lustre/ptlrpc/pinger.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 static int suppress_pings;
  52 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  53
  54 struct mutex pinger_mutex;
  55 static CFS_LIST_HEAD(pinger_imports);
  56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  57
  58 int ptlrpc_pinger_suppress_pings()
  59 {
  60         return suppress_pings;
  61 }
  62 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  63
  64 struct ptlrpc_request *
  65 ptlrpc_prep_ping(struct obd_import *imp)
  66 {
  67         struct ptlrpc_request *req;
  68
  69         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  70                                         LUSTRE_OBD_VERSION, OBD_PING);
  71         if (req) {
  72                 ptlrpc_request_set_replen(req);
  73                 req->rq_no_resend = req->rq_no_delay = 1;
  74         }
  75         return req;
  76 }
  77
  78 int ptlrpc_obd_ping(struct obd_device *obd)
  79 {
  80         int rc;
  81         struct ptlrpc_request *req;
  82         ENTRY;
  83
  84         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  85         if (req == NULL)
  86                 RETURN(-ENOMEM);
  87
  88         req->rq_send_state = LUSTRE_IMP_FULL;
  89
  90         rc = ptlrpc_queue_wait(req);
  91
  92         ptlrpc_req_finished(req);
  93
  94         RETURN(rc);
  95 }
  96 EXPORT_SYMBOL(ptlrpc_obd_ping);
  97
  98 int ptlrpc_ping(struct obd_import *imp)
  99 {
 100         struct ptlrpc_request *req;
 101         ENTRY;
 102
 103         req = ptlrpc_prep_ping(imp);
 104         if (req == NULL) {
 105                 CERROR("OOM trying to ping %s->%s\n",
 106                        imp->imp_obd->obd_uuid.uuid,
 107                        obd2cli_tgt(imp->imp_obd));
 108                 RETURN(-ENOMEM);
 109         }
 110
 111         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 112                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 113         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 114
 115         RETURN(0);
 116 }
 117
 118 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 119 {
 120 #ifdef ENABLE_PINGER
 121         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 122         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 123                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 124                                   AT_OFF ? 0 :
 125                                   at_get(&imp->imp_at.iat_net_latency));
 126                 time = min(time, dtime);
 127         }
 128         imp->imp_next_ping = cfs_time_shift(time);
 129 #endif /* ENABLE_PINGER */
 130 }
 131
 132 void ptlrpc_ping_import_soon(struct obd_import *imp)
 133 {
 134         imp->imp_next_ping = cfs_time_current();
 135 }
 136
 137 static inline int imp_is_deactive(struct obd_import *imp)
 138 {
 139         return (imp->imp_deactive ||
 140                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 141 }
 142
 143 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 144 {
 145         if (imp->imp_server_timeout)
 146                 return cfs_time_shift(obd_timeout / 2);
 147         else
 148                 return cfs_time_shift(obd_timeout);
 149 }
 150
 151 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 152 static cfs_time_t suspend_wakeup_time = 0;
 153
 154 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 155 {
 156         struct timeout_item *item;
 157         cfs_time_t timeout = PING_INTERVAL;
 158
 159         /* The timeout list is a increase order sorted list */
 160         mutex_lock(&pinger_mutex);
 161         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 162                 int ti_timeout = item->ti_timeout;
 163                 if (timeout > ti_timeout)
 164                         timeout = ti_timeout;
 165                 break;
 166         }
 167         mutex_unlock(&pinger_mutex);
 168
 169         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 170                                          cfs_time_current());
 171 }
 172
 173 static cfs_waitq_t suspend_timeouts_waitq;
 174
 175 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 176 {
 177         return suspend_wakeup_time;
 178 }
 179
 180 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 181 {
 182         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 183 #if 0
 184         if (imp->imp_no_timeout)
 185                 return;
 186         imp->imp_no_timeout = 1;
 187         cfs_atomic_inc(&suspend_timeouts);
 188         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 189                cfs_atomic_read(&suspend_timeouts));
 190 #endif
 191 }
 192
 193 void ptlrpc_activate_timeouts(struct obd_import *imp)
 194 {
 195         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 196 #if 0
 197         if (!imp->imp_no_timeout)
 198                 return;
 199         imp->imp_no_timeout = 0;
 200         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 201         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 202                 suspend_wakeup_time = cfs_time_current();
 203                 cfs_waitq_signal(&suspend_timeouts_waitq);
 204         }
 205         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 206                cfs_atomic_read(&suspend_timeouts));
 207 #endif
 208 }
 209
 210 int ptlrpc_check_suspend(void)
 211 {
 212         if (cfs_atomic_read(&suspend_timeouts))
