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