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         ENTRY;
  46
  47         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
  48                                         OBD_PING);
  49         if (req == NULL) {
  50                 CERROR("OOM trying to ping %s->%s\n",
  51                        imp->imp_obd->obd_uuid.uuid,
  52                        obd2cli_tgt(imp->imp_obd));
  53                 RETURN(-ENOMEM);
  54         }
  55
  56         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  57                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
  58         req->rq_no_resend = req->rq_no_delay = 1;
  59         ptlrpc_request_set_replen(req);
  60         req->rq_timeout = PING_INTERVAL;
  61         ptlrpcd_add_req(req);
  62
  63         RETURN(0);
  64 }
  65
  66 void ptlrpc_update_next_ping(struct obd_import *imp)
  67 {
  68 #ifdef ENABLE_PINGER
  69         imp->imp_next_ping = cfs_time_shift(
  70                                 (imp->imp_state == LUSTRE_IMP_DISCON ?
  71                                  RECONNECT_INTERVAL : PING_INTERVAL));
  72 #endif /* ENABLE_PINGER */
  73 }
  74
  75 void ptlrpc_ping_import_soon(struct obd_import *imp)
  76 {
  77         imp->imp_next_ping = cfs_time_current();
  78 }
  79
  80 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
  81 {
  82         if (imp->imp_server_timeout)
  83                 return cfs_time_shift(obd_timeout / 2);
  84         else
  85                 return cfs_time_shift(obd_timeout);
  86 }
  87
  88 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
  89 static cfs_time_t suspend_wakeup_time = 0;
  90
  91 #ifdef __KERNEL__
  92 static wait_queue_head_t suspend_timeouts_waitq;
  93 #endif
  94
  95 cfs_time_t ptlrpc_suspend_wakeup_time(void)
  96 {
  97         return suspend_wakeup_time;
  98 }
  99
 100 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 101 {
 102         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 103 #if 0
 104         if (imp->imp_no_timeout)
 105                 return;
 106         imp->imp_no_timeout = 1;
 107         atomic_inc(&suspend_timeouts);
 108         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
 109 #endif
 110 }
 111
 112 void ptlrpc_activate_timeouts(struct obd_import *imp)
 113 {
 114         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 115 #if 0
 116         if (!imp->imp_no_timeout)
 117                 return;
 118         imp->imp_no_timeout = 0;
 119         LASSERT(atomic_read(&suspend_timeouts) > 0);
 120         if (atomic_dec_and_test(&suspend_timeouts)) {
 121                 suspend_wakeup_time = cfs_time_current();
 122                 wake_up(&suspend_timeouts_waitq);
 123         }
 124         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
 125 #endif
 126 }
 127
 128 int ptlrpc_check_suspend(void)
 129 {
 130         if (atomic_read(&suspend_timeouts))
 131                 return 1;
 132         return 0;
 133 }
 134
 135 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 136 {
 137         struct l_wait_info lwi;
 138
 139         if (atomic_read(&suspend_timeouts)) {
 140                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 141                           atomic_read(&suspend_timeouts));
 142                 lwi = LWI_INTR(NULL, NULL);
 143                 l_wait_event(suspend_timeouts_waitq,
 144                              atomic_read(&suspend_timeouts) == 0, &lwi);
 145                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 146                 return 1;
 147         }
 148         return 0;
 149 }
 150
 151 #ifdef __KERNEL__
 152
 153 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 154                                          unsigned long this_ping)
 155 {
 156         int force, level;
 157
 158         spin_lock(&imp->imp_lock);
 159         level = imp->imp_state;
 160         force = imp->imp_force_verify;
 161         if (force)
 162                 imp->imp_force_verify = 0;
 163         spin_unlock(&imp->imp_lock);
 164
 165         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 166                "level %s/%u force %u deactive %u pingable %u\n",
 167                ptlrpc_import_state_name(level), level,
 168                force, imp->imp_deactive, imp->imp_pingable);
 169
 170         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 171                              this_ping) && force == 0)
 172                 return;
 173
 174         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 175                 /* wait at least a timeout before trying recovery again */
 176                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 177                 ptlrpc_initiate_recovery(imp);
 178         } else if (level != LUSTRE_IMP_FULL ||
