lustre/ptlrpc/pinger.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  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 int suppress_pings;
  52 EXPORT_SYMBOL(suppress_pings);
  53 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  54
  55 struct mutex pinger_mutex;
  56 static CFS_LIST_HEAD(pinger_imports);
  57 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  58
  59 struct ptlrpc_request *
  60 ptlrpc_prep_ping(struct obd_import *imp)
  61 {
  62         struct ptlrpc_request *req;
  63
  64         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  65                                         LUSTRE_OBD_VERSION, OBD_PING);
  66         if (req) {
  67                 ptlrpc_request_set_replen(req);
  68                 req->rq_no_resend = req->rq_no_delay = 1;
  69         }
  70         return req;
  71 }
  72
  73 int ptlrpc_obd_ping(struct obd_device *obd)
  74 {
  75         int rc;
  76         struct ptlrpc_request *req;
  77         ENTRY;
  78
  79         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  80         if (req == NULL)
  81                 RETURN(-ENOMEM);
  82
  83         req->rq_send_state = LUSTRE_IMP_FULL;
  84
  85         rc = ptlrpc_queue_wait(req);
  86
  87         ptlrpc_req_finished(req);
  88
  89         RETURN(rc);
  90 }
  91 EXPORT_SYMBOL(ptlrpc_obd_ping);
  92
  93 int ptlrpc_ping(struct obd_import *imp)
  94 {
  95         struct ptlrpc_request *req;
  96         ENTRY;
  97
  98         req = ptlrpc_prep_ping(imp);
  99         if (req == NULL) {
 100                 CERROR("OOM trying to ping %s->%s\n",
 101                        imp->imp_obd->obd_uuid.uuid,
 102                        obd2cli_tgt(imp->imp_obd));
 103                 RETURN(-ENOMEM);
 104         }
 105
 106         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 107                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 108         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 109
 110         RETURN(0);
 111 }
 112
 113 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 114 {
 115 #ifdef ENABLE_PINGER
 116         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 117         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 118                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 119                                   AT_OFF ? 0 :
 120                                   at_get(&imp->imp_at.iat_net_latency));
 121                 time = min(time, dtime);
 122         }
 123         imp->imp_next_ping = cfs_time_shift(time);
 124 #endif /* ENABLE_PINGER */
 125 }
 126
 127 void ptlrpc_ping_import_soon(struct obd_import *imp)
 128 {
 129         imp->imp_next_ping = cfs_time_current();
 130 }
 131
 132 static inline int imp_is_deactive(struct obd_import *imp)
 133 {
 134         return (imp->imp_deactive ||
 135                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 136 }
 137
 138 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 139 {
 140         if (imp->imp_server_timeout)
 141                 return cfs_time_shift(obd_timeout / 2);
 142         else
 143                 return cfs_time_shift(obd_timeout);
 144 }
 145
 146 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 147 static cfs_time_t suspend_wakeup_time = 0;
 148
 149 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 150 {
 151         struct timeout_item *item;
 152         cfs_time_t timeout = PING_INTERVAL;
 153
 154         /* The timeout list is a increase order sorted list */
 155         mutex_lock(&pinger_mutex);
 156         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 157                 int ti_timeout = item->ti_timeout;
 158                 if (timeout > ti_timeout)
 159                         timeout = ti_timeout;
 160                 break;
 161         }
 162         mutex_unlock(&pinger_mutex);
 163
 164         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 165                                          cfs_time_current());
 166 }
 167
 168 static cfs_waitq_t suspend_timeouts_waitq;
 169
 170 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 171 {
 172         return suspend_wakeup_time;
 173 }
 174
 175 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 176 {
 177         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 178 #if 0
 179         if (imp->imp_no_timeout)
 180                 return;
 181         imp->imp_no_timeout = 1;
 182         cfs_atomic_inc(&suspend_timeouts);
 183         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 184                cfs_atomic_read(&suspend_timeouts));
 185 #endif
 186 }
 187
 188 void ptlrpc_activate_timeouts(struct obd_import *imp)
 189 {
 190         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 191 #if 0
 192         if (!imp->imp_no_timeout)
 193                 return;
 194         imp->imp_no_timeout = 0;
 195         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 196         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 197                 suspend_wakeup_time = cfs_time_current();
 198                 cfs_waitq_signal(&suspend_timeouts_waitq);
 199         }
 200         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 201                cfs_atomic_read(&suspend_timeouts));