 213                 return 1;
 214         return 0;
 215 }
 216
 217 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 218 {
 219         struct l_wait_info lwi;
 220
 221         if (cfs_atomic_read(&suspend_timeouts)) {
 222                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 223                           cfs_atomic_read(&suspend_timeouts));
 224                 lwi = LWI_INTR(NULL, NULL);
 225                 l_wait_event(suspend_timeouts_waitq,
 226                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 227                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 228                 return 1;
 229         }
 230         return 0;
 231 }
 232
 233 #ifdef __KERNEL__
 234
 235 static bool ir_up;
 236
 237 void ptlrpc_pinger_ir_up(void)
 238 {
 239         CDEBUG(D_HA, "IR up\n");
 240         ir_up = true;
 241 }
 242 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 243
 244 void ptlrpc_pinger_ir_down(void)
 245 {
 246         CDEBUG(D_HA, "IR down\n");
 247         ir_up = false;
 248 }
 249 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 250
 251 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 252                                          unsigned long this_ping)
 253 {
 254         int level;
 255         int force;
 256         int force_next;
 257         int suppress;
 258
 259         spin_lock(&imp->imp_lock);
 260
 261         level = imp->imp_state;
 262         force = imp->imp_force_verify;
 263         force_next = imp->imp_force_next_verify;
 264         /*
 265          * This will be used below only if the import is "FULL".
 266          */
 267         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 268
 269         imp->imp_force_verify = 0;
 270
 271         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 272             !force) {
 273                 spin_unlock(&imp->imp_lock);
 274                 return;
 275         }
 276
 277         imp->imp_force_next_verify = 0;
 278
 279         spin_unlock(&imp->imp_lock);
 280
 281         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 282                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 283                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 284                ptlrpc_import_state_name(level), level, force, force_next,
 285                imp->imp_deactive, imp->imp_pingable, suppress);
 286
 287         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 288                 /* wait for a while before trying recovery again */
 289                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 290                 if (!imp->imp_no_pinger_recover)
 291                         ptlrpc_initiate_recovery(imp);
 292         } else if (level != LUSTRE_IMP_FULL ||
 293                    imp->imp_obd->obd_no_recov ||
 294                    imp_is_deactive(imp)) {
 295                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 296                        "or recovery disabled: %s)\n",
 297                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 298                        ptlrpc_import_state_name(level));
 299         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 300                 ptlrpc_ping(imp);
 301         }
 302 }
 303
 304 static int ptlrpc_pinger_main(void *arg)
 305 {
 306         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 307         ENTRY;
 308
 309         cfs_daemonize(thread->t_name);
 310
 311         /* Record that the thread is running */
 312         thread_set_flags(thread, SVC_RUNNING);
 313         cfs_waitq_signal(&thread->t_ctl_waitq);
 314
 315         /* And now, loop forever, pinging as needed. */
 316         while (1) {
 317                 cfs_time_t this_ping = cfs_time_current();
 318                 struct l_wait_info lwi;
 319                 cfs_duration_t time_to_next_wake;
 320                 struct timeout_item *item;
 321                 cfs_list_t *iter;
 322
 323                 mutex_lock(&pinger_mutex);
 324                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 325                         item->ti_cb(item, item->ti_cb_data);
 326                 }
 327                 cfs_list_for_each(iter, &pinger_imports) {
 328                         struct obd_import *imp =
 329                                 cfs_list_entry(iter, struct obd_import,
 330                                                imp_pinger_chain);
 331
 332                         ptlrpc_pinger_process_import(imp, this_ping);
 333                         /* obd_timeout might have changed */
 334                         if (imp->imp_pingable && imp->imp_next_ping &&
 335                             cfs_time_after(imp->imp_next_ping,
 336                                            cfs_time_add(this_ping,
 337                                                         cfs_time_seconds(PING_INTERVAL))))
 338                                 ptlrpc_update_next_ping(imp, 0);
 339                 }
 340                 mutex_unlock(&pinger_mutex);
 341                 /* update memory usage info */
 342                 obd_update_maxusage();
 343
 344                 /* Wait until the next ping time, or until we're stopped. */
 345                 time_to_next_wake = pinger_check_timeout(this_ping);
 346                 /* The ping sent by ptlrpc_send_rpc may get sent out
 347                    say .01 second after this.