 179                    imp->imp_obd->obd_no_recov ||
 180                    imp->imp_deactive) {
 181                 CDEBUG(D_HA, "not pinging %s (in recovery "
 182                        " or recovery disabled: %s)\n",
 183                        obd2cli_tgt(imp->imp_obd),
 184                        ptlrpc_import_state_name(level));
 185         } else if (imp->imp_pingable || force) {
 186                 ptlrpc_ping(imp);
 187         }
 188 }
 189
 190 static int ptlrpc_pinger_main(void *arg)
 191 {
 192         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 193         struct ptlrpc_thread *thread = data->thread;
 194         ENTRY;
 195
 196         cfs_daemonize(data->name);
 197
 198         /* Record that the thread is running */
 199         thread->t_flags = SVC_RUNNING;
 200         cfs_waitq_signal(&thread->t_ctl_waitq);
 201
 202         /* And now, loop forever, pinging as needed. */
 203         while (1) {
 204                 cfs_time_t this_ping = cfs_time_current();
 205                 struct l_wait_info lwi;
 206                 cfs_duration_t time_to_next_ping;
 207                 struct list_head *iter;
 208
 209                 mutex_down(&pinger_sem);
 210                 list_for_each(iter, &pinger_imports) {
 211                         struct obd_import *imp =
 212                                 list_entry(iter, struct obd_import,
 213                                            imp_pinger_chain);
 214
 215                         ptlrpc_pinger_process_import(imp, this_ping);
 216                         /* obd_timeout might have changed */
 217                         if (imp->imp_pingable && imp->imp_next_ping &&
 218                             cfs_time_after(imp->imp_next_ping,
 219                                            cfs_time_add(this_ping,
 220                                                         cfs_time_seconds(PING_INTERVAL))))
 221                                 ptlrpc_update_next_ping(imp);
 222                 }
 223                 mutex_up(&pinger_sem);
 224                 /* update memory usage info */
 225                 obd_update_maxusage();
 226
 227                 /* Wait until the next ping time, or until we're stopped. */
 228                 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
 229                                                  cfs_time_seconds(PING_INTERVAL)),
 230                                                  cfs_time_current());
 231
 232                 /* The ping sent by ptlrpc_send_rpc may get sent out
 233                    say .01 second after this.
 234                    ptlrpc_pinger_eending_on_import will then set the
 235                    next ping time to next_ping + .01 sec, which means
 236                    we will SKIP the next ping at next_ping, and the
 237                    ping will get sent 2 timeouts from now!  Beware. */
 238                 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 239                                time_to_next_ping,
 240                                cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
 241                 if (time_to_next_ping > 0) {
 242                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_ping, cfs_time_seconds(1)),
 243                                           NULL, NULL);
 244                         l_wait_event(thread->t_ctl_waitq,
 245                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 246                                      &lwi);
 247                         if (thread->t_flags & SVC_STOPPING) {
 248                                 thread->t_flags &= ~SVC_STOPPING;
 249                                 EXIT;
 250                                 break;
 251                         } else if (thread->t_flags & SVC_EVENT) {
 252                                 /* woken after adding import to reset timer */
 253                                 thread->t_flags &= ~SVC_EVENT;
 254                         }
 255                 }
 256         }
 257
 258         thread->t_flags = SVC_STOPPED;
 259         cfs_waitq_signal(&thread->t_ctl_waitq);
 260
 261         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 262         return 0;
 263 }
 264
 265 static struct ptlrpc_thread *pinger_thread = NULL;
 266
 267 int ptlrpc_start_pinger(void)
 268 {
 269         struct l_wait_info lwi = { 0 };
 270         struct ptlrpc_svc_data d;
 271         int rc;
 272 #ifndef ENABLE_PINGER
 273         return 0;
 274 #endif
 275         ENTRY;
 276
 277         if (pinger_thread != NULL)
 278                 RETURN(-EALREADY);
 279
 280         OBD_ALLOC_PTR(pinger_thread);
 281         if (pinger_thread == NULL)
 282                 RETURN(-ENOMEM);
 283         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 284         cfs_waitq_init(&suspend_timeouts_waitq);