 202 #endif
 203 }
 204
 205 int ptlrpc_check_suspend(void)
 206 {
 207         if (cfs_atomic_read(&suspend_timeouts))
 208                 return 1;
 209         return 0;
 210 }
 211
 212 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 213 {
 214         struct l_wait_info lwi;
 215
 216         if (cfs_atomic_read(&suspend_timeouts)) {
 217                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 218                           cfs_atomic_read(&suspend_timeouts));
 219                 lwi = LWI_INTR(NULL, NULL);
 220                 l_wait_event(suspend_timeouts_waitq,
 221                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 222                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 223                 return 1;
 224         }
 225         return 0;
 226 }
 227
 228 #ifdef __KERNEL__
 229
 230 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 231                                          unsigned long this_ping)
 232 {
 233         int level;
 234         int force;
 235         int force_next;
 236         int suppress;
 237
 238         spin_lock(&imp->imp_lock);
 239
 240         level = imp->imp_state;
 241         force = imp->imp_force_verify;
 242         force_next = imp->imp_force_next_verify;
 243         /*
 244          * This will be used below only if the import is "FULL".
 245          */
 246         suppress = !!(imp->imp_connect_data.ocd_connect_flags &
 247                       OBD_CONNECT_PINGLESS);
 248
 249         imp->imp_force_verify = 0;
 250
 251         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 252             !force) {
 253                 spin_unlock(&imp->imp_lock);
 254                 return;
 255         }
 256
 257         imp->imp_force_next_verify = 0;
 258
 259         spin_unlock(&imp->imp_lock);
 260
 261         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 262                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 263                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 264                ptlrpc_import_state_name(level), level, force, force_next,
 265                imp->imp_deactive, imp->imp_pingable, suppress);
 266
 267         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 268                 /* wait for a while before trying recovery again */
 269                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 270                 if (!imp->imp_no_pinger_recover)
 271                         ptlrpc_initiate_recovery(imp);
 272         } else if (level != LUSTRE_IMP_FULL ||
 273                    imp->imp_obd->obd_no_recov ||
 274                    imp_is_deactive(imp)) {
 275                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 276                        "or recovery disabled: %s)\n",
 277                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 278                        ptlrpc_import_state_name(level));
 279         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 280                 ptlrpc_ping(imp);
 281         }
 282 }
 283
 284 static int ptlrpc_pinger_main(void *arg)
 285 {
 286         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 287         ENTRY;
 288
 289         cfs_daemonize(thread->t_name);
 290
 291         /* Record that the thread is running */
 292         thread_set_flags(thread, SVC_RUNNING);
 293         cfs_waitq_signal(&thread->t_ctl_waitq);
 294
 295         /* And now, loop forever, pinging as needed. */
 296         while (1) {
 297                 cfs_time_t this_ping = cfs_time_current();
 298                 struct l_wait_info lwi;
 299                 cfs_duration_t time_to_next_wake;
 300                 struct timeout_item *item;
 301                 cfs_list_t *iter;
 302
 303                 mutex_lock(&pinger_mutex);
 304                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 305                         item->ti_cb(item, item->ti_cb_data);
 306                 }
 307                 cfs_list_for_each(iter, &pinger_imports) {
 308                         struct obd_import *imp =
 309                                 cfs_list_entry(iter, struct obd_import,
 310                                                imp_pinger_chain);
 311
 312                         ptlrpc_pinger_process_import(imp, this_ping);
 313                         /* obd_timeout might have changed */
 314                         if (imp->imp_pingable && imp->imp_next_ping &&
 315                             cfs_time_after(imp->imp_next_ping,
 316                                            cfs_time_add(this_ping,
 317                                                         cfs_time_seconds(PING_INTERVAL))))
 318                                 ptlrpc_update_next_ping(imp, 0);
 319                 }
 320                 mutex_unlock(&pinger_mutex);
 321                 /* update memory usage info */
 322                 obd_update_maxusage();
 323
 324                 /* Wait until the next ping time, or until we're stopped. */
 325                 time_to_next_wake = pinger_check_timeout(this_ping);
 326                 /* The ping sent by ptlrpc_send_rpc may get sent out
 327                    say .01 second after this.
