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, 2012, 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 CFS_LIST_HEAD(pinger_imports);
  56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  57
  58 int ptlrpc_pinger_suppress_pings()
  59 {
  60         return suppress_pings;
  61 }
  62 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  63
  64 struct ptlrpc_request *
  65 ptlrpc_prep_ping(struct obd_import *imp)
  66 {
  67         struct ptlrpc_request *req;
  68
  69         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  70                                         LUSTRE_OBD_VERSION, OBD_PING);
  71         if (req) {
  72                 ptlrpc_request_set_replen(req);
  73                 req->rq_no_resend = req->rq_no_delay = 1;
  74         }
  75         return req;
  76 }
  77
  78 int ptlrpc_obd_ping(struct obd_device *obd)
  79 {
  80         int rc;
  81         struct ptlrpc_request *req;
  82         ENTRY;
  83
  84         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  85         if (req == NULL)
  86                 RETURN(-ENOMEM);
  87
  88         req->rq_send_state = LUSTRE_IMP_FULL;
  89
  90         rc = ptlrpc_queue_wait(req);
  91
  92         ptlrpc_req_finished(req);
  93
  94         RETURN(rc);
  95 }
  96 EXPORT_SYMBOL(ptlrpc_obd_ping);
  97
  98 int ptlrpc_ping(struct obd_import *imp)
  99 {
 100         struct ptlrpc_request *req;
 101         ENTRY;
 102
 103         req = ptlrpc_prep_ping(imp);
 104         if (req == NULL) {
 105                 CERROR("OOM trying to ping %s->%s\n",
 106                        imp->imp_obd->obd_uuid.uuid,
 107                        obd2cli_tgt(imp->imp_obd));
 108                 RETURN(-ENOMEM);
 109         }
 110
 111         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 112                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 113         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 114
 115         RETURN(0);
 116 }
 117
 118 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 119 {
 120 #ifdef ENABLE_PINGER
 121         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 122         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 123                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 124                                   AT_OFF ? 0 :
 125                                   at_get(&imp->imp_at.iat_net_latency));
 126                 time = min(time, dtime);
 127         }
 128         imp->imp_next_ping = cfs_time_shift(time);
 129 #endif /* ENABLE_PINGER */
 130 }
 131
 132 void ptlrpc_ping_import_soon(struct obd_import *imp)
 133 {
 134         imp->imp_next_ping = cfs_time_current();
 135 }
 136
 137 static inline int imp_is_deactive(struct obd_import *imp)
 138 {
 139         return (imp->imp_deactive ||
 140                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 141 }
 142
 143 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 144 {
 145         if (imp->imp_server_timeout)
 146                 return cfs_time_shift(obd_timeout / 2);
 147         else
 148                 return cfs_time_shift(obd_timeout);
 149 }
 150
 151 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 152 static cfs_time_t suspend_wakeup_time = 0;
 153
 154 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 155 {
 156         struct timeout_item *item;
 157         cfs_time_t timeout = PING_INTERVAL;
 158
 159         /* The timeout list is a increase order sorted list */
 160         mutex_lock(&pinger_mutex);
 161         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 162                 int ti_timeout = item->ti_timeout;
 163                 if (timeout > ti_timeout)
 164                         timeout = ti_timeout;
 165                 break;
 166         }
 167         mutex_unlock(&pinger_mutex);
 168
 169         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 170                                          cfs_time_current());
 171 }
 172
 173 static cfs_waitq_t suspend_timeouts_waitq;
 174
 175 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 176 {
 177         return suspend_wakeup_time;
 178 }
 179
 180 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 181 {
 182         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 183 #if 0
 184         if (imp->imp_no_timeout)
 185                 return;
 186         imp->imp_no_timeout = 1;
 187         cfs_atomic_inc(&suspend_timeouts);
 188         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 189                cfs_atomic_read(&suspend_timeouts));
 190 #endif
 191 }
 192
 193 void ptlrpc_activate_timeouts(struct obd_import *imp)
 194 {
 195         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 196 #if 0
 197         if (!imp->imp_no_timeout)
 198                 return;
 199         imp->imp_no_timeout = 0;
 200         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 201         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 202                 suspend_wakeup_time = cfs_time_current();
 203                 cfs_waitq_signal(&suspend_timeouts_waitq);
 204         }
 205         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 206                cfs_atomic_read(&suspend_timeouts));