 348                    ptlrpc_pinger_sending_on_import will then set the
 349                    next ping time to next_ping + .01 sec, which means
 350                    we will SKIP the next ping at next_ping, and the
 351                    ping will get sent 2 timeouts from now!  Beware. */
 352                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 353                        CFS_TIME_T")\n", time_to_next_wake,
 354                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 355                 if (time_to_next_wake > 0) {
 356                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 357                                                 time_to_next_wake,
 358                                                 cfs_time_seconds(1)),
 359                                           NULL, NULL);
 360                         l_wait_event(thread->t_ctl_waitq,
 361                                      thread_is_stopping(thread) ||
 362                                      thread_is_event(thread),
 363                                      &lwi);
 364                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 365                                 EXIT;
 366                                 break;
 367                         } else {
 368                                 /* woken after adding import to reset timer */
 369                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 370                         }
 371                 }
 372         }
 373
 374         thread_set_flags(thread, SVC_STOPPED);
 375         cfs_waitq_signal(&thread->t_ctl_waitq);
 376
 377         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 378         return 0;
 379 }
 380
 381 static struct ptlrpc_thread *pinger_thread = NULL;
 382
 383 int ptlrpc_start_pinger(void)
 384 {
 385         struct l_wait_info lwi = { 0 };
 386         int rc;
 387 #ifndef ENABLE_PINGER
 388         return 0;
 389 #endif
 390         ENTRY;
 391
 392         if (pinger_thread != NULL)
 393                 RETURN(-EALREADY);
 394
 395         OBD_ALLOC_PTR(pinger_thread);
 396         if (pinger_thread == NULL)
 397                 RETURN(-ENOMEM);
 398         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 399         cfs_waitq_init(&suspend_timeouts_waitq);
 400
 401         strcpy(pinger_thread->t_name, "ll_ping");
 402
 403         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 404          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 405         rc = cfs_create_thread(ptlrpc_pinger_main,
 406                                pinger_thread, CFS_DAEMON_FLAGS);
 407         if (rc < 0) {
 408                 CERROR("cannot start thread: %d\n", rc);
 409                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 410                 pinger_thread = NULL;
 411                 RETURN(rc);
 412         }
 413         l_wait_event(pinger_thread->t_ctl_waitq,
 414                      thread_is_running(pinger_thread), &lwi);
 415
 416         if (suppress_pings)
 417                 CWARN("Pings will be suppressed at the request of the "
 418                       "administrator.  The configuration shall meet the "
 419                       "additional requirements described in the manual.  "
 420                       "(Search for the \"suppress_pings\" kernel module "
 421                       "parameter.)\n");
 422
 423         RETURN(0);
 424 }
 425
 426 int ptlrpc_pinger_remove_timeouts(void);
 427
 428 int ptlrpc_stop_pinger(void)
 429 {
 430         struct l_wait_info lwi = { 0 };
 431         int rc = 0;
 432 #ifndef ENABLE_PINGER
 433         return 0;
 434 #endif
 435         ENTRY;
 436
 437         if (pinger_thread == NULL)
 438                 RETURN(-EALREADY);
 439
 440         ptlrpc_pinger_remove_timeouts();
 441         mutex_lock(&pinger_mutex);
 442         thread_set_flags(pinger_thread, SVC_STOPPING);
 443         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 444         mutex_unlock(&pinger_mutex);
 445
 446         l_wait_event(pinger_thread->t_ctl_waitq,
 447                      thread_is_stopped(pinger_thread), &lwi);
 448
 449         OBD_FREE_PTR(pinger_thread);
 450         pinger_thread = NULL;
 451         RETURN(rc);
 452 }
 453
 454 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 455 {
 456         ptlrpc_update_next_ping(imp, 0);
 457 }
 458 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 459
 460 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 461 {
 462         ptlrpc_update_next_ping(imp, 1);
 463         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 464         /*
 465          * Avoid reading stale imp_connect_data.  When not sure if pings are
 466          * expected or not on next connection, we assume they are not and force
 467          * one anyway to guarantee the chance of updating
 468          * imp_peer_committed_transno.
