lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 struct semaphore pinger_sem;
  52 static CFS_LIST_HEAD(pinger_imports);
  53
  54 int ptlrpc_ping(struct obd_import *imp)
  55 {
  56         struct ptlrpc_request *req;
  57         ENTRY;
  58
  59         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
  60                                         OBD_PING);
  61         if (req == NULL) {
  62                 CERROR("OOM trying to ping %s->%s\n",
  63                        imp->imp_obd->obd_uuid.uuid,
  64                        obd2cli_tgt(imp->imp_obd));
  65                 RETURN(-ENOMEM);
  66         }
  67
  68         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  69                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
  70         req->rq_no_resend = req->rq_no_delay = 1;
  71         ptlrpc_request_set_replen(req);
  72         ptlrpcd_add_req(req, PSCOPE_OTHER);
  73
  74         RETURN(0);
  75 }
  76
  77 void ptlrpc_update_next_ping(struct obd_import *imp)
  78 {
  79 #ifdef ENABLE_PINGER
  80         int time = PING_INTERVAL;
  81         if (imp->imp_state == LUSTRE_IMP_DISCON) {
  82                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
  83                                   AT_OFF ? 0 :
  84                                   at_get(&imp->imp_at.iat_net_latency));
  85                 time = min(time, dtime);
  86         }
  87         imp->imp_next_ping = cfs_time_shift(time);
  88 #endif /* ENABLE_PINGER */
  89 }
  90
  91 void ptlrpc_ping_import_soon(struct obd_import *imp)
  92 {
  93         imp->imp_next_ping = cfs_time_current();
  94 }
  95
  96 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
  97 {
  98         if (imp->imp_server_timeout)
  99                 return cfs_time_shift(obd_timeout / 2);
 100         else
 101                 return cfs_time_shift(obd_timeout);
 102 }
 103
 104 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
 105 static cfs_time_t suspend_wakeup_time = 0;
 106
 107 #ifdef __KERNEL__
 108 static wait_queue_head_t suspend_timeouts_waitq;
 109 #endif
 110
 111 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 112 {
 113         return suspend_wakeup_time;
 114 }
 115
 116 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 117 {
 118         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 119 #if 0
 120         if (imp->imp_no_timeout)
 121                 return;
 122         imp->imp_no_timeout = 1;
 123         atomic_inc(&suspend_timeouts);
 124         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
 125 #endif
 126 }
 127
 128 void ptlrpc_activate_timeouts(struct obd_import *imp)
 129 {
 130         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 131 #if 0
 132         if (!imp->imp_no_timeout)
 133                 return;
 134         imp->imp_no_timeout = 0;
 135         LASSERT(atomic_read(&suspend_timeouts) > 0);
 136         if (atomic_dec_and_test(&suspend_timeouts)) {
 137                 suspend_wakeup_time = cfs_time_current();
 138                 wake_up(&suspend_timeouts_waitq);
 139         }
 140         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
 141 #endif
 142 }
 143
 144 int ptlrpc_check_suspend(void)
 145 {
 146         if (atomic_read(&suspend_timeouts))
 147                 return 1;
 148         return 0;
 149 }
 150
 151 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 152 {
 153         struct l_wait_info lwi;
 154
 155         if (atomic_read(&suspend_timeouts)) {
 156                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 157                           atomic_read(&suspend_timeouts));
 158                 lwi = LWI_INTR(NULL, NULL);
 159                 l_wait_event(suspend_timeouts_waitq,
 160                              atomic_read(&suspend_timeouts) == 0, &lwi);
 161                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 162                 return 1;
 163         }
 164         return 0;
 165 }
 166
 167 #ifdef __KERNEL__
 168
 169 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 170                                          unsigned long this_ping)
 171 {
 172         int force, level;
 173
 174         spin_lock(&imp->imp_lock);
 175         level = imp->imp_state;
 176         force = imp->imp_force_verify;
 177         if (force)
 178                 imp->imp_force_verify = 0;
 179         spin_unlock(&imp->imp_lock);
 180
 181         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 182                "level %s/%u force %u deactive %u pingable %u\n",
 183                ptlrpc_import_state_name(level), level,
 184                force, imp->imp_deactive, imp->imp_pingable);
 185
 186         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 187                              this_ping) && force == 0)
 188                 return;
 189
