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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/ptlrpc/pinger.c
  39  *
  40  * Portal-RPC reconnection and replay operations, for use in recovery.
  41  */
  42
  43 #ifndef __KERNEL__
  44 #include <liblustre.h>
  45 #else
  46 #define DEBUG_SUBSYSTEM S_RPC
  47 #endif
  48
  49 #include <obd_support.h>
  50 #include <obd_class.h>
  51 #include "ptlrpc_internal.h"
  52
  53 cfs_mutex_t pinger_mutex;
  54 static CFS_LIST_HEAD(pinger_imports);
  55 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  56 struct ptlrpc_request *
  57 ptlrpc_prep_ping(struct obd_import *imp)
  58 {
  59         struct ptlrpc_request *req;
  60
  61         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  62                                         LUSTRE_OBD_VERSION, OBD_PING);
  63         if (req) {
  64                 ptlrpc_request_set_replen(req);
  65                 req->rq_no_resend = req->rq_no_delay = 1;
  66         }
  67         return req;
  68 }
  69
  70 int ptlrpc_obd_ping(struct obd_device *obd)
  71 {
  72         int rc;
  73         struct ptlrpc_request *req;
  74         ENTRY;
  75
  76         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  77         if (req == NULL)
  78                 RETURN(-ENOMEM);
  79
  80         req->rq_send_state = LUSTRE_IMP_FULL;
  81
  82         rc = ptlrpc_queue_wait(req);
  83
  84         ptlrpc_req_finished(req);
  85
  86         RETURN(rc);
  87 }
  88 EXPORT_SYMBOL(ptlrpc_obd_ping);
  89
  90 int ptlrpc_ping(struct obd_import *imp)
  91 {
  92         struct ptlrpc_request *req;
  93         ENTRY;
  94
  95         req = ptlrpc_prep_ping(imp);
  96         if (req == NULL) {
  97                 CERROR("OOM trying to ping %s->%s\n",
  98                        imp->imp_obd->obd_uuid.uuid,
  99                        obd2cli_tgt(imp->imp_obd));
 100                 RETURN(-ENOMEM);
 101         }
 102
 103         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 104                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 105         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 106
 107         RETURN(0);
 108 }
 109
 110 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 111 {
 112 #ifdef ENABLE_PINGER
 113         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 114         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 115                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 116                                   AT_OFF ? 0 :
 117                                   at_get(&imp->imp_at.iat_net_latency));
 118                 time = min(time, dtime);
 119         }
 120         imp->imp_next_ping = cfs_time_shift(time);
 121 #endif /* ENABLE_PINGER */
 122 }
 123
 124 void ptlrpc_ping_import_soon(struct obd_import *imp)
 125 {
 126         imp->imp_next_ping = cfs_time_current();
 127 }
 128
 129 static inline int imp_is_deactive(struct obd_import *imp)
 130 {
 131         return (imp->imp_deactive ||
 132                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 133 }
 134
 135 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 136 {
 137         if (imp->imp_server_timeout)
 138                 return cfs_time_shift(obd_timeout / 2);
 139         else
 140                 return cfs_time_shift(obd_timeout);
 141 }
 142
 143 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 144 static cfs_time_t suspend_wakeup_time = 0;
 145
 146 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 147 {
 148         struct timeout_item *item;
 149         cfs_time_t timeout = PING_INTERVAL;
 150
 151         /* The timeout list is a increase order sorted list */
 152         cfs_mutex_lock(&pinger_mutex);
 153         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 154                 int ti_timeout = item->ti_timeout;
 155                 if (timeout > ti_timeout)
 156                         timeout = ti_timeout;
 157                 break;
 158         }
 159         cfs_mutex_unlock(&pinger_mutex);
 160
 161         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 162                                          cfs_time_current());
 163 }
 164
 165 static cfs_waitq_t suspend_timeouts_waitq;
 166
 167 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 168 {
 169         return suspend_wakeup_time;
 170 }
 171
 172 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 173 {
 174         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 175 #if 0
 176         if (imp->imp_no_timeout)
 177                 return;
 178         imp->imp_no_timeout = 1;
 179         cfs_atomic_inc(&suspend_timeouts);
 180         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 181                cfs_atomic_read(&suspend_timeouts));
 182 #endif
 183 }
 184
 185 void ptlrpc_activate_timeouts(struct obd_import *imp)
 186 {
 187         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 188 #if 0
 189         if (!imp->imp_no_timeout)
 190                 return;
 191         imp->imp_no_timeout = 0;
