lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Portal-RPC reconnection and replay operations, for use in recovery.
   5  *
   6  *  Copyright (c) 2003 Cluster File Systems, Inc.
   7  *   Authors: Phil Schwan <phil@clusterfs.com>
   8  *            Mike Shaver <shaver@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef __KERNEL__
  30 #include <liblustre.h>
  31 #else
  32 #define DEBUG_SUBSYSTEM S_RPC
  33 #endif
  34
  35 #include <obd_support.h>
  36 #include <obd_class.h>
  37 #include "ptlrpc_internal.h"
  38
  39 struct semaphore pinger_sem;
  40 static struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
  41
  42 int ptlrpc_ping(struct obd_import *imp)
  43 {
  44         struct ptlrpc_request *req;
  45         int rc = 0;
  46         ENTRY;
  47
  48         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
  49         if (req) {
  50                 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  51                           imp->imp_obd->obd_uuid.uuid,
  52                           obd2cli_tgt(imp->imp_obd));
  53                 req->rq_no_resend = req->rq_no_delay = 1;
  54                 ptlrpc_req_set_repsize(req, 1, NULL);
  55                 req->rq_timeout = PING_INTERVAL;
  56                 ptlrpcd_add_req(req);
  57         } else {
  58                 CERROR("OOM trying to ping %s->%s\n",
  59                        imp->imp_obd->obd_uuid.uuid,
  60                        obd2cli_tgt(imp->imp_obd));
  61                 rc = -ENOMEM;
  62         }
  63
  64         RETURN(rc);
  65 }
  66
  67 static void ptlrpc_update_next_ping(struct obd_import *imp)
  68 {
  69         imp->imp_next_ping = cfs_time_shift(
  70                                 (imp->imp_state == LUSTRE_IMP_DISCON ?
  71                                  RECONNECT_INTERVAL : PING_INTERVAL));
  72 }
  73
  74 void ptlrpc_ping_import_soon(struct obd_import *imp)
  75 {
  76         imp->imp_next_ping = cfs_time_current();
  77 }
  78
  79 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
  80 {
  81         if (imp->imp_server_timeout)
  82                 return cfs_time_shift(obd_timeout / 2);
  83         else
  84                 return cfs_time_shift(obd_timeout);
  85 }
  86
  87 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
  88 static cfs_time_t suspend_wakeup_time = 0;
  89
  90 #ifdef __KERNEL__
  91 static wait_queue_head_t suspend_timeouts_waitq;
  92 #endif
  93
  94 cfs_time_t ptlrpc_suspend_wakeup_time(void)
  95 {
  96         return suspend_wakeup_time;
  97 }
  98
  99 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 100 {
 101         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 102 #if 0
 103         if (imp->imp_no_timeout)
 104                 return;
 105         imp->imp_no_timeout = 1;
 106         atomic_inc(&suspend_timeouts);
 107         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
 108 #endif
 109 }
 110
 111 void ptlrpc_activate_timeouts(struct obd_import *imp)
 112 {
 113         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 114 #if 0
 115         if (!imp->imp_no_timeout)
 116                 return;
 117         imp->imp_no_timeout = 0;
 118         LASSERT(atomic_read(&suspend_timeouts) > 0);
 119         if (atomic_dec_and_test(&suspend_timeouts)) {
 120                 suspend_wakeup_time = cfs_time_current();
 121                 wake_up(&suspend_timeouts_waitq);
 122         }
 123         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
 124 #endif
 125 }
 126
 127 int ptlrpc_check_suspend(void)
 128 {
 129         if (atomic_read(&suspend_timeouts))
 130                 return 1;
 131         return 0;
 132 }
 133
 134 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 135 {
 136         struct l_wait_info lwi;
 137
 138         if (atomic_read(&suspend_timeouts)) {
 139                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 140                           atomic_read(&suspend_timeouts));
 141                 lwi = LWI_INTR(NULL, NULL);
 142                 l_wait_event(suspend_timeouts_waitq,
 143                              atomic_read(&suspend_timeouts) == 0, &lwi);
 144                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 145                 return 1;
 146         }
 147         return 0;
 148 }
 149
 150 #ifdef __KERNEL__
 151
 152 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 153                                          unsigned long this_ping)
 154 {
 155         int force, level;
 156
 157         spin_lock(&imp->imp_lock);
 158         level = imp->imp_state;
 159         force = imp->imp_force_verify;
 160         if (force)
 161                 imp->imp_force_verify = 0;
 162         spin_unlock(&imp->imp_lock);
 163
 164         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 165                "level %s/%u force %u deactive %u pingable %u\n",