 285
 286         d.name = "ll_ping";
 287         d.thread = pinger_thread;
 288
 289         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 290          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 291         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 292         if (rc < 0) {
 293                 CERROR("cannot start thread: %d\n", rc);
 294                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 295                 pinger_thread = NULL;
 296                 RETURN(rc);
 297         }
 298         l_wait_event(pinger_thread->t_ctl_waitq,
 299                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 300
 301         RETURN(0);
 302 }
 303
 304 int ptlrpc_stop_pinger(void)
 305 {
 306         struct l_wait_info lwi = { 0 };
 307         int rc = 0;
 308 #ifndef ENABLE_PINGER
 309         return 0;
 310 #endif
 311         ENTRY;
 312
 313         if (pinger_thread == NULL)
 314                 RETURN(-EALREADY);
 315         mutex_down(&pinger_sem);
 316         pinger_thread->t_flags = SVC_STOPPING;
 317         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 318         mutex_up(&pinger_sem);
 319
 320         l_wait_event(pinger_thread->t_ctl_waitq,
 321                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 322
 323         OBD_FREE_PTR(pinger_thread);
 324         pinger_thread = NULL;
 325         RETURN(rc);
 326 }
 327
 328 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 329 {
 330         ptlrpc_update_next_ping(imp);
 331 }
 332
 333 int ptlrpc_pinger_add_import(struct obd_import *imp)
 334 {
 335         ENTRY;
 336         if (!list_empty(&imp->imp_pinger_chain))
 337                 RETURN(-EALREADY);
 338
 339         mutex_down(&pinger_sem);
 340         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 341                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 342         /* if we add to pinger we want recovery on this import */
 343         imp->imp_obd->obd_no_recov = 0;
 344         ptlrpc_update_next_ping(imp);
 345         /* XXX sort, blah blah */
 346         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 347         class_import_get(imp);
 348
 349         ptlrpc_pinger_wake_up();
 350         mutex_up(&pinger_sem);
 351
 352         RETURN(0);
 353 }
 354
 355 int ptlrpc_pinger_del_import(struct obd_import *imp)
 356 {
 357         ENTRY;
 358         if (list_empty(&imp->imp_pinger_chain))
 359                 RETURN(-ENOENT);
 360
 361         mutex_down(&pinger_sem);
 362         list_del_init(&imp->imp_pinger_chain);
 363         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 364                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 365         /* if we remove from pinger we don't want recovery on this import */
 366         imp->imp_obd->obd_no_recov = 1;
 367         class_import_put(imp);
 368         mutex_up(&pinger_sem);
 369         RETURN(0);
 370 }
 371
 372 void ptlrpc_pinger_wake_up()
 373 {
 374 #ifdef ENABLE_PINGER
 375         pinger_thread->t_flags |= SVC_EVENT;
 376         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 377 #endif
 378 }
 379
 380 /* Ping evictor thread */
 381 #define PET_READY     1
 382 #define PET_TERMINATE 2
 383
 384 static int               pet_refcount = 0;
 385 static int               pet_state;
 386 static wait_queue_head_t pet_waitq;
 387 static struct obd_export *pet_exp = NULL;
 388 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 389
 390 int ping_evictor_wake(struct obd_export *exp)
 391 {
 392         spin_lock(&pet_lock);
 393         if (pet_exp || (pet_state != PET_READY)) {
 394                 /* eventually the new obd will call here again. */
 395                 spin_unlock(&pet_lock);
 396                 return 1;
 397         }
 398
 399         /* We have to make sure the obd isn't destroyed between now and when
 400          * the ping evictor runs.  We'll take a reference here, and drop it
 401          * when we finish in the evictor.  We don't really care about this
 402          * export in particular; we just need one to keep the obd alive. */
 403         pet_exp = class_export_get(exp);
 404         spin_unlock(&pet_lock);
 405
 406         wake_up(&pet_waitq);
 407         return 0;
 408 }
 409
 410 static int ping_evictor_main(void *arg)
 411 {
 412         struct obd_device *obd;
 413         struct obd_export *exp;
 414         struct l_wait_info lwi = { 0 };
 415         time_t expire_time;
 416         ENTRY;
 417
 418         ptlrpc_daemonize("ll_evictor");
 419
 420         CDEBUG(D_HA, "Starting Ping Evictor\n");
 421         pet_exp = NULL;
 422         pet_state = PET_READY;
 423         while (1) {
 424                 l_wait_event(pet_waitq, pet_exp ||
