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 struct list_head pinger_imports = CFS_LIST_HEAD_INIT(pinger_imports);
  53 static struct list_head timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  54
  55 struct ptlrpc_request *
  56 ptlrpc_prep_ping(struct obd_import *imp)
  57 {
  58         struct ptlrpc_request *req;
  59
  60         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
  61                               OBD_PING, 1, NULL, NULL);
  62         if (req) {
  63                 ptlrpc_req_set_repsize(req, 1, NULL);
  64                 req->rq_no_resend = req->rq_no_delay = 1;
  65         }
  66         return req;
  67 }
  68
  69 int ptlrpc_obd_ping(struct obd_device *obd)
  70 {
  71         int rc;
  72         struct ptlrpc_request *req;
  73         ENTRY;
  74
  75         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  76         if (req == NULL)
  77                 RETURN(-ENOMEM);
  78
  79         req->rq_send_state = LUSTRE_IMP_FULL;
  80
  81         rc = ptlrpc_queue_wait(req);
  82
  83         ptlrpc_req_finished(req);
  84
  85         RETURN(rc);
  86 }
  87 EXPORT_SYMBOL(ptlrpc_obd_ping);
  88
  89 int ptlrpc_ping(struct obd_import *imp)
  90 {
  91         struct ptlrpc_request *req;
  92         int rc = 0;
  93         ENTRY;
  94
  95         req = ptlrpc_prep_ping(imp);
  96         if (req) {
  97                 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  98                           imp->imp_obd->obd_uuid.uuid,
  99                           obd2cli_tgt(imp->imp_obd));
 100                 ptlrpcd_add_req(req);
 101         } else {
 102                 CERROR("OOM trying to ping %s->%s\n",
 103                        imp->imp_obd->obd_uuid.uuid,
 104                        obd2cli_tgt(imp->imp_obd));
 105                 rc = -ENOMEM;
 106         }
 107
 108         RETURN(rc);
 109 }
 110 EXPORT_SYMBOL(ptlrpc_ping);
 111
 112 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 113 {
 114 #ifdef ENABLE_PINGER
 115         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 116         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 117                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 118                                   AT_OFF ? 0 :
 119                                   at_get(&imp->imp_at.iat_net_latency));
 120                 time = min(time, dtime);
 121         }
 122         imp->imp_next_ping = cfs_time_shift(time);
 123 #endif /* ENABLE_PINGER */
 124 }
 125
 126 void ptlrpc_ping_import_soon(struct obd_import *imp)
 127 {
 128         imp->imp_next_ping = cfs_time_current();
 129 }
 130
 131 static inline int imp_is_deactive(struct obd_import *imp)
 132 {
 133         return (imp->imp_deactive ||
 134                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 135 }
 136
 137 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 138 {
 139         struct timeout_item *item;
 140         cfs_time_t timeout = PING_INTERVAL;
 141
 142         /* The timeout list is a increase order sorted list */
 143         mutex_down(&pinger_sem);
 144         list_for_each_entry(item, &timeout_list, ti_chain) {
 145                 int ti_timeout = item->ti_timeout;
 146                 if (timeout > ti_timeout)
 147                          timeout = ti_timeout;
 148                 break;
 149         }
 150         mutex_up(&pinger_sem);
 151
 152         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 153                                          cfs_time_current());
 154 }
 155
 156 #ifdef __KERNEL__
 157 static int ptlrpc_pinger_main(void *arg)
 158 {
 159         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 160         struct ptlrpc_thread *thread = data->thread;
 161         ENTRY;
 162
 163         cfs_daemonize(data->name);
 164
 165         /* Record that the thread is running */
 166         thread->t_flags = SVC_RUNNING;
 167         cfs_waitq_signal(&thread->t_ctl_waitq);
 168
 169         /* And now, loop forever, pinging as needed. */
 170         while (1) {
 171                 cfs_time_t this_ping = cfs_time_current();
 172                 struct l_wait_info lwi;
 173                 cfs_duration_t time_to_next_wake;
 174                 struct timeout_item *item;