 207 #endif
 208 }
 209
 210 int ptlrpc_check_suspend(void)
 211 {
 212         if (cfs_atomic_read(&suspend_timeouts))
 213                 return 1;
 214         return 0;
 215 }
 216
 217 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 218 {
 219         struct l_wait_info lwi;
 220
 221         if (cfs_atomic_read(&suspend_timeouts)) {
 222                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 223                           cfs_atomic_read(&suspend_timeouts));
 224                 lwi = LWI_INTR(NULL, NULL);
 225                 l_wait_event(suspend_timeouts_waitq,
 226                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 227                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 228                 return 1;
 229         }
 230         return 0;
 231 }
 232
 233 #ifdef __KERNEL__
 234
 235 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 236                                          unsigned long this_ping)
 237 {
 238         int level;
 239         int force;
 240         int force_next;
 241         int suppress;
 242
 243         spin_lock(&imp->imp_lock);
 244
 245         level = imp->imp_state;
 246         force = imp->imp_force_verify;
 247         force_next = imp->imp_force_next_verify;
 248         /*
 249          * This will be used below only if the import is "FULL".
 250          */
 251         suppress = !!(imp->imp_connect_data.ocd_connect_flags &
 252                       OBD_CONNECT_PINGLESS);
 253
 254         imp->imp_force_verify = 0;
 255
 256         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 257             !force) {
 258                 spin_unlock(&imp->imp_lock);
 259                 return;
 260         }
 261
 262         imp->imp_force_next_verify = 0;
 263
 264         spin_unlock(&imp->imp_lock);
 265
 266         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 267                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 268                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 269                ptlrpc_import_state_name(level), level, force, force_next,
 270                imp->imp_deactive, imp->imp_pingable, suppress);
 271
 272         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 273                 /* wait for a while before trying recovery again */
 274                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 275                 if (!imp->imp_no_pinger_recover)
 276                         ptlrpc_initiate_recovery(imp);
 277         } else if (level != LUSTRE_IMP_FULL ||
 278                    imp->imp_obd->obd_no_recov ||
 279                    imp_is_deactive(imp)) {
 280                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 281                        "or recovery disabled: %s)\n",
 282                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 283                        ptlrpc_import_state_name(level));
 284         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 285                 ptlrpc_ping(imp);
 286         }
 287 }
 288
 289 static int ptlrpc_pinger_main(void *arg)
 290 {
 291         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 292         ENTRY;
 293
 294         cfs_daemonize(thread->t_name);
 295
 296         /* Record that the thread is running */
 297         thread_set_flags(thread, SVC_RUNNING);
 298         cfs_waitq_signal(&thread->t_ctl_waitq);
 299
 300         /* And now, loop forever, pinging as needed. */
 301         while (1) {
 302                 cfs_time_t this_ping = cfs_time_current();
 303                 struct l_wait_info lwi;
 304                 cfs_duration_t time_to_next_wake;
 305                 struct timeout_item *item;
 306                 cfs_list_t *iter;
 307
 308                 mutex_lock(&pinger_mutex);
 309                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 310                         item->ti_cb(item, item->ti_cb_data);
 311                 }
 312                 cfs_list_for_each(iter, &pinger_imports) {
 313                         struct obd_import *imp =
 314                                 cfs_list_entry(iter, struct obd_import,
 315                                                imp_pinger_chain);
 316
 317                         ptlrpc_pinger_process_import(imp, this_ping);
 318                         /* obd_timeout might have changed */
 319                         if (imp->imp_pingable && imp->imp_next_ping &&
 320                             cfs_time_after(imp->imp_next_ping,
 321                                            cfs_time_add(this_ping,
 322                                                         cfs_time_seconds(PING_INTERVAL))))
 323                                 ptlrpc_update_next_ping(imp, 0);
 324                 }
 325                 mutex_unlock(&pinger_mutex);
 326                 /* update memory usage info */
 327                 obd_update_maxusage();
 328
 329                 /* Wait until the next ping time, or until we're stopped. */
 330                 time_to_next_wake = pinger_check_timeout(this_ping);
 331                 /* The ping sent by ptlrpc_send_rpc may get sent out
 332                    say .01 second after this.
