lustre/ptlrpc/pinger.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 static int suppress_pings;
  52 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  53
  54 struct mutex pinger_mutex;
  55 static 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 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 152 {
 153         struct timeout_item *item;
 154         cfs_time_t timeout = PING_INTERVAL;
 155
 156         /* The timeout list is a increase order sorted list */
 157         mutex_lock(&pinger_mutex);
 158         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 159                 int ti_timeout = item->ti_timeout;
 160                 if (timeout > ti_timeout)
 161                         timeout = ti_timeout;
 162                 break;
 163         }
 164         mutex_unlock(&pinger_mutex);
 165
 166         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 167                                          cfs_time_current());
 168 }
 169
 170 #ifdef __KERNEL__
 171
 172 static bool ir_up;
 173
 174 void ptlrpc_pinger_ir_up(void)
 175 {
 176         CDEBUG(D_HA, "IR up\n");
 177         ir_up = true;
 178 }
 179 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 180
 181 void ptlrpc_pinger_ir_down(void)
 182 {
 183         CDEBUG(D_HA, "IR down\n");
 184         ir_up = false;
 185 }
 186 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 187
 188 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 189                                          unsigned long this_ping)
 190 {
 191         int level;
 192         int force;
 193         int force_next;
 194         int suppress;
 195
 196         spin_lock(&imp->imp_lock);
 197
 198         level = imp->imp_state;
 199         force = imp->imp_force_verify;
 200         force_next = imp->imp_force_next_verify;
 201         /*
 202          * This will be used below only if the import is "FULL".
 203          */
 204         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 205
 206         imp->imp_force_verify = 0;
 207
 208         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 209             !force) {
 210                 spin_unlock(&imp->imp_lock);
 211                 return;
 212         }
 213
 214         imp->imp_force_next_verify = 0;
 215
 216         spin_unlock(&imp->imp_lock);
 217
 218         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 219                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 220                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 221                ptlrpc_import_state_name(level), level, force, force_next,
 222                imp->imp_deactive, imp->imp_pingable, suppress);
 223
 224         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 225                 /* wait for a while before trying recovery again */
 226                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 227                 if (!imp->imp_no_pinger_recover)
 228                         ptlrpc_initiate_recovery(imp);
 229         } else if (level != LUSTRE_IMP_FULL ||
 230                    imp->imp_obd->obd_no_recov ||
 231                    imp_is_deactive(imp)) {
 232                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 233                        "or recovery disabled: %s)\n",
 234                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 235                        ptlrpc_import_state_name(level));
 236                 if (force) {
 237                         spin_lock(&imp->imp_lock);
 238                         imp->imp_force_verify = 1;
 239                         spin_unlock(&imp->imp_lock);
 240                 }
 241         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 242                 ptlrpc_ping(imp);
 243         }
 244 }
 245
 246 static int ptlrpc_pinger_main(void *arg)
 247 {
 248         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 249         ENTRY;
 250
 251         /* Record that the thread is running */
 252         thread_set_flags(thread, SVC_RUNNING);
 253         wake_up(&thread->t_ctl_waitq);
 254
 255         /* And now, loop forever, pinging as needed. */
 256         while (1) {
 257                 cfs_time_t this_ping = cfs_time_current();
 258                 struct l_wait_info lwi;
 259                 cfs_duration_t time_to_next_wake;
 260                 struct timeout_item *item;
 261                 cfs_list_t *iter;
 262
 263                 mutex_lock(&pinger_mutex);
 264                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 265                         item->ti_cb(item, item->ti_cb_data);
 266                 }
 267                 cfs_list_for_each(iter, &pinger_imports) {
 268                         struct obd_import *imp =
 269                                 cfs_list_entry(iter, struct obd_import,
 270                                                imp_pinger_chain);
 271
 272                         ptlrpc_pinger_process_import(imp, this_ping);
 273                         /* obd_timeout might have changed */
 274                         if (imp->imp_pingable && imp->imp_next_ping &&
 275                             cfs_time_after(imp->imp_next_ping,
 276                                            cfs_time_add(this_ping,
 277                                                         cfs_time_seconds(PING_INTERVAL))))
 278                                 ptlrpc_update_next_ping(imp, 0);
 279                 }
 280                 mutex_unlock(&pinger_mutex);
 281                 /* update memory usage info */
 282                 obd_update_maxusage();
 283
 284                 /* Wait until the next ping time, or until we're stopped. */
 285                 time_to_next_wake = pinger_check_timeout(this_ping);
 286                 /* The ping sent by ptlrpc_send_rpc may get sent out
 287                    say .01 second after this.
