lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 struct semaphore pinger_sem;
  52 static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
  53 static struct list_head timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  54
  55 struct ptlrpc_request *
  56 ptlrpc_prep_ping(struct obd_import *imp)
  57 {
  58         struct ptlrpc_request *req;
  59
  60         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
  61                               OBD_PING, 1, NULL, NULL);
  62         if (req) {
  63                 ptlrpc_req_set_repsize(req, 1, NULL);
  64                 req->rq_no_resend = req->rq_no_delay = 1;
  65         }
  66         return req;
  67 }
  68
  69 int ptlrpc_obd_ping(struct obd_device *obd)
  70 {
  71         int rc;
  72         struct ptlrpc_request *req;
  73         ENTRY;
  74
  75         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  76         if (req == NULL)
  77                 RETURN(-ENOMEM);
  78
  79         req->rq_send_state = LUSTRE_IMP_FULL;
  80
  81         rc = ptlrpc_queue_wait(req);
  82
  83         ptlrpc_req_finished(req);
  84
  85         RETURN(rc);
  86 }
  87 EXPORT_SYMBOL(ptlrpc_obd_ping);
  88
  89 int ptlrpc_ping(struct obd_import *imp)
  90 {
  91         struct ptlrpc_request *req;
  92         int rc = 0;
  93         ENTRY;
  94
  95         req = ptlrpc_prep_ping(imp);
  96         if (req) {
  97                 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  98                           imp->imp_obd->obd_uuid.uuid,
  99                           obd2cli_tgt(imp->imp_obd));
 100
 101                 /* To quickly detect server failure ping timeouts must be
 102                  * kept small.  Therefore we must override/ignore the server
 103                  * rpc completion estimate which may be very large since
 104                  * it includes non-ping service times.  The right long term
 105                  * fix will be to add a per-server (not per-service) thread
 106                  * in order to reduce the number of pings in the system in
 107                  * general (see bug 12471). */
 108                 if (!AT_OFF) {
 109                         req->rq_timeout = PING_SVC_TIMEOUT +
 110                                           at_get(&imp->imp_at.iat_net_latency);
 111                         lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
 112                 }
 113
 114                 ptlrpcd_add_req(req);
 115         } else {
 116                 CERROR("OOM trying to ping %s->%s\n",
 117                        imp->imp_obd->obd_uuid.uuid,
 118                        obd2cli_tgt(imp->imp_obd));
 119                 rc = -ENOMEM;
 120         }
 121
 122         RETURN(rc);
 123 }
 124 EXPORT_SYMBOL(ptlrpc_ping);
 125
 126 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 127 {
 128 #ifdef ENABLE_PINGER
 129         cfs_time_t delay, dtime, ctime = cfs_time_current();
 130
 131         if (imp->imp_state == LUSTRE_IMP_DISCON ||
 132             imp->imp_state == LUSTRE_IMP_CONNECTING) {
 133                 /* In the disconnected case aggressively reconnect, for
 134                  * this request the AT service timeout will be set to
 135                  * INITIAL_CONNECT_TIMEOUT.  To ensure the request times
 136                  * out before we send another we add one extra second. */
 137                 dtime = cfs_time_seconds(max_t(int, CONNECTION_SWITCH_MIN,
 138                                 AT_OFF ? 0 : INITIAL_CONNECT_TIMEOUT + 1 +
 139                                 at_get(&imp->imp_at.iat_net_latency)));
 140         } else {
 141                 /* In the common case we want to cluster the pings at
 142                  * at regular intervals to minimize system noise. */
 143                 delay = cfs_time_seconds(soon ? PING_INTERVAL_SHORT :
 144                                          PING_INTERVAL);
 145                 dtime = delay - (ctime % delay);
 146         }
 147         /* May harmlessly race with ptlrpc_update_next_ping() */
 148         imp->imp_next_ping = cfs_time_add(ctime, dtime);
 149
 150         CDEBUG(D_INFO, "Setting %s next ping to "CFS_TIME_T" ("CFS_TIME_T")\n",
 151                obd2cli_tgt(imp->imp_obd), imp->imp_next_ping, dtime);
 152
 153 #endif /* ENABLE_PINGER */
 154 }
 155
 156 void ptlrpc_ping_import_soon(struct obd_import *imp)
 157 {
 158         imp->imp_next_ping = cfs_time_current();
 159 }
 160
