lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 struct semaphore pinger_sem;
  52 static CFS_LIST_HEAD(pinger_imports);
  53 static struct list_head timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  54 struct ptlrpc_request *
  55 ptlrpc_prep_ping(struct obd_import *imp)
  56 {
  57         struct ptlrpc_request *req;
  58
  59         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  60                                         LUSTRE_OBD_VERSION, OBD_PING);
  61         if (req) {
  62                 ptlrpc_request_set_replen(req);
  63                 req->rq_no_resend = req->rq_no_delay = 1;
  64         }
  65         return req;
  66 }
  67
  68 int ptlrpc_obd_ping(struct obd_device *obd)
  69 {
  70         int rc;
  71         struct ptlrpc_request *req;
  72         ENTRY;
  73
  74         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  75         if (req == NULL)
  76                 RETURN(-ENOMEM);
  77
  78         req->rq_send_state = LUSTRE_IMP_FULL;
  79
  80         rc = ptlrpc_queue_wait(req);
  81
  82         ptlrpc_req_finished(req);
  83
  84         RETURN(rc);
  85 }
  86 EXPORT_SYMBOL(ptlrpc_obd_ping);
  87
  88 int ptlrpc_ping(struct obd_import *imp)
  89 {
  90         struct ptlrpc_request *req;
  91         ENTRY;
  92
  93         req = ptlrpc_prep_ping(imp);
  94         if (req == NULL) {
  95                 CERROR("OOM trying to ping %s->%s\n",
  96                        imp->imp_obd->obd_uuid.uuid,
  97                        obd2cli_tgt(imp->imp_obd));
  98                 RETURN(-ENOMEM);
  99         }
 100
 101         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 102                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 103         ptlrpcd_add_req(req, PSCOPE_OTHER);
 104
 105         RETURN(0);
 106 }
 107
 108 void ptlrpc_update_next_ping(struct obd_import *imp)
 109 {
 110 #ifdef ENABLE_PINGER
 111         int time = PING_INTERVAL;
 112         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 113                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 114                                   AT_OFF ? 0 :
 115                                   at_get(&imp->imp_at.iat_net_latency));
 116                 time = min(time, dtime);
 117         }
 118         imp->imp_next_ping = cfs_time_shift(time);
 119 #endif /* ENABLE_PINGER */
 120 }
 121
 122 void ptlrpc_ping_import_soon(struct obd_import *imp)
 123 {
 124         imp->imp_next_ping = cfs_time_current();
 125 }
 126
 127 static inline int imp_is_deactive(struct obd_import *imp)
 128 {
 129         return (imp->imp_deactive ||
 130                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 131 }
 132
 133 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 134 {
 135         if (imp->imp_server_timeout)
 136                 return cfs_time_shift(obd_timeout / 2);
 137         else
 138                 return cfs_time_shift(obd_timeout);
 139 }
 140
 141 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
 142 static cfs_time_t suspend_wakeup_time = 0;
 143
 144 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 145 {
 146         struct timeout_item *item;
 147         cfs_time_t timeout = PING_INTERVAL;
 148
 149         /* The timeout list is a increase order sorted list */
 150         mutex_down(&pinger_sem);
 151         list_for_each_entry(item, &timeout_list, ti_chain) {
 152                 int ti_timeout = item->ti_timeout;
 153                 if (timeout > ti_timeout)
 154                          timeout = ti_timeout;
 155                 break;
 156         }
 157         mutex_up(&pinger_sem);
 158
 159         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 160                                          cfs_time_current());
 161 }
 162
 163 #ifdef __KERNEL__
 164 static wait_queue_head_t suspend_timeouts_waitq;
 165 #endif
 166
 167 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 168 {
 169         return suspend_wakeup_time;
 170 }
 171
 172 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 173 {
 174         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 175 #if 0
 176         if (imp->imp_no_timeout)
 177                 return;
 178         imp->imp_no_timeout = 1;
 179         atomic_inc(&suspend_timeouts);
 180         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
 181 #endif
 182 }
 183
 184 void ptlrpc_activate_timeouts(struct obd_import *imp)
 185 {
 186         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 187 #if 0
 188         if (!imp->imp_no_timeout)
 189                 return;
 190         imp->imp_no_timeout = 0;
 191         LASSERT(atomic_read(&suspend_timeouts) > 0);