 469          */
 470         if (imp->imp_state != LUSTRE_IMP_FULL ||
 471             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 472                 imp->imp_force_next_verify = 1;
 473 }
 474
 475 int ptlrpc_pinger_add_import(struct obd_import *imp)
 476 {
 477         ENTRY;
 478         if (!cfs_list_empty(&imp->imp_pinger_chain))
 479                 RETURN(-EALREADY);
 480
 481         mutex_lock(&pinger_mutex);
 482         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 483                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 484         /* if we add to pinger we want recovery on this import */
 485         imp->imp_obd->obd_no_recov = 0;
 486         ptlrpc_update_next_ping(imp, 0);
 487         /* XXX sort, blah blah */
 488         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 489         class_import_get(imp);
 490
 491         ptlrpc_pinger_wake_up();
 492         mutex_unlock(&pinger_mutex);
 493
 494         RETURN(0);
 495 }
 496 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 497
 498 int ptlrpc_pinger_del_import(struct obd_import *imp)
 499 {
 500         ENTRY;
 501         if (cfs_list_empty(&imp->imp_pinger_chain))
 502                 RETURN(-ENOENT);
 503
 504         mutex_lock(&pinger_mutex);
 505         cfs_list_del_init(&imp->imp_pinger_chain);
 506         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 507                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 508         /* if we remove from pinger we don't want recovery on this import */
 509         imp->imp_obd->obd_no_recov = 1;
 510         class_import_put(imp);
 511         mutex_unlock(&pinger_mutex);
 512         RETURN(0);
 513 }
 514 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 515
 516 /**
 517  * Register a timeout callback to the pinger list, and the callback will
 518  * be called when timeout happens.
 519  */
 520 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 521                                         timeout_cb_t cb, void *data)
 522 {
 523         struct timeout_item *ti;
 524
 525         OBD_ALLOC_PTR(ti);
 526         if (!ti)
 527                 return(NULL);
 528
 529         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 530         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 531         ti->ti_timeout = time;
 532         ti->ti_event = event;
 533         ti->ti_cb = cb;
 534         ti->ti_cb_data = data;
 535
 536         return ti;
 537 }
 538
 539 /**
 540  * Register timeout event on the the pinger thread.
 541  * Note: the timeout list is an sorted list with increased timeout value.
 542  */
 543 static struct timeout_item*
 544 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 545                                timeout_cb_t cb, void *data)
 546 {
 547         struct timeout_item *item, *tmp;
 548
 549         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 550
 551         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 552                 if (item->ti_event == event)
 553                         goto out;
 554
 555         item = ptlrpc_new_timeout(time, event, cb, data);
 556         if (item) {
 557                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 558                         if (tmp->ti_timeout < time) {
 559                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 560                                 goto out;
 561                         }
 562                 }
 563                 cfs_list_add(&item->ti_chain, &timeout_list);
 564         }
 565 out:
 566         return item;
 567 }
 568
 569 /* Add a client_obd to the timeout event list, when timeout(@time)
 570  * happens, the callback(@cb) will be called.