 161 static inline int imp_is_deactive(struct obd_import *imp)
 162 {
 163         return (imp->imp_deactive ||
 164                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 165 }
 166
 167 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 168 {
 169         struct timeout_item *item;
 170         cfs_time_t timeout = PING_INTERVAL;
 171
 172         /* The timeout list is a increase order sorted list */
 173         mutex_down(&pinger_sem);
 174         list_for_each_entry(item, &timeout_list, ti_chain) {
 175                 int ti_timeout = item->ti_timeout;
 176                 if (timeout > ti_timeout)
 177                          timeout = ti_timeout;
 178                 break;
 179         }
 180         mutex_up(&pinger_sem);
 181
 182         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 183                                          cfs_time_current());
 184 }
 185
 186 #ifdef __KERNEL__
 187 static int ptlrpc_pinger_main(void *arg)
 188 {
 189         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 190         struct ptlrpc_thread *thread = data->thread;
 191         ENTRY;
 192
 193         cfs_daemonize(data->name);
 194
 195         /* Record that the thread is running */
 196         thread->t_flags = SVC_RUNNING;
 197         cfs_waitq_signal(&thread->t_ctl_waitq);
 198
 199         /* And now, loop forever, pinging as needed. */
 200         while (1) {
 201                 cfs_time_t this_ping = cfs_time_current();
 202                 struct l_wait_info lwi;
 203                 cfs_duration_t time_to_next_wake;
 204                 cfs_time_t time_of_next_wake;
 205                 struct timeout_item *item;
 206                 struct list_head *iter;
 207
 208                 time_to_next_wake = cfs_time_seconds(PING_INTERVAL);
 209                 time_of_next_wake = cfs_time_shift(PING_INTERVAL);
 210
 211                 mutex_down(&pinger_sem);
 212                 list_for_each_entry(item, &timeout_list, ti_chain) {
 213                         item->ti_cb(item, item->ti_cb_data);
 214                 }
 215                 list_for_each(iter, &pinger_imports) {
 216                         struct obd_import *imp =
 217                                 list_entry(iter, struct obd_import,
 218                                            imp_pinger_chain);
 219                         int force, level;
 220
 221                         spin_lock(&imp->imp_lock);
 222                         level = imp->imp_state;
 223                         force = imp->imp_force_verify;
 224                         imp->imp_force_verify = 0;
 225                         spin_unlock(&imp->imp_lock);
 226
 227                         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_RPCTRACE,
 228                                "level %s/%u force %u deactive %u pingable %u\n",
 229                                ptlrpc_import_state_name(level), level,
 230                                force, imp->imp_deactive, imp->imp_pingable);
 231
 232                         /* Include any ping which misses the deadline by up to
 233                          * 1/10 of a second.  The pings are designed to clump
 234                          * and this helps ensure the entire batch gets sent
 235                          * promptly, which minimizes system noise from pings */
 236
 237                         if (force ||
 238                             cfs_time_aftereq(this_ping, imp->imp_next_ping -
 239                                              (cfs_time_seconds(1) + 9) / 10)) {
 240                                 if (level == LUSTRE_IMP_DISCON &&
 241                                     !imp_is_deactive(imp)) {
 242                                         ptlrpc_update_next_ping(imp, 0);
 243                                         ptlrpc_initiate_recovery(imp);
 244                                 } else if (level != LUSTRE_IMP_FULL ||
 245                                          imp->imp_obd->obd_no_recov ||
 246                                          imp_is_deactive(imp)) {
 247                                         CDEBUG(D_HA, "not pinging %s "
 248                                                "(in recovery: %s or recovery "
 249                                                "disabled: %u/%u)\n",
 250                                                obd2cli_tgt(imp->imp_obd),
 251                                                ptlrpc_import_state_name(level),