 192         if (atomic_dec_and_test(&suspend_timeouts)) {
 193                 suspend_wakeup_time = cfs_time_current();
 194                 wake_up(&suspend_timeouts_waitq);
 195         }
 196         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
 197 #endif
 198 }
 199
 200 int ptlrpc_check_suspend(void)
 201 {
 202         if (atomic_read(&suspend_timeouts))
 203                 return 1;
 204         return 0;
 205 }
 206
 207 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 208 {
 209         struct l_wait_info lwi;
 210
 211         if (atomic_read(&suspend_timeouts)) {
 212                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 213                           atomic_read(&suspend_timeouts));
 214                 lwi = LWI_INTR(NULL, NULL);
 215                 l_wait_event(suspend_timeouts_waitq,
 216                              atomic_read(&suspend_timeouts) == 0, &lwi);
 217                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 218                 return 1;
 219         }
 220         return 0;
 221 }
 222
 223 #ifdef __KERNEL__
 224
 225 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 226                                          unsigned long this_ping)
 227 {
 228         int force, level;
 229
 230         spin_lock(&imp->imp_lock);
 231         level = imp->imp_state;
 232         force = imp->imp_force_verify;
 233         if (force)
 234                 imp->imp_force_verify = 0;
 235         spin_unlock(&imp->imp_lock);
 236
 237         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 238                "level %s/%u force %u deactive %u pingable %u\n",
 239                ptlrpc_import_state_name(level), level,
 240                force, imp->imp_deactive, imp->imp_pingable);
 241
 242         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 243                              this_ping) && force == 0)
 244                 return;
 245
 246         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 247                 /* wait at least a timeout before trying recovery again */
 248                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 249                 ptlrpc_initiate_recovery(imp);
 250         } else if (level != LUSTRE_IMP_FULL ||
 251                    imp->imp_obd->obd_no_recov ||
 252                    imp_is_deactive(imp)) {
 253                 CDEBUG(D_HA, "not pinging %s (in recovery "
 254                        " or recovery disabled: %s)\n",
 255                        obd2cli_tgt(imp->imp_obd),
 256                        ptlrpc_import_state_name(level));
 257         } else if (imp->imp_pingable || force) {
 258                 ptlrpc_ping(imp);
 259         }
 260 }
 261
 262 static int ptlrpc_pinger_main(void *arg)
 263 {
 264         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 265         struct ptlrpc_thread *thread = data->thread;
 266         ENTRY;
 267
 268         cfs_daemonize(data->name);
 269
 270         /* Record that the thread is running */
 271         thread->t_flags = SVC_RUNNING;
 272         cfs_waitq_signal(&thread->t_ctl_waitq);
 273
 274         /* And now, loop forever, pinging as needed. */
 275         while (1) {
 276                 cfs_time_t this_ping = cfs_time_current();
 277                 struct l_wait_info lwi;
 278                 cfs_duration_t time_to_next_wake;
 279                 struct timeout_item *item;
 280                 struct list_head *iter;
 281
 282                 mutex_down(&pinger_sem);
 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                         struct obd_import *imp =
 288                                 list_entry(iter, struct obd_import,
 289                                            imp_pinger_chain);
 290
 291                         ptlrpc_pinger_process_import(imp, this_ping);
 292                         /* obd_timeout might have changed */
 293                         if (imp->imp_pingable && imp->imp_next_ping &&
 294                             cfs_time_after(imp->imp_next_ping,
 295                                            cfs_time_add(this_ping,
 296                                                         cfs_time_seconds(PING_INTERVAL))))
 297                                 ptlrpc_update_next_ping(imp);
 298                 }
 299                 mutex_up(&pinger_sem);
 300                 /* update memory usage info */
 301                 obd_update_maxusage();
 302
 303                 /* Wait until the next ping time, or until we're stopped. */
 304                 time_to_next_wake = pinger_check_timeout(this_ping);
 305                 /* The ping sent by ptlrpc_send_rpc may get sent out
 306                    say .01 second after this.