 333                    ptlrpc_pinger_sending_on_import will then set the
 334                    next ping time to next_ping + .01 sec, which means
 335                    we will SKIP the next ping at next_ping, and the
 336                    ping will get sent 2 timeouts from now!  Beware. */
 337                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 338                        CFS_TIME_T")\n", time_to_next_wake,
 339                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 340                 if (time_to_next_wake > 0) {
 341                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 342                                                 time_to_next_wake,
 343                                                 cfs_time_seconds(1)),
 344                                           NULL, NULL);
 345                         l_wait_event(thread->t_ctl_waitq,
 346                                      thread_is_stopping(thread) ||
 347                                      thread_is_event(thread),
 348                                      &lwi);
 349                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 350                                 EXIT;
 351                                 break;
 352                         } else {
 353                                 /* woken after adding import to reset timer */
 354                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 355                         }
 356                 }
 357         }
 358
 359         thread_set_flags(thread, SVC_STOPPED);
 360         cfs_waitq_signal(&thread->t_ctl_waitq);
 361
 362         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 363         return 0;
 364 }
 365
 366 static struct ptlrpc_thread *pinger_thread = NULL;
 367
 368 int ptlrpc_start_pinger(void)
 369 {
 370         struct l_wait_info lwi = { 0 };
 371         int rc;
 372 #ifndef ENABLE_PINGER
 373         return 0;
 374 #endif
 375         ENTRY;
 376
 377         if (pinger_thread != NULL)
 378                 RETURN(-EALREADY);
 379
 380         OBD_ALLOC_PTR(pinger_thread);
 381         if (pinger_thread == NULL)
 382                 RETURN(-ENOMEM);
 383         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 384         cfs_waitq_init(&suspend_timeouts_waitq);
 385
 386         strcpy(pinger_thread->t_name, "ll_ping");
 387
 388         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 389          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 390         rc = cfs_create_thread(ptlrpc_pinger_main,
 391                                pinger_thread, CFS_DAEMON_FLAGS);
 392         if (rc < 0) {
 393                 CERROR("cannot start thread: %d\n", rc);
 394                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 395                 pinger_thread = NULL;
 396                 RETURN(rc);
 397         }
 398         l_wait_event(pinger_thread->t_ctl_waitq,
 399                      thread_is_running(pinger_thread), &lwi);
 400
 401         if (suppress_pings)
 402                 CWARN("Pings will be suppressed at the request of the "
 403                       "administrator.  The configuration shall meet the "
 404                       "additional requirements described in the manual.  "
 405                       "(Search for the \"suppress_pings\" kernel module "
 406                       "parameter.)\n");
 407
 408         RETURN(0);
 409 }
 410
 411 int ptlrpc_pinger_remove_timeouts(void);
 412
 413 int ptlrpc_stop_pinger(void)
 414 {
 415         struct l_wait_info lwi = { 0 };
 416         int rc = 0;
 417 #ifndef ENABLE_PINGER
 418         return 0;
 419 #endif
 420         ENTRY;
 421
 422         if (pinger_thread == NULL)
 423                 RETURN(-EALREADY);
 424
 425         ptlrpc_pinger_remove_timeouts();
 426         mutex_lock(&pinger_mutex);
 427         thread_set_flags(pinger_thread, SVC_STOPPING);
 428         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 429         mutex_unlock(&pinger_mutex);
 430
 431         l_wait_event(pinger_thread->t_ctl_waitq,
 432                      thread_is_stopped(pinger_thread), &lwi);
 433
 434         OBD_FREE_PTR(pinger_thread);
 435         pinger_thread = NULL;
 436         RETURN(rc);
 437 }
 438
 439 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 440 {
 441         ptlrpc_update_next_ping(imp, 0);
 442 }
 443 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 444
 445 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 446 {
 447         ptlrpc_update_next_ping(imp, 1);
 448         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 449         /*
 450          * Avoid reading stale imp_connect_data.  When not sure if pings are
 451          * expected or not on next connection, we assume they are not and force
 452          * one anyway to guarantee the chance of updating
 453          * imp_peer_committed_transno.