 190         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 191                 /* wait at least a timeout before trying recovery again */
 192                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 193                 ptlrpc_initiate_recovery(imp);
 194         } else if (level != LUSTRE_IMP_FULL ||
 195                    imp->imp_obd->obd_no_recov ||
 196                    imp->imp_deactive) {
 197                 CDEBUG(D_HA, "not pinging %s (in recovery "
 198                        " or recovery disabled: %s)\n",
 199                        obd2cli_tgt(imp->imp_obd),
 200                        ptlrpc_import_state_name(level));
 201         } else if (imp->imp_pingable || force) {
 202                 ptlrpc_ping(imp);
 203         }
 204 }
 205
 206 static int ptlrpc_pinger_main(void *arg)
 207 {
 208         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 209         struct ptlrpc_thread *thread = data->thread;
 210         ENTRY;
 211
 212         cfs_daemonize(data->name);
 213
 214         /* Record that the thread is running */
 215         thread->t_flags = SVC_RUNNING;
 216         cfs_waitq_signal(&thread->t_ctl_waitq);
 217
 218         /* And now, loop forever, pinging as needed. */
 219         while (1) {
 220                 cfs_time_t this_ping = cfs_time_current();
 221                 struct l_wait_info lwi;
 222                 cfs_duration_t time_to_next_ping;
 223                 struct list_head *iter;
 224
 225                 mutex_down(&pinger_sem);
 226                 list_for_each(iter, &pinger_imports) {
 227                         struct obd_import *imp =
 228                                 list_entry(iter, struct obd_import,
 229                                            imp_pinger_chain);
 230
 231                         ptlrpc_pinger_process_import(imp, this_ping);
 232                         /* obd_timeout might have changed */
 233                         if (imp->imp_pingable && imp->imp_next_ping &&
 234                             cfs_time_after(imp->imp_next_ping,
 235                                            cfs_time_add(this_ping,
 236                                                         cfs_time_seconds(PING_INTERVAL))))
 237                                 ptlrpc_update_next_ping(imp);
 238                 }
 239                 mutex_up(&pinger_sem);
 240                 /* update memory usage info */
 241                 obd_update_maxusage();
 242
 243                 /* Wait until the next ping time, or until we're stopped. */
 244                 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
 245                                                cfs_time_seconds(PING_INTERVAL)),
 246                                                cfs_time_current());
 247
 248                 /* The ping sent by ptlrpc_send_rpc may get sent out
 249                    say .01 second after this.
 250                    ptlrpc_pinger_sending_on_import will then set the
 251                    next ping time to next_ping + .01 sec, which means
 252                    we will SKIP the next ping at next_ping, and the
 253                    ping will get sent 2 timeouts from now!  Beware. */
 254                 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 255                                time_to_next_ping,
 256                                cfs_time_add(this_ping,
 257                                             cfs_time_seconds(PING_INTERVAL)));
 258                 if (time_to_next_ping > 0) {
 259                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 260                                                 time_to_next_ping,
 261                                                 cfs_time_seconds(1)),
 262                                           NULL, NULL);
 263                         l_wait_event(thread->t_ctl_waitq,
 264                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 265                                      &lwi);
 266                         if (thread->t_flags & SVC_STOPPING) {
 267                                 thread->t_flags &= ~SVC_STOPPING;
 268                                 EXIT;
 269                                 break;
 270                         } else if (thread->t_flags & SVC_EVENT) {
 271                                 /* woken after adding import to reset timer */
 272                                 thread->t_flags &= ~SVC_EVENT;
 273                         }
 274                 }
 275         }
 276
 277         thread->t_flags = SVC_STOPPED;
 278         cfs_waitq_signal(&thread->t_ctl_waitq);
 279
 280         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 281         return 0;
 282 }
 283
 284 static struct ptlrpc_thread *pinger_thread = NULL;
 285
 286 int ptlrpc_start_pinger(void)
 287 {
 288         struct l_wait_info lwi = { 0 };
 289         struct ptlrpc_svc_data d;
 290         int rc;
 291 #ifndef ENABLE_PINGER
 292         return 0;