 252                                                imp->imp_deactive,
 253                                                imp->imp_obd->obd_no_recov);
 254                                 } else if (imp->imp_pingable || force) {
 255                                                 ptlrpc_ping(imp);
 256                                                 /* ptlrpc_pinger_sending_on_import()
 257                                                  * will asynch update imp_next_ping
 258                                                  * so it must not be used below to
 259                                                  * calculate minimum wait time. */
 260                                                 continue;
 261                                 }
 262                         } else {
 263                                 if (!imp->imp_pingable)
 264                                         continue;
 265                                 CDEBUG(D_INFO,
 266                                        "don't need to ping %s ("CFS_TIME_T
 267                                        " > "CFS_TIME_T")\n",
 268                                        obd2cli_tgt(imp->imp_obd),
 269                                        imp->imp_next_ping, this_ping);
 270                         }
 271
 272                         /* Wait time until next ping, or until we stopped. */
 273                         if (cfs_time_before(imp->imp_next_ping,
 274                                             time_of_next_wake)) {
 275                                 time_of_next_wake = imp->imp_next_ping;
 276                                 time_to_next_wake = max_t(cfs_duration_t,
 277                                         cfs_time_seconds(1),
 278                                         cfs_time_sub(time_of_next_wake,
 279                                                      cfs_time_current()));
 280                         }
 281                 }
 282                 mutex_up(&pinger_sem);
 283                 obd_update_maxusage();
 284                 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 285                                time_to_next_wake, time_of_next_wake);
 286
 287                 if (time_to_next_wake > 0) {
 288                         lwi = LWI_TIMEOUT(time_to_next_wake, NULL, NULL);
 289                         l_wait_event(thread->t_ctl_waitq,
 290                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 291                                      &lwi);
 292                         if (thread->t_flags & SVC_STOPPING) {
 293                                 thread->t_flags &= ~SVC_STOPPING;
 294                                 EXIT;
 295                                 break;
 296                         } else if (thread->t_flags & SVC_EVENT) {
 297                                 /* woken after adding import to reset timer */
 298                                 thread->t_flags &= ~SVC_EVENT;
 299                         }
 300                 }
 301         }
 302
 303         thread->t_flags = SVC_STOPPED;
 304         cfs_waitq_signal(&thread->t_ctl_waitq);
 305
 306         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 307         return 0;
 308 }
 309
 310 static struct ptlrpc_thread *pinger_thread = NULL;
 311
 312 int ptlrpc_start_pinger(void)
 313 {
 314         struct l_wait_info lwi = { 0 };
 315         struct ptlrpc_svc_data d;
 316         int rc;
 317 #ifndef ENABLE_PINGER
 318         return 0;
 319 #endif
 320         ENTRY;
 321
 322         if (pinger_thread != NULL)
 323                 RETURN(-EALREADY);
 324
 325         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 326         if (pinger_thread == NULL)
 327                 RETURN(-ENOMEM);
 328         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 329
 330         d.name = "ll_ping";
 331         d.thread = pinger_thread;
 332
 333         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 334          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 335         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 336         if (rc < 0) {
 337                 CERROR("cannot start thread: %d\n", rc);
 338                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 339                 pinger_thread = NULL;
 340                 RETURN(rc);
 341         }
 342         l_wait_event(pinger_thread->t_ctl_waitq,
 343                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 344
 345         RETURN(0);