 328                    ptlrpc_pinger_sending_on_import will then set the
 329                    next ping time to next_ping + .01 sec, which means
 330                    we will SKIP the next ping at next_ping, and the
 331                    ping will get sent 2 timeouts from now!  Beware. */
 332                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 333                        CFS_TIME_T")\n", time_to_next_wake,
 334                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 335                 if (time_to_next_wake > 0) {
 336                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 337                                                 time_to_next_wake,
 338                                                 cfs_time_seconds(1)),
 339                                           NULL, NULL);
 340                         l_wait_event(thread->t_ctl_waitq,
 341                                      thread_is_stopping(thread) ||
 342                                      thread_is_event(thread),
 343                                      &lwi);
 344                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 345                                 EXIT;
 346                                 break;
 347                         } else {
 348                                 /* woken after adding import to reset timer */
 349                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 350                         }
 351                 }
 352         }
 353
 354         thread_set_flags(thread, SVC_STOPPED);
 355         cfs_waitq_signal(&thread->t_ctl_waitq);
 356
 357         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 358         return 0;
 359 }
 360
 361 static struct ptlrpc_thread *pinger_thread = NULL;
 362
 363 int ptlrpc_start_pinger(void)
 364 {
 365         struct l_wait_info lwi = { 0 };
 366         int rc;
 367 #ifndef ENABLE_PINGER
 368         return 0;
 369 #endif
 370         ENTRY;
 371
 372         if (pinger_thread != NULL)
 373                 RETURN(-EALREADY);
 374
 375         OBD_ALLOC_PTR(pinger_thread);
 376         if (pinger_thread == NULL)
 377                 RETURN(-ENOMEM);
 378         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 379         cfs_waitq_init(&suspend_timeouts_waitq);
 380
 381         strcpy(pinger_thread->t_name, "ll_ping");
 382
 383         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 384          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 385         rc = cfs_create_thread(ptlrpc_pinger_main,
 386                                pinger_thread, CFS_DAEMON_FLAGS);
 387         if (rc < 0) {
 388                 CERROR("cannot start thread: %d\n", rc);
 389                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 390                 pinger_thread = NULL;
 391                 RETURN(rc);
 392         }
 393         l_wait_event(pinger_thread->t_ctl_waitq,
 394                      thread_is_running(pinger_thread), &lwi);
 395
 396         if (suppress_pings)
 397                 CWARN("Pings will be suppressed at the request of the "
 398                       "administrator.  The configuration shall meet the "
 399                       "additional requirements described in the manual.  "
 400                       "(Search for the \"suppress_pings\" kernel module "
 401                       "parameter.)\n");
 402
 403         RETURN(0);
 404 }
 405
 406 int ptlrpc_pinger_remove_timeouts(void);
 407
 408 int ptlrpc_stop_pinger(void)
 409 {
 410         struct l_wait_info lwi = { 0 };
 411         int rc = 0;
 412 #ifndef ENABLE_PINGER
 413         return 0;
 414 #endif
 415         ENTRY;
 416
 417         if (pinger_thread == NULL)
 418                 RETURN(-EALREADY);
 419
 420         ptlrpc_pinger_remove_timeouts();
 421         mutex_lock(&pinger_mutex);
 422         thread_set_flags(pinger_thread, SVC_STOPPING);
 423         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 424         mutex_unlock(&pinger_mutex);
 425
 426         l_wait_event(pinger_thread->t_ctl_waitq,
 427                      thread_is_stopped(pinger_thread), &lwi);
 428
 429         OBD_FREE_PTR(pinger_thread);
 430         pinger_thread = NULL;
 431         RETURN(rc);
 432 }
 433
 434 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 435 {
 436         ptlrpc_update_next_ping(imp, 0);
 437 }
 438 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 439
 440 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 441 {
 442         ptlrpc_update_next_ping(imp, 1);
 443         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 444         /*
 445          * Avoid reading stale imp_connect_data.  When not sure if pings are
 446          * expected or not on next connection, we assume they are not and force
 447          * one anyway to guarantee the chance of updating
 448          * imp_peer_committed_transno.