 192         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 193         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 194                 suspend_wakeup_time = cfs_time_current();
 195                 cfs_waitq_signal(&suspend_timeouts_waitq);
 196         }
 197         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 198                cfs_atomic_read(&suspend_timeouts));
 199 #endif
 200 }
 201
 202 int ptlrpc_check_suspend(void)
 203 {
 204         if (cfs_atomic_read(&suspend_timeouts))
 205                 return 1;
 206         return 0;
 207 }
 208
 209 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 210 {
 211         struct l_wait_info lwi;
 212
 213         if (cfs_atomic_read(&suspend_timeouts)) {
 214                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 215                           cfs_atomic_read(&suspend_timeouts));
 216                 lwi = LWI_INTR(NULL, NULL);
 217                 l_wait_event(suspend_timeouts_waitq,
 218                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 219                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 220                 return 1;
 221         }
 222         return 0;
 223 }
 224
 225 #ifdef __KERNEL__
 226
 227 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 228                                          unsigned long this_ping)
 229 {
 230         int force, level;
 231
 232         cfs_spin_lock(&imp->imp_lock);
 233         level = imp->imp_state;
 234         force = imp->imp_force_verify;
 235         if (force)
 236                 imp->imp_force_verify = 0;
 237         cfs_spin_unlock(&imp->imp_lock);
 238
 239         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 240                "level %s/%u force %u deactive %u pingable %u\n",
 241                ptlrpc_import_state_name(level), level,
 242                force, imp->imp_deactive, imp->imp_pingable);
 243
 244         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 245                              this_ping) && force == 0)
 246                 return;
 247
 248         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 249                 /* wait for a while before trying recovery again */
 250                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 251                 if (!imp->imp_no_pinger_recover)
 252                         ptlrpc_initiate_recovery(imp);
 253         } else if (level != LUSTRE_IMP_FULL ||
 254                    imp->imp_obd->obd_no_recov ||
 255                    imp_is_deactive(imp)) {
 256                 CDEBUG(D_HA, "not pinging %s (in recovery "
 257                        " or recovery disabled: %s)\n",
 258                        obd2cli_tgt(imp->imp_obd),
 259                        ptlrpc_import_state_name(level));
 260         } else if (imp->imp_pingable || force) {
 261                 ptlrpc_ping(imp);
 262         }
 263 }
 264
 265 static int ptlrpc_pinger_main(void *arg)
 266 {
 267         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 268         struct ptlrpc_thread *thread = data->thread;
 269         ENTRY;
 270
 271         cfs_daemonize(data->name);
 272
 273         /* Record that the thread is running */
 274         thread_set_flags(thread, SVC_RUNNING);
 275         cfs_waitq_signal(&thread->t_ctl_waitq);
 276
 277         /* And now, loop forever, pinging as needed. */
 278         while (1) {
 279                 cfs_time_t this_ping = cfs_time_current();
 280                 struct l_wait_info lwi;
 281                 cfs_duration_t time_to_next_wake;
 282                 struct timeout_item *item;
 283                 cfs_list_t *iter;
 284
 285                 cfs_mutex_lock(&pinger_mutex);
 286                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 287                         item->ti_cb(item, item->ti_cb_data);
 288                 }
 289                 cfs_list_for_each(iter, &pinger_imports) {
 290                         struct obd_import *imp =
 291                                 cfs_list_entry(iter, struct obd_import,
 292                                                imp_pinger_chain);
 293
 294                         ptlrpc_pinger_process_import(imp, this_ping);
 295                         /* obd_timeout might have changed */
 296                         if (imp->imp_pingable && imp->imp_next_ping &&
 297                             cfs_time_after(imp->imp_next_ping,
 298                                            cfs_time_add(this_ping,
 299                                                         cfs_time_seconds(PING_INTERVAL))))
 300                                 ptlrpc_update_next_ping(imp, 0);
 301                 }
 302                 cfs_mutex_unlock(&pinger_mutex);
 303                 /* update memory usage info */
 304                 obd_update_maxusage();
 305
 306                 /* Wait until the next ping time, or until we're stopped. */
 307                 time_to_next_wake = pinger_check_timeout(this_ping);
 308                 /* The ping sent by ptlrpc_send_rpc may get sent out
 309                    say .01 second after this.