 175                 struct list_head *iter;
 176
 177                 mutex_down(&pinger_sem);
 178                 list_for_each_entry(item, &timeout_list, ti_chain) {
 179                         item->ti_cb(item, item->ti_cb_data);
 180                 }
 181                 list_for_each(iter, &pinger_imports) {
 182                         struct obd_import *imp =
 183                                 list_entry(iter, struct obd_import,
 184                                            imp_pinger_chain);
 185                         int force, level;
 186
 187                         spin_lock(&imp->imp_lock);
 188                         level = imp->imp_state;
 189                         force = imp->imp_force_verify;
 190                         imp->imp_force_verify = 0;
 191                         spin_unlock(&imp->imp_lock);
 192
 193                         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_RPCTRACE,
 194                                "level %s/%u force %u deactive %u pingable %u\n",
 195                                ptlrpc_import_state_name(level), level,
 196                                force, imp->imp_deactive, imp->imp_pingable);
 197
 198                         if (force ||
 199                             /* if the next ping is within, say, 5 jiffies from
 200                                now, go ahead and ping. See note below. */
 201                             cfs_time_aftereq(this_ping,
 202                                              imp->imp_next_ping - 5 * CFS_TICK)) {
 203                                 if (level == LUSTRE_IMP_DISCON &&
 204                                     !imp_is_deactive(imp)) {
 205                                         /* wait at least a timeout before
 206                                            trying recovery again. */
 207                                         imp->imp_next_ping = cfs_time_shift(obd_timeout);
 208                                         ptlrpc_initiate_recovery(imp);
 209                                 } else if (level != LUSTRE_IMP_FULL ||
 210                                          imp->imp_obd->obd_no_recov ||
 211                                          imp_is_deactive(imp)) {
 212                                         CDEBUG(D_HA, "not pinging %s "
 213                                                "(in recovery: %s or recovery "
 214                                                "disabled: %u/%u)\n",
 215                                                obd2cli_tgt(imp->imp_obd),
 216                                                ptlrpc_import_state_name(level),
 217                                                imp->imp_deactive,
 218                                                imp->imp_obd->obd_no_recov);
 219                                 } else if (imp->imp_pingable || force) {
 220                                                 ptlrpc_ping(imp);
 221                                 }
 222                         } else {
 223                                 if (!imp->imp_pingable)
 224                                         continue;
 225                                 CDEBUG(D_INFO,
 226                                        "don't need to ping %s ("CFS_TIME_T
 227                                        " > "CFS_TIME_T")\n",
 228                                        obd2cli_tgt(imp->imp_obd),
 229                                        imp->imp_next_ping, this_ping);
 230                         }
 231
 232                         /* obd_timeout might have changed */
 233                         if (cfs_time_after(imp->imp_next_ping,
 234                                            cfs_time_add(this_ping,
 235                                                         cfs_time_seconds(PING_INTERVAL))))
 236                                 ptlrpc_update_next_ping(imp, 0);
 237                 }
 238                 mutex_up(&pinger_sem);
 239                 /* update memory usage info */
 240                 obd_update_maxusage();
 241
 242                 /* Wait until the next ping time, or until we're stopped. */
 243                 time_to_next_wake = pinger_check_timeout(this_ping);
 244                 /* The ping sent by ptlrpc_send_rpc may get sent out
 245                    say .01 second after this.