 307                    ptlrpc_pinger_sending_on_import will then set the
 308                    next ping time to next_ping + .01 sec, which means
 309                    we will SKIP the next ping at next_ping, and the
 310                    ping will get sent 2 timeouts from now!  Beware. */
 311                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 312                                 time_to_next_wake,
 313                                 cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
 314                 if (time_to_next_wake > 0) {
 315                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_wake, cfs_time_seconds(1)),
 316                                             NULL, NULL);
 317                         l_wait_event(thread->t_ctl_waitq,
 318                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 319                                      &lwi);
 320                         if (thread->t_flags & SVC_STOPPING) {
 321                                 thread->t_flags &= ~SVC_STOPPING;
 322                                 EXIT;
 323                                 break;
 324                         } else if (thread->t_flags & SVC_EVENT) {
 325                                 /* woken after adding import to reset timer */
 326                                 thread->t_flags &= ~SVC_EVENT;
 327                         }
 328                 }
 329         }
 330
 331         thread->t_flags = SVC_STOPPED;
 332         cfs_waitq_signal(&thread->t_ctl_waitq);
 333
 334         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 335         return 0;
 336 }
 337
 338 static struct ptlrpc_thread *pinger_thread = NULL;
 339
 340 int ptlrpc_start_pinger(void)
 341 {
 342         struct l_wait_info lwi = { 0 };
 343         struct ptlrpc_svc_data d;
 344         int rc;
 345 #ifndef ENABLE_PINGER
 346         return 0;
 347 #endif
 348         ENTRY;
 349
 350         if (pinger_thread != NULL)
 351                 RETURN(-EALREADY);
 352
 353         OBD_ALLOC_PTR(pinger_thread);
 354         if (pinger_thread == NULL)
 355                 RETURN(-ENOMEM);
 356         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 357         cfs_waitq_init(&suspend_timeouts_waitq);
 358
 359         d.name = "ll_ping";
 360         d.thread = pinger_thread;
 361
 362         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 363          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 364         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 365         if (rc < 0) {
 366                 CERROR("cannot start thread: %d\n", rc);
 367                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 368                 pinger_thread = NULL;
 369                 RETURN(rc);
 370         }
 371         l_wait_event(pinger_thread->t_ctl_waitq,
 372                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 373
 374         RETURN(0);
 375 }
 376
 377 int ptlrpc_pinger_remove_timeouts(void);
 378
 379 int ptlrpc_stop_pinger(void)
 380 {
 381         struct l_wait_info lwi = { 0 };
 382         int rc = 0;
 383 #ifndef ENABLE_PINGER
 384         return 0;
 385 #endif
 386         ENTRY;
 387
 388         if (pinger_thread == NULL)
 389                 RETURN(-EALREADY);
 390
 391         ptlrpc_pinger_remove_timeouts();
 392         mutex_down(&pinger_sem);
 393         pinger_thread->t_flags = SVC_STOPPING;
 394         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 395         mutex_up(&pinger_sem);
 396
 397         l_wait_event(pinger_thread->t_ctl_waitq,
 398                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 399
 400         OBD_FREE_PTR(pinger_thread);
 401         pinger_thread = NULL;
 402         RETURN(rc);
 403 }
 404
 405 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 406 {
 407         ptlrpc_update_next_ping(imp);
 408 }
 409
 410 int ptlrpc_pinger_add_import(struct obd_import *imp)
 411 {
 412         ENTRY;
 413         if (!list_empty(&imp->imp_pinger_chain))
 414                 RETURN(-EALREADY);
 415
 416         mutex_down(&pinger_sem);
 417         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 418                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 419         /* if we add to pinger we want recovery on this import */
 420         imp->imp_obd->obd_no_recov = 0;
 421         ptlrpc_update_next_ping(imp);
 422         /* XXX sort, blah blah */
 423         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 424         class_import_get(imp);
 425
 426         ptlrpc_pinger_wake_up();
 427         mutex_up(&pinger_sem);
 428
 429         RETURN(0);
 430 }
 431
 432 int ptlrpc_pinger_del_import(struct obd_import *imp)
 433 {
 434         ENTRY;
 435         if (list_empty(&imp->imp_pinger_chain))
 436                 RETURN(-ENOENT);
 437
 438         mutex_down(&pinger_sem);
 439         list_del_init(&imp->imp_pinger_chain);
 440         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 441                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 442         /* if we remove from pinger we don't want recovery on this import */
 443         imp->imp_obd->obd_no_recov = 1;
 444         class_import_put(imp);
 445         mutex_up(&pinger_sem);
 446         RETURN(0);
 447 }
 448
 449 /**
 450  * Register a timeout callback to the pinger list, and the callback will
 451  * be called when timeout happens.