 571  */
 572 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 573                               timeout_cb_t cb, void *data,
 574                               cfs_list_t *obd_list)
 575 {
 576         struct timeout_item *ti;
 577
 578         mutex_lock(&pinger_mutex);
 579         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 580         if (!ti) {
 581                 mutex_unlock(&pinger_mutex);
 582                 return (-EINVAL);
 583         }
 584         cfs_list_add(obd_list, &ti->ti_obd_list);
 585         mutex_unlock(&pinger_mutex);
 586         return 0;
 587 }
 588 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 589
 590 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 591                               enum timeout_event event)
 592 {
 593         struct timeout_item *ti = NULL, *item;
 594
 595         if (cfs_list_empty(obd_list))
 596                 return 0;
 597         mutex_lock(&pinger_mutex);
 598         cfs_list_del_init(obd_list);
 599         /**
 600          * If there are no obd attached to the timeout event
 601          * list, remove this timeout event from the pinger
 602          */
 603         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 604                 if (item->ti_event == event) {
 605                         ti = item;
 606                         break;
 607                 }
 608         }
 609         LASSERTF(ti != NULL, "ti is NULL ! \n");
 610         if (cfs_list_empty(&ti->ti_obd_list)) {
 611                 cfs_list_del(&ti->ti_chain);
 612                 OBD_FREE_PTR(ti);
 613         }
 614         mutex_unlock(&pinger_mutex);
 615         return 0;
 616 }
 617 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 618
 619 int ptlrpc_pinger_remove_timeouts(void)
 620 {
 621         struct timeout_item *item, *tmp;
 622
 623         mutex_lock(&pinger_mutex);
 624         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 625                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 626                 cfs_list_del(&item->ti_chain);
 627                 OBD_FREE_PTR(item);
 628         }
 629         mutex_unlock(&pinger_mutex);
 630         return 0;
 631 }
 632
 633 void ptlrpc_pinger_wake_up()
 634 {
 635 #ifdef ENABLE_PINGER
 636         thread_add_flags(pinger_thread, SVC_EVENT);
 637         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 638 #endif
 639 }
 640
 641 /* Ping evictor thread */
 642 #define PET_READY     1
 643 #define PET_TERMINATE 2
 644
 645 static int               pet_refcount = 0;
 646 static int               pet_state;
 647 static cfs_waitq_t       pet_waitq;
 648 CFS_LIST_HEAD(pet_list);
 649 static DEFINE_SPINLOCK(pet_lock);
 650
 651 int ping_evictor_wake(struct obd_export *exp)
 652 {
 653         struct obd_device *obd;
 654
 655         spin_lock(&pet_lock);
 656         if (pet_state != PET_READY) {
 657                 /* eventually the new obd will call here again. */
 658                 spin_unlock(&pet_lock);
 659                 return 1;
 660         }
 661
 662         obd = class_exp2obd(exp);
 663         if (cfs_list_empty(&obd->obd_evict_list)) {
 664                 class_incref(obd, "evictor", obd);
 665                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 666         }
 667         spin_unlock(&pet_lock);
 668
 669         cfs_waitq_signal(&pet_waitq);
 670         return 0;
 671 }
 672
 673 static int ping_evictor_main(void *arg)
 674 {
 675         struct obd_device *obd;
 676         struct obd_export *exp;
 677         struct l_wait_info lwi = { 0 };
 678         time_t expire_time;
 679         ENTRY;
 680
 681         cfs_daemonize_ctxt("ll_evictor");
 682
 683         CDEBUG(D_HA, "Starting Ping Evictor\n");
 684         pet_state = PET_READY;
 685         while (1) {
 686                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 687                              (pet_state == PET_TERMINATE), &lwi);
 688
 689                 /* loop until all obd's will be removed */
 690                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 691                         break;
 692
 693                 /* we only get here if pet_exp != NULL, and the end of this
 694                  * loop is the only place which sets it NULL again, so lock
 695                  * is not strictly necessary. */
 696                 spin_lock(&pet_lock);
 697                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 698                                      obd_evict_list);
 699                 spin_unlock(&pet_lock);
 700
 701                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 702
 703                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 704                        obd->obd_name, expire_time);
 705
 706                 /* Exports can't be deleted out of the list while we hold
 707                  * the obd lock (class_unlink_export), which means we can't
 708                  * lose the last ref on the export.  If they've already been