 449          */
 450         if (imp->imp_state != LUSTRE_IMP_FULL ||
 451             imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_PINGLESS)
 452                 imp->imp_force_next_verify = 1;
 453 }
 454
 455 int ptlrpc_pinger_add_import(struct obd_import *imp)
 456 {
 457         ENTRY;
 458         if (!cfs_list_empty(&imp->imp_pinger_chain))
 459                 RETURN(-EALREADY);
 460
 461         mutex_lock(&pinger_mutex);
 462         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 463                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 464         /* if we add to pinger we want recovery on this import */
 465         imp->imp_obd->obd_no_recov = 0;
 466         ptlrpc_update_next_ping(imp, 0);
 467         /* XXX sort, blah blah */
 468         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 469         class_import_get(imp);
 470
 471         ptlrpc_pinger_wake_up();
 472         mutex_unlock(&pinger_mutex);
 473
 474         RETURN(0);
 475 }
 476 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 477
 478 int ptlrpc_pinger_del_import(struct obd_import *imp)
 479 {
 480         ENTRY;
 481         if (cfs_list_empty(&imp->imp_pinger_chain))
 482                 RETURN(-ENOENT);
 483
 484         mutex_lock(&pinger_mutex);
 485         cfs_list_del_init(&imp->imp_pinger_chain);
 486         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 487                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 488         /* if we remove from pinger we don't want recovery on this import */
 489         imp->imp_obd->obd_no_recov = 1;
 490         class_import_put(imp);
 491         mutex_unlock(&pinger_mutex);
 492         RETURN(0);
 493 }
 494 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 495
 496 /**
 497  * Register a timeout callback to the pinger list, and the callback will
 498  * be called when timeout happens.
 499  */
 500 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 501                                         timeout_cb_t cb, void *data)
 502 {
 503         struct timeout_item *ti;
 504
 505         OBD_ALLOC_PTR(ti);
 506         if (!ti)
 507                 return(NULL);
 508
 509         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 510         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 511         ti->ti_timeout = time;
 512         ti->ti_event = event;
 513         ti->ti_cb = cb;
 514         ti->ti_cb_data = data;
 515
 516         return ti;
 517 }
 518
 519 /**
 520  * Register timeout event on the the pinger thread.
 521  * Note: the timeout list is an sorted list with increased timeout value.
 522  */
 523 static struct timeout_item*
 524 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 525                                timeout_cb_t cb, void *data)
 526 {
 527         struct timeout_item *item, *tmp;
 528
 529         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 530
 531         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 532                 if (item->ti_event == event)
 533                         goto out;
 534
 535         item = ptlrpc_new_timeout(time, event, cb, data);
 536         if (item) {
 537                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 538                         if (tmp->ti_timeout < time) {
 539                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 540                                 goto out;
 541                         }
 542                 }
 543                 cfs_list_add(&item->ti_chain, &timeout_list);
 544         }
 545 out:
 546         return item;
 547 }
 548
 549 /* Add a client_obd to the timeout event list, when timeout(@time)
 550  * happens, the callback(@cb) will be called.