 166                ptlrpc_import_state_name(level), level,
 167                force, imp->imp_deactive, imp->imp_pingable);
 168
 169         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 170                              this_ping) && force == 0)
 171                 return;
 172
 173         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 174                 /* wait at least a timeout before trying recovery again */
 175                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 176                 ptlrpc_initiate_recovery(imp);
 177         } else if (level != LUSTRE_IMP_FULL ||
 178                    imp->imp_obd->obd_no_recov ||
 179                    imp->imp_deactive) {
 180                 CDEBUG(D_HA, "not pinging %s (in recovery "
 181                        " or recovery disabled: %s)\n",
 182                        obd2cli_tgt(imp->imp_obd),
 183                        ptlrpc_import_state_name(level));
 184         } else if (imp->imp_pingable || force) {
 185                 ptlrpc_ping(imp);
 186         }
 187 }
 188
 189 static int ptlrpc_pinger_main(void *arg)
 190 {
 191         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 192         struct ptlrpc_thread *thread = data->thread;
 193         ENTRY;
 194
 195         cfs_daemonize(data->name);
 196
 197         /* Record that the thread is running */
 198         thread->t_flags = SVC_RUNNING;
 199         cfs_waitq_signal(&thread->t_ctl_waitq);
 200
 201         /* And now, loop forever, pinging as needed. */
 202         while (1) {
 203                 cfs_time_t this_ping = cfs_time_current();
 204                 struct l_wait_info lwi;
 205                 cfs_duration_t time_to_next_ping;
 206                 struct list_head *iter;
 207
 208                 mutex_down(&pinger_sem);
 209                 list_for_each(iter, &pinger_imports) {
 210                         struct obd_import *imp =
 211                                 list_entry(iter, struct obd_import,
 212                                            imp_pinger_chain);
 213
 214                         ptlrpc_pinger_process_import(imp, this_ping);
 215                         /* obd_timeout might have changed */
 216                         if (imp->imp_pingable && imp->imp_next_ping &&
 217                             cfs_time_after(imp->imp_next_ping,
 218                                            cfs_time_add(this_ping,
 219                                                         cfs_time_seconds(PING_INTERVAL))))
 220                                 ptlrpc_update_next_ping(imp);
 221                 }
 222                 mutex_up(&pinger_sem);
 223
 224                 /* Wait until the next ping time, or until we're stopped. */
 225                 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
 226                                                  cfs_time_seconds(PING_INTERVAL)),
 227                                                  cfs_time_current());
 228
 229                 /* The ping sent by ptlrpc_send_rpc may get sent out
 230                    say .01 second after this.
 231                    ptlrpc_pinger_eending_on_import will then set the
 232                    next ping time to next_ping + .01 sec, which means
 233                    we will SKIP the next ping at next_ping, and the
 234                    ping will get sent 2 timeouts from now!  Beware. */
 235                 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 236                                time_to_next_ping,
 237                                cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
 238                 if (time_to_next_ping > 0) {
 239                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_ping, cfs_time_seconds(1)),
 240                                           NULL, NULL);
 241                         l_wait_event(thread->t_ctl_waitq,
 242                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 243                                      &lwi);
 244                         if (thread->t_flags & SVC_STOPPING) {
 245                                 thread->t_flags &= ~SVC_STOPPING;
 246                                 EXIT;
 247                                 break;
 248                         } else if (thread->t_flags & SVC_EVENT) {
 249                                 /* woken after adding import to reset timer */
 250                                 thread->t_flags &= ~SVC_EVENT;
 251                         }
 252                 }
 253         }
 254
 255         thread->t_flags = SVC_STOPPED;
 256         cfs_waitq_signal(&thread->t_ctl_waitq);
 257
 258         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 259         return 0;
 260 }
 261
 262 static struct ptlrpc_thread *pinger_thread = NULL;
 263
 264 int ptlrpc_start_pinger(void)
 265 {
 266         struct l_wait_info lwi = { 0 };