 310                    ptlrpc_pinger_sending_on_import will then set the
 311                    next ping time to next_ping + .01 sec, which means
 312                    we will SKIP the next ping at next_ping, and the
 313                    ping will get sent 2 timeouts from now!  Beware. */
 314                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 315                        CFS_TIME_T")\n", time_to_next_wake,
 316                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 317                 if (time_to_next_wake > 0) {
 318                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 319                                                 time_to_next_wake,
 320                                                 cfs_time_seconds(1)),
 321                                           NULL, NULL);
 322                         l_wait_event(thread->t_ctl_waitq,
 323                                      thread_is_stopping(thread) ||
 324                                      thread_is_event(thread),
 325                                      &lwi);
 326                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 327                                 EXIT;
 328                                 break;
 329                         } else {
 330                                 /* woken after adding import to reset timer */
 331                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 332                         }
 333                 }
 334         }
 335
 336         thread_set_flags(thread, SVC_STOPPED);
 337         cfs_waitq_signal(&thread->t_ctl_waitq);
 338
 339         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 340         return 0;
 341 }
 342
 343 static struct ptlrpc_thread *pinger_thread = NULL;
 344
 345 int ptlrpc_start_pinger(void)
 346 {
 347         struct l_wait_info lwi = { 0 };
 348         struct ptlrpc_svc_data d;
 349         int rc;
 350 #ifndef ENABLE_PINGER
 351         return 0;
 352 #endif
 353         ENTRY;
 354
 355         if (pinger_thread != NULL)
 356                 RETURN(-EALREADY);
 357
 358         OBD_ALLOC_PTR(pinger_thread);
 359         if (pinger_thread == NULL)
 360                 RETURN(-ENOMEM);
 361         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 362         cfs_waitq_init(&suspend_timeouts_waitq);
 363
 364         d.name = "ll_ping";
 365         d.thread = pinger_thread;
 366
 367         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 368          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 369         rc = cfs_create_thread(ptlrpc_pinger_main, &d, CFS_DAEMON_FLAGS);
 370         if (rc < 0) {
 371                 CERROR("cannot start thread: %d\n", rc);
 372                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 373                 pinger_thread = NULL;
 374                 RETURN(rc);
 375         }
 376         l_wait_event(pinger_thread->t_ctl_waitq,
 377                      thread_is_running(pinger_thread), &lwi);
 378
 379         RETURN(0);
 380 }
 381
 382 int ptlrpc_pinger_remove_timeouts(void);
 383
 384 int ptlrpc_stop_pinger(void)
 385 {
 386         struct l_wait_info lwi = { 0 };
 387         int rc = 0;
 388 #ifndef ENABLE_PINGER
 389         return 0;
 390 #endif
 391         ENTRY;
 392
 393         if (pinger_thread == NULL)
 394                 RETURN(-EALREADY);
 395
 396         ptlrpc_pinger_remove_timeouts();
 397         cfs_mutex_lock(&pinger_mutex);
 398         thread_set_flags(pinger_thread, SVC_STOPPING);
 399         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 400         cfs_mutex_unlock(&pinger_mutex);
 401
 402         l_wait_event(pinger_thread->t_ctl_waitq,
 403                      thread_is_stopped(pinger_thread), &lwi);
 404