 288                    ptlrpc_pinger_sending_on_import will then set the
 289                    next ping time to next_ping + .01 sec, which means
 290                    we will SKIP the next ping at next_ping, and the
 291                    ping will get sent 2 timeouts from now!  Beware. */
 292                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 293                        CFS_TIME_T")\n", time_to_next_wake,
 294                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 295                 if (time_to_next_wake > 0) {
 296                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 297                                                 time_to_next_wake,
 298                                                 cfs_time_seconds(1)),
 299                                           NULL, NULL);
 300                         l_wait_event(thread->t_ctl_waitq,
 301                                      thread_is_stopping(thread) ||
 302                                      thread_is_event(thread),
 303                                      &lwi);
 304                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 305                                 EXIT;
 306                                 break;
 307                         } else {
 308                                 /* woken after adding import to reset timer */
 309                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 310                         }
 311                 }
 312         }
 313
 314         thread_set_flags(thread, SVC_STOPPED);
 315         wake_up(&thread->t_ctl_waitq);
 316
 317         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
 318         return 0;
 319 }
 320
 321 static struct ptlrpc_thread pinger_thread;
 322
 323 int ptlrpc_start_pinger(void)
 324 {
 325         struct l_wait_info lwi = { 0 };
 326         struct task_struct *task;
 327         int rc;
 328 #ifndef ENABLE_PINGER
 329         return 0;
 330 #endif
 331         ENTRY;
 332
 333         if (!thread_is_init(&pinger_thread) &&
 334             !thread_is_stopped(&pinger_thread))
 335                 RETURN(-EALREADY);
 336
 337         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 338
 339         strcpy(pinger_thread.t_name, "ll_ping");
 340
 341         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 342          * just drop the VM and FILES in kthread_run() right away. */
 343         task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
 344                            pinger_thread.t_name);
 345         if (IS_ERR(task)) {
 346                 rc = PTR_ERR(task);
 347                 CERROR("cannot start pinger thread: rc = %d\n", rc);
 348                 RETURN(rc);
 349         }
 350
 351         l_wait_event(pinger_thread.t_ctl_waitq,
 352                      thread_is_running(&pinger_thread), &lwi);
 353
 354         if (suppress_pings)
 355                 CWARN("Pings will be suppressed at the request of the "
 356                       "administrator.  The configuration shall meet the "
 357                       "additional requirements described in the manual.  "
 358                       "(Search for the \"suppress_pings\" kernel module "
 359                       "parameter.)\n");
 360
 361         RETURN(0);
 362 }
 363
 364 int ptlrpc_pinger_remove_timeouts(void);
 365
 366 int ptlrpc_stop_pinger(void)
 367 {
 368         struct l_wait_info lwi = { 0 };
 369 #ifndef ENABLE_PINGER
 370         return 0;
 371 #endif
 372         ENTRY;
 373
 374         if (thread_is_init(&pinger_thread) ||
 375             thread_is_stopped(&pinger_thread))
 376                 RETURN(-EALREADY);
 377
 378         ptlrpc_pinger_remove_timeouts();
 379
 380         thread_set_flags(&pinger_thread, SVC_STOPPING);
 381         wake_up(&pinger_thread.t_ctl_waitq);
 382
 383         l_wait_event(pinger_thread.t_ctl_waitq,
 384                      thread_is_stopped(&pinger_thread), &lwi);
 385         RETURN(0);
 386 }
 387
 388 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 389 {
 390         ptlrpc_update_next_ping(imp, 0);
 391 }
 392 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 393
 394 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 395 {
 396         ptlrpc_update_next_ping(imp, 1);
 397         assert_spin_locked(&imp->imp_lock);
 398         /*
 399          * Avoid reading stale imp_connect_data.  When not sure if pings are
 400          * expected or not on next connection, we assume they are not and force
 401          * one anyway to guarantee the chance of updating
 402          * imp_peer_committed_transno.