 709                  * removed from the list, we won't find them here. */
 710                 spin_lock(&obd->obd_dev_lock);
 711                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 712                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 713                                              struct obd_export,
 714                                              exp_obd_chain_timed);
 715                         if (expire_time > exp->exp_last_request_time) {
 716                                 class_export_get(exp);
 717                                 spin_unlock(&obd->obd_dev_lock);
 718                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 719                                               " (at %s) in %ld seconds. I think"
 720                                               " it's dead, and I am evicting"
 721                                               " it. exp %p, cur %ld expire %ld"
 722                                               " last %ld\n",
 723                                               obd->obd_name,
 724                                               obd_uuid2str(&exp->exp_client_uuid),
 725                                               obd_export_nid2str(exp),
 726                                               (long)(cfs_time_current_sec() -
 727                                                      exp->exp_last_request_time),
 728                                               exp, (long)cfs_time_current_sec(),
 729                                               (long)expire_time,
 730                                               (long)exp->exp_last_request_time);
 731                                 CDEBUG(D_HA, "Last request was at %ld\n",
 732                                        exp->exp_last_request_time);
 733                                 class_fail_export(exp);
 734                                 class_export_put(exp);
 735                                 spin_lock(&obd->obd_dev_lock);
 736                         } else {
 737                                 /* List is sorted, so everyone below is ok */
 738                                 break;
 739                         }
 740                 }
 741                 spin_unlock(&obd->obd_dev_lock);
 742
 743                 spin_lock(&pet_lock);
 744                 cfs_list_del_init(&obd->obd_evict_list);
 745                 spin_unlock(&pet_lock);
 746
 747                 class_decref(obd, "evictor", obd);
 748         }
 749         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 750
 751         RETURN(0);
 752 }
 753
 754 void ping_evictor_start(void)
 755 {
 756         int rc;
 757
 758         if (++pet_refcount > 1)
 759                 return;
 760
 761         cfs_waitq_init(&pet_waitq);
 762
 763         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 764         if (rc < 0) {
 765                 pet_refcount--;
 766                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 767         }
 768 }
 769 EXPORT_SYMBOL(ping_evictor_start);
 770
 771 void ping_evictor_stop(void)
 772 {
 773         if (--pet_refcount > 0)
 774                 return;
 775
 776         pet_state = PET_TERMINATE;
 777         cfs_waitq_signal(&pet_waitq);
 778 }
 779 EXPORT_SYMBOL(ping_evictor_stop);
 780 #else /* !__KERNEL__ */
 781
 782 /* XXX
 783  * the current implementation of pinger in liblustre is not optimized
 784  */
 785
 786 #ifdef ENABLE_PINGER
 787 static struct pinger_data {
 788         int             pd_recursion;
 789         cfs_time_t      pd_this_ping;   /* jiffies */
 790         cfs_time_t      pd_next_ping;   /* jiffies */
 791         struct ptlrpc_request_set *pd_set;
 792 } pinger_args;
 793
 794 static int pinger_check_rpcs(void *arg)
 795 {
 796         cfs_time_t curtime = cfs_time_current();
 797         struct ptlrpc_request *req;
 798         struct ptlrpc_request_set *set;
 799         cfs_list_t *iter;
 800         struct obd_import *imp;
 801         struct pinger_data *pd = &pinger_args;
 802         int rc;
 803
 804         /* prevent recursion */
 805         if (pd->pd_recursion++) {
 806                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 807                 LASSERT(pd->pd_set);
 808                 pd->pd_recursion--;
 809                 return 0;
 810         }
 811
 812         /* have we reached ping point? */
 813         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 814                 pd->pd_recursion--;
 815                 return 0;
 816         }
 817
 818         /* if we have rpc_set already, continue processing it */
 819         if (pd->pd_set) {
 820                 LASSERT(pd->pd_this_ping);
 821                 set = pd->pd_set;
 822                 goto do_check_set;
 823         }
 824
 825         pd->pd_this_ping = curtime;
 826         pd->pd_set = ptlrpc_prep_set();
 827         if (pd->pd_set == NULL)
 828                 goto out;
 829         set = pd->pd_set;
 830
 831         /* add rpcs into set */