 346 }
 347
 348 int ptlrpc_pinger_remove_timeouts(void);
 349
 350 int ptlrpc_stop_pinger(void)
 351 {
 352         struct l_wait_info lwi = { 0 };
 353         int rc = 0;
 354 #ifndef ENABLE_PINGER
 355         return 0;
 356 #endif
 357         ENTRY;
 358
 359         if (pinger_thread == NULL)
 360                 RETURN(-EALREADY);
 361
 362         ptlrpc_pinger_remove_timeouts();
 363         mutex_down(&pinger_sem);
 364         pinger_thread->t_flags = SVC_STOPPING;
 365         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 366         mutex_up(&pinger_sem);
 367
 368         l_wait_event(pinger_thread->t_ctl_waitq,
 369                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 370
 371         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 372         pinger_thread = NULL;
 373         RETURN(rc);
 374 }
 375
 376 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 377 {
 378         ptlrpc_update_next_ping(imp, 0);
 379 }
 380
 381 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 382 {
 383         ptlrpc_update_next_ping(imp, 1);
 384 }
 385
 386 int ptlrpc_pinger_add_import(struct obd_import *imp)
 387 {
 388         ENTRY;
 389         if (!list_empty(&imp->imp_pinger_chain))
 390                 RETURN(-EALREADY);
 391
 392         mutex_down(&pinger_sem);
 393         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 394                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 395         /* if we add to pinger we want recovery on this import */
 396         imp->imp_obd->obd_no_recov = 0;
 397
 398         ptlrpc_update_next_ping(imp, 0);
 399         /* XXX sort, blah blah */
 400         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 401         class_import_get(imp);
 402
 403         ptlrpc_pinger_wake_up();
 404         mutex_up(&pinger_sem);
 405
 406         RETURN(0);
 407 }
 408
 409 int ptlrpc_pinger_del_import(struct obd_import *imp)
 410 {
 411         ENTRY;
 412         if (list_empty(&imp->imp_pinger_chain))
 413                 RETURN(-ENOENT);
 414
 415         mutex_down(&pinger_sem);
 416         list_del_init(&imp->imp_pinger_chain);
 417         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 418                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 419         /* if we remove from pinger we don't want recovery on this import */
 420         imp->imp_obd->obd_no_recov = 1;
 421         class_import_put(imp);
 422         mutex_up(&pinger_sem);
 423         RETURN(0);
 424 }
 425
 426 /**
 427  * Register a timeout callback to the pinger list, and the callback will
 428  * be called when timeout happens.
 429  */
 430 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 431                                         timeout_cb_t cb, void *data)
 432 {
 433         struct timeout_item *ti;
 434
 435         OBD_ALLOC_PTR(ti);
 436         if (!ti)
 437                 return(NULL);
 438
 439         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 440         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 441         ti->ti_timeout = time;
 442         ti->ti_event = event;
 443         ti->ti_cb = cb;
 444         ti->ti_cb_data = data;
 445
 446         return ti;
 447 }
 448
 449 /**
 450  * Register timeout event on the the pinger thread.
 451  * Note: the timeout list is an sorted list with increased timeout value.
 452  */
 453 static struct timeout_item*
 454 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 455                                timeout_cb_t cb, void *data)
 456 {
 457         struct timeout_item *item, *tmp;
 458
 459         LASSERT_SEM_LOCKED(&pinger_sem);
 460
 461         list_for_each_entry(item, &timeout_list, ti_chain)
 462                 if (item->ti_event == event)
 463                         goto out;
 464
 465         item = ptlrpc_new_timeout(time, event, cb, data);
 466         if (item) {
 467                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 468                         if (tmp->ti_timeout < time) {
 469                                 list_add(&item->ti_chain, &tmp->ti_chain);
 470                                 goto out;
 471                         }
 472                 }
 473                 list_add(&item->ti_chain, &timeout_list);
 474         }
 475 out:
 476         return item;
 477 }
 478
 479 /* Add a client_obd to the timeout event list, when timeout(@time)
 480  * happens, the callback(@cb) will be called.