 551  */
 552 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 553                               timeout_cb_t cb, void *data,
 554                               cfs_list_t *obd_list)
 555 {
 556         struct timeout_item *ti;
 557
 558         mutex_lock(&pinger_mutex);
 559         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 560         if (!ti) {
 561                 mutex_unlock(&pinger_mutex);
 562                 return (-EINVAL);
 563         }
 564         cfs_list_add(obd_list, &ti->ti_obd_list);
 565         mutex_unlock(&pinger_mutex);
 566         return 0;
 567 }
 568 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 569
 570 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 571                               enum timeout_event event)
 572 {
 573         struct timeout_item *ti = NULL, *item;
 574
 575         if (cfs_list_empty(obd_list))
 576                 return 0;
 577         mutex_lock(&pinger_mutex);
 578         cfs_list_del_init(obd_list);
 579         /**
 580          * If there are no obd attached to the timeout event
 581          * list, remove this timeout event from the pinger
 582          */
 583         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 584                 if (item->ti_event == event) {
 585                         ti = item;
 586                         break;
 587                 }
 588         }
 589         LASSERTF(ti != NULL, "ti is NULL ! \n");
 590         if (cfs_list_empty(&ti->ti_obd_list)) {
 591                 cfs_list_del(&ti->ti_chain);
 592                 OBD_FREE_PTR(ti);
 593         }
 594         mutex_unlock(&pinger_mutex);
 595         return 0;
 596 }
 597 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 598
 599 int ptlrpc_pinger_remove_timeouts(void)
 600 {
 601         struct timeout_item *item, *tmp;
 602
 603         mutex_lock(&pinger_mutex);
 604         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 605                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 606                 cfs_list_del(&item->ti_chain);
 607                 OBD_FREE_PTR(item);
 608         }
 609         mutex_unlock(&pinger_mutex);
 610         return 0;
 611 }
 612
 613 void ptlrpc_pinger_wake_up()
 614 {
 615 #ifdef ENABLE_PINGER
 616         thread_add_flags(pinger_thread, SVC_EVENT);
 617         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 618 #endif
 619 }
 620
 621 /* Ping evictor thread */
 622 #define PET_READY     1
 623 #define PET_TERMINATE 2
 624
 625 static int               pet_refcount = 0;
 626 static int               pet_state;
 627 static cfs_waitq_t       pet_waitq;
 628 CFS_LIST_HEAD(pet_list);
 629 static DEFINE_SPINLOCK(pet_lock);
 630
 631 int ping_evictor_wake(struct obd_export *exp)
 632 {
 633         struct obd_device *obd;
 634
 635         spin_lock(&pet_lock);
 636         if (pet_state != PET_READY) {
 637                 /* eventually the new obd will call here again. */
 638                 spin_unlock(&pet_lock);
 639                 return 1;
 640         }
 641
 642         obd = class_exp2obd(exp);
 643         if (cfs_list_empty(&obd->obd_evict_list)) {
 644                 class_incref(obd, "evictor", obd);
 645                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 646         }
 647         spin_unlock(&pet_lock);
 648
 649         cfs_waitq_signal(&pet_waitq);
 650         return 0;
 651 }
 652
 653 static int ping_evictor_main(void *arg)
 654 {
 655         struct obd_device *obd;
 656         struct obd_export *exp;
 657         struct l_wait_info lwi = { 0 };
 658         time_t expire_time;
 659         ENTRY;
 660
 661         cfs_daemonize_ctxt("ll_evictor");
 662
 663         CDEBUG(D_HA, "Starting Ping Evictor\n");
 664         pet_state = PET_READY;
 665         while (1) {
 666                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 667                              (pet_state == PET_TERMINATE), &lwi);
 668
 669                 /* loop until all obd's will be removed */
 670                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 671                         break;
 672
 673                 /* we only get here if pet_exp != NULL, and the end of this
 674                  * loop is the only place which sets it NULL again, so lock
 675                  * is not strictly necessary. */
 676                 spin_lock(&pet_lock);
 677                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 678                                      obd_evict_list);
 679                 spin_unlock(&pet_lock);
 680
 681                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 682
 683                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 684                        obd->obd_name, expire_time);
 685
 686                 /* Exports can't be deleted out of the list while we hold
 687                  * the obd lock (class_unlink_export), which means we can't
 688                  * lose the last ref on the export.  If they've already been
