
 998                 /*XXX only recover for the initial connection */
 999                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1000                     imp->imp_state == LUSTRE_IMP_DISCON &&
1001                     !imp_is_deactive(imp))
1002 #endif
1003                         ptlrpc_initiate_recovery(imp);
1004                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1005                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1006                                      "state %d, deactive %d\n",
1007                                      imp->imp_obd->obd_uuid.uuid,
1008                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1009                                      imp_is_deactive(imp));
1010         }
1011         EXIT;
1012 #endif
1013 }
1014 #endif /* !__KERNEL__ */