 481  */
 482 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 483                               timeout_cb_t cb, void *data,
 484                               struct list_head *obd_list)
 485 {
 486         struct timeout_item *ti;
 487
 488         mutex_down(&pinger_sem);
 489         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 490         if (!ti) {
 491                 mutex_up(&pinger_sem);
 492                 return (-EINVAL);
 493         }
 494         list_add(obd_list, &ti->ti_obd_list);
 495         mutex_up(&pinger_sem);
 496         return 0;
 497 }
 498
 499 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 500                               enum timeout_event event)
 501 {
 502         struct timeout_item *ti = NULL, *item;
 503
 504         if (list_empty(obd_list))
 505                 return 0;
 506         mutex_down(&pinger_sem);
 507         list_del_init(obd_list);
 508         /**
 509          * If there are no obd attached to the timeout event
 510          * list, remove this timeout event from the pinger
 511          */
 512         list_for_each_entry(item, &timeout_list, ti_chain) {
 513                 if (item->ti_event == event) {
 514                         ti = item;
 515                         break;
 516                 }
 517         }
 518         LASSERTF(ti != NULL, "ti is NULL ! \n");
 519         if (list_empty(&ti->ti_obd_list)) {
 520                 list_del(&ti->ti_chain);
 521                 OBD_FREE_PTR(ti);
 522         }
 523         mutex_up(&pinger_sem);
 524         return 0;
 525 }
 526
 527 int ptlrpc_pinger_remove_timeouts(void)
 528 {
 529         struct timeout_item *item, *tmp;
 530
 531         mutex_down(&pinger_sem);
 532         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 533                 LASSERT(list_empty(&item->ti_obd_list));
 534                 list_del(&item->ti_chain);
 535                 OBD_FREE_PTR(item);
 536         }
 537         mutex_up(&pinger_sem);
 538         return 0;
 539 }
 540
 541 void ptlrpc_pinger_wake_up()
 542 {
 543 #ifdef ENABLE_PINGER
 544         pinger_thread->t_flags |= SVC_EVENT;
 545         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 546 #endif
 547 }
 548
 549 /* Ping evictor thread */
 550 #define PET_READY     1
 551 #define PET_TERMINATE 2
 552
 553 static int               pet_refcount = 0;
 554 static int               pet_state;
 555 static wait_queue_head_t pet_waitq;
 556 static struct obd_export *pet_exp = NULL;
 557 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 558
 559 int ping_evictor_wake(struct obd_export *exp)
 560 {
 561         spin_lock(&pet_lock);
 562         if (pet_exp || (pet_state != PET_READY)) {
 563                 /* eventually the new obd will call here again. */
 564                 spin_unlock(&pet_lock);
 565                 return 1;
 566         }
 567
 568         /* We have to make sure the obd isn't destroyed between now and when
 569          * the ping evictor runs.  We'll take a reference here, and drop it
 570          * when we finish in the evictor.  We don't really care about this
 571          * export in particular; we just need one to keep the obd alive. */
 572         pet_exp = class_export_get(exp);
 573         spin_unlock(&pet_lock);
 574
 575         wake_up(&pet_waitq);
 576         return 0;
 577 }
 578
 579 static int ping_evictor_main(void *arg)
 580 {
 581         struct obd_device *obd;
 582         struct obd_export *exp;
 583         struct l_wait_info lwi = { 0 };
 584         time_t expire_time;
 585         ENTRY;
 586
 587         cfs_daemonize_ctxt("ll_evictor");
 588
 589         CDEBUG(D_HA, "Starting Ping Evictor\n");
 590         pet_exp = NULL;
 591         pet_state = PET_READY;
 592         while (1) {
 593                 l_wait_event(pet_waitq, pet_exp ||
 594                              (pet_state == PET_TERMINATE), &lwi);
 595                 if (pet_state == PET_TERMINATE)
 596                         break;
 597
 598                 /* we only get here if pet_exp != NULL, and the end of this
 599                  * loop is the only place which sets it NULL again, so lock
 600                  * is not strictly necessary. */
 601                 spin_lock(&pet_lock);
 602                 obd = pet_exp->exp_obd;
 603                 spin_unlock(&pet_lock);
 604
 605                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 606
 607                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 608                        obd->obd_name, expire_time);
 609
 610                 /* Exports can't be deleted out of the list while we hold