 246                    ptlrpc_pinger_eending_on_import will then set the
 247                    next ping time to next_ping + .01 sec, which means
 248                    we will SKIP the next ping at next_ping, and the
 249                    ping will get sent 2 timeouts from now!  Beware. */
 250                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 251                                time_to_next_wake,
 252                                cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
 253                 if (time_to_next_wake > 0) {
 254                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_wake, cfs_time_seconds(1)),
 255                                           NULL, NULL);
 256                         l_wait_event(thread->t_ctl_waitq,
 257                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 258                                      &lwi);
 259                         if (thread->t_flags & SVC_STOPPING) {
 260                                 thread->t_flags &= ~SVC_STOPPING;
 261                                 EXIT;
 262                                 break;
 263                         } else if (thread->t_flags & SVC_EVENT) {
 264                                 /* woken after adding import to reset timer */
 265                                 thread->t_flags &= ~SVC_EVENT;
 266                         }
 267                 }
 268         }
 269
 270         thread->t_flags = SVC_STOPPED;
 271         cfs_waitq_signal(&thread->t_ctl_waitq);
 272
 273         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 274         return 0;
 275 }
 276
 277 static struct ptlrpc_thread *pinger_thread = NULL;
 278
 279 int ptlrpc_start_pinger(void)
 280 {
 281         struct l_wait_info lwi = { 0 };
 282         struct ptlrpc_svc_data d;
 283         int rc;
 284 #ifndef ENABLE_PINGER
 285         return 0;
 286 #endif
 287         ENTRY;
 288
 289         if (pinger_thread != NULL)
 290                 RETURN(-EALREADY);
 291
 292         OBD_ALLOC(pinger_thread, sizeof(*pinger_thread));
 293         if (pinger_thread == NULL)
 294                 RETURN(-ENOMEM);
 295         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 296
 297         d.name = "ll_ping";
 298         d.thread = pinger_thread;
 299
 300         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 301          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 302         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 303         if (rc < 0) {
 304                 CERROR("cannot start thread: %d\n", rc);
 305                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 306                 pinger_thread = NULL;
 307                 RETURN(rc);
 308         }
 309         l_wait_event(pinger_thread->t_ctl_waitq,
 310                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 311
 312         RETURN(0);
 313 }
 314
 315 int ptlrpc_pinger_remove_timeouts(void);
 316
 317 int ptlrpc_stop_pinger(void)
 318 {
 319         struct l_wait_info lwi = { 0 };
 320         int rc = 0;
 321 #ifndef ENABLE_PINGER
 322         return 0;
 323 #endif
 324         ENTRY;
 325
 326         if (pinger_thread == NULL)
 327                 RETURN(-EALREADY);
 328
 329         ptlrpc_pinger_remove_timeouts();
 330         mutex_down(&pinger_sem);
 331         pinger_thread->t_flags = SVC_STOPPING;
 332         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 333         mutex_up(&pinger_sem);
 334
 335         l_wait_event(pinger_thread->t_ctl_waitq,
 336                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 337
 338         OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 339         pinger_thread = NULL;
 340         RETURN(rc);
 341 }
 342
 343 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 344 {
 345         ptlrpc_update_next_ping(imp, 0);
 346 }
 347
 348 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 349 {
 350         ptlrpc_update_next_ping(imp, 1);
 351 }
 352
 353 int ptlrpc_pinger_add_import(struct obd_import *imp)
 354 {
 355         ENTRY;
 356         if (!list_empty(&imp->imp_pinger_chain))
 357                 RETURN(-EALREADY);
 358
 359         mutex_down(&pinger_sem);
 360         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 361                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 362         /* if we add to pinger we want recovery on this import */
 363         imp->imp_obd->obd_no_recov = 0;
 364
 365         ptlrpc_update_next_ping(imp, 0);
 366         /* XXX sort, blah blah */
 367         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 368         class_import_get(imp);
 369
 370         ptlrpc_pinger_wake_up();
 371         mutex_up(&pinger_sem);
 372
 373         RETURN(0);
 374 }
 375
 376 int ptlrpc_pinger_del_import(struct obd_import *imp)
 377 {
 378         ENTRY;
 379         if (list_empty(&imp->imp_pinger_chain))
 380                 RETURN(-ENOENT);
 381
 382         mutex_down(&pinger_sem);
 383         list_del_init(&imp->imp_pinger_chain);
 384         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 385                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 386         /* if we remove from pinger we don't want recovery on this import */
 387         imp->imp_obd->obd_no_recov = 1;
 388         class_import_put(imp);
 389         mutex_up(&pinger_sem);
 390         RETURN(0);
 391 }
 392
 393 /**
 394  * Register a timeout callback to the pinger list, and the callback will
 395  * be called when timeout happens.