 452  */
 453 struct timeout_item* ptlrpc_new_timeout(int time, 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         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 463         CFS_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;
 481         struct timeout_item *ti = NULL;
 482
 483         LASSERT_SEM_LOCKED(&pinger_sem);
 484         list_for_each_entry_reverse(item, &timeout_list, ti_chain) {
 485                 if (item->ti_event == event) {
 486                         ti = item;
 487                         break;
 488                 }
 489                 if (item->ti_timeout < ti->ti_timeout) {
 490                         ti = ptlrpc_new_timeout(time, event, cb, data);
 491                         if (!ti) {
 492                                 ti = ERR_PTR(-ENOMEM);
 493                                 break;
 494                         }
 495                         list_add(&ti->ti_chain, &item->ti_chain);
 496                 }
 497         }
 498         if (!ti) {
 499                 ti = ptlrpc_new_timeout(time, event, cb, data);
 500                 if (ti)
 501                         list_add(&ti->ti_chain, &timeout_list);
 502         }
 503
 504         return ti;
 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_down(&pinger_sem);
 516         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 517         if (!ti) {
 518                 mutex_up(&pinger_sem);
 519                 return (-EINVAL);
 520         }
 521         list_add(obd_list, &ti->ti_obd_list);
 522         mutex_up(&pinger_sem);
 523         return 0;
 524 }
 525
 526 int ptlrpc_del_timeout_client(struct list_head *obd_list)
 527 {
 528         mutex_down(&pinger_sem);
 529         list_del_init(obd_list);
 530         mutex_up(&pinger_sem);
 531         return 0;
 532 }
 533
 534 int ptlrpc_pinger_remove_timeouts(void)
 535 {
 536         struct timeout_item *item, *tmp;
 537
 538         mutex_down(&pinger_sem);
 539         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 540                 LASSERT(list_empty(&item->ti_obd_list));
 541                 list_del(&item->ti_chain);
 542                 OBD_FREE_PTR(item);
 543         }
 544         mutex_up(&pinger_sem);
 545         return 0;
 546 }
 547
 548 void ptlrpc_pinger_wake_up()
 549 {
 550 #ifdef ENABLE_PINGER
 551         pinger_thread->t_flags |= SVC_EVENT;
 552         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 553 #endif
 554 }
 555
 556 /* Ping evictor thread */
 557 #define PET_READY     1
 558 #define PET_TERMINATE 2
 559
 560 static int               pet_refcount = 0;
 561 static int               pet_state;
 562 static wait_queue_head_t pet_waitq;
 563 static struct obd_export *pet_exp = NULL;
 564 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 565
 566 int ping_evictor_wake(struct obd_export *exp)
 567 {
 568         spin_lock(&pet_lock);
 569         if (pet_exp || (pet_state != PET_READY)) {
 570                 /* eventually the new obd will call here again. */
 571                 spin_unlock(&pet_lock);
 572                 return 1;
 573         }
 574
 575         /* We have to make sure the obd isn't destroyed between now and when
 576          * the ping evictor runs.  We'll take a reference here, and drop it
 577          * when we finish in the evictor.  We don't really care about this
 578          * export in particular; we just need one to keep the obd alive. */
 579         pet_exp = class_export_get(exp);
 580         spin_unlock(&pet_lock);
 581
 582         wake_up(&pet_waitq);
 583         return 0;
 584 }
 585
 586 static int ping_evictor_main(void *arg)
 587 {
 588         struct obd_device *obd;
 589         struct obd_export *exp;
 590         struct l_wait_info lwi = { 0 };
 591         time_t expire_time;
 592         ENTRY;
 593
 594         ptlrpc_daemonize("ll_evictor");
 595
 596         CDEBUG(D_HA, "Starting Ping Evictor\n");
 597         pet_exp = NULL;
 598         pet_state = PET_READY;
 599         while (1) {
 600                 l_wait_event(pet_waitq, pet_exp ||
 601                              (pet_state == PET_TERMINATE), &lwi);
 602                 if (pet_state == PET_TERMINATE)
 603                         break;
 604
 605                 /* we only get here if pet_exp != NULL, and the end of this
 606                  * loop is the only place which sets it NULL again, so lock
 607                  * is not strictly necessary. */
 608                 spin_lock(&pet_lock);
 609                 obd = pet_exp->exp_obd;
 610                 spin_unlock(&pet_lock);
 611
 612                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 613
 614                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 615                        obd->obd_name, expire_time);
 616