 403          */
 404         if (imp->imp_state != LUSTRE_IMP_FULL ||
 405             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 406                 imp->imp_force_next_verify = 1;
 407 }
 408
 409 int ptlrpc_pinger_add_import(struct obd_import *imp)
 410 {
 411         ENTRY;
 412         if (!cfs_list_empty(&imp->imp_pinger_chain))
 413                 RETURN(-EALREADY);
 414
 415         mutex_lock(&pinger_mutex);
 416         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 417                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 418         /* if we add to pinger we want recovery on this import */
 419         imp->imp_obd->obd_no_recov = 0;
 420         ptlrpc_update_next_ping(imp, 0);
 421         /* XXX sort, blah blah */
 422         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 423         class_import_get(imp);
 424
 425         ptlrpc_pinger_wake_up();
 426         mutex_unlock(&pinger_mutex);
 427
 428         RETURN(0);
 429 }
 430 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 431
 432 int ptlrpc_pinger_del_import(struct obd_import *imp)
 433 {
 434         ENTRY;
 435         if (cfs_list_empty(&imp->imp_pinger_chain))
 436                 RETURN(-ENOENT);
 437
 438         mutex_lock(&pinger_mutex);
 439         cfs_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_unlock(&pinger_mutex);
 446         RETURN(0);
 447 }
 448 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 449
 450 /**
 451  * Register a timeout callback to the pinger list, and the callback will
 452  * be called when timeout happens.
 453  */
 454 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 455                                         timeout_cb_t cb, void *data)
 456 {
 457         struct timeout_item *ti;
 458
 459         OBD_ALLOC_PTR(ti);
 460         if (!ti)
 461                 return(NULL);
 462
 463         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 464         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 465         ti->ti_timeout = time;
 466         ti->ti_event = event;
 467         ti->ti_cb = cb;
 468         ti->ti_cb_data = data;
 469
 470         return ti;
 471 }
 472
 473 /**
 474  * Register timeout event on the the pinger thread.
 475  * Note: the timeout list is an sorted list with increased timeout value.
 476  */
 477 static struct timeout_item*
 478 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 479                                timeout_cb_t cb, void *data)
 480 {
 481         struct timeout_item *item, *tmp;
 482
 483         LASSERT(mutex_is_locked(&pinger_mutex));
 484
 485         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 486                 if (item->ti_event == event)
 487                         goto out;
 488
 489         item = ptlrpc_new_timeout(time, event, cb, data);
 490         if (item) {
 491                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 492                         if (tmp->ti_timeout < time) {
 493                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 494                                 goto out;
 495                         }
 496                 }
 497                 cfs_list_add(&item->ti_chain, &timeout_list);
 498         }
 499 out:
 500         return item;
 501 }
 502
 503 /* Add a client_obd to the timeout event list, when timeout(@time)
 504  * happens, the callback(@cb) will be called.