 425                              (pet_state == PET_TERMINATE), &lwi);
 426                 if (pet_state == PET_TERMINATE)
 427                         break;
 428
 429                 /* we only get here if pet_exp != NULL, and the end of this
 430                  * loop is the only place which sets it NULL again, so lock
 431                  * is not strictly necessary. */
 432                 spin_lock(&pet_lock);
 433                 obd = pet_exp->exp_obd;
 434                 spin_unlock(&pet_lock);
 435
 436                 expire_time = cfs_time_current_sec() - (3 * obd_timeout / 2);
 437
 438                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 439                        obd->obd_name, expire_time);
 440
 441                 /* Exports can't be deleted out of the list while we hold
 442                  * the obd lock (class_unlink_export), which means we can't
 443                  * lose the last ref on the export.  If they've already been
 444                  * removed from the list, we won't find them here. */
 445                 spin_lock(&obd->obd_dev_lock);
 446                 while (!list_empty(&obd->obd_exports_timed)) {
 447                         exp = list_entry(obd->obd_exports_timed.next,
 448                                          struct obd_export,exp_obd_chain_timed);
 449                         if (expire_time > exp->exp_last_request_time) {
 450                                 class_export_get(exp);
 451                                 spin_unlock(&obd->obd_dev_lock);
 452                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 453                                               " (at %s) in %ld seconds. I think"
 454                                               " it's dead, and I am evicting"
 455                                               " it. exp %p, cur %ld expire %ld"
 456                                               " last %ld\n",
 457                                               obd->obd_name,
 458                                               obd_uuid2str(&exp->exp_client_uuid),
 459                                               obd_export_nid2str(exp),
 460                                               (long)(cfs_time_current_sec() -
 461                                                      exp->exp_last_request_time),
 462                                               exp, (long)cfs_time_current_sec(),
 463                                               (long)expire_time,
 464                                               (long)exp->exp_last_request_time);
 465                                 CDEBUG(D_HA, "Last request was at %ld\n",
 466                                        exp->exp_last_request_time);
 467                                 class_fail_export(exp);
 468                                 class_export_put(exp);
 469                                 spin_lock(&obd->obd_dev_lock);
 470                         } else {
 471                                 /* List is sorted, so everyone below is ok */
 472                                 break;
 473                         }
 474                 }
 475                 spin_unlock(&obd->obd_dev_lock);
 476
 477                 class_export_put(pet_exp);
 478
 479                 spin_lock(&pet_lock);
 480                 pet_exp = NULL;
 481                 spin_unlock(&pet_lock);
 482         }
 483         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 484
 485         RETURN(0);
 486 }
 487
 488 void ping_evictor_start(void)
 489 {
 490         int rc;
 491
 492         if (++pet_refcount > 1)
 493                 return;
 494
 495         init_waitqueue_head(&pet_waitq);
 496
 497         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 498         if (rc < 0) {
 499                 pet_refcount--;
 500                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 501         }
 502 }
 503 EXPORT_SYMBOL(ping_evictor_start);
 504
 505 void ping_evictor_stop(void)
 506 {
 507         if (--pet_refcount > 0)
 508                 return;
 509
 510         pet_state = PET_TERMINATE;
 511         wake_up(&pet_waitq);
 512 }
 513 EXPORT_SYMBOL(ping_evictor_stop);
 514 #else /* !__KERNEL__ */
 515
 516 /* XXX
 517  * the current implementation of pinger in liblustre is not optimized
 518  */
 519
 520 #ifdef ENABLE_PINGER
 521 static struct pinger_data {
 522         int             pd_recursion;
 523         cfs_time_t      pd_this_ping;   /* jiffies */
 524         cfs_time_t      pd_next_ping;   /* jiffies */
 525         struct ptlrpc_request_set *pd_set;
 526 } pinger_args;
 527
 528 static int pinger_check_rpcs(void *arg)
 529 {
 530         cfs_time_t curtime = cfs_time_current();
 531         struct ptlrpc_request *req;
 532         struct ptlrpc_request_set *set;
 533         struct list_head *iter;