 832         mutex_lock(&pinger_mutex);
 833         cfs_list_for_each(iter, &pinger_imports) {
 834                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 835                                                         imp_pinger_chain);
 836                 int generation, level;
 837
 838                 if (cfs_time_aftereq(pd->pd_this_ping,
 839                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 840                         /* Add a ping. */
 841                         spin_lock(&imp->imp_lock);
 842                         generation = imp->imp_generation;
 843                         level = imp->imp_state;
 844                         spin_unlock(&imp->imp_lock);
 845
 846                         if (level != LUSTRE_IMP_FULL) {
 847                                 CDEBUG(D_HA,
 848                                        "not pinging %s (in recovery)\n",
 849                                        obd2cli_tgt(imp->imp_obd));
 850                                 continue;
 851                         }
 852
 853                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 854                                                         LUSTRE_OBD_VERSION,
 855                                                         OBD_PING);
 856                         if (req == NULL) {
 857                                 CERROR("OOM trying to ping %s->%s\n",
 858                                        imp->imp_obd->obd_uuid.uuid,
 859                                        obd2cli_tgt(imp->imp_obd));
 860                                 break;
 861                         }
 862
 863                         req->rq_no_resend = 1;
 864                         ptlrpc_request_set_replen(req);
 865                         req->rq_send_state = LUSTRE_IMP_FULL;
 866                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 867                         req->rq_import_generation = generation;
 868                         ptlrpc_set_add_req(set, req);
 869                 } else {
 870                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 871                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 872                                imp->imp_next_ping, pd->pd_this_ping);
 873                 }
 874         }
 875         pd->pd_this_ping = curtime;
 876         mutex_unlock(&pinger_mutex);
 877
 878         /* Might be empty, that's OK. */
 879         if (cfs_atomic_read(&set->set_remaining) == 0)
 880                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 881
 882         cfs_list_for_each(iter, &set->set_requests) {
 883                 struct ptlrpc_request *req =
 884                         cfs_list_entry(iter, struct ptlrpc_request,
 885                                        rq_set_chain);
 886                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 887                           req->rq_import->imp_obd->obd_uuid.uuid,
 888                           obd2cli_tgt(req->rq_import->imp_obd));
 889                 (void)ptl_send_rpc(req, 0);
 890         }
 891
 892 do_check_set:
 893         rc = ptlrpc_check_set(NULL, set);
 894
 895         /* not finished, and we are not expired, simply return */
 896         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 897                                             cfs_time_seconds(PING_INTERVAL)))) {
 898                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 899                 pd->pd_recursion--;
 900                 return 0;
 901         }
 902
 903         /* Expire all the requests that didn't come back. */
 904         mutex_lock(&pinger_mutex);
 905         cfs_list_for_each(iter, &set->set_requests) {
 906                 req = cfs_list_entry(iter, struct ptlrpc_request,
 907                                      rq_set_chain);
 908
 909                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 910                         continue;
 911
 912                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 913                        req);
 914
 915                 /* This will also unregister reply. */
 916                 ptlrpc_expire_one_request(req, 0);
 917
 918                 /* We're done with this req, let's finally move it to complete
 919                  * phase and take care of inflights. */
 920                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 921                 imp = req->rq_import;
 922                 spin_lock(&imp->imp_lock);
 923                 if (!cfs_list_empty(&req->rq_list)) {
 924                         cfs_list_del_init(&req->rq_list);
 925                         cfs_atomic_dec(&imp->imp_inflight);
 926                 }
 927                 spin_unlock(&imp->imp_lock);
 928                 cfs_atomic_dec(&set->set_remaining);
 929         }
 930         mutex_unlock(&pinger_mutex);
 931
 932         ptlrpc_set_destroy(set);
 933         pd->pd_set = NULL;
 934
 935 out:
 936         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 937                                         cfs_time_seconds(PING_INTERVAL));