 405         OBD_FREE_PTR(pinger_thread);
 406         pinger_thread = NULL;
 407         RETURN(rc);
 408 }
 409
 410 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 411 {
 412         ptlrpc_update_next_ping(imp, 0);
 413 }
 414
 415 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 416 {
 417         ptlrpc_update_next_ping(imp, 1);
 418 }
 419
 420 int ptlrpc_pinger_add_import(struct obd_import *imp)
 421 {
 422         ENTRY;
 423         if (!cfs_list_empty(&imp->imp_pinger_chain))
 424                 RETURN(-EALREADY);
 425
 426         cfs_mutex_lock(&pinger_mutex);
 427         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 428                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 429         /* if we add to pinger we want recovery on this import */
 430         imp->imp_obd->obd_no_recov = 0;
 431         ptlrpc_update_next_ping(imp, 0);
 432         /* XXX sort, blah blah */
 433         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 434         class_import_get(imp);
 435
 436         ptlrpc_pinger_wake_up();
 437         cfs_mutex_unlock(&pinger_mutex);
 438
 439         RETURN(0);
 440 }
 441
 442 int ptlrpc_pinger_del_import(struct obd_import *imp)
 443 {
 444         ENTRY;
 445         if (cfs_list_empty(&imp->imp_pinger_chain))
 446                 RETURN(-ENOENT);
 447
 448         cfs_mutex_lock(&pinger_mutex);
 449         cfs_list_del_init(&imp->imp_pinger_chain);
 450         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 451                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 452         /* if we remove from pinger we don't want recovery on this import */
 453         imp->imp_obd->obd_no_recov = 1;
 454         class_import_put(imp);
 455         cfs_mutex_unlock(&pinger_mutex);
 456         RETURN(0);
 457 }
 458
 459 /**
 460  * Register a timeout callback to the pinger list, and the callback will
 461  * be called when timeout happens.
 462  */
 463 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 464                                         timeout_cb_t cb, void *data)
 465 {
 466         struct timeout_item *ti;
 467
 468         OBD_ALLOC_PTR(ti);
 469         if (!ti)
 470                 return(NULL);
 471
 472         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 473         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 474         ti->ti_timeout = time;
 475         ti->ti_event = event;
 476         ti->ti_cb = cb;
 477         ti->ti_cb_data = data;
 478
 479         return ti;
 480 }
 481
 482 /**
 483  * Register timeout event on the the pinger thread.
 484  * Note: the timeout list is an sorted list with increased timeout value.
 485  */
 486 static struct timeout_item*
 487 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 488                                timeout_cb_t cb, void *data)
 489 {
 490         struct timeout_item *item, *tmp;
 491
 492         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 493
 494         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 495                 if (item->ti_event == event)
 496                         goto out;
 497
 498         item = ptlrpc_new_timeout(time, event, cb, data);
 499         if (item) {
 500                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 501                         if (tmp->ti_timeout < time) {
 502                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 503                                 goto out;
 504                         }
 505                 }
 506                 cfs_list_add(&item->ti_chain, &timeout_list);
 507         }
 508 out:
 509         return item;
 510 }
 511
 512 /* Add a client_obd to the timeout event list, when timeout(@time)
 513  * happens, the callback(@cb) will be called.
