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