 396  */
 397 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 398                                         timeout_cb_t cb, void *data)
 399 {
 400         struct timeout_item *ti;
 401
 402         OBD_ALLOC_PTR(ti);
 403         if (!ti)
 404                 return(NULL);
 405
 406         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 407         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 408         ti->ti_timeout = time;
 409         ti->ti_event = event;
 410         ti->ti_cb = cb;
 411         ti->ti_cb_data = data;
 412
 413         return ti;
 414 }
 415
 416 /**
 417  * Register timeout event on the the pinger thread.
 418  * Note: the timeout list is an sorted list with increased timeout value.
 419  */
 420 static struct timeout_item*
 421 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 422                                timeout_cb_t cb, void *data)
 423 {
 424         struct timeout_item *item, *tmp;
 425
 426         LASSERT_SEM_LOCKED(&pinger_sem);
 427
 428         list_for_each_entry(item, &timeout_list, ti_chain)
 429                 if (item->ti_event == event)
 430                         goto out;
 431
 432         item = ptlrpc_new_timeout(time, event, cb, data);
 433         if (item) {
 434                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 435                         if (tmp->ti_timeout < time) {
 436                                 list_add(&item->ti_chain, &tmp->ti_chain);
 437                                 goto out;
 438                         }
 439                 }
 440                 list_add(&item->ti_chain, &timeout_list);
 441         }
 442 out:
 443         return item;
 444 }
 445
 446 /* Add a client_obd to the timeout event list, when timeout(@time)
 447  * happens, the callback(@cb) will be called.
 448  */
 449 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 450                               timeout_cb_t cb, void *data,
 451                               struct list_head *obd_list)
 452 {
 453         struct timeout_item *ti;
 454
 455         mutex_down(&pinger_sem);
 456         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 457         if (!ti) {
 458                 mutex_up(&pinger_sem);
 459                 return (-EINVAL);
 460         }
 461         list_add(obd_list, &ti->ti_obd_list);
 462         mutex_up(&pinger_sem);
 463         return 0;
 464 }
 465
 466 int ptlrpc_del_timeout_client(struct list_head *obd_list)
 467 {
 468         mutex_down(&pinger_sem);
 469         list_del_init(obd_list);
 470         mutex_up(&pinger_sem);
 471         return 0;
 472 }
 473
 474 int ptlrpc_pinger_remove_timeouts(void)
 475 {
 476         struct timeout_item *item, *tmp;
 477
 478         mutex_down(&pinger_sem);
 479         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 480                 LASSERT(list_empty(&item->ti_obd_list));
 481                 list_del(&item->ti_chain);
 482                 OBD_FREE_PTR(item);
 483         }
 484         mutex_up(&pinger_sem);
 485         return 0;
 486 }
 487
 488 void ptlrpc_pinger_wake_up()
 489 {
 490 #ifdef ENABLE_PINGER
 491         pinger_thread->t_flags |= SVC_EVENT;
 492         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 493 #endif
 494 }
 495
 496 /* Ping evictor thread */
 497 #define PET_READY     1
 498 #define PET_TERMINATE 2
 499
 500 static int               pet_refcount = 0;
 501 static int               pet_state;
 502 static wait_queue_head_t pet_waitq;
 503 static struct obd_export *pet_exp = NULL;
 504 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 505
 506 int ping_evictor_wake(struct obd_export *exp)
 507 {
 508         spin_lock(&pet_lock);
 509         if (pet_exp || (pet_state != PET_READY)) {
 510                 /* eventually the new obd will call here again. */
 511                 spin_unlock(&pet_lock);
 512                 return 1;
 513         }
 514
 515         /* We have to make sure the obd isn't destroyed between now and when
 516          * the ping evictor runs.  We'll take a reference here, and drop it
 517          * when we finish in the evictor.  We don't really care about this
 518          * export in particular; we just need one to keep the obd alive. */
 519         pet_exp = class_export_get(exp);
 520         spin_unlock(&pet_lock);
 521
 522         wake_up(&pet_waitq);
 523         return 0;
 524 }
 525
 526 static int ping_evictor_main(void *arg)
 527 {
 528         struct obd_device *obd;
 529         struct obd_export *exp;
 530         struct l_wait_info lwi = { 0 };
 531         time_t expire_time;
 532         ENTRY;
 533
 534         ptlrpc_daemonize("ll_evictor");
 535
 536         CDEBUG(D_HA, "Starting Ping Evictor\n");
 537         pet_exp = NULL;
 538         pet_state = PET_READY;
 539         while (1) {
 540                 l_wait_event(pet_waitq, pet_exp ||