 514  */
 515 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 516                               timeout_cb_t cb, void *data,
 517                               cfs_list_t *obd_list)
 518 {
 519         struct timeout_item *ti;
 520
 521         cfs_mutex_lock(&pinger_mutex);
 522         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 523         if (!ti) {
 524                 cfs_mutex_unlock(&pinger_mutex);
 525                 return (-EINVAL);
 526         }
 527         cfs_list_add(obd_list, &ti->ti_obd_list);
 528         cfs_mutex_unlock(&pinger_mutex);
 529         return 0;
 530 }
 531
 532 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 533                               enum timeout_event event)
 534 {
 535         struct timeout_item *ti = NULL, *item;
 536
 537         if (cfs_list_empty(obd_list))
 538                 return 0;
 539         cfs_mutex_lock(&pinger_mutex);
 540         cfs_list_del_init(obd_list);
 541         /**
 542          * If there are no obd attached to the timeout event
 543          * list, remove this timeout event from the pinger
 544          */
 545         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 546                 if (item->ti_event == event) {
 547                         ti = item;
 548                         break;
 549                 }
 550         }
 551         LASSERTF(ti != NULL, "ti is NULL ! \n");
 552         if (cfs_list_empty(&ti->ti_obd_list)) {
 553                 cfs_list_del(&ti->ti_chain);
 554                 OBD_FREE_PTR(ti);
 555         }
 556         cfs_mutex_unlock(&pinger_mutex);
 557         return 0;
 558 }
 559
 560 int ptlrpc_pinger_remove_timeouts(void)
 561 {
 562         struct timeout_item *item, *tmp;
 563
 564         cfs_mutex_lock(&pinger_mutex);
 565         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 566                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 567                 cfs_list_del(&item->ti_chain);
 568                 OBD_FREE_PTR(item);
 569         }
 570         cfs_mutex_unlock(&pinger_mutex);
 571         return 0;
 572 }
 573
 574 void ptlrpc_pinger_wake_up()
 575 {
 576 #ifdef ENABLE_PINGER
 577         thread_add_flags(pinger_thread, SVC_EVENT);
 578         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 579 #endif
 580 }
 581
 582 /* Ping evictor thread */
 583 #define PET_READY     1
 584 #define PET_TERMINATE 2
 585
 586 static int               pet_refcount = 0;
 587 static int               pet_state;
 588 static cfs_waitq_t       pet_waitq;
 589 CFS_LIST_HEAD(pet_list);
 590 static cfs_spinlock_t    pet_lock = CFS_SPIN_LOCK_UNLOCKED;
 591
 592 int ping_evictor_wake(struct obd_export *exp)
 593 {
 594         struct obd_device *obd;
 595
 596         cfs_spin_lock(&pet_lock);
 597         if (pet_state != PET_READY) {
 598                 /* eventually the new obd will call here again. */
 599                 cfs_spin_unlock(&pet_lock);
 600                 return 1;
 601         }
 602
 603         obd = class_exp2obd(exp);
 604         if (cfs_list_empty(&obd->obd_evict_list)) {
 605                 class_incref(obd, "evictor", obd);
 606                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 607         }
 608         cfs_spin_unlock(&pet_lock);
 609
 610         cfs_waitq_signal(&pet_waitq);
 611         return 0;
 612 }
 613
 614 static int ping_evictor_main(void *arg)
 615 {
 616         struct obd_device *obd;
 617         struct obd_export *exp;
 618         struct l_wait_info lwi = { 0 };
 619         time_t expire_time;
 620         ENTRY;
 621
 622         cfs_daemonize_ctxt("ll_evictor");
 623
 624         CDEBUG(D_HA, "Starting Ping Evictor\n");
 625         pet_state = PET_READY;
 626         while (1) {
 627                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 628                              (pet_state == PET_TERMINATE), &lwi);
 629
 630                 /* loop until all obd's will be removed */
 631                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 632                         break;
 633
 634                 /* we only get here if pet_exp != NULL, and the end of this
 635                  * loop is the only place which sets it NULL again, so lock
 636                  * is not strictly necessary. */
 637                 cfs_spin_lock(&pet_lock);
 638                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 639                                      obd_evict_list);
 640                 cfs_spin_unlock(&pet_lock);
 641
 642                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 643
 644                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 645                        obd->obd_name, expire_time);
 646
 647                 /* Exports can't be deleted out of the list while we hold
 648                  * the obd lock (class_unlink_export), which means we can't
 649                  * lose the last ref on the export.  If they've already been
 650                  * removed from the list, we won't find them here. */