 534         struct pinger_data *pd = &pinger_args;
 535         int rc;
 536
 537         /* prevent recursion */
 538         if (pd->pd_recursion++) {
 539                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 540                 LASSERT(pd->pd_set);
 541                 pd->pd_recursion--;
 542                 return 0;
 543         }
 544
 545         /* have we reached ping point? */
 546         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 547                 pd->pd_recursion--;
 548                 return 0;
 549         }
 550
 551         /* if we have rpc_set already, continue processing it */
 552         if (pd->pd_set) {
 553                 LASSERT(pd->pd_this_ping);
 554                 set = pd->pd_set;
 555                 goto do_check_set;
 556         }
 557
 558         pd->pd_this_ping = curtime;
 559         pd->pd_set = ptlrpc_prep_set();
 560         if (pd->pd_set == NULL)
 561                 goto out;
 562         set = pd->pd_set;
 563
 564         /* add rpcs into set */
 565         mutex_down(&pinger_sem);
 566         list_for_each(iter, &pinger_imports) {
 567                 struct obd_import *imp =
 568                         list_entry(iter, struct obd_import, imp_pinger_chain);
 569                 int generation, level;
 570
 571                 if (cfs_time_aftereq(pd->pd_this_ping,
 572                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 573                         /* Add a ping. */
 574                         spin_lock(&imp->imp_lock);
 575                         generation = imp->imp_generation;
 576                         level = imp->imp_state;
 577                         spin_unlock(&imp->imp_lock);
 578
 579                         if (level != LUSTRE_IMP_FULL) {
 580                                 CDEBUG(D_HA,
 581                                        "not pinging %s (in recovery)\n",
 582                                        obd2cli_tgt(imp->imp_obd));
 583                                 continue;
 584                         }
 585
 586                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 587                                                         LUSTRE_OBD_VERSION,
 588                                                         OBD_PING);
 589                         if (req == NULL) {
 590                                 CERROR("OOM trying to ping %s->%s\n",
 591                                        imp->imp_obd->obd_uuid.uuid,
 592                                        obd2cli_tgt(imp->imp_obd));
 593                                 break;
 594                         }
 595
 596                         req->rq_no_resend = 1;
 597                         ptlrpc_request_set_replen(req);
 598                         req->rq_send_state = LUSTRE_IMP_FULL;
 599                         req->rq_phase = RQ_PHASE_RPC;
 600                         req->rq_import_generation = generation;
 601                         ptlrpc_set_add_req(set, req);
 602                 } else {
 603                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 604                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 605                                imp->imp_next_ping, pd->pd_this_ping);
 606                 }
 607         }
 608         pd->pd_this_ping = curtime;
 609         mutex_up(&pinger_sem);
 610
 611         /* Might be empty, that's OK. */
 612         if (set->set_remaining == 0)
 613                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 614
 615         list_for_each(iter, &set->set_requests) {
 616                 struct ptlrpc_request *req =
 617                         list_entry(iter, struct ptlrpc_request,
 618                                    rq_set_chain);
 619                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 620                           req->rq_import->imp_obd->obd_uuid.uuid,
 621                           obd2cli_tgt(req->rq_import->imp_obd));
 622                 (void)ptl_send_rpc(req, 0);
 623         }
 624
 625 do_check_set:
 626         rc = ptlrpc_check_set(set);
 627
 628         /* not finished, and we are not expired, simply return */
 629         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 630                                             cfs_time_seconds(PING_INTERVAL)))) {
 631                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 632                 pd->pd_recursion--;
 633                 return 0;
 634         }
 635
 636         /* Expire all the requests that didn't come back. */
 637         mutex_down(&pinger_sem);
 638         list_for_each(iter, &set->set_requests) {
 639                 req = list_entry(iter, struct ptlrpc_request,
 640                                  rq_set_chain);
 641