 293 #endif
 294         ENTRY;
 295
 296         if (pinger_thread != NULL)
 297                 RETURN(-EALREADY);
 298
 299         OBD_ALLOC_PTR(pinger_thread);
 300         if (pinger_thread == NULL)
 301                 RETURN(-ENOMEM);
 302         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 303         cfs_waitq_init(&suspend_timeouts_waitq);
 304
 305         d.name = "ll_ping";
 306         d.thread = pinger_thread;
 307
 308         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 309          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 310         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 311         if (rc < 0) {
 312                 CERROR("cannot start thread: %d\n", rc);
 313                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 314                 pinger_thread = NULL;
 315                 RETURN(rc);
 316         }
 317         l_wait_event(pinger_thread->t_ctl_waitq,
 318                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 319
 320         RETURN(0);
 321 }
 322
 323 int ptlrpc_stop_pinger(void)
 324 {
 325         struct l_wait_info lwi = { 0 };
 326         int rc = 0;
 327 #ifndef ENABLE_PINGER
 328         return 0;
 329 #endif
 330         ENTRY;
 331
 332         if (pinger_thread == NULL)
 333                 RETURN(-EALREADY);
 334         mutex_down(&pinger_sem);
 335         pinger_thread->t_flags = SVC_STOPPING;
 336         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 337         mutex_up(&pinger_sem);
 338
 339         l_wait_event(pinger_thread->t_ctl_waitq,
 340                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 341
 342         OBD_FREE_PTR(pinger_thread);
 343         pinger_thread = NULL;
 344         RETURN(rc);
 345 }
 346
 347 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 348 {
 349         ptlrpc_update_next_ping(imp);
 350 }
 351
 352 int ptlrpc_pinger_add_import(struct obd_import *imp)
 353 {
 354         ENTRY;
 355         if (!list_empty(&imp->imp_pinger_chain))
 356                 RETURN(-EALREADY);
 357
 358         mutex_down(&pinger_sem);
 359         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 360                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 361         /* if we add to pinger we want recovery on this import */
 362         imp->imp_obd->obd_no_recov = 0;
 363         ptlrpc_update_next_ping(imp);
 364         /* XXX sort, blah blah */
 365         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 366         class_import_get(imp);
 367
 368         ptlrpc_pinger_wake_up();
 369         mutex_up(&pinger_sem);
 370
 371         RETURN(0);
 372 }
 373
 374 int ptlrpc_pinger_del_import(struct obd_import *imp)
 375 {
 376         ENTRY;
 377         if (list_empty(&imp->imp_pinger_chain))
 378                 RETURN(-ENOENT);
 379
 380         mutex_down(&pinger_sem);
 381         list_del_init(&imp->imp_pinger_chain);
 382         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 383                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 384         /* if we remove from pinger we don't want recovery on this import */
 385         imp->imp_obd->obd_no_recov = 1;
 386         class_import_put(imp);
 387         mutex_up(&pinger_sem);
 388         RETURN(0);
 389 }
 390
 391 void ptlrpc_pinger_wake_up()
 392 {
 393 #ifdef ENABLE_PINGER
 394         pinger_thread->t_flags |= SVC_EVENT;
 395         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 396 #endif
 397 }
 398
 399 /* Ping evictor thread */
 400 #define PET_READY     1
 401 #define PET_TERMINATE 2
 402
 403 static int               pet_refcount = 0;
 404 static int               pet_state;
 405 static wait_queue_head_t pet_waitq;
 406 static struct obd_export *pet_exp = NULL;
 407 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 408
 409 int ping_evictor_wake(struct obd_export *exp)
 410 {
 411         spin_lock(&pet_lock);
 412         if (pet_exp || (pet_state != PET_READY)) {
 413                 /* eventually the new obd will call here again. */
 414                 spin_unlock(&pet_lock);
 415                 return 1;
 416         }
 417
 418         /* We have to make sure the obd isn't destroyed between now and when
 419          * the ping evictor runs.  We'll take a reference here, and drop it
 420          * when we finish in the evictor.  We don't really care about this
 421          * export in particular; we just need one to keep the obd alive. */
 422         pet_exp = class_export_get(exp);
 423         spin_unlock(&pet_lock);
 424
 425         wake_up(&pet_waitq);
 426         return 0;
 427 }
 428
 429 static int ping_evictor_main(void *arg)
 430 {
 431         struct obd_device *obd;
 432         struct obd_export *exp;
 433         struct l_wait_info lwi = { 0 };
 434         time_t expire_time;
 435         ENTRY;
 436
 437         ptlrpc_daemonize("ll_evictor");