 651                 cfs_spin_lock(&obd->obd_dev_lock);
 652                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 653                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 654                                              struct obd_export,
 655                                              exp_obd_chain_timed);
 656                         if (expire_time > exp->exp_last_request_time) {
 657                                 class_export_get(exp);
 658                                 cfs_spin_unlock(&obd->obd_dev_lock);
 659                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 660                                               " (at %s) in %ld seconds. I think"
 661                                               " it's dead, and I am evicting"
 662                                               " it. exp %p, cur %ld expire %ld"
 663                                               " last %ld\n",
 664                                               obd->obd_name,
 665                                               obd_uuid2str(&exp->exp_client_uuid),
 666                                               obd_export_nid2str(exp),
 667                                               (long)(cfs_time_current_sec() -
 668                                                      exp->exp_last_request_time),
 669                                               exp, (long)cfs_time_current_sec(),
 670                                               (long)expire_time,
 671                                               (long)exp->exp_last_request_time);
 672                                 CDEBUG(D_HA, "Last request was at %ld\n",
 673                                        exp->exp_last_request_time);
 674                                 class_fail_export(exp);
 675                                 class_export_put(exp);
 676                                 cfs_spin_lock(&obd->obd_dev_lock);
 677                         } else {
 678                                 /* List is sorted, so everyone below is ok */
 679                                 break;
 680                         }
 681                 }
 682                 cfs_spin_unlock(&obd->obd_dev_lock);
 683
 684                 cfs_spin_lock(&pet_lock);
 685                 cfs_list_del_init(&obd->obd_evict_list);
 686                 cfs_spin_unlock(&pet_lock);
 687
 688                 class_decref(obd, "evictor", obd);
 689         }
 690         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 691
 692         RETURN(0);
 693 }
 694
 695 void ping_evictor_start(void)
 696 {
 697         int rc;
 698
 699         if (++pet_refcount > 1)
 700                 return;
 701
 702         cfs_waitq_init(&pet_waitq);
 703
 704         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 705         if (rc < 0) {
 706                 pet_refcount--;
 707                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 708         }
 709 }
 710 EXPORT_SYMBOL(ping_evictor_start);
 711
 712 void ping_evictor_stop(void)
 713 {
 714         if (--pet_refcount > 0)
 715                 return;
 716
 717         pet_state = PET_TERMINATE;
 718         cfs_waitq_signal(&pet_waitq);
 719 }
 720 EXPORT_SYMBOL(ping_evictor_stop);
 721 #else /* !__KERNEL__ */
 722
 723 /* XXX
 724  * the current implementation of pinger in liblustre is not optimized
 725  */
 726
 727 #ifdef ENABLE_PINGER
 728 static struct pinger_data {
 729         int             pd_recursion;
 730         cfs_time_t      pd_this_ping;   /* jiffies */
 731         cfs_time_t      pd_next_ping;   /* jiffies */
 732         struct ptlrpc_request_set *pd_set;
 733 } pinger_args;
 734
 735 static int pinger_check_rpcs(void *arg)
 736 {
 737         cfs_time_t curtime = cfs_time_current();
 738         struct ptlrpc_request *req;
 739         struct ptlrpc_request_set *set;
 740         cfs_list_t *iter;
 741         struct obd_import *imp;
 742         struct pinger_data *pd = &pinger_args;
 743         int rc;
 744
 745         /* prevent recursion */
 746         if (pd->pd_recursion++) {
 747                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 748                 LASSERT(pd->pd_set);
 749                 pd->pd_recursion--;
 750                 return 0;
 751         }
 752
 753         /* have we reached ping point? */
 754         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 755                 pd->pd_recursion--;
 756                 return 0;
 757         }
 758
 759         /* if we have rpc_set already, continue processing it */
 760         if (pd->pd_set) {
 761                 LASSERT(pd->pd_this_ping);
 762                 set = pd->pd_set;
 763                 goto do_check_set;
 764         }
 765
 766         pd->pd_this_ping = curtime;
 767         pd->pd_set = ptlrpc_prep_set();
 768         if (pd->pd_set == NULL)
 769                 goto out;
 770         set = pd->pd_set;
 771
 772         /* add rpcs into set */
 773         cfs_mutex_lock(&pinger_mutex);