 617                 /* Exports can't be deleted out of the list while we hold
 618                  * the obd lock (class_unlink_export), which means we can't
 619                  * lose the last ref on the export.  If they've already been
 620                  * removed from the list, we won't find them here. */
 621                 spin_lock(&obd->obd_dev_lock);
 622                 while (!list_empty(&obd->obd_exports_timed)) {
 623                         exp = list_entry(obd->obd_exports_timed.next,
 624                                          struct obd_export,exp_obd_chain_timed);
 625                         if (expire_time > exp->exp_last_request_time) {
 626                                 class_export_get(exp);
 627                                 spin_unlock(&obd->obd_dev_lock);
 628                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 629                                               " (at %s) in %ld seconds. I think"
 630                                               " it's dead, and I am evicting"
 631                                               " it. exp %p, cur %ld expire %ld"
 632                                               " last %ld\n",
 633                                               obd->obd_name,
 634                                               obd_uuid2str(&exp->exp_client_uuid),
 635                                               obd_export_nid2str(exp),
 636                                               (long)(cfs_time_current_sec() -
 637                                                      exp->exp_last_request_time),
 638                                               exp, (long)cfs_time_current_sec(),
 639                                               (long)expire_time,
 640                                               (long)exp->exp_last_request_time);
 641                                 CDEBUG(D_HA, "Last request was at %ld\n",
 642                                        exp->exp_last_request_time);
 643                                 class_fail_export(exp);
 644                                 class_export_put(exp);
 645                                 spin_lock(&obd->obd_dev_lock);
 646                         } else {
 647                                 /* List is sorted, so everyone below is ok */
 648                                 break;
 649                         }
 650                 }
 651                 spin_unlock(&obd->obd_dev_lock);
 652
 653                 class_export_put(pet_exp);
 654
 655                 spin_lock(&pet_lock);
 656                 pet_exp = NULL;
 657                 spin_unlock(&pet_lock);
 658         }
 659         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 660
 661         RETURN(0);
 662 }
 663
 664 void ping_evictor_start(void)
 665 {
 666         int rc;
 667
 668         if (++pet_refcount > 1)
 669                 return;
 670
 671         init_waitqueue_head(&pet_waitq);
 672
 673         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 674         if (rc < 0) {
 675                 pet_refcount--;
 676                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 677         }
 678 }
 679 EXPORT_SYMBOL(ping_evictor_start);
 680
 681 void ping_evictor_stop(void)
 682 {
 683         if (--pet_refcount > 0)
 684                 return;
 685
 686         pet_state = PET_TERMINATE;
 687         wake_up(&pet_waitq);
 688 }
 689 EXPORT_SYMBOL(ping_evictor_stop);
 690 #else /* !__KERNEL__ */
 691
 692 /* XXX
 693  * the current implementation of pinger in liblustre is not optimized
 694  */
 695
 696 #ifdef ENABLE_PINGER
 697 static struct pinger_data {
 698         int             pd_recursion;
 699         cfs_time_t      pd_this_ping;   /* jiffies */
 700         cfs_time_t      pd_next_ping;   /* jiffies */
 701         struct ptlrpc_request_set *pd_set;
 702 } pinger_args;
 703
 704 static int pinger_check_rpcs(void *arg)
 705 {
 706         cfs_time_t curtime = cfs_time_current();
 707         struct ptlrpc_request *req;
 708         struct ptlrpc_request_set *set;
 709         struct list_head *iter;
 710         struct obd_import *imp;
 711         struct pinger_data *pd = &pinger_args;
 712         int rc;
 713
 714         /* prevent recursion */
 715         if (pd->pd_recursion++) {
 716                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 717                 LASSERT(pd->pd_set);
 718                 pd->pd_recursion--;
 719                 return 0;
 720         }
 721
 722         /* have we reached ping point? */
 723         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 724                 pd->pd_recursion--;
 725                 return 0;
 726         }
 727
 728         /* if we have rpc_set already, continue processing it */
 729         if (pd->pd_set) {
 730                 LASSERT(pd->pd_this_ping);