 438
 439         CDEBUG(D_HA, "Starting Ping Evictor\n");
 440         pet_exp = NULL;
 441         pet_state = PET_READY;
 442         while (1) {
 443                 l_wait_event(pet_waitq, pet_exp ||
 444                              (pet_state == PET_TERMINATE), &lwi);
 445                 if (pet_state == PET_TERMINATE)
 446                         break;
 447
 448                 /* we only get here if pet_exp != NULL, and the end of this
 449                  * loop is the only place which sets it NULL again, so lock
 450                  * is not strictly necessary. */
 451                 spin_lock(&pet_lock);
 452                 obd = pet_exp->exp_obd;
 453                 spin_unlock(&pet_lock);
 454
 455                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 456
 457                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 458                        obd->obd_name, expire_time);
 459
 460                 /* Exports can't be deleted out of the list while we hold
 461                  * the obd lock (class_unlink_export), which means we can't
 462                  * lose the last ref on the export.  If they've already been
 463                  * removed from the list, we won't find them here. */
 464                 spin_lock(&obd->obd_dev_lock);
 465                 while (!list_empty(&obd->obd_exports_timed)) {
 466                         exp = list_entry(obd->obd_exports_timed.next,
 467                                          struct obd_export,exp_obd_chain_timed);
 468                         if (expire_time > exp->exp_last_request_time) {
 469                                 class_export_get(exp);
 470                                 spin_unlock(&obd->obd_dev_lock);
 471                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 472                                               " (at %s) in %ld seconds. I think"
 473                                               " it's dead, and I am evicting"
 474                                               " it. exp %p, cur %ld expire %ld"
 475                                               " last %ld\n",
 476                                               obd->obd_name,
 477                                               obd_uuid2str(&exp->exp_client_uuid),
 478                                               obd_export_nid2str(exp),
 479                                               (long)(cfs_time_current_sec() -
 480                                                      exp->exp_last_request_time),
 481                                               exp, (long)cfs_time_current_sec(),
 482                                               (long)expire_time,
 483                                               (long)exp->exp_last_request_time);
 484                                 CDEBUG(D_HA, "Last request was at %ld\n",
 485                                        exp->exp_last_request_time);
 486                                 class_fail_export(exp);
 487                                 class_export_put(exp);
 488                                 spin_lock(&obd->obd_dev_lock);
 489                         } else {
 490                                 /* List is sorted, so everyone below is ok */
 491                                 break;
 492                         }
 493                 }
 494                 spin_unlock(&obd->obd_dev_lock);
 495
 496                 class_export_put(pet_exp);
 497
 498                 spin_lock(&pet_lock);
 499                 pet_exp = NULL;
 500                 spin_unlock(&pet_lock);
 501         }
 502         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 503
 504         RETURN(0);
 505 }
 506
 507 void ping_evictor_start(void)
 508 {
 509         int rc;
 510
 511         if (++pet_refcount > 1)
 512                 return;
 513
 514         init_waitqueue_head(&pet_waitq);
 515
 516         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 517         if (rc < 0) {
 518                 pet_refcount--;
 519                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 520         }
 521 }
 522 EXPORT_SYMBOL(ping_evictor_start);
 523
 524 void ping_evictor_stop(void)
 525 {
 526         if (--pet_refcount > 0)
 527                 return;
 528
 529         pet_state = PET_TERMINATE;
 530         wake_up(&pet_waitq);
 531 }
 532 EXPORT_SYMBOL(ping_evictor_stop);
 533 #else /* !__KERNEL__ */
 534
 535 /* XXX
 536  * the current implementation of pinger in liblustre is not optimized
 537  */
 538
 539 #ifdef ENABLE_PINGER
 540 static struct pinger_data {
 541         int             pd_recursion;
 542         cfs_time_t      pd_this_ping;   /* jiffies */
 543         cfs_time_t      pd_next_ping;   /* jiffies */
 544         struct ptlrpc_request_set *pd_set;
 545 } pinger_args;
 546
 547 static int pinger_check_rpcs(void *arg)
 548 {
 549         cfs_time_t curtime = cfs_time_current();