 774         cfs_list_for_each(iter, &pinger_imports) {
 775                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 776                                                         imp_pinger_chain);
 777                 int generation, level;
 778
 779                 if (cfs_time_aftereq(pd->pd_this_ping,
 780                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 781                         /* Add a ping. */
 782                         cfs_spin_lock(&imp->imp_lock);
 783                         generation = imp->imp_generation;
 784                         level = imp->imp_state;
 785                         cfs_spin_unlock(&imp->imp_lock);
 786
 787                         if (level != LUSTRE_IMP_FULL) {
 788                                 CDEBUG(D_HA,
 789                                        "not pinging %s (in recovery)\n",
 790                                        obd2cli_tgt(imp->imp_obd));
 791                                 continue;
 792                         }
 793
 794                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 795                                                         LUSTRE_OBD_VERSION,
 796                                                         OBD_PING);
 797                         if (req == NULL) {
 798                                 CERROR("OOM trying to ping %s->%s\n",
 799                                        imp->imp_obd->obd_uuid.uuid,
 800                                        obd2cli_tgt(imp->imp_obd));
 801                                 break;
 802                         }
 803
 804                         req->rq_no_resend = 1;
 805                         ptlrpc_request_set_replen(req);
 806                         req->rq_send_state = LUSTRE_IMP_FULL;
 807                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 808                         req->rq_import_generation = generation;
 809                         ptlrpc_set_add_req(set, req);
 810                 } else {
 811                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 812                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 813                                imp->imp_next_ping, pd->pd_this_ping);
 814                 }
 815         }
 816         pd->pd_this_ping = curtime;
 817         cfs_mutex_unlock(&pinger_mutex);
 818
 819         /* Might be empty, that's OK. */
 820         if (cfs_atomic_read(&set->set_remaining) == 0)
 821                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 822
 823         cfs_list_for_each(iter, &set->set_requests) {
 824                 struct ptlrpc_request *req =
 825                         cfs_list_entry(iter, struct ptlrpc_request,
 826                                        rq_set_chain);
 827                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 828                           req->rq_import->imp_obd->obd_uuid.uuid,
 829                           obd2cli_tgt(req->rq_import->imp_obd));
 830                 (void)ptl_send_rpc(req, 0);
 831         }
 832
 833 do_check_set:
 834         rc = ptlrpc_check_set(NULL, set);
 835
 836         /* not finished, and we are not expired, simply return */
 837         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 838                                             cfs_time_seconds(PING_INTERVAL)))) {
 839                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 840                 pd->pd_recursion--;
 841                 return 0;
 842         }
 843
 844         /* Expire all the requests that didn't come back. */
 845         cfs_mutex_lock(&pinger_mutex);
 846         cfs_list_for_each(iter, &set->set_requests) {
 847                 req = cfs_list_entry(iter, struct ptlrpc_request,
 848                                      rq_set_chain);
 849
 850                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 851                         continue;
 852
 853                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 854                        req);
 855
 856                 /* This will also unregister reply. */
 857                 ptlrpc_expire_one_request(req, 0);
 858
 859                 /* We're done with this req, let's finally move it to complete
 860                  * phase and take care of inflights. */
 861                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 862                 imp = req->rq_import;
 863                 cfs_spin_lock(&imp->imp_lock);
 864                 if (!cfs_list_empty(&req->rq_list)) {
 865                         cfs_list_del_init(&req->rq_list);
 866                         cfs_atomic_dec(&imp->imp_inflight);
 867                 }
 868                 cfs_spin_unlock(&imp->imp_lock);
 869                 cfs_atomic_dec(&set->set_remaining);
 870         }
 871         cfs_mutex_unlock(&pinger_mutex);
 872
 873         ptlrpc_set_destroy(set);
 874         pd->pd_set = NULL;
 875
 876 out:
 877         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 878                                         cfs_time_seconds(PING_INTERVAL));