 611                  * the obd lock (class_unlink_export), which means we can't
 612                  * lose the last ref on the export.  If they've already been
 613                  * removed from the list, we won't find them here. */
 614                 spin_lock(&obd->obd_dev_lock);
 615                 while (!list_empty(&obd->obd_exports_timed)) {
 616                         exp = list_entry(obd->obd_exports_timed.next,
 617                                          struct obd_export,exp_obd_chain_timed);
 618                         if (expire_time > exp->exp_last_request_time) {
 619                                 class_export_get(exp);
 620                                 spin_unlock(&obd->obd_dev_lock);
 621                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 622                                               " (at %s) in %ld seconds. I think"
 623                                               " it's dead, and I am evicting"
 624                                               " it.\n", obd->obd_name,
 625                                               obd_uuid2str(&exp->exp_client_uuid),
 626                                               obd_export_nid2str(exp),
 627                                               (long)(cfs_time_current_sec() -
 628                                                      exp->exp_last_request_time));
 629                                 CDEBUG(D_HA, "Last request was at %ld\n",
 630                                        exp->exp_last_request_time);
 631                                 class_fail_export(exp);
 632                                 class_export_put(exp);
 633                                 spin_lock(&obd->obd_dev_lock);
 634                         } else {
 635                                 /* List is sorted, so everyone below is ok */
 636                                 break;
 637                         }
 638                 }
 639                 spin_unlock(&obd->obd_dev_lock);
 640
 641                 class_export_put(pet_exp);
 642
 643                 spin_lock(&pet_lock);
 644                 pet_exp = NULL;
 645                 spin_unlock(&pet_lock);
 646         }
 647         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 648
 649         RETURN(0);
 650 }
 651
 652 void ping_evictor_start(void)
 653 {
 654         int rc;
 655
 656         if (++pet_refcount > 1)
 657                 return;
 658
 659         init_waitqueue_head(&pet_waitq);
 660
 661         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 662         if (rc < 0) {
 663                 pet_refcount--;
 664                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 665         }
 666 }
 667 EXPORT_SYMBOL(ping_evictor_start);
 668
 669 void ping_evictor_stop(void)
 670 {
 671         if (--pet_refcount > 0)
 672                 return;
 673
 674         pet_state = PET_TERMINATE;
 675         wake_up(&pet_waitq);
 676 }
 677 EXPORT_SYMBOL(ping_evictor_stop);
 678 #else /* !__KERNEL__ */
 679
 680 /* XXX
 681  * the current implementation of pinger in liblustre is not optimized
 682  */
 683
 684 #ifdef ENABLE_PINGER
 685 static struct pinger_data {
 686         int             pd_recursion;
 687         cfs_time_t      pd_this_ping;   /* jiffies */
 688         cfs_time_t      pd_next_ping;   /* jiffies */
 689         struct ptlrpc_request_set *pd_set;
 690 } pinger_args;
 691
 692 static int pinger_check_rpcs(void *arg)
 693 {
 694         cfs_time_t curtime = cfs_time_current();
 695         struct ptlrpc_request *req;
 696         struct ptlrpc_request_set *set;
 697         struct list_head *iter;
 698         struct obd_import *imp;
 699         struct pinger_data *pd = &pinger_args;
 700         int rc;
 701
 702         /* prevent recursion */
 703         if (pd->pd_recursion++) {
 704                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 705                 LASSERT(pd->pd_set);
 706                 pd->pd_recursion--;
 707                 return 0;
 708         }
 709
 710         /* have we reached ping point? */
 711         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 712                 pd->pd_recursion--;
 713                 return 0;
 714         }
 715
 716         /* if we have rpc_set already, continue processing it */
 717         if (pd->pd_set) {
 718                 LASSERT(pd->pd_this_ping);
 719                 set = pd->pd_set;
 720                 goto do_check_set;
 721         }
 722
 723         pd->pd_this_ping = curtime;
 724         pd->pd_set = ptlrpc_prep_set();
 725         if (pd->pd_set == NULL)
 726                 goto out;