 541                              (pet_state == PET_TERMINATE), &lwi);
 542                 if (pet_state == PET_TERMINATE)
 543                         break;
 544
 545                 /* we only get here if pet_exp != NULL, and the end of this
 546                  * loop is the only place which sets it NULL again, so lock
 547                  * is not strictly necessary. */
 548                 spin_lock(&pet_lock);
 549                 obd = pet_exp->exp_obd;
 550                 spin_unlock(&pet_lock);
 551
 552                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 553
 554                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 555                        obd->obd_name, expire_time);
 556
 557                 /* Exports can't be deleted out of the list while we hold
 558                  * the obd lock (class_unlink_export), which means we can't
 559                  * lose the last ref on the export.  If they've already been
 560                  * removed from the list, we won't find them here. */
 561                 spin_lock(&obd->obd_dev_lock);
 562                 while (!list_empty(&obd->obd_exports_timed)) {
 563                         exp = list_entry(obd->obd_exports_timed.next,
 564                                          struct obd_export,exp_obd_chain_timed);
 565                         if (expire_time > exp->exp_last_request_time) {
 566                                 class_export_get(exp);
 567                                 spin_unlock(&obd->obd_dev_lock);
 568                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 569                                               " (at %s) in %ld seconds. I think"
 570                                               " it's dead, and I am evicting"
 571                                               " it.\n", obd->obd_name,
 572                                               obd_uuid2str(&exp->exp_client_uuid),
 573                                               obd_export_nid2str(exp),
 574                                               (long)(cfs_time_current_sec() -
 575                                                      exp->exp_last_request_time));
 576                                 CDEBUG(D_HA, "Last request was at %ld\n",
 577                                        exp->exp_last_request_time);
 578                                 class_fail_export(exp);
 579                                 class_export_put(exp);
 580                                 spin_lock(&obd->obd_dev_lock);
 581                         } else {
 582                                 /* List is sorted, so everyone below is ok */
 583                                 break;
 584                         }
 585                 }
 586                 spin_unlock(&obd->obd_dev_lock);
 587
 588                 class_export_put(pet_exp);
 589
 590                 spin_lock(&pet_lock);
 591                 pet_exp = NULL;
 592                 spin_unlock(&pet_lock);
 593         }
 594         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 595
 596         RETURN(0);
 597 }
 598
 599 void ping_evictor_start(void)
 600 {
 601         int rc;
 602
 603         if (++pet_refcount > 1)
 604                 return;
 605
 606         init_waitqueue_head(&pet_waitq);
 607
 608         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 609         if (rc < 0) {
 610                 pet_refcount--;
 611                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 612         }
 613 }
 614 EXPORT_SYMBOL(ping_evictor_start);
 615
 616 void ping_evictor_stop(void)
 617 {
 618         if (--pet_refcount > 0)
 619                 return;
 620
 621         pet_state = PET_TERMINATE;
 622         wake_up(&pet_waitq);
 623 }
 624 EXPORT_SYMBOL(ping_evictor_stop);
 625 #else /* !__KERNEL__ */
 626
 627 /* XXX
 628  * the current implementation of pinger in liblustre is not optimized
 629  */
 630
 631 #ifdef ENABLE_PINGER
 632 static struct pinger_data {
 633         int             pd_recursion;
 634         cfs_time_t      pd_this_ping;   /* jiffies */
 635         cfs_time_t      pd_next_ping;   /* jiffies */
 636         struct ptlrpc_request_set *pd_set;
 637 } pinger_args;
 638
 639 static int pinger_check_rpcs(void *arg)
 640 {
 641         cfs_time_t curtime = cfs_time_current();
 642         struct ptlrpc_request *req;
 643         struct ptlrpc_request_set *set;
 644         struct list_head *iter;
 645         struct obd_import *imp;
 646         struct pinger_data *pd = &pinger_args;
 647         int rc;
 648
 649         /* prevent recursion */
 650         if (pd->pd_recursion++) {
 651                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 652                 LASSERT(pd->pd_set);
 653                 pd->pd_recursion--;
 654                 return 0;
 655         }
 656