 879         pd->pd_this_ping = 0; /* XXX for debug */
 880
 881         CDEBUG(D_INFO, "finished a round ping\n");
 882         pd->pd_recursion--;
 883         return 0;
 884 }
 885
 886 static void *pinger_callback = NULL;
 887 #endif /* ENABLE_PINGER */
 888
 889 int ptlrpc_start_pinger(void)
 890 {
 891 #ifdef ENABLE_PINGER
 892         memset(&pinger_args, 0, sizeof(pinger_args));
 893         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 894                                                            &pinger_check_rpcs,
 895                                                            &pinger_args);
 896 #endif
 897         return 0;
 898 }
 899
 900 int ptlrpc_stop_pinger(void)
 901 {
 902 #ifdef ENABLE_PINGER
 903         if (pinger_callback)
 904                 liblustre_deregister_wait_callback(pinger_callback);
 905 #endif
 906         return 0;
 907 }
 908
 909 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 910 {
 911 #ifdef ENABLE_PINGER
 912         cfs_mutex_lock(&pinger_mutex);
 913         ptlrpc_update_next_ping(imp, 0);
 914         if (pinger_args.pd_set == NULL &&
 915             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 916                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 917                         imp->imp_next_ping, cfs_time_current());
 918                 pinger_args.pd_next_ping = imp->imp_next_ping;
 919         }
 920         cfs_mutex_unlock(&pinger_mutex);
 921 #endif
 922 }
 923
 924 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 925 {
 926 #ifdef ENABLE_PINGER
 927         cfs_mutex_lock(&pinger_mutex);
 928         ptlrpc_update_next_ping(imp, 1);
 929         if (pinger_args.pd_set == NULL &&
 930             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 931                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 932                         imp->imp_next_ping, cfs_time_current());
 933                 pinger_args.pd_next_ping = imp->imp_next_ping;
 934         }
 935         cfs_mutex_unlock(&pinger_mutex);
 936 #endif
 937 }
 938
 939 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 940                               timeout_cb_t cb, void *data,
 941                               cfs_list_t *obd_list)
 942 {
 943         return 0;
 944 }
 945
 946 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 947                               enum timeout_event event)
 948 {
 949         return 0;
 950 }
 951
 952 int ptlrpc_pinger_add_import(struct obd_import *imp)
 953 {
 954         ENTRY;
 955         if (!cfs_list_empty(&imp->imp_pinger_chain))
 956                 RETURN(-EALREADY);
 957
 958         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 959                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 960         ptlrpc_pinger_sending_on_import(imp);
 961
 962         cfs_mutex_lock(&pinger_mutex);
 963         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 964         class_import_get(imp);
 965         cfs_mutex_unlock(&pinger_mutex);
 966
 967         RETURN(0);
 968 }
 969
 970 int ptlrpc_pinger_del_import(struct obd_import *imp)
 971 {
 972         ENTRY;
 973         if (cfs_list_empty(&imp->imp_pinger_chain))
 974                 RETURN(-ENOENT);
 975
 976         cfs_mutex_lock(&pinger_mutex);
 977         cfs_list_del_init(&imp->imp_pinger_chain);
 978         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 979                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 980         class_import_put(imp);
 981         cfs_mutex_unlock(&pinger_mutex);
 982         RETURN(0);
 983 }
 984
 985 void ptlrpc_pinger_wake_up()
 986 {
 987 #ifdef ENABLE_PINGER
 988         /* XXX force pinger to run, if needed */
 989         struct obd_import *imp;
 990         ENTRY;
 991         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 992                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 993                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 994 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 995                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 996                     !imp_is_deactive(imp))
 997 #else
 998                 /*XXX only recover for the initial connection */
 999                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1000                     imp->imp_state == LUSTRE_IMP_DISCON &&
1001                     !imp_is_deactive(imp))
1002 #endif
1003                         ptlrpc_initiate_recovery(imp);
1004                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1005                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1006                                      "state %d, deactive %d\n",
1007                                      imp->imp_obd->obd_uuid.uuid,
1008                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1009                                      imp_is_deactive(imp));
1010         }
1011         EXIT;
1012 #endif
1013 }
1014 #endif /* !__KERNEL__ */