 727         set = pd->pd_set;
 728
 729         /* add rpcs into set */
 730         mutex_down(&pinger_sem);
 731         list_for_each(iter, &pinger_imports) {
 732                 struct obd_import *imp =
 733                         list_entry(iter, struct obd_import, imp_pinger_chain);
 734                 int generation, level;
 735
 736                 /* Include any ping within 1/10 of a second of the deadline */
 737                 if (cfs_time_aftereq(pd->pd_this_ping, imp->imp_next_ping -
 738                                      (cfs_time_seconds(1) + 9) / 10)) {
 739                         /* Add a ping. */
 740                         spin_lock(&imp->imp_lock);
 741                         generation = imp->imp_generation;
 742                         level = imp->imp_state;
 743                         spin_unlock(&imp->imp_lock);
 744
 745                         if (level != LUSTRE_IMP_FULL) {
 746                                 CDEBUG(D_HA,
 747                                        "not pinging %s (in recovery)\n",
 748                                        obd2cli_tgt(imp->imp_obd));
 749                                 continue;
 750                         }
 751
 752                         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
 753                                               1, NULL, NULL);
 754                         if (!req) {
 755                                 CERROR("out of memory\n");
 756                                 break;
 757                         }
 758                         req->rq_no_resend = 1;
 759                         ptlrpc_req_set_repsize(req, 1, NULL);
 760                         req->rq_send_state = LUSTRE_IMP_FULL;
 761                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 762                         req->rq_import_generation = generation;
 763                         ptlrpc_set_add_req(set, req);
 764                 } else {
 765                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 766                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 767                                imp->imp_next_ping, pd->pd_this_ping);
 768                 }
 769         }
 770         pd->pd_this_ping = curtime;
 771         mutex_up(&pinger_sem);
 772
 773         /* Might be empty, that's OK. */
 774         if (set->set_remaining == 0)
 775                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 776
 777         list_for_each(iter, &set->set_requests) {
 778                 struct ptlrpc_request *req =
 779                         list_entry(iter, struct ptlrpc_request,
 780                                    rq_set_chain);
 781                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 782                           req->rq_import->imp_obd->obd_uuid.uuid,
 783                           obd2cli_tgt(req->rq_import->imp_obd));
 784                 (void)ptl_send_rpc(req, 0);
 785         }
 786
 787 do_check_set:
 788         rc = ptlrpc_check_set(set);
 789
 790         /* not finished, and we are not expired, simply return */
 791         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 792                                             cfs_time_seconds(PING_INTERVAL)))) {
 793                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 794                 pd->pd_recursion--;
 795                 return 0;
 796         }
 797
 798         /* Expire all the requests that didn't come back. */
 799         mutex_down(&pinger_sem);
 800         list_for_each(iter, &set->set_requests) {
 801                 req = list_entry(iter, struct ptlrpc_request,
 802                                  rq_set_chain);
 803
 804                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 805                         continue;
 806
 807                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 808                        req);
 809
 810                 /* This will also unregister reply. */
 811                 ptlrpc_expire_one_request(req, 0);
 812
 813                 /* We're done with this req, let's finally move it to complete
 814                  * phase and take care of inflights. */
 815                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 816                 imp = req->rq_import;
 817                 spin_lock(&imp->imp_lock);
 818                 if (!list_empty(&req->rq_list)) {
 819                         list_del_init(&req->rq_list);
 820                         atomic_dec(&imp->imp_inflight);
 821                 }
 822                 spin_unlock(&imp->imp_lock);
 823                 set->set_remaining--;