 267         struct ptlrpc_svc_data d;
 268         int rc;
 269 #ifndef ENABLE_PINGER
 270         return 0;
 271 #endif
 272         ENTRY;
 273
 274         if (pinger_thread != NULL)
 275                 RETURN(-EALREADY);
 276
 277         OBD_ALLOC_PTR(pinger_thread);
 278         if (pinger_thread == NULL)
 279                 RETURN(-ENOMEM);
 280         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 281         cfs_waitq_init(&suspend_timeouts_waitq);
 282
 283         d.name = "ll_ping";
 284         d.thread = pinger_thread;
 285
 286         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 287          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 288         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 289         if (rc < 0) {
 290                 CERROR("cannot start thread: %d\n", rc);
 291                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 292                 pinger_thread = NULL;
 293                 RETURN(rc);
 294         }
 295         l_wait_event(pinger_thread->t_ctl_waitq,
 296                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 297
 298         RETURN(0);
 299 }
 300
 301 int ptlrpc_stop_pinger(void)
 302 {
 303         struct l_wait_info lwi = { 0 };
 304         int rc = 0;
 305 #ifndef ENABLE_PINGER
 306         return 0;
 307 #endif
 308         ENTRY;
 309
 310         if (pinger_thread == NULL)
 311                 RETURN(-EALREADY);
 312         mutex_down(&pinger_sem);
 313         pinger_thread->t_flags = SVC_STOPPING;
 314         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 315         mutex_up(&pinger_sem);
 316
 317         l_wait_event(pinger_thread->t_ctl_waitq,
 318                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 319
 320         OBD_FREE_PTR(pinger_thread);
 321         pinger_thread = NULL;
 322         RETURN(rc);
 323 }
 324
 325 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 326 {
 327         ptlrpc_update_next_ping(imp);
 328 }
 329
 330 int ptlrpc_pinger_add_import(struct obd_import *imp)
 331 {
 332         ENTRY;
 333         if (!list_empty(&imp->imp_pinger_chain))
 334                 RETURN(-EALREADY);
 335
 336         mutex_down(&pinger_sem);
 337         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 338                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 339         /* if we add to pinger we want recovery on this import */
 340         imp->imp_obd->obd_no_recov = 0;
 341         ptlrpc_update_next_ping(imp);
 342         /* XXX sort, blah blah */
 343         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 344         class_import_get(imp);
 345
 346         ptlrpc_pinger_wake_up();
 347         mutex_up(&pinger_sem);
 348
 349         RETURN(0);
 350 }
 351
 352 int ptlrpc_pinger_del_import(struct obd_import *imp)
 353 {
 354         ENTRY;
 355         if (list_empty(&imp->imp_pinger_chain))
 356                 RETURN(-ENOENT);
 357
 358         mutex_down(&pinger_sem);
 359         list_del_init(&imp->imp_pinger_chain);
 360         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 361                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 362         /* if we remove from pinger we don't want recovery on this import */
 363         imp->imp_obd->obd_no_recov = 1;
 364         class_import_put(imp);
 365         mutex_up(&pinger_sem);
 366         RETURN(0);
 367 }
 368
 369 void ptlrpc_pinger_wake_up()
 370 {
 371 #ifdef ENABLE_PINGER
 372         pinger_thread->t_flags |= SVC_EVENT;
 373         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 374 #endif
 375 }
 376
 377 /* Ping evictor thread */
 378 #define PET_READY     1
 379 #define PET_TERMINATE 2
 380
 381 static int               pet_refcount = 0;
 382 static int               pet_state;
 383 static wait_queue_head_t pet_waitq;
 384 static struct obd_export *pet_exp = NULL;
 385 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 386
 387 int ping_evictor_wake(struct obd_export *exp)
 388 {
 389         spin_lock(&pet_lock);
 390         if (pet_exp) {
 391                 /* eventually the new obd will call here again. */
 392                 spin_unlock(&pet_lock);
 393                 return 1;
 394         }
 395
 396         /* We have to make sure the obd isn't destroyed between now and when
 397          * the ping evictor runs.  We'll take a reference here, and drop it
 398          * when we finish in the evictor.  We don't really care about this
 399          * export in particular; we just need one to keep the obd alive. */
 400         pet_exp = class_export_get(exp);
 401         spin_unlock(&pet_lock);
 402
 403         wake_up(&pet_waitq);
 404         return 0;
 405 }
 406
 407 static int ping_evictor_main(void *arg)
 408 {
 409         struct obd_device *obd;
 410         struct obd_export *exp;
 411         struct l_wait_info lwi = { 0 };