 689                  * removed from the list, we won't find them here. */
 690                 spin_lock(&obd->obd_dev_lock);
 691                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 692                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 693                                              struct obd_export,
 694                                              exp_obd_chain_timed);
 695                         if (expire_time > exp->exp_last_request_time) {
 696                                 class_export_get(exp);
 697                                 spin_unlock(&obd->obd_dev_lock);
 698                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 699                                               " (at %s) in %ld seconds. I think"
 700                                               " it's dead, and I am evicting"
 701                                               " it. exp %p, cur %ld expire %ld"
 702                                               " last %ld\n",
 703                                               obd->obd_name,
 704                                               obd_uuid2str(&exp->exp_client_uuid),
 705                                               obd_export_nid2str(exp),
 706                                               (long)(cfs_time_current_sec() -
 707                                                      exp->exp_last_request_time),
 708                                               exp, (long)cfs_time_current_sec(),
 709                                               (long)expire_time,
 710                                               (long)exp->exp_last_request_time);
 711                                 CDEBUG(D_HA, "Last request was at %ld\n",
 712                                        exp->exp_last_request_time);
 713                                 class_fail_export(exp);
 714                                 class_export_put(exp);
 715                                 spin_lock(&obd->obd_dev_lock);
 716                         } else {
 717                                 /* List is sorted, so everyone below is ok */
 718                                 break;
 719                         }
 720                 }
 721                 spin_unlock(&obd->obd_dev_lock);
 722
 723                 spin_lock(&pet_lock);
 724                 cfs_list_del_init(&obd->obd_evict_list);
 725                 spin_unlock(&pet_lock);
 726
 727                 class_decref(obd, "evictor", obd);
 728         }
 729         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 730
 731         RETURN(0);
 732 }
 733
 734 void ping_evictor_start(void)
 735 {
 736         int rc;
 737
 738         if (++pet_refcount > 1)
 739                 return;
 740
 741         cfs_waitq_init(&pet_waitq);
 742
 743         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 744         if (rc < 0) {
 745                 pet_refcount--;
 746                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 747         }
 748 }
 749 EXPORT_SYMBOL(ping_evictor_start);
 750
 751 void ping_evictor_stop(void)
 752 {
 753         if (--pet_refcount > 0)
 754                 return;
 755
 756         pet_state = PET_TERMINATE;
 757         cfs_waitq_signal(&pet_waitq);
 758 }
 759 EXPORT_SYMBOL(ping_evictor_stop);
 760 #else /* !__KERNEL__ */
 761
 762 /* XXX
 763  * the current implementation of pinger in liblustre is not optimized
 764  */
 765
 766 #ifdef ENABLE_PINGER
 767 static struct pinger_data {
 768         int             pd_recursion;
 769         cfs_time_t      pd_this_ping;   /* jiffies */
 770         cfs_time_t      pd_next_ping;   /* jiffies */
 771         struct ptlrpc_request_set *pd_set;
 772 } pinger_args;
 773
 774 static int pinger_check_rpcs(void *arg)
 775 {
 776         cfs_time_t curtime = cfs_time_current();
 777         struct ptlrpc_request *req;
 778         struct ptlrpc_request_set *set;
 779         cfs_list_t *iter;
 780         struct obd_import *imp;
 781         struct pinger_data *pd = &pinger_args;
 782         int rc;
 783
 784         /* prevent recursion */
 785         if (pd->pd_recursion++) {
 786                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 787                 LASSERT(pd->pd_set);
 788                 pd->pd_recursion--;
 789                 return 0;
 790         }
 791
 792         /* have we reached ping point? */
 793         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 794                 pd->pd_recursion--;
 795                 return 0;
 796         }
 797
 798         /* if we have rpc_set already, continue processing it */
 799         if (pd->pd_set) {
 800                 LASSERT(pd->pd_this_ping);
 801                 set = pd->pd_set;
 802                 goto do_check_set;
 803         }
 804
 805         pd->pd_this_ping = curtime;
 806         pd->pd_set = ptlrpc_prep_set();
 807         if (pd->pd_set == NULL)
 808                 goto out;
 809         set = pd->pd_set;
 810
 811         /* add rpcs into set */
