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