 505  */
 506 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 507                               timeout_cb_t cb, void *data,
 508                               cfs_list_t *obd_list)
 509 {
 510         struct timeout_item *ti;
 511
 512         mutex_lock(&pinger_mutex);
 513         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 514         if (!ti) {
 515                 mutex_unlock(&pinger_mutex);
 516                 return (-EINVAL);
 517         }
 518         cfs_list_add(obd_list, &ti->ti_obd_list);
 519         mutex_unlock(&pinger_mutex);
 520         return 0;
 521 }
 522 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 523
 524 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 525                               enum timeout_event event)
 526 {
 527         struct timeout_item *ti = NULL, *item;
 528
 529         if (cfs_list_empty(obd_list))
 530                 return 0;
 531         mutex_lock(&pinger_mutex);
 532         cfs_list_del_init(obd_list);
 533         /**
 534          * If there are no obd attached to the timeout event
 535          * list, remove this timeout event from the pinger
 536          */
 537         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 538                 if (item->ti_event == event) {
 539                         ti = item;
 540                         break;
 541                 }
 542         }
 543         LASSERTF(ti != NULL, "ti is NULL ! \n");
 544         if (cfs_list_empty(&ti->ti_obd_list)) {
 545                 cfs_list_del(&ti->ti_chain);
 546                 OBD_FREE_PTR(ti);
 547         }
 548         mutex_unlock(&pinger_mutex);
 549         return 0;
 550 }
 551 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 552
 553 int ptlrpc_pinger_remove_timeouts(void)
 554 {
 555         struct timeout_item *item, *tmp;
 556
 557         mutex_lock(&pinger_mutex);
 558         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 559                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 560                 cfs_list_del(&item->ti_chain);
 561                 OBD_FREE_PTR(item);
 562         }
 563         mutex_unlock(&pinger_mutex);
 564         return 0;
 565 }
 566
 567 void ptlrpc_pinger_wake_up()
 568 {
 569 #ifdef ENABLE_PINGER
 570         thread_add_flags(&pinger_thread, SVC_EVENT);
 571         wake_up(&pinger_thread.t_ctl_waitq);
 572 #endif
 573 }
 574
 575 /* Ping evictor thread */
 576 #define PET_READY     1
 577 #define PET_TERMINATE 2
 578
 579 static int               pet_refcount = 0;
 580 static int               pet_state;
 581 static wait_queue_head_t       pet_waitq;
 582 CFS_LIST_HEAD(pet_list);
 583 static DEFINE_SPINLOCK(pet_lock);
 584
 585 int ping_evictor_wake(struct obd_export *exp)
 586 {
 587         struct obd_device *obd;
 588
 589         spin_lock(&pet_lock);
 590         if (pet_state != PET_READY) {
 591                 /* eventually the new obd will call here again. */
 592                 spin_unlock(&pet_lock);
 593                 return 1;
 594         }
 595
 596         obd = class_exp2obd(exp);
 597         if (cfs_list_empty(&obd->obd_evict_list)) {
 598                 class_incref(obd, "evictor", obd);
 599                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 600         }
 601         spin_unlock(&pet_lock);
 602
 603         wake_up(&pet_waitq);
 604         return 0;
 605 }
 606
 607 static int ping_evictor_main(void *arg)
 608 {
 609         struct obd_device *obd;
 610         struct obd_export *exp;
 611         struct l_wait_info lwi = { 0 };
 612         time_t expire_time;
 613         ENTRY;
 614
 615         unshare_fs_struct();
 616
 617         CDEBUG(D_HA, "Starting Ping Evictor\n");
 618         pet_state = PET_READY;
 619         while (1) {
 620                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 621                              (pet_state == PET_TERMINATE), &lwi);
 622
 623                 /* loop until all obd's will be removed */
 624                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 625                         break;
 626
 627                 /* we only get here if pet_exp != NULL, and the end of this
 628                  * loop is the only place which sets it NULL again, so lock
 629                  * is not strictly necessary. */
 630                 spin_lock(&pet_lock);
 631                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 632                                      obd_evict_list);
 633                 spin_unlock(&pet_lock);
 634
 635                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 636
 637                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 638                        obd->obd_name, expire_time);
 639
 640                 /* Exports can't be deleted out of the list while we hold
 641                  * the obd lock (class_unlink_export), which means we can't
 642                  * lose the last ref on the export.  If they've already been
 643                  * removed from the list, we won't find them here. */
