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