 657         /* have we reached ping point? */
 658         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 659                 pd->pd_recursion--;
 660                 return 0;
 661         }
 662
 663         /* if we have rpc_set already, continue processing it */
 664         if (pd->pd_set) {
 665                 LASSERT(pd->pd_this_ping);
 666                 set = pd->pd_set;
 667                 goto do_check_set;
 668         }
 669
 670         pd->pd_this_ping = curtime;
 671         pd->pd_set = ptlrpc_prep_set();
 672         if (pd->pd_set == NULL)
 673                 goto out;
 674         set = pd->pd_set;
 675
 676         /* add rpcs into set */
 677         mutex_down(&pinger_sem);
 678         list_for_each(iter, &pinger_imports) {
 679                 struct obd_import *imp =
 680                         list_entry(iter, struct obd_import, imp_pinger_chain);
 681                 int generation, level;
 682
 683                 if (cfs_time_aftereq(pd->pd_this_ping,
 684                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 685                         /* Add a ping. */
 686                         spin_lock(&imp->imp_lock);
 687                         generation = imp->imp_generation;
 688                         level = imp->imp_state;
 689                         spin_unlock(&imp->imp_lock);
 690
 691                         if (level != LUSTRE_IMP_FULL) {
 692                                 CDEBUG(D_HA,
 693                                        "not pinging %s (in recovery)\n",
 694                                        obd2cli_tgt(imp->imp_obd));
 695                                 continue;
 696                         }
 697
 698                         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING,
 699                                               1, NULL, NULL);
 700                         if (!req) {
 701                                 CERROR("out of memory\n");
 702                                 break;
 703                         }
 704                         req->rq_no_resend = 1;
 705                         ptlrpc_req_set_repsize(req, 1, NULL);
 706                         req->rq_send_state = LUSTRE_IMP_FULL;
 707                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 708                         req->rq_import_generation = generation;
 709                         ptlrpc_set_add_req(set, req);
 710                 } else {
 711                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 712                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 713                                imp->imp_next_ping, pd->pd_this_ping);
 714                 }
 715         }
 716         pd->pd_this_ping = curtime;
 717         mutex_up(&pinger_sem);
 718
 719         /* Might be empty, that's OK. */
 720         if (set->set_remaining == 0)
 721                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 722
 723         list_for_each(iter, &set->set_requests) {
 724                 struct ptlrpc_request *req =
 725                         list_entry(iter, struct ptlrpc_request,
 726                                    rq_set_chain);
 727                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 728                           req->rq_import->imp_obd->obd_uuid.uuid,
 729                           obd2cli_tgt(req->rq_import->imp_obd));
 730                 (void)ptl_send_rpc(req, 0);
 731         }
 732
 733 do_check_set:
 734         rc = ptlrpc_check_set(set);
 735
 736         /* not finished, and we are not expired, simply return */
 737         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 738                                             cfs_time_seconds(PING_INTERVAL)))) {
 739                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 740                 pd->pd_recursion--;
 741                 return 0;
 742         }
 743
 744         /* Expire all the requests that didn't come back. */
 745         mutex_down(&pinger_sem);
 746         list_for_each(iter, &set->set_requests) {
 747                 req = list_entry(iter, struct ptlrpc_request,
 748                                  rq_set_chain);
 749
 750                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 751                         continue;
 752
 753                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 754                        req);
 755
 756                 /* This will also unregister reply. */
 757                 ptlrpc_expire_one_request(req, 0);
 758
 759                 /* We're done with this req, let's finally move it to complete
 760                  * phase and take care of inflights. */
 761                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 762                 imp = req->rq_import;
 763                 spin_lock(&imp->imp_lock);