 454          */
 455         if (imp->imp_state != LUSTRE_IMP_FULL ||
 456             imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_PINGLESS)
 457                 imp->imp_force_next_verify = 1;
 458 }
 459
 460 int ptlrpc_pinger_add_import(struct obd_import *imp)
 461 {
 462         ENTRY;
 463         if (!cfs_list_empty(&imp->imp_pinger_chain))
 464                 RETURN(-EALREADY);
 465
 466         mutex_lock(&pinger_mutex);
 467         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 468                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 469         /* if we add to pinger we want recovery on this import */
 470         imp->imp_obd->obd_no_recov = 0;
 471         ptlrpc_update_next_ping(imp, 0);
 472         /* XXX sort, blah blah */
 473         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 474         class_import_get(imp);
 475
 476         ptlrpc_pinger_wake_up();
 477         mutex_unlock(&pinger_mutex);
 478
 479         RETURN(0);
 480 }
 481 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 482
 483 int ptlrpc_pinger_del_import(struct obd_import *imp)
 484 {
 485         ENTRY;
 486         if (cfs_list_empty(&imp->imp_pinger_chain))
 487                 RETURN(-ENOENT);
 488
 489         mutex_lock(&pinger_mutex);
 490         cfs_list_del_init(&imp->imp_pinger_chain);
 491         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 492                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 493         /* if we remove from pinger we don't want recovery on this import */
 494         imp->imp_obd->obd_no_recov = 1;
 495         class_import_put(imp);
 496         mutex_unlock(&pinger_mutex);
 497         RETURN(0);
 498 }
 499 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 500
 501 /**
 502  * Register a timeout callback to the pinger list, and the callback will
 503  * be called when timeout happens.
 504  */
 505 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 506                                         timeout_cb_t cb, void *data)
 507 {
 508         struct timeout_item *ti;
 509
 510         OBD_ALLOC_PTR(ti);
 511         if (!ti)
 512                 return(NULL);
 513
 514         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 515         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 516         ti->ti_timeout = time;
 517         ti->ti_event = event;
 518         ti->ti_cb = cb;
 519         ti->ti_cb_data = data;
 520
 521         return ti;
 522 }
 523
 524 /**
 525  * Register timeout event on the the pinger thread.
 526  * Note: the timeout list is an sorted list with increased timeout value.
 527  */
 528 static struct timeout_item*
 529 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 530                                timeout_cb_t cb, void *data)
 531 {
 532         struct timeout_item *item, *tmp;
 533
 534         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 535
 536         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 537                 if (item->ti_event == event)
 538                         goto out;
 539
 540         item = ptlrpc_new_timeout(time, event, cb, data);
 541         if (item) {
 542                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 543                         if (tmp->ti_timeout < time) {
 544                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 545                                 goto out;
 546                         }
 547                 }
 548                 cfs_list_add(&item->ti_chain, &timeout_list);
 549         }
 550 out:
 551         return item;
 552 }
 553
 554 /* Add a client_obd to the timeout event list, when timeout(@time)
 555  * happens, the callback(@cb) will be called.
