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