 731                 set = pd->pd_set;
 732                 goto do_check_set;
 733         }
 734
 735         pd->pd_this_ping = curtime;
 736         pd->pd_set = ptlrpc_prep_set();
 737         if (pd->pd_set == NULL)
 738                 goto out;
 739         set = pd->pd_set;
 740
 741         /* add rpcs into set */
 742         mutex_down(&pinger_sem);
 743         list_for_each(iter, &pinger_imports) {
 744                 struct obd_import *imp =
 745                         list_entry(iter, struct obd_import, imp_pinger_chain);
 746                 int generation, level;
 747
 748                 if (cfs_time_aftereq(pd->pd_this_ping,
 749                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 750                         /* Add a ping. */
 751                         spin_lock(&imp->imp_lock);
 752                         generation = imp->imp_generation;
 753                         level = imp->imp_state;
 754                         spin_unlock(&imp->imp_lock);
 755
 756                         if (level != LUSTRE_IMP_FULL) {
 757                                 CDEBUG(D_HA,
 758                                        "not pinging %s (in recovery)\n",
 759                                        obd2cli_tgt(imp->imp_obd));
 760                                 continue;
 761                         }
 762
 763                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 764                                                         LUSTRE_OBD_VERSION,
 765                                                         OBD_PING);
 766                         if (req == NULL) {
 767                                 CERROR("OOM trying to ping %s->%s\n",
 768                                        imp->imp_obd->obd_uuid.uuid,
 769                                        obd2cli_tgt(imp->imp_obd));
 770                                 break;
 771                         }
 772
 773                         req->rq_no_resend = 1;
 774                         ptlrpc_request_set_replen(req);
 775                         req->rq_send_state = LUSTRE_IMP_FULL;
 776                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 777                         req->rq_import_generation = generation;
 778                         ptlrpc_set_add_req(set, req);
 779                 } else {
 780                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 781                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 782                                imp->imp_next_ping, pd->pd_this_ping);
 783                 }
 784         }
 785         pd->pd_this_ping = curtime;
 786         mutex_up(&pinger_sem);
 787
 788         /* Might be empty, that's OK. */
 789         if (set->set_remaining == 0)
 790                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 791
 792         list_for_each(iter, &set->set_requests) {
 793                 struct ptlrpc_request *req =
 794                         list_entry(iter, struct ptlrpc_request,
 795                                    rq_set_chain);
 796                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 797                           req->rq_import->imp_obd->obd_uuid.uuid,
 798                           obd2cli_tgt(req->rq_import->imp_obd));
 799                 (void)ptl_send_rpc(req, 0);
 800         }
 801
 802 do_check_set:
 803         rc = ptlrpc_check_set(NULL, set);
 804
 805         /* not finished, and we are not expired, simply return */
 806         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 807                                             cfs_time_seconds(PING_INTERVAL)))) {
 808                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 809                 pd->pd_recursion--;
 810                 return 0;
 811         }
 812
 813         /* Expire all the requests that didn't come back. */
 814         mutex_down(&pinger_sem);
 815         list_for_each(iter, &set->set_requests) {
 816                 req = list_entry(iter, struct ptlrpc_request,
 817                                  rq_set_chain);
 818
 819                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 820                         continue;
 821
 822                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 823                        req);
 824
 825                 /* This will also unregister reply. */
 826                 ptlrpc_expire_one_request(req, 0);
 827
 828                 /* We're done with this req, let's finally move it to complete
 829                  * phase and take care of inflights. */
 830                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 831                 imp = req->rq_import;
 832                 spin_lock(&imp->imp_lock);
 833                 if (!list_empty(&req->rq_list)) {