 644                 spin_lock(&obd->obd_dev_lock);
 645                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 646                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 647                                              struct obd_export,
 648                                              exp_obd_chain_timed);
 649                         if (expire_time > exp->exp_last_request_time) {
 650                                 class_export_get(exp);
 651                                 spin_unlock(&obd->obd_dev_lock);
 652                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 653                                               " (at %s) in %ld seconds. I think"
 654                                               " it's dead, and I am evicting"
 655                                               " it. exp %p, cur %ld expire %ld"
 656                                               " last %ld\n",
 657                                               obd->obd_name,
 658                                               obd_uuid2str(&exp->exp_client_uuid),
 659                                               obd_export_nid2str(exp),
 660                                               (long)(cfs_time_current_sec() -
 661                                                      exp->exp_last_request_time),
 662                                               exp, (long)cfs_time_current_sec(),
 663                                               (long)expire_time,
 664                                               (long)exp->exp_last_request_time);
 665                                 CDEBUG(D_HA, "Last request was at %ld\n",
 666                                        exp->exp_last_request_time);
 667                                 class_fail_export(exp);
 668                                 class_export_put(exp);
 669                                 spin_lock(&obd->obd_dev_lock);
 670                         } else {
 671                                 /* List is sorted, so everyone below is ok */
 672                                 break;
 673                         }
 674                 }
 675                 spin_unlock(&obd->obd_dev_lock);
 676
 677                 spin_lock(&pet_lock);
 678                 cfs_list_del_init(&obd->obd_evict_list);
 679                 spin_unlock(&pet_lock);
 680
 681                 class_decref(obd, "evictor", obd);
 682         }
 683         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 684
 685         RETURN(0);
 686 }
 687
 688 void ping_evictor_start(void)
 689 {
 690         struct task_struct *task;
 691
 692         if (++pet_refcount > 1)
 693                 return;
 694
 695         init_waitqueue_head(&pet_waitq);
 696
 697         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 698         if (IS_ERR(task)) {
 699                 pet_refcount--;
 700                 CERROR("Cannot start ping evictor thread: %ld\n",
 701                         PTR_ERR(task));
 702         }
 703 }
 704 EXPORT_SYMBOL(ping_evictor_start);
 705
 706 void ping_evictor_stop(void)
 707 {
 708         if (--pet_refcount > 0)
 709                 return;
 710
 711         pet_state = PET_TERMINATE;
 712         wake_up(&pet_waitq);
 713 }
 714 EXPORT_SYMBOL(ping_evictor_stop);
 715 #else /* !__KERNEL__ */
 716
 717 /* XXX
 718  * the current implementation of pinger in liblustre is not optimized
 719  */
 720
 721 #ifdef ENABLE_PINGER
 722 static struct pinger_data {
 723         int             pd_recursion;
 724         cfs_time_t      pd_this_ping;   /* jiffies */
 725         cfs_time_t      pd_next_ping;   /* jiffies */
 726         struct ptlrpc_request_set *pd_set;
 727 } pinger_args;
 728
 729 static int pinger_check_rpcs(void *arg)
 730 {
 731         cfs_time_t curtime = cfs_time_current();
 732         struct ptlrpc_request *req;
 733         struct ptlrpc_request_set *set;
 734         cfs_list_t *iter;
 735         struct obd_import *imp;
 736         struct pinger_data *pd = &pinger_args;
 737         int rc;
 738
 739         /* prevent recursion */
 740         if (pd->pd_recursion++) {
 741                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 742                 LASSERT(pd->pd_set);
 743                 pd->pd_recursion--;
 744                 return 0;
 745         }
 746
 747         /* have we reached ping point? */
 748         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 749                 pd->pd_recursion--;
 750                 return 0;
 751         }
 752
 753         /* if we have rpc_set already, continue processing it */
 754         if (pd->pd_set) {
 755                 LASSERT(pd->pd_this_ping);
 756                 set = pd->pd_set;
 757                 goto do_check_set;
 758         }
 759
 760         pd->pd_this_ping = curtime;
 761         pd->pd_set = ptlrpc_prep_set();
 762         if (pd->pd_set == NULL)
 763                 goto out;
 764         set = pd->pd_set;
 765
 766         /* add rpcs into set */