 812         mutex_lock(&pinger_mutex);
 813         cfs_list_for_each(iter, &pinger_imports) {
 814                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 815                                                         imp_pinger_chain);
 816                 int generation, level;
 817
 818                 if (cfs_time_aftereq(pd->pd_this_ping,
 819                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 820                         /* Add a ping. */
 821                         spin_lock(&imp->imp_lock);
 822                         generation = imp->imp_generation;
 823                         level = imp->imp_state;
 824                         spin_unlock(&imp->imp_lock);
 825
 826                         if (level != LUSTRE_IMP_FULL) {
 827                                 CDEBUG(D_HA,
 828                                        "not pinging %s (in recovery)\n",
 829                                        obd2cli_tgt(imp->imp_obd));
 830                                 continue;
 831                         }
 832
 833                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 834                                                         LUSTRE_OBD_VERSION,
 835                                                         OBD_PING);
 836                         if (req == NULL) {
 837                                 CERROR("OOM trying to ping %s->%s\n",
 838                                        imp->imp_obd->obd_uuid.uuid,
 839                                        obd2cli_tgt(imp->imp_obd));
 840                                 break;
 841                         }
 842
 843                         req->rq_no_resend = 1;
 844                         ptlrpc_request_set_replen(req);
 845                         req->rq_send_state = LUSTRE_IMP_FULL;
 846                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 847                         req->rq_import_generation = generation;
 848                         ptlrpc_set_add_req(set, req);
 849                 } else {
 850                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 851                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 852                                imp->imp_next_ping, pd->pd_this_ping);
 853                 }
 854         }
 855         pd->pd_this_ping = curtime;
 856         mutex_unlock(&pinger_mutex);
 857
 858         /* Might be empty, that's OK. */
 859         if (cfs_atomic_read(&set->set_remaining) == 0)
 860                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 861
 862         cfs_list_for_each(iter, &set->set_requests) {
 863                 struct ptlrpc_request *req =
 864                         cfs_list_entry(iter, struct ptlrpc_request,
 865                                        rq_set_chain);
 866                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 867                           req->rq_import->imp_obd->obd_uuid.uuid,
 868                           obd2cli_tgt(req->rq_import->imp_obd));
 869                 (void)ptl_send_rpc(req, 0);
 870         }
 871
 872 do_check_set:
 873         rc = ptlrpc_check_set(NULL, set);
 874
 875         /* not finished, and we are not expired, simply return */
 876         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 877                                             cfs_time_seconds(PING_INTERVAL)))) {
 878                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 879                 pd->pd_recursion--;
 880                 return 0;
 881         }
 882
 883         /* Expire all the requests that didn't come back. */
 884         mutex_lock(&pinger_mutex);
 885         cfs_list_for_each(iter, &set->set_requests) {
 886                 req = cfs_list_entry(iter, struct ptlrpc_request,
 887                                      rq_set_chain);
 888
 889                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 890                         continue;
 891
 892                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 893                        req);
 894
 895                 /* This will also unregister reply. */
 896                 ptlrpc_expire_one_request(req, 0);
 897
 898                 /* We're done with this req, let's finally move it to complete
 899                  * phase and take care of inflights. */
 900                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 901                 imp = req->rq_import;
 902                 spin_lock(&imp->imp_lock);
 903                 if (!cfs_list_empty(&req->rq_list)) {
 904                         cfs_list_del_init(&req->rq_list);
 905                         cfs_atomic_dec(&imp->imp_inflight);
 906                 }
 907                 spin_unlock(&imp->imp_lock);
 908                 cfs_atomic_dec(&set->set_remaining);
 909         }
 910         mutex_unlock(&pinger_mutex);
 911
 912         ptlrpc_set_destroy(set);
 913         pd->pd_set = NULL;
 914
 915 out:
 916         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 917                                         cfs_time_seconds(PING_INTERVAL));