 412         time_t expire_time;
 413         ENTRY;
 414
 415         ptlrpc_daemonize("ll_evictor");
 416
 417         CDEBUG(D_HA, "Starting Ping Evictor\n");
 418         pet_exp = NULL;
 419         pet_state = PET_READY;
 420         while (1) {
 421                 l_wait_event(pet_waitq, pet_exp ||
 422                              (pet_state == PET_TERMINATE), &lwi);
 423                 if (pet_state == PET_TERMINATE)
 424                         break;
 425
 426                 /* we only get here if pet_exp != NULL, and the end of this
 427                  * loop is the only place which sets it NULL again, so lock
 428                  * is not strictly necessary. */
 429                 spin_lock(&pet_lock);
 430                 obd = pet_exp->exp_obd;
 431                 spin_unlock(&pet_lock);
 432
 433                 expire_time = CURRENT_SECONDS - (3 * obd_timeout / 2);
 434
 435                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 436                        obd->obd_name, expire_time);
 437
 438                 /* Exports can't be deleted out of the list while we hold
 439                  * the obd lock (class_unlink_export), which means we can't
 440                  * lose the last ref on the export.  If they've already been
 441                  * removed from the list, we won't find them here. */
 442                 spin_lock(&obd->obd_dev_lock);
 443                 while (!list_empty(&obd->obd_exports_timed)) {
 444                         exp = list_entry(obd->obd_exports_timed.next,
 445                                          struct obd_export,exp_obd_chain_timed);
 446                         if (expire_time > exp->exp_last_request_time) {
 447                                 class_export_get(exp);
 448                                 spin_unlock(&obd->obd_dev_lock);
 449                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 450                                               " (at %s) in %ld seconds. I think"
 451                                               " it's dead, and I am evicting"
 452                                               " it. exp %p, cur %ld expire %ld"
 453                                               " last %ld\n",
 454                                               obd->obd_name,
 455                                               obd_uuid2str(&exp->exp_client_uuid),
 456                                               obd_export_nid2str(exp),
 457                                               (long)(CURRENT_SECONDS -
 458                                                      exp->exp_last_request_time),
 459                                               exp, (long)CURRENT_SECONDS,
 460                                               (long)expire_time,
 461                                               (long)exp->exp_last_request_time);
 462                                 CDEBUG(D_HA, "Last request was at %ld\n",
 463                                        exp->exp_last_request_time);
 464                                 class_fail_export(exp);
 465                                 class_export_put(exp);
 466                                 spin_lock(&obd->obd_dev_lock);
 467                         } else {
 468                                 /* List is sorted, so everyone below is ok */
 469                                 break;
 470                         }
 471                 }
 472                 spin_unlock(&obd->obd_dev_lock);
 473
 474                 class_export_put(pet_exp);
 475
 476                 spin_lock(&pet_lock);
 477                 pet_exp = NULL;
 478                 spin_unlock(&pet_lock);
 479         }
 480         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 481
 482         RETURN(0);
 483 }
 484
 485 void ping_evictor_start(void)
 486 {
 487         int rc;
 488
 489         if (++pet_refcount > 1)
 490                 return;
 491
 492         init_waitqueue_head(&pet_waitq);
 493
 494         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 495         if (rc < 0) {
 496                 pet_refcount--;
 497                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 498         }
 499 }
 500 EXPORT_SYMBOL(ping_evictor_start);
 501
 502 void ping_evictor_stop(void)
 503 {
 504         if (--pet_refcount > 0)
 505                 return;
 506
 507         pet_state = PET_TERMINATE;
 508         wake_up(&pet_waitq);
 509 }
 510 EXPORT_SYMBOL(ping_evictor_stop);
 511 #else /* !__KERNEL__ */
 512
 513 /* XXX
 514  * the current implementation of pinger in liblustre is not optimized
 515  */
 516
 517 #ifdef ENABLE_PINGER
 518 static struct pinger_data {
 519         int             pd_recursion;
 520         cfs_time_t      pd_this_ping;   /* jiffies */
 521         cfs_time_t      pd_next_ping;   /* jiffies */
 522         struct ptlrpc_request_set *pd_set;