 767         mutex_lock(&pinger_mutex);
 768         cfs_list_for_each(iter, &pinger_imports) {
 769                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 770                                                         imp_pinger_chain);
 771                 int generation, level;
 772
 773                 if (cfs_time_aftereq(pd->pd_this_ping,
 774                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 775                         /* Add a ping. */
 776                         spin_lock(&imp->imp_lock);
 777                         generation = imp->imp_generation;
 778                         level = imp->imp_state;
 779                         spin_unlock(&imp->imp_lock);
 780
 781                         if (level != LUSTRE_IMP_FULL) {
 782                                 CDEBUG(D_HA,
 783                                        "not pinging %s (in recovery)\n",
 784                                        obd2cli_tgt(imp->imp_obd));
 785                                 continue;
 786                         }
 787
 788                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 789                                                         LUSTRE_OBD_VERSION,
 790                                                         OBD_PING);
 791                         if (req == NULL) {
 792                                 CERROR("OOM trying to ping %s->%s\n",
 793                                        imp->imp_obd->obd_uuid.uuid,
 794                                        obd2cli_tgt(imp->imp_obd));
 795                                 break;
 796                         }
 797
 798                         req->rq_no_resend = 1;
 799                         ptlrpc_request_set_replen(req);
 800                         req->rq_send_state = LUSTRE_IMP_FULL;
 801                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 802                         req->rq_import_generation = generation;
 803                         ptlrpc_set_add_req(set, req);
 804                 } else {
 805                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 806                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 807                                imp->imp_next_ping, pd->pd_this_ping);
 808                 }
 809         }
 810         pd->pd_this_ping = curtime;
 811         mutex_unlock(&pinger_mutex);
 812
 813         /* Might be empty, that's OK. */
 814         if (atomic_read(&set->set_remaining) == 0)
 815                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 816
 817         cfs_list_for_each(iter, &set->set_requests) {
 818                 struct ptlrpc_request *req =
 819                         cfs_list_entry(iter, struct ptlrpc_request,
 820                                        rq_set_chain);
 821                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 822                           req->rq_import->imp_obd->obd_uuid.uuid,
 823                           obd2cli_tgt(req->rq_import->imp_obd));
 824                 (void)ptl_send_rpc(req, 0);
 825         }
 826
 827 do_check_set:
 828         rc = ptlrpc_check_set(NULL, set);
 829
 830         /* not finished, and we are not expired, simply return */
 831         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 832                                             cfs_time_seconds(PING_INTERVAL)))) {
 833                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 834                 pd->pd_recursion--;
 835                 return 0;
 836         }
 837
 838         /* Expire all the requests that didn't come back. */
 839         mutex_lock(&pinger_mutex);
 840         cfs_list_for_each(iter, &set->set_requests) {
 841                 req = cfs_list_entry(iter, struct ptlrpc_request,
 842                                      rq_set_chain);
 843
 844                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 845                         continue;
 846
 847                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 848                        req);
 849
 850                 /* This will also unregister reply. */
 851                 ptlrpc_expire_one_request(req, 0);
 852
 853                 /* We're done with this req, let's finally move it to complete
 854                  * phase and take care of inflights. */
 855                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 856                 imp = req->rq_import;
 857                 spin_lock(&imp->imp_lock);
 858                 if (!cfs_list_empty(&req->rq_list)) {
 859                         cfs_list_del_init(&req->rq_list);
 860                         atomic_dec(&imp->imp_inflight);
 861                 }
 862                 spin_unlock(&imp->imp_lock);
 863                 atomic_dec(&set->set_remaining);
 864         }
 865         mutex_unlock(&pinger_mutex);
 866
 867         ptlrpc_set_destroy(set);
 868         pd->pd_set = NULL;
 869
 870 out:
 871         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 872                                         cfs_time_seconds(PING_INTERVAL));