 556  */
 557 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 558                               timeout_cb_t cb, void *data,
 559                               cfs_list_t *obd_list)
 560 {
 561         struct timeout_item *ti;
 562
 563         mutex_lock(&pinger_mutex);
 564         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 565         if (!ti) {
 566                 mutex_unlock(&pinger_mutex);
 567                 return (-EINVAL);
 568         }
 569         cfs_list_add(obd_list, &ti->ti_obd_list);
 570         mutex_unlock(&pinger_mutex);
 571         return 0;
 572 }
 573 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 574
 575 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 576                               enum timeout_event event)
 577 {
 578         struct timeout_item *ti = NULL, *item;
 579
 580         if (cfs_list_empty(obd_list))
 581                 return 0;
 582         mutex_lock(&pinger_mutex);
 583         cfs_list_del_init(obd_list);
 584         /**
 585          * If there are no obd attached to the timeout event
 586          * list, remove this timeout event from the pinger
 587          */
 588         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 589                 if (item->ti_event == event) {
 590                         ti = item;
 591                         break;
 592                 }
 593         }
 594         LASSERTF(ti != NULL, "ti is NULL ! \n");
 595         if (cfs_list_empty(&ti->ti_obd_list)) {
 596                 cfs_list_del(&ti->ti_chain);
 597                 OBD_FREE_PTR(ti);
 598         }
 599         mutex_unlock(&pinger_mutex);
 600         return 0;
 601 }
 602 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 603
 604 int ptlrpc_pinger_remove_timeouts(void)
 605 {
 606         struct timeout_item *item, *tmp;
 607
 608         mutex_lock(&pinger_mutex);
 609         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 610                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 611                 cfs_list_del(&item->ti_chain);
 612                 OBD_FREE_PTR(item);
 613         }
 614         mutex_unlock(&pinger_mutex);
 615         return 0;
 616 }
 617
 618 void ptlrpc_pinger_wake_up()
 619 {
 620 #ifdef ENABLE_PINGER
 621         thread_add_flags(pinger_thread, SVC_EVENT);
 622         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 623 #endif
 624 }
 625
 626 /* Ping evictor thread */
 627 #define PET_READY     1
 628 #define PET_TERMINATE 2
 629
 630 static int               pet_refcount = 0;
 631 static int               pet_state;
 632 static cfs_waitq_t       pet_waitq;
 633 CFS_LIST_HEAD(pet_list);
 634 static DEFINE_SPINLOCK(pet_lock);
 635
 636 int ping_evictor_wake(struct obd_export *exp)
 637 {
 638         struct obd_device *obd;
 639
 640         spin_lock(&pet_lock);
 641         if (pet_state != PET_READY) {
 642                 /* eventually the new obd will call here again. */
 643                 spin_unlock(&pet_lock);
 644                 return 1;
 645         }
 646
 647         obd = class_exp2obd(exp);
 648         if (cfs_list_empty(&obd->obd_evict_list)) {
 649                 class_incref(obd, "evictor", obd);
 650                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 651         }
 652         spin_unlock(&pet_lock);
 653
 654         cfs_waitq_signal(&pet_waitq);
 655         return 0;
 656 }
 657
 658 static int ping_evictor_main(void *arg)
 659 {
 660         struct obd_device *obd;
 661         struct obd_export *exp;
 662         struct l_wait_info lwi = { 0 };
 663         time_t expire_time;
 664         ENTRY;
 665
 666         cfs_daemonize_ctxt("ll_evictor");
 667
 668         CDEBUG(D_HA, "Starting Ping Evictor\n");
 669         pet_state = PET_READY;
 670         while (1) {
 671                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 672                              (pet_state == PET_TERMINATE), &lwi);
 673
 674                 /* loop until all obd's will be removed */
 675                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 676                         break;
 677
 678                 /* we only get here if pet_exp != NULL, and the end of this
 679                  * loop is the only place which sets it NULL again, so lock
 680                  * is not strictly necessary. */
 681                 spin_lock(&pet_lock);
 682                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 683                                      obd_evict_list);
 684                 spin_unlock(&pet_lock);
 685
 686                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 687
 688                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 689                        obd->obd_name, expire_time);
 690
 691                 /* Exports can't be deleted out of the list while we hold
 692                  * the obd lock (class_unlink_export), which means we can't
 693                  * lose the last ref on the export.  If they've already been