 938         pd->pd_this_ping = 0; /* XXX for debug */
 939
 940         CDEBUG(D_INFO, "finished a round ping\n");
 941         pd->pd_recursion--;
 942         return 0;
 943 }
 944
 945 static void *pinger_callback = NULL;
 946 #endif /* ENABLE_PINGER */
 947
 948 int ptlrpc_start_pinger(void)
 949 {
 950 #ifdef ENABLE_PINGER
 951         memset(&pinger_args, 0, sizeof(pinger_args));
 952         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 953                                                            &pinger_check_rpcs,
 954                                                            &pinger_args);
 955 #endif
 956         return 0;
 957 }
 958
 959 int ptlrpc_stop_pinger(void)
 960 {
 961 #ifdef ENABLE_PINGER
 962         if (pinger_callback)
 963                 liblustre_deregister_wait_callback(pinger_callback);
 964 #endif
 965         return 0;
 966 }
 967
 968 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 969 {
 970 #ifdef ENABLE_PINGER
 971         mutex_lock(&pinger_mutex);
 972         ptlrpc_update_next_ping(imp, 0);
 973         if (pinger_args.pd_set == NULL &&
 974             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 975                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 976                         imp->imp_next_ping, cfs_time_current());
 977                 pinger_args.pd_next_ping = imp->imp_next_ping;
 978         }
 979         mutex_unlock(&pinger_mutex);
 980 #endif
 981 }
 982
 983 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 984 {
 985 #ifdef ENABLE_PINGER
 986         mutex_lock(&pinger_mutex);
 987         ptlrpc_update_next_ping(imp, 1);
 988         if (pinger_args.pd_set == NULL &&
 989             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 990                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 991                         imp->imp_next_ping, cfs_time_current());
 992                 pinger_args.pd_next_ping = imp->imp_next_ping;
 993         }
 994         mutex_unlock(&pinger_mutex);
 995 #endif
 996 }
 997
 998 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 999                               timeout_cb_t cb, void *data,
1000                               cfs_list_t *obd_list)
1001 {
1002         return 0;
1003 }
1004
1005 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
1006                               enum timeout_event event)
1007 {
1008         return 0;
1009 }
1010
1011 int ptlrpc_pinger_add_import(struct obd_import *imp)
1012 {
1013         ENTRY;
1014         if (!cfs_list_empty(&imp->imp_pinger_chain))
1015                 RETURN(-EALREADY);
1016
1017         CDEBUG(D_HA, "adding pingable import %s->%s\n",
1018                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1019         ptlrpc_pinger_sending_on_import(imp);
1020
1021         mutex_lock(&pinger_mutex);
1022         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1023         class_import_get(imp);
1024         mutex_unlock(&pinger_mutex);
1025
1026         RETURN(0);
1027 }
1028
1029 int ptlrpc_pinger_del_import(struct obd_import *imp)
1030 {
1031         ENTRY;
1032         if (cfs_list_empty(&imp->imp_pinger_chain))
1033                 RETURN(-ENOENT);
1034
1035         mutex_lock(&pinger_mutex);
1036         cfs_list_del_init(&imp->imp_pinger_chain);
1037         CDEBUG(D_HA, "removing pingable import %s->%s\n",
1038                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1039         class_import_put(imp);
1040         mutex_unlock(&pinger_mutex);
1041         RETURN(0);
1042 }
1043
1044 void ptlrpc_pinger_wake_up()
1045 {
1046 #ifdef ENABLE_PINGER
1047         /* XXX force pinger to run, if needed */
1048         struct obd_import *imp;
1049         ENTRY;
1050         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1051                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1052                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1053 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1054                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1055                     !imp_is_deactive(imp))
1056 #else
1057                 /*XXX only recover for the initial connection */
1058                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1059                     imp->imp_state == LUSTRE_IMP_DISCON &&
1060                     !imp_is_deactive(imp))
1061 #endif
1062                         ptlrpc_initiate_recovery(imp);
1063                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1064                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1065                                      "state %d, deactive %d\n",
1066                                      imp->imp_obd->obd_uuid.uuid,
1067                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1068                                      imp_is_deactive(imp));
1069         }
1070         EXIT;
1071 #endif
1072 }
1073 #endif /* !__KERNEL__ */