 824         }
 825         mutex_up(&pinger_sem);
 826
 827         ptlrpc_set_destroy(set);
 828         pd->pd_set = NULL;
 829
 830 out:
 831         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 832                                         cfs_time_seconds(PING_INTERVAL));
 833         pd->pd_this_ping = 0; /* XXX for debug */
 834
 835         CDEBUG(D_INFO, "finished a round ping\n");
 836         pd->pd_recursion--;
 837         return 0;
 838 }
 839
 840 static void *pinger_callback = NULL;
 841 #endif /* ENABLE_PINGER */
 842
 843 int ptlrpc_start_pinger(void)
 844 {
 845 #ifdef ENABLE_PINGER
 846         memset(&pinger_args, 0, sizeof(pinger_args));
 847         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 848                                                            &pinger_check_rpcs,
 849                                                            &pinger_args);
 850 #endif
 851         return 0;
 852 }
 853
 854 int ptlrpc_stop_pinger(void)
 855 {
 856 #ifdef ENABLE_PINGER
 857         if (pinger_callback)
 858                 liblustre_deregister_wait_callback(pinger_callback);
 859 #endif
 860         return 0;
 861 }
 862
 863 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 864 {
 865 #ifdef ENABLE_PINGER
 866         mutex_down(&pinger_sem);
 867         ptlrpc_update_next_ping(imp, 0);
 868         if (pinger_args.pd_set == NULL &&
 869             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 870                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 871                         imp->imp_next_ping, cfs_time_current());
 872                 pinger_args.pd_next_ping = imp->imp_next_ping;
 873         }
 874         mutex_up(&pinger_sem);
 875 #endif
 876 }
 877
 878 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 879                               timeout_cb_t cb, void *data,
 880                               struct list_head *obd_list)
 881 {
 882         return 0;
 883 }
 884
 885 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 886                               enum timeout_event event)
 887 {
 888         return 0;
 889 }
 890
 891 int ptlrpc_pinger_add_import(struct obd_import *imp)
 892 {
 893         ENTRY;
 894         if (!list_empty(&imp->imp_pinger_chain))
 895                 RETURN(-EALREADY);
 896
 897         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 898                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 899         ptlrpc_pinger_sending_on_import(imp);
 900
 901         mutex_down(&pinger_sem);
 902         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 903         class_import_get(imp);
 904         mutex_up(&pinger_sem);
 905
 906         RETURN(0);
 907 }
 908
 909 int ptlrpc_pinger_del_import(struct obd_import *imp)
 910 {
 911         ENTRY;
 912         if (list_empty(&imp->imp_pinger_chain))
 913                 RETURN(-ENOENT);
 914
 915         mutex_down(&pinger_sem);
 916         list_del_init(&imp->imp_pinger_chain);
 917         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 918                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 919         class_import_put(imp);
 920         mutex_up(&pinger_sem);
 921         RETURN(0);
 922 }
 923
 924 void ptlrpc_pinger_wake_up()
 925 {
 926 #ifdef ENABLE_PINGER
 927         ENTRY;
 928         /* XXX force pinger to run, if needed */
 929         struct obd_import *imp;
 930         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 931                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 932                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 933 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 934                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 935                     !imp_is_deactive(imp))
 936 #else
 937                 /*XXX only recover for the initial connection */
 938                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 939                     imp->imp_state == LUSTRE_IMP_DISCON &&
 940                     !imp_is_deactive(imp))
 941 #endif
 942                         ptlrpc_initiate_recovery(imp);
 943                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 944                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 945                                      "state %d, deactive %d\n",
 946                                      imp->imp_obd->obd_uuid.uuid,
 947                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 948                                      imp_is_deactive(imp));
 949         }
 950 #endif
 951         EXIT;
 952 }
 953 #endif /* !__KERNEL__ */