 694                  * removed from the list, we won't find them here. */
 695                 spin_lock(&obd->obd_dev_lock);
 696                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 697                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 698                                              struct obd_export,
 699                                              exp_obd_chain_timed);
 700                         if (expire_time > exp->exp_last_request_time) {
 701                                 class_export_get(exp);
 702                                 spin_unlock(&obd->obd_dev_lock);
 703                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 704                                               " (at %s) in %ld seconds. I think"
 705                                               " it's dead, and I am evicting"
 706                                               " it. exp %p, cur %ld expire %ld"
 707                                               " last %ld\n",
 708                                               obd->obd_name,
 709                                               obd_uuid2str(&exp->exp_client_uuid),
 710                                               obd_export_nid2str(exp),
 711                                               (long)(cfs_time_current_sec() -
 712                                                      exp->exp_last_request_time),
 713                                               exp, (long)cfs_time_current_sec(),
 714                                               (long)expire_time,
 715                                               (long)exp->exp_last_request_time);
 716                                 CDEBUG(D_HA, "Last request was at %ld\n",
 717                                        exp->exp_last_request_time);
 718                                 class_fail_export(exp);
 719                                 class_export_put(exp);
 720                                 spin_lock(&obd->obd_dev_lock);
 721                         } else {
 722                                 /* List is sorted, so everyone below is ok */
 723                                 break;
 724                         }
 725                 }
 726                 spin_unlock(&obd->obd_dev_lock);
 727
 728                 spin_lock(&pet_lock);
 729                 cfs_list_del_init(&obd->obd_evict_list);
 730                 spin_unlock(&pet_lock);
 731
 732                 class_decref(obd, "evictor", obd);
 733         }
 734         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 735
 736         RETURN(0);
 737 }
 738
 739 void ping_evictor_start(void)
 740 {
 741         int rc;
 742
 743         if (++pet_refcount > 1)
 744                 return;
 745
 746         cfs_waitq_init(&pet_waitq);
 747
 748         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 749         if (rc < 0) {
 750                 pet_refcount--;
 751                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 752         }
 753 }
 754 EXPORT_SYMBOL(ping_evictor_start);
 755
 756 void ping_evictor_stop(void)
 757 {
 758         if (--pet_refcount > 0)
 759                 return;
 760
 761         pet_state = PET_TERMINATE;
 762         cfs_waitq_signal(&pet_waitq);
 763 }
 764 EXPORT_SYMBOL(ping_evictor_stop);
 765 #else /* !__KERNEL__ */
 766
 767 /* XXX
 768  * the current implementation of pinger in liblustre is not optimized
 769  */
 770
 771 #ifdef ENABLE_PINGER
 772 static struct pinger_data {
 773         int             pd_recursion;
 774         cfs_time_t      pd_this_ping;   /* jiffies */
 775         cfs_time_t      pd_next_ping;   /* jiffies */
 776         struct ptlrpc_request_set *pd_set;
 777 } pinger_args;
 778
 779 static int pinger_check_rpcs(void *arg)
 780 {
 781         cfs_time_t curtime = cfs_time_current();
 782         struct ptlrpc_request *req;
 783         struct ptlrpc_request_set *set;
 784         cfs_list_t *iter;
 785         struct obd_import *imp;
 786         struct pinger_data *pd = &pinger_args;
 787         int rc;
 788
 789         /* prevent recursion */
 790         if (pd->pd_recursion++) {
 791                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 792                 LASSERT(pd->pd_set);
 793                 pd->pd_recursion--;
 794                 return 0;
 795         }
 796
 797         /* have we reached ping point? */
 798         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 799                 pd->pd_recursion--;
 800                 return 0;
 801         }
 802
 803         /* if we have rpc_set already, continue processing it */
 804         if (pd->pd_set) {
 805                 LASSERT(pd->pd_this_ping);
 806                 set = pd->pd_set;
 807                 goto do_check_set;
 808         }
 809
 810         pd->pd_this_ping = curtime;
 811         pd->pd_set = ptlrpc_prep_set();
 812         if (pd->pd_set == NULL)
 813                 goto out;
 814         set = pd->pd_set;
 815
 816         /* add rpcs into set */