 918         pd->pd_this_ping = 0; /* XXX for debug */
 919
 920         CDEBUG(D_INFO, "finished a round ping\n");
 921         pd->pd_recursion--;
 922         return 0;
 923 }
 924
 925 static void *pinger_callback = NULL;
 926 #endif /* ENABLE_PINGER */
 927
 928 int ptlrpc_start_pinger(void)
 929 {
 930 #ifdef ENABLE_PINGER
 931         memset(&pinger_args, 0, sizeof(pinger_args));
 932         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 933                                                            &pinger_check_rpcs,
 934                                                            &pinger_args);
 935 #endif
 936         return 0;
 937 }
 938
 939 int ptlrpc_stop_pinger(void)
 940 {
 941 #ifdef ENABLE_PINGER
 942         if (pinger_callback)
 943                 liblustre_deregister_wait_callback(pinger_callback);
 944 #endif
 945         return 0;
 946 }
 947
 948 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 949 {
 950 #ifdef ENABLE_PINGER
 951         mutex_lock(&pinger_mutex);
 952         ptlrpc_update_next_ping(imp, 0);
 953         if (pinger_args.pd_set == NULL &&
 954             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 955                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 956                         imp->imp_next_ping, cfs_time_current());
 957                 pinger_args.pd_next_ping = imp->imp_next_ping;
 958         }
 959         mutex_unlock(&pinger_mutex);
 960 #endif
 961 }
 962
 963 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 964 {
 965 #ifdef ENABLE_PINGER
 966         mutex_lock(&pinger_mutex);
 967         ptlrpc_update_next_ping(imp, 1);
 968         if (pinger_args.pd_set == NULL &&
 969             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 970                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 971                         imp->imp_next_ping, cfs_time_current());
 972                 pinger_args.pd_next_ping = imp->imp_next_ping;
 973         }
 974         mutex_unlock(&pinger_mutex);
 975 #endif
 976 }
 977
 978 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 979                               timeout_cb_t cb, void *data,
 980                               cfs_list_t *obd_list)
 981 {
 982         return 0;
 983 }
 984
 985 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 986                               enum timeout_event event)
 987 {
 988         return 0;
 989 }
 990
 991 int ptlrpc_pinger_add_import(struct obd_import *imp)
 992 {
 993         ENTRY;
 994         if (!cfs_list_empty(&imp->imp_pinger_chain))
 995                 RETURN(-EALREADY);
 996
 997         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 998                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 999         ptlrpc_pinger_sending_on_import(imp);
1000
1001         mutex_lock(&pinger_mutex);
1002         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1003         class_import_get(imp);
1004         mutex_unlock(&pinger_mutex);
1005
1006         RETURN(0);
1007 }
1008
1009 int ptlrpc_pinger_del_import(struct obd_import *imp)
1010 {
1011         ENTRY;
1012         if (cfs_list_empty(&imp->imp_pinger_chain))
1013                 RETURN(-ENOENT);
1014
1015         mutex_lock(&pinger_mutex);
1016         cfs_list_del_init(&imp->imp_pinger_chain);
1017         CDEBUG(D_HA, "removing pingable import %s->%s\n",
1018                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1019         class_import_put(imp);
1020         mutex_unlock(&pinger_mutex);
1021         RETURN(0);
1022 }
1023
1024 void ptlrpc_pinger_wake_up()
1025 {
1026 #ifdef ENABLE_PINGER
1027         /* XXX force pinger to run, if needed */
1028         struct obd_import *imp;
1029         ENTRY;
1030         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1031                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1032                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1033 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1034                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1035                     !imp_is_deactive(imp))
1036 #else
1037                 /*XXX only recover for the initial connection */
1038                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1039                     imp->imp_state == LUSTRE_IMP_DISCON &&
1040                     !imp_is_deactive(imp))
1041 #endif
1042                         ptlrpc_initiate_recovery(imp);
1043                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1044                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1045                                      "state %d, deactive %d\n",
1046                                      imp->imp_obd->obd_uuid.uuid,
1047                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1048                                      imp_is_deactive(imp));
1049         }
1050         EXIT;
1051 #endif
1052 }
1053 #endif /* !__KERNEL__ */