 834                         list_del_init(&req->rq_list);
 835                         atomic_dec(&imp->imp_inflight);
 836                 }
 837                 spin_unlock(&imp->imp_lock);
 838                 set->set_remaining--;
 839         }
 840         mutex_up(&pinger_sem);
 841
 842         ptlrpc_set_destroy(set);
 843         pd->pd_set = NULL;
 844
 845 out:
 846         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 847                                         cfs_time_seconds(PING_INTERVAL));
 848         pd->pd_this_ping = 0; /* XXX for debug */
 849
 850         CDEBUG(D_INFO, "finished a round ping\n");
 851         pd->pd_recursion--;
 852         return 0;
 853 }
 854
 855 static void *pinger_callback = NULL;
 856 #endif /* ENABLE_PINGER */
 857
 858 int ptlrpc_start_pinger(void)
 859 {
 860 #ifdef ENABLE_PINGER
 861         memset(&pinger_args, 0, sizeof(pinger_args));
 862         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 863                                                            &pinger_check_rpcs,
 864                                                            &pinger_args);
 865 #endif
 866         return 0;
 867 }
 868
 869 int ptlrpc_stop_pinger(void)
 870 {
 871 #ifdef ENABLE_PINGER
 872         if (pinger_callback)
 873                 liblustre_deregister_wait_callback(pinger_callback);
 874 #endif
 875         return 0;
 876 }
 877
 878 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 879 {
 880 #ifdef ENABLE_PINGER
 881         mutex_down(&pinger_sem);
 882         ptlrpc_update_next_ping(imp);
 883         if (pinger_args.pd_set == NULL &&
 884             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 885                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 886                         imp->imp_next_ping, cfs_time_current());
 887                 pinger_args.pd_next_ping = imp->imp_next_ping;
 888         }
 889         mutex_up(&pinger_sem);
 890 #endif
 891 }
 892
 893 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 894                               timeout_cb_t cb, void *data,
 895                               struct list_head *obd_list)
 896 {
 897         return 0;
 898 }
 899
 900 int ptlrpc_del_timeout_client(struct list_head *obd_list)
 901 {
 902         return 0;
 903 }
 904
 905 int ptlrpc_pinger_add_import(struct obd_import *imp)
 906 {
 907         ENTRY;
 908         if (!list_empty(&imp->imp_pinger_chain))
 909                 RETURN(-EALREADY);
 910
 911         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 912                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 913         ptlrpc_pinger_sending_on_import(imp);
 914
 915         mutex_down(&pinger_sem);
 916         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 917         class_import_get(imp);
 918         mutex_up(&pinger_sem);
 919
 920         RETURN(0);
 921 }
 922
 923 int ptlrpc_pinger_del_import(struct obd_import *imp)
 924 {
 925         ENTRY;
 926         if (list_empty(&imp->imp_pinger_chain))
 927                 RETURN(-ENOENT);
 928
 929         mutex_down(&pinger_sem);
 930         list_del_init(&imp->imp_pinger_chain);
 931         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 932                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 933         class_import_put(imp);
 934         mutex_up(&pinger_sem);
 935         RETURN(0);
 936 }
 937
 938 void ptlrpc_pinger_wake_up()
 939 {
 940 #ifdef ENABLE_PINGER
 941         /* XXX force pinger to run, if needed */
 942         struct obd_import *imp;
 943         ENTRY;
 944         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 945                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 946                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 947 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 948                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 949                     !imp_is_deactive(imp))
 950 #else
 951                 /*XXX only recover for the initial connection */
 952                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 953                     imp->imp_state == LUSTRE_IMP_DISCON &&
 954                     !imp_is_deactive(imp))
 955 #endif
 956                         ptlrpc_initiate_recovery(imp);
 957                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 958                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 959                                      "state %d, deactive %d\n",
 960                                      imp->imp_obd->obd_uuid.uuid,
 961                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 962                                      imp_is_deactive(imp));
 963         }
 964         EXIT;
 965 #endif
 966 }
 967 #endif /* !__KERNEL__ */