 817         mutex_lock(&pinger_mutex);
 818         cfs_list_for_each(iter, &pinger_imports) {
 819                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 820                                                         imp_pinger_chain);
 821                 int generation, level;
 822
 823                 if (cfs_time_aftereq(pd->pd_this_ping,
 824                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 825                         /* Add a ping. */
 826                         spin_lock(&imp->imp_lock);
 827                         generation = imp->imp_generation;
 828                         level = imp->imp_state;
 829                         spin_unlock(&imp->imp_lock);
 830
 831                         if (level != LUSTRE_IMP_FULL) {
 832                                 CDEBUG(D_HA,
 833                                        "not pinging %s (in recovery)\n",
 834                                        obd2cli_tgt(imp->imp_obd));
 835                                 continue;
 836                         }
 837
 838                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 839                                                         LUSTRE_OBD_VERSION,
 840                                                         OBD_PING);
 841                         if (req == NULL) {
 842                                 CERROR("OOM trying to ping %s->%s\n",
 843                                        imp->imp_obd->obd_uuid.uuid,
 844                                        obd2cli_tgt(imp->imp_obd));
 845                                 break;
 846                         }
 847
 848                         req->rq_no_resend = 1;
 849                         ptlrpc_request_set_replen(req);
 850                         req->rq_send_state = LUSTRE_IMP_FULL;
 851                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 852                         req->rq_import_generation = generation;
 853                         ptlrpc_set_add_req(set, req);
 854                 } else {
 855                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 856                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 857                                imp->imp_next_ping, pd->pd_this_ping);
 858                 }
 859         }
 860         pd->pd_this_ping = curtime;
 861         mutex_unlock(&pinger_mutex);
 862
 863         /* Might be empty, that's OK. */
 864         if (cfs_atomic_read(&set->set_remaining) == 0)
 865                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 866
 867         cfs_list_for_each(iter, &set->set_requests) {
 868                 struct ptlrpc_request *req =
 869                         cfs_list_entry(iter, struct ptlrpc_request,
 870                                        rq_set_chain);
 871                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 872                           req->rq_import->imp_obd->obd_uuid.uuid,
 873                           obd2cli_tgt(req->rq_import->imp_obd));
 874                 (void)ptl_send_rpc(req, 0);
 875         }
 876
 877 do_check_set:
 878         rc = ptlrpc_check_set(NULL, set);
 879
 880         /* not finished, and we are not expired, simply return */
 881         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 882                                             cfs_time_seconds(PING_INTERVAL)))) {
 883                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 884                 pd->pd_recursion--;
 885                 return 0;
 886         }
 887
 888         /* Expire all the requests that didn't come back. */
 889         mutex_lock(&pinger_mutex);
 890         cfs_list_for_each(iter, &set->set_requests) {
 891                 req = cfs_list_entry(iter, struct ptlrpc_request,
 892                                      rq_set_chain);
 893
 894                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 895                         continue;
 896
 897                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 898                        req);
 899
 900                 /* This will also unregister reply. */
 901                 ptlrpc_expire_one_request(req, 0);
 902
 903                 /* We're done with this req, let's finally move it to complete
 904                  * phase and take care of inflights. */
 905                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 906                 imp = req->rq_import;
 907                 spin_lock(&imp->imp_lock);
 908                 if (!cfs_list_empty(&req->rq_list)) {
 909                         cfs_list_del_init(&req->rq_list);
 910                         cfs_atomic_dec(&imp->imp_inflight);
 911                 }
 912                 spin_unlock(&imp->imp_lock);
 913                 cfs_atomic_dec(&set->set_remaining);
 914         }
 915         mutex_unlock(&pinger_mutex);
 916
 917         ptlrpc_set_destroy(set);
 918         pd->pd_set = NULL;
 919
 920 out:
 921         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 922                                         cfs_time_seconds(PING_INTERVAL));
