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