 523 } pinger_args;
 524
 525 static int pinger_check_rpcs(void *arg)
 526 {
 527         cfs_time_t curtime = cfs_time_current();
 528         struct ptlrpc_request *req;
 529         struct ptlrpc_request_set *set;
 530         struct list_head *iter;
 531         struct pinger_data *pd = &pinger_args;
 532         int rc;
 533
 534         /* prevent recursion */
 535         if (pd->pd_recursion++) {
 536                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 537                 LASSERT(pd->pd_set);
 538                 pd->pd_recursion--;
 539                 return 0;
 540         }
 541
 542         /* have we reached ping point? */
 543         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 544                 pd->pd_recursion--;
 545                 return 0;
 546         }
 547
 548         /* if we have rpc_set already, continue processing it */
 549         if (pd->pd_set) {
 550                 LASSERT(pd->pd_this_ping);
 551                 set = pd->pd_set;
 552                 goto do_check_set;
 553         }
 554
 555         pd->pd_this_ping = curtime;
 556         pd->pd_set = ptlrpc_prep_set();
 557         if (pd->pd_set == NULL)
 558                 goto out;
 559         set = pd->pd_set;
 560
 561         /* add rpcs into set */
 562         mutex_down(&pinger_sem);
 563         list_for_each(iter, &pinger_imports) {
 564                 struct obd_import *imp =
 565                         list_entry(iter, struct obd_import, imp_pinger_chain);
 566                 int generation, level;
 567
 568                 if (cfs_time_aftereq(pd->pd_this_ping,
 569                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 570                         /* Add a ping. */
 571                         spin_lock(&imp->imp_lock);
 572                         generation = imp->imp_generation;
 573                         level = imp->imp_state;
 574                         spin_unlock(&imp->imp_lock);
 575
 576                         if (level != LUSTRE_IMP_FULL) {
 577                                 CDEBUG(D_HA,
 578                                        "not pinging %s (in recovery)\n",
 579                                        obd2cli_tgt(imp->imp_obd));
 580                                 continue;
 581                         }
 582
 583                         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
 584                                               1, NULL, NULL);
 585                         if (!req) {
 586                                 CERROR("out of memory\n");
 587                                 break;
 588                         }
 589                         req->rq_no_resend = 1;
 590                         ptlrpc_req_set_repsize(req, 1, NULL);
 591                         req->rq_send_state = LUSTRE_IMP_FULL;
 592                         req->rq_phase = RQ_PHASE_RPC;
 593                         req->rq_import_generation = generation;
 594                         ptlrpc_set_add_req(set, req);
 595                 } else {
 596                         CDEBUG(D_HA, "don't need to ping %s ("CFS_TIME_T" > "
 597                                CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 598                                imp->imp_next_ping, pd->pd_this_ping);
 599                 }
 600         }
 601         pd->pd_this_ping = curtime;
 602         mutex_up(&pinger_sem);
 603
 604         /* Might be empty, that's OK. */
 605         if (set->set_remaining == 0)
 606                 CDEBUG(D_HA, "nothing to ping\n");
 607
 608         list_for_each(iter, &set->set_requests) {
 609                 struct ptlrpc_request *req =
 610                         list_entry(iter, struct ptlrpc_request,
 611                                    rq_set_chain);
 612                 DEBUG_REQ(D_HA, req, "pinging %s->%s",
 613                           req->rq_import->imp_obd->obd_uuid.uuid,
 614                           obd2cli_tgt(req->rq_import->imp_obd));
 615                 (void)ptl_send_rpc(req, 0);
 616         }
 617
 618 do_check_set:
 619         rc = ptlrpc_check_set(set);
 620
 621         /* not finished, and we are not expired, simply return */
 622         if (!rc && cfs_time_before(curtime,
 623                                    cfs_time_add(pd->pd_this_ping,
 624                                                 cfs_time_seconds(PING_INTERVAL)))) {
 625                 CDEBUG(D_HA, "not finished, but also not expired\n");
 626                 pd->pd_recursion--;
 627                 return 0;
 628         }
 629
 630         /* Expire all the requests that didn't come back. */
 631         mutex_down(&pinger_sem);
 632         list_for_each(iter, &set->set_requests) {
 633                 req = list_entry(iter, struct ptlrpc_request,
 634                                  rq_set_chain);
 635