 923         pd->pd_this_ping = 0; /* XXX for debug */
 924
 925         CDEBUG(D_INFO, "finished a round ping\n");
 926         pd->pd_recursion--;
 927         return 0;
 928 }
 929
 930 static void *pinger_callback = NULL;
 931 #endif /* ENABLE_PINGER */
 932
 933 int ptlrpc_start_pinger(void)
 934 {
 935 #ifdef ENABLE_PINGER
 936         memset(&pinger_args, 0, sizeof(pinger_args));
 937         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 938                                                            &pinger_check_rpcs,
 939                                                            &pinger_args);
 940 #endif
 941         return 0;
 942 }
 943
 944 int ptlrpc_stop_pinger(void)
 945 {
 946 #ifdef ENABLE_PINGER
 947         if (pinger_callback)
 948                 liblustre_deregister_wait_callback(pinger_callback);
 949 #endif
 950         return 0;
 951 }
 952
 953 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 954 {
 955 #ifdef ENABLE_PINGER
 956         mutex_lock(&pinger_mutex);
 957         ptlrpc_update_next_ping(imp, 0);
 958         if (pinger_args.pd_set == NULL &&
 959             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 960                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 961                         imp->imp_next_ping, cfs_time_current());
 962                 pinger_args.pd_next_ping = imp->imp_next_ping;
 963         }
 964         mutex_unlock(&pinger_mutex);
 965 #endif
 966 }
 967
 968 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 969 {
 970 #ifdef ENABLE_PINGER
 971         mutex_lock(&pinger_mutex);
 972         ptlrpc_update_next_ping(imp, 1);
 973         if (pinger_args.pd_set == NULL &&
 974             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 975                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 976                         imp->imp_next_ping, cfs_time_current());
 977                 pinger_args.pd_next_ping = imp->imp_next_ping;
 978         }
 979         mutex_unlock(&pinger_mutex);
 980 #endif
 981 }
 982
 983 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 984                               timeout_cb_t cb, void *data,
 985                               cfs_list_t *obd_list)
 986 {
 987         return 0;
 988 }
 989
 990 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 991                               enum timeout_event event)
 992 {
 993         return 0;
 994 }
 995
 996 int ptlrpc_pinger_add_import(struct obd_import *imp)
 997 {
 998         ENTRY;
 999         if (!cfs_list_empty(&imp->imp_pinger_chain))
1000                 RETURN(-EALREADY);
1001
1002         CDEBUG(D_HA, "adding pingable import %s->%s\n",
1003                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1004         ptlrpc_pinger_sending_on_import(imp);
1005
1006         mutex_lock(&pinger_mutex);
1007         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1008         class_import_get(imp);
1009         mutex_unlock(&pinger_mutex);
1010
1011         RETURN(0);
1012 }
1013
1014 int ptlrpc_pinger_del_import(struct obd_import *imp)
1015 {
1016         ENTRY;
1017         if (cfs_list_empty(&imp->imp_pinger_chain))
1018                 RETURN(-ENOENT);
1019
1020         mutex_lock(&pinger_mutex);
1021         cfs_list_del_init(&imp->imp_pinger_chain);
1022         CDEBUG(D_HA, "removing pingable import %s->%s\n",
1023                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1024         class_import_put(imp);
1025         mutex_unlock(&pinger_mutex);
1026         RETURN(0);
1027 }
1028
1029 void ptlrpc_pinger_wake_up()
1030 {
1031 #ifdef ENABLE_PINGER
1032         /* XXX force pinger to run, if needed */
1033         struct obd_import *imp;
1034         ENTRY;
1035         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1036                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1037                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1038 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1039                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1040                     !imp_is_deactive(imp))
1041 #else
1042                 /*XXX only recover for the initial connection */
1043                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1044                     imp->imp_state == LUSTRE_IMP_DISCON &&
1045                     !imp_is_deactive(imp))
1046 #endif
1047                         ptlrpc_initiate_recovery(imp);
1048                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1049                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1050                                      "state %d, deactive %d\n",
1051                                      imp->imp_obd->obd_uuid.uuid,
1052                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1053                                      imp_is_deactive(imp));
1054         }
1055         EXIT;
1056 #endif
1057 }
1058 #endif /* !__KERNEL__ */