 550         struct ptlrpc_request *req;
 551         struct ptlrpc_request_set *set;
 552         struct list_head *iter;
 553         struct pinger_data *pd = &pinger_args;
 554         int rc;
 555
 556         /* prevent recursion */
 557         if (pd->pd_recursion++) {
 558                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 559                 LASSERT(pd->pd_set);
 560                 pd->pd_recursion--;
 561                 return 0;
 562         }
 563
 564         /* have we reached ping point? */
 565         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 566                 pd->pd_recursion--;
 567                 return 0;
 568         }
 569
 570         /* if we have rpc_set already, continue processing it */
 571         if (pd->pd_set) {
 572                 LASSERT(pd->pd_this_ping);
 573                 set = pd->pd_set;
 574                 goto do_check_set;
 575         }
 576
 577         pd->pd_this_ping = curtime;
 578         pd->pd_set = ptlrpc_prep_set();
 579         if (pd->pd_set == NULL)
 580                 goto out;
 581         set = pd->pd_set;
 582
 583         /* add rpcs into set */
 584         mutex_down(&pinger_sem);
 585         list_for_each(iter, &pinger_imports) {
 586                 struct obd_import *imp =
 587                         list_entry(iter, struct obd_import, imp_pinger_chain);
 588                 int generation, level;
 589
 590                 if (cfs_time_aftereq(pd->pd_this_ping,
 591                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 592                         /* Add a ping. */
 593                         spin_lock(&imp->imp_lock);
 594                         generation = imp->imp_generation;
 595                         level = imp->imp_state;
 596                         spin_unlock(&imp->imp_lock);
 597
 598                         if (level != LUSTRE_IMP_FULL) {
 599                                 CDEBUG(D_HA,
 600                                        "not pinging %s (in recovery)\n",
 601                                        obd2cli_tgt(imp->imp_obd));
 602                                 continue;
 603                         }
 604
 605                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 606                                                         LUSTRE_OBD_VERSION,
 607                                                         OBD_PING);
 608                         if (req == NULL) {
 609                                 CERROR("OOM trying to ping %s->%s\n",
 610                                        imp->imp_obd->obd_uuid.uuid,
 611                                        obd2cli_tgt(imp->imp_obd));
 612                                 break;
 613                         }
 614
 615                         req->rq_no_resend = 1;
 616                         ptlrpc_request_set_replen(req);
 617                         req->rq_send_state = LUSTRE_IMP_FULL;
 618                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 619                         req->rq_import_generation = generation;
 620                         ptlrpc_set_add_req(set, req);
 621                 } else {
 622                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 623                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 624                                imp->imp_next_ping, pd->pd_this_ping);
 625                 }
 626         }
 627         pd->pd_this_ping = curtime;
 628         mutex_up(&pinger_sem);
 629
 630         /* Might be empty, that's OK. */
 631         if (set->set_remaining == 0)
 632                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 633
 634         list_for_each(iter, &set->set_requests) {
 635                 struct ptlrpc_request *req =
 636                         list_entry(iter, struct ptlrpc_request,
 637                                    rq_set_chain);
 638                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 639                           req->rq_import->imp_obd->obd_uuid.uuid,
 640                           obd2cli_tgt(req->rq_import->imp_obd));
 641                 (void)ptl_send_rpc(req, 0);
 642         }
 643
 644 do_check_set:
 645         rc = ptlrpc_check_set(NULL, set);
 646
 647         /* not finished, and we are not expired, simply return */
 648         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 649                                             cfs_time_seconds(PING_INTERVAL)))) {
 650                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 651                 pd->pd_recursion--;
 652                 return 0;
 653         }
 654
 655         /* Expire all the requests that didn't come back. */
 656         mutex_down(&pinger_sem);
 657         list_for_each(iter, &set->set_requests) {
 658                 req = list_entry(iter, struct ptlrpc_request,
 659                                  rq_set_chain);