 873         pd->pd_this_ping = 0; /* XXX for debug */
 874
 875         CDEBUG(D_INFO, "finished a round ping\n");
 876         pd->pd_recursion--;
 877         return 0;
 878 }
 879
 880 static void *pinger_callback = NULL;
 881 #endif /* ENABLE_PINGER */
 882
 883 int ptlrpc_start_pinger(void)
 884 {
 885 #ifdef ENABLE_PINGER
 886         memset(&pinger_args, 0, sizeof(pinger_args));
 887         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 888                                                            &pinger_check_rpcs,
 889                                                            &pinger_args);
 890 #endif
 891         return 0;
 892 }
 893
 894 int ptlrpc_stop_pinger(void)
 895 {
 896 #ifdef ENABLE_PINGER
 897         if (pinger_callback)
 898                 liblustre_deregister_wait_callback(pinger_callback);
 899 #endif
 900         return 0;
 901 }
 902
 903 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 904 {
 905 #ifdef ENABLE_PINGER
 906         mutex_lock(&pinger_mutex);
 907         ptlrpc_update_next_ping(imp, 0);
 908         if (pinger_args.pd_set == NULL &&
 909             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 910                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 911                         imp->imp_next_ping, cfs_time_current());
 912                 pinger_args.pd_next_ping = imp->imp_next_ping;
 913         }
 914         mutex_unlock(&pinger_mutex);
 915 #endif
 916 }
 917
 918 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 919 {
 920 #ifdef ENABLE_PINGER
 921         mutex_lock(&pinger_mutex);
 922         ptlrpc_update_next_ping(imp, 1);
 923         if (pinger_args.pd_set == NULL &&
 924             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 925                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 926                         imp->imp_next_ping, cfs_time_current());
 927                 pinger_args.pd_next_ping = imp->imp_next_ping;
 928         }
 929         mutex_unlock(&pinger_mutex);
 930 #endif
 931 }
 932
 933 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 934                               timeout_cb_t cb, void *data,
 935                               cfs_list_t *obd_list)
 936 {
 937         return 0;
 938 }
 939
 940 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 941                               enum timeout_event event)
 942 {
 943         return 0;
 944 }
 945
 946 int ptlrpc_pinger_add_import(struct obd_import *imp)
 947 {
 948         ENTRY;
 949         if (!cfs_list_empty(&imp->imp_pinger_chain))
 950                 RETURN(-EALREADY);
 951
 952         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 953                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 954         ptlrpc_pinger_sending_on_import(imp);
 955
 956         mutex_lock(&pinger_mutex);
 957         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 958         class_import_get(imp);
 959         mutex_unlock(&pinger_mutex);
 960
 961         RETURN(0);
 962 }
 963
 964 int ptlrpc_pinger_del_import(struct obd_import *imp)
 965 {
 966         ENTRY;
 967         if (cfs_list_empty(&imp->imp_pinger_chain))
 968                 RETURN(-ENOENT);
 969
 970         mutex_lock(&pinger_mutex);
 971         cfs_list_del_init(&imp->imp_pinger_chain);
 972         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 973                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 974         class_import_put(imp);
 975         mutex_unlock(&pinger_mutex);
 976         RETURN(0);
 977 }
 978
 979 void ptlrpc_pinger_wake_up()
 980 {
 981 #ifdef ENABLE_PINGER
 982         /* XXX force pinger to run, if needed */
 983         struct obd_import *imp;
 984         ENTRY;
 985         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 986                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 987                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 988 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 989                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 990                     !imp_is_deactive(imp))
 991 #else
 992                 /*XXX only recover for the initial connection */
 993                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 994                     imp->imp_state == LUSTRE_IMP_DISCON &&
 995                     !imp_is_deactive(imp))
 996 #endif
 997                         ptlrpc_initiate_recovery(imp);
 998                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 999                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1000                                      "state %d, deactive %d\n",
1001                                      imp->imp_obd->obd_uuid.uuid,
1002                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1003                                      imp_is_deactive(imp));
1004         }
1005         EXIT;
1006 #endif
1007 }
1008 #endif /* !__KERNEL__ */