 642                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 643                         continue;
 644
 645                 req->rq_phase = RQ_PHASE_COMPLETE;
 646                 atomic_dec(&req->rq_import->imp_inflight);
 647                 set->set_remaining--;
 648                 /* If it was disconnected, don't sweat it. */
 649                 if (list_empty(&req->rq_import->imp_pinger_chain)) {
 650                         ptlrpc_unregister_reply(req);
 651                         continue;
 652                 }
 653
 654                 CDEBUG(D_RPCTRACE, "pinger initiate expire_one_request\n");
 655                 ptlrpc_expire_one_request(req);
 656         }
 657         mutex_up(&pinger_sem);
 658
 659         ptlrpc_set_destroy(set);
 660         pd->pd_set = NULL;
 661
 662 out:
 663         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 664                                         cfs_time_seconds(PING_INTERVAL));
 665         pd->pd_this_ping = 0; /* XXX for debug */
 666
 667         CDEBUG(D_INFO, "finished a round ping\n");
 668         pd->pd_recursion--;
 669         return 0;
 670 }
 671
 672 static void *pinger_callback = NULL;
 673 #endif /* ENABLE_PINGER */
 674
 675 int ptlrpc_start_pinger(void)
 676 {
 677 #ifdef ENABLE_PINGER
 678         memset(&pinger_args, 0, sizeof(pinger_args));
 679         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 680                                                            &pinger_check_rpcs,
 681                                                            &pinger_args);
 682 #endif
 683         return 0;
 684 }
 685
 686 int ptlrpc_stop_pinger(void)
 687 {
 688 #ifdef ENABLE_PINGER
 689         if (pinger_callback)
 690                 liblustre_deregister_wait_callback(pinger_callback);
 691 #endif
 692         return 0;
 693 }
 694
 695 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 696 {
 697 #ifdef ENABLE_PINGER
 698         mutex_down(&pinger_sem);
 699         ptlrpc_update_next_ping(imp);
 700         if (pinger_args.pd_set == NULL &&
 701             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 702                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 703                         imp->imp_next_ping, cfs_time_current());
 704                 pinger_args.pd_next_ping = imp->imp_next_ping;
 705         }
 706         mutex_up(&pinger_sem);
 707 #endif
 708 }
 709
 710 int ptlrpc_pinger_add_import(struct obd_import *imp)
 711 {
 712         ENTRY;
 713         if (!list_empty(&imp->imp_pinger_chain))
 714                 RETURN(-EALREADY);
 715
 716         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 717                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 718         ptlrpc_pinger_sending_on_import(imp);
 719
 720         mutex_down(&pinger_sem);
 721         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 722         class_import_get(imp);
 723         mutex_up(&pinger_sem);
 724
 725         RETURN(0);
 726 }
 727
 728 int ptlrpc_pinger_del_import(struct obd_import *imp)
 729 {
 730         ENTRY;
 731         if (list_empty(&imp->imp_pinger_chain))
 732                 RETURN(-ENOENT);
 733
 734         mutex_down(&pinger_sem);
 735         list_del_init(&imp->imp_pinger_chain);
 736         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 737                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 738         class_import_put(imp);
 739         mutex_up(&pinger_sem);
 740         RETURN(0);
 741 }
 742
 743 void ptlrpc_pinger_wake_up()
 744 {
 745 #ifdef ENABLE_PINGER
 746         /* XXX force pinger to run, if needed */
 747         struct obd_import *imp;
 748         ENTRY;
 749         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 750                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 751                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 752 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 753                 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 754 #else
 755                 /*XXX only recover for the initial connection */
 756                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 757                     imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 758 #endif
 759                         ptlrpc_initiate_recovery(imp);
 760                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 761                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 762                                      "state %d, deactive %d\n",
 763                                      imp->imp_obd->obd_uuid.uuid,
 764                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 765                                      imp->imp_deactive);
 766         }
 767         EXIT;
 768 #endif
 769 }
 770 #endif /* !__KERNEL__ */