 660
 661                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 662                         continue;
 663
 664                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 665                 atomic_dec(&req->rq_import->imp_inflight);
 666                 set->set_remaining--;
 667                 /* If it was disconnected, don't sweat it. */
 668                 if (list_empty(&req->rq_import->imp_pinger_chain)) {
 669                         ptlrpc_unregister_reply(req, 0);
 670                         continue;
 671                 }
 672
 673                 CDEBUG(D_RPCTRACE, "pinger initiate expire_one_request\n");
 674                 ptlrpc_expire_one_request(req, 0);
 675         }
 676         mutex_up(&pinger_sem);
 677
 678         ptlrpc_set_destroy(set);
 679         pd->pd_set = NULL;
 680
 681 out:
 682         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 683                                         cfs_time_seconds(PING_INTERVAL));
 684         pd->pd_this_ping = 0; /* XXX for debug */
 685
 686         CDEBUG(D_INFO, "finished a round ping\n");
 687         pd->pd_recursion--;
 688         return 0;
 689 }
 690
 691 static void *pinger_callback = NULL;
 692 #endif /* ENABLE_PINGER */
 693
 694 int ptlrpc_start_pinger(void)
 695 {
 696 #ifdef ENABLE_PINGER
 697         memset(&pinger_args, 0, sizeof(pinger_args));
 698         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 699                                                            &pinger_check_rpcs,
 700                                                            &pinger_args);
 701 #endif
 702         return 0;
 703 }
 704
 705 int ptlrpc_stop_pinger(void)
 706 {
 707 #ifdef ENABLE_PINGER
 708         if (pinger_callback)
 709                 liblustre_deregister_wait_callback(pinger_callback);
 710 #endif
 711         return 0;
 712 }
 713
 714 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 715 {
 716 #ifdef ENABLE_PINGER
 717         mutex_down(&pinger_sem);
 718         ptlrpc_update_next_ping(imp);
 719         if (pinger_args.pd_set == NULL &&
 720             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 721                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 722                         imp->imp_next_ping, cfs_time_current());
 723                 pinger_args.pd_next_ping = imp->imp_next_ping;
 724         }
 725         mutex_up(&pinger_sem);
 726 #endif
 727 }
 728
 729 int ptlrpc_pinger_add_import(struct obd_import *imp)
 730 {
 731         ENTRY;
 732         if (!list_empty(&imp->imp_pinger_chain))
 733                 RETURN(-EALREADY);
 734
 735         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 736                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 737         ptlrpc_pinger_sending_on_import(imp);
 738
 739         mutex_down(&pinger_sem);
 740         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 741         class_import_get(imp);
 742         mutex_up(&pinger_sem);
 743
 744         RETURN(0);
 745 }
 746
 747 int ptlrpc_pinger_del_import(struct obd_import *imp)
 748 {
 749         ENTRY;
 750         if (list_empty(&imp->imp_pinger_chain))
 751                 RETURN(-ENOENT);
 752
 753         mutex_down(&pinger_sem);
 754         list_del_init(&imp->imp_pinger_chain);
 755         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 756                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 757         class_import_put(imp);
 758         mutex_up(&pinger_sem);
 759         RETURN(0);
 760 }
 761
 762 void ptlrpc_pinger_wake_up()
 763 {
 764 #ifdef ENABLE_PINGER
 765         /* XXX force pinger to run, if needed */
 766         struct obd_import *imp;
 767         ENTRY;
 768         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 769                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 770                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 771 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 772                 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 773 #else
 774                 /*XXX only recover for the initial connection */
 775                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 776                     imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 777 #endif
 778                         ptlrpc_initiate_recovery(imp);
 779                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 780                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 781                                      "state %d, deactive %d\n",
 782                                      imp->imp_obd->obd_uuid.uuid,
 783                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 784                                      imp->imp_deactive);
 785         }
 786         EXIT;
 787 #endif
 788 }
 789 #endif /* !__KERNEL__ */