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