 636                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 637                         continue;
 638
 639                 req->rq_phase = RQ_PHASE_COMPLETE;
 640                 atomic_dec(&req->rq_import->imp_inflight);
 641                 set->set_remaining--;
 642                 /* If it was disconnected, don't sweat it. */
 643                 if (list_empty(&req->rq_import->imp_pinger_chain)) {
 644                         ptlrpc_unregister_reply(req);
 645                         continue;
 646                 }
 647
 648                 CDEBUG(D_HA, "pinger initiate expire_one_request\n");
 649                 ptlrpc_expire_one_request(req);
 650         }
 651         mutex_up(&pinger_sem);
 652
 653         ptlrpc_set_destroy(set);
 654         pd->pd_set = NULL;
 655
 656 out:
 657         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 658                                         cfs_time_seconds(PING_INTERVAL));
 659         pd->pd_this_ping = 0; /* XXX for debug */
 660
 661         CDEBUG(D_HA, "finished a round ping\n");
 662         pd->pd_recursion--;
 663         return 0;
 664 }
 665
 666 static void *pinger_callback = NULL;
 667 #endif /* ENABLE_PINGER */
 668
 669 int ptlrpc_start_pinger(void)
 670 {
 671 #ifdef ENABLE_PINGER
 672         memset(&pinger_args, 0, sizeof(pinger_args));
 673         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 674                                                            &pinger_check_rpcs,
 675                                                            &pinger_args);
 676 #endif
 677         return 0;
 678 }
 679
 680 int ptlrpc_stop_pinger(void)
 681 {
 682 #ifdef ENABLE_PINGER
 683         if (pinger_callback)
 684                 liblustre_deregister_wait_callback(pinger_callback);
 685 #endif
 686         return 0;
 687 }
 688
 689 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 690 {
 691 #ifdef ENABLE_PINGER
 692         mutex_down(&pinger_sem);
 693         ptlrpc_update_next_ping(imp);
 694         if (pinger_args.pd_set == NULL &&
 695             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 696                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 697                         imp->imp_next_ping, cfs_time_current());
 698                 pinger_args.pd_next_ping = imp->imp_next_ping;
 699         }
 700         mutex_up(&pinger_sem);
 701 #endif
 702 }
 703
 704 int ptlrpc_pinger_add_import(struct obd_import *imp)
 705 {
 706         ENTRY;
 707         if (!list_empty(&imp->imp_pinger_chain))
 708                 RETURN(-EALREADY);
 709
 710         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 711                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 712         ptlrpc_pinger_sending_on_import(imp);
 713
 714         mutex_down(&pinger_sem);
 715         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 716         class_import_get(imp);
 717         mutex_up(&pinger_sem);
 718
 719         RETURN(0);
 720 }
 721
 722 int ptlrpc_pinger_del_import(struct obd_import *imp)
 723 {
 724         ENTRY;
 725         if (list_empty(&imp->imp_pinger_chain))
 726                 RETURN(-ENOENT);
 727
 728         mutex_down(&pinger_sem);
 729         list_del_init(&imp->imp_pinger_chain);
 730         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 731                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 732         class_import_put(imp);
 733         mutex_up(&pinger_sem);
 734         RETURN(0);
 735 }
 736
 737 void ptlrpc_pinger_wake_up()
 738 {
 739 #ifdef ENABLE_PINGER
 740         /* XXX force pinger to run, if needed */
 741         struct obd_import *imp;
 742         ENTRY;
 743         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 744                 CDEBUG(D_HA, "Checking that we need to do anything about import"
 745                              " %s->%s\n", imp->imp_obd->obd_uuid.uuid,
 746                              obd2cli_tgt(imp->imp_obd));
 747 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 748                 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 749 #else
 750                 /*XXX only recover for the initial connection */
 751                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 752                     imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 753 #endif
 754                         ptlrpc_initiate_recovery(imp);
 755                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 756                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 757                                      "state %d, deactive %d\n",
 758                                      imp->imp_obd->obd_uuid.uuid,
 759                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 760                                      imp->imp_deactive);
 761         }
 762         EXIT;
 763 #endif
 764 }
 765 #endif /* !__KERNEL__ */