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