 764                 if (!list_empty(&req->rq_list)) {
 765                         list_del_init(&req->rq_list);
 766                         atomic_dec(&imp->imp_inflight);
 767                 }
 768                 spin_unlock(&imp->imp_lock);
 769                 set->set_remaining--;
 770         }
 771         mutex_up(&pinger_sem);
 772
 773         ptlrpc_set_destroy(set);
 774         pd->pd_set = NULL;
 775
 776 out:
 777         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 778                                         cfs_time_seconds(PING_INTERVAL));
 779         pd->pd_this_ping = 0; /* XXX for debug */
 780
 781         CDEBUG(D_INFO, "finished a round ping\n");
 782         pd->pd_recursion--;
 783         return 0;
 784 }
 785
 786 static void *pinger_callback = NULL;
 787 #endif /* ENABLE_PINGER */
 788
 789 int ptlrpc_start_pinger(void)
 790 {
 791 #ifdef ENABLE_PINGER
 792         memset(&pinger_args, 0, sizeof(pinger_args));
 793         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 794                                                            &pinger_check_rpcs,
 795                                                            &pinger_args);
 796 #endif
 797         return 0;
 798 }
 799
 800 int ptlrpc_stop_pinger(void)
 801 {
 802 #ifdef ENABLE_PINGER
 803         if (pinger_callback)
 804                 liblustre_deregister_wait_callback(pinger_callback);
 805 #endif
 806         return 0;
 807 }
 808
 809 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 810 {
 811 #ifdef ENABLE_PINGER
 812         mutex_down(&pinger_sem);
 813         ptlrpc_update_next_ping(imp, 0);
 814         if (pinger_args.pd_set == NULL &&
 815             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 816                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 817                         imp->imp_next_ping, cfs_time_current());
 818                 pinger_args.pd_next_ping = imp->imp_next_ping;
 819         }
 820         mutex_up(&pinger_sem);
 821 #endif
 822 }
 823
 824 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 825                               timeout_cb_t cb, void *data,
 826                               struct list_head *obd_list)
 827 {
 828         return 0;
 829 }
 830
 831 int ptlrpc_del_timeout_client(struct list_head *obd_list)
 832 {
 833         return 0;
 834 }
 835
 836 int ptlrpc_pinger_add_import(struct obd_import *imp)
 837 {
 838         ENTRY;
 839         if (!list_empty(&imp->imp_pinger_chain))
 840                 RETURN(-EALREADY);
 841
 842         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 843                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 844         ptlrpc_pinger_sending_on_import(imp);
 845
 846         mutex_down(&pinger_sem);
 847         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 848         class_import_get(imp);
 849         mutex_up(&pinger_sem);
 850
 851         RETURN(0);
 852 }
 853
 854 int ptlrpc_pinger_del_import(struct obd_import *imp)
 855 {
 856         ENTRY;
 857         if (list_empty(&imp->imp_pinger_chain))
 858                 RETURN(-ENOENT);
 859
 860         mutex_down(&pinger_sem);
 861         list_del_init(&imp->imp_pinger_chain);
 862         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 863                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 864         class_import_put(imp);
 865         mutex_up(&pinger_sem);
 866         RETURN(0);
 867 }
 868
 869 void ptlrpc_pinger_wake_up()
 870 {
 871 #ifdef ENABLE_PINGER
 872         ENTRY;
 873         /* XXX force pinger to run, if needed */
 874         struct obd_import *imp;
 875         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 876                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 877                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 878 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 879                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 880                     !imp_is_deactive(imp))
 881 #else
 882                 /*XXX only recover for the initial connection */
 883                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 884                     imp->imp_state == LUSTRE_IMP_DISCON &&
 885                     !imp_is_deactive(imp))
 886 #endif
 887                         ptlrpc_initiate_recovery(imp);
 888                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 889                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 890                                      "state %d, deactive %d\n",
 891                                      imp->imp_obd->obd_uuid.uuid,
 892                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 893                                      imp_is_deactive(imp));
 894         }
 895 #endif
 896         EXIT;
 897 }
 898 #endif /* !__KERNEL__ */