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 cfs_mutex_t 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         cfs_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         cfs_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         cfs_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         cfs_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_svc_data *data = (struct ptlrpc_svc_data *)arg;
 266         struct ptlrpc_thread *thread = data->thread;
 267         ENTRY;
 268
 269         cfs_daemonize(data->name);
 270
 271         /* Record that the thread is running */
 272         thread_set_flags(thread, SVC_RUNNING);
 273         cfs_waitq_signal(&thread->t_ctl_waitq);
 274
 275         /* And now, loop forever, pinging as needed. */
 276         while (1) {
 277                 cfs_time_t this_ping = cfs_time_current();
 278                 struct l_wait_info lwi;
 279                 cfs_duration_t time_to_next_wake;
 280                 struct timeout_item *item;
 281                 cfs_list_t *iter;
 282
 283                 cfs_mutex_lock(&pinger_mutex);
 284                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 285                         item->ti_cb(item, item->ti_cb_data);
 286                 }
 287                 cfs_list_for_each(iter, &pinger_imports) {
 288                         struct obd_import *imp =
 289                                 cfs_list_entry(iter, struct obd_import,
 290                                                imp_pinger_chain);
 291
 292                         ptlrpc_pinger_process_import(imp, this_ping);
 293                         /* obd_timeout might have changed */
 294                         if (imp->imp_pingable && imp->imp_next_ping &&
 295                             cfs_time_after(imp->imp_next_ping,
 296                                            cfs_time_add(this_ping,
 297                                                         cfs_time_seconds(PING_INTERVAL))))
 298                                 ptlrpc_update_next_ping(imp, 0);
 299                 }
 300                 cfs_mutex_unlock(&pinger_mutex);
 301                 /* update memory usage info */
 302                 obd_update_maxusage();
 303
 304                 /* Wait until the next ping time, or until we're stopped. */
 305                 time_to_next_wake = pinger_check_timeout(this_ping);
 306                 /* The ping sent by ptlrpc_send_rpc may get sent out
 307                    say .01 second after this.
 308                    ptlrpc_pinger_sending_on_import will then set the
 309                    next ping time to next_ping + .01 sec, which means
 310                    we will SKIP the next ping at next_ping, and the
 311                    ping will get sent 2 timeouts from now!  Beware. */
 312                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 313                        CFS_TIME_T")\n", time_to_next_wake,
 314                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 315                 if (time_to_next_wake > 0) {
 316                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 317                                                 time_to_next_wake,
 318                                                 cfs_time_seconds(1)),
 319                                           NULL, NULL);
 320                         l_wait_event(thread->t_ctl_waitq,
 321                                      thread_is_stopping(thread) ||
 322                                      thread_is_event(thread),
 323                                      &lwi);
 324                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 325                                 EXIT;
 326                                 break;
 327                         } else {
 328                                 /* woken after adding import to reset timer */
 329                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 330                         }
 331                 }
 332         }
 333
 334         thread_set_flags(thread, SVC_STOPPED);
 335         cfs_waitq_signal(&thread->t_ctl_waitq);
 336
 337         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 338         return 0;
 339 }
 340
 341 static struct ptlrpc_thread *pinger_thread = NULL;
 342
 343 int ptlrpc_start_pinger(void)
 344 {
 345         struct l_wait_info lwi = { 0 };
 346         struct ptlrpc_svc_data d;
 347         int rc;
 348 #ifndef ENABLE_PINGER
 349         return 0;
 350 #endif
 351         ENTRY;
 352
 353         if (pinger_thread != NULL)
 354                 RETURN(-EALREADY);
 355
 356         OBD_ALLOC_PTR(pinger_thread);
 357         if (pinger_thread == NULL)
 358                 RETURN(-ENOMEM);
 359         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 360         cfs_waitq_init(&suspend_timeouts_waitq);
 361
 362         d.name = "ll_ping";
 363         d.thread = pinger_thread;
 364
 365         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 366          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 367         rc = cfs_create_thread(ptlrpc_pinger_main, &d, CFS_DAEMON_FLAGS);
 368         if (rc < 0) {
 369                 CERROR("cannot start thread: %d\n", rc);
 370                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 371                 pinger_thread = NULL;
 372                 RETURN(rc);
 373         }
 374         l_wait_event(pinger_thread->t_ctl_waitq,
 375                      thread_is_running(pinger_thread), &lwi);
 376
 377         RETURN(0);
 378 }
 379
 380 int ptlrpc_pinger_remove_timeouts(void);
 381
 382 int ptlrpc_stop_pinger(void)
 383 {
 384         struct l_wait_info lwi = { 0 };
 385         int rc = 0;
 386 #ifndef ENABLE_PINGER
 387         return 0;
 388 #endif
 389         ENTRY;
 390
 391         if (pinger_thread == NULL)
 392                 RETURN(-EALREADY);
 393
 394         ptlrpc_pinger_remove_timeouts();
 395         cfs_mutex_lock(&pinger_mutex);
 396         thread_set_flags(pinger_thread, SVC_STOPPING);
 397         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 398         cfs_mutex_unlock(&pinger_mutex);
 399
 400         l_wait_event(pinger_thread->t_ctl_waitq,
 401                      thread_is_stopped(pinger_thread), &lwi);
 402
 403         OBD_FREE_PTR(pinger_thread);
 404         pinger_thread = NULL;
 405         RETURN(rc);
 406 }
 407
 408 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 409 {
 410         ptlrpc_update_next_ping(imp, 0);
 411 }
 412
 413 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 414 {
 415         ptlrpc_update_next_ping(imp, 1);
 416 }
 417
 418 int ptlrpc_pinger_add_import(struct obd_import *imp)
 419 {
 420         ENTRY;
 421         if (!cfs_list_empty(&imp->imp_pinger_chain))
 422                 RETURN(-EALREADY);
 423
 424         cfs_mutex_lock(&pinger_mutex);
 425         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 426                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 427         /* if we add to pinger we want recovery on this import */
 428         imp->imp_obd->obd_no_recov = 0;
 429         ptlrpc_update_next_ping(imp, 0);
 430         /* XXX sort, blah blah */
 431         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 432         class_import_get(imp);
 433
 434         ptlrpc_pinger_wake_up();
 435         cfs_mutex_unlock(&pinger_mutex);
 436
 437         RETURN(0);
 438 }
 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         cfs_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         cfs_mutex_unlock(&pinger_mutex);
 454         RETURN(0);
 455 }
 456
 457 /**
 458  * Register a timeout callback to the pinger list, and the callback will
 459  * be called when timeout happens.
 460  */
 461 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 462                                         timeout_cb_t cb, void *data)
 463 {
 464         struct timeout_item *ti;
 465
 466         OBD_ALLOC_PTR(ti);
 467         if (!ti)
 468                 return(NULL);
 469
 470         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 471         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 472         ti->ti_timeout = time;
 473         ti->ti_event = event;
 474         ti->ti_cb = cb;
 475         ti->ti_cb_data = data;
 476
 477         return ti;
 478 }
 479
 480 /**
 481  * Register timeout event on the the pinger thread.
 482  * Note: the timeout list is an sorted list with increased timeout value.
 483  */
 484 static struct timeout_item*
 485 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 486                                timeout_cb_t cb, void *data)
 487 {
 488         struct timeout_item *item, *tmp;
 489
 490         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 491
 492         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 493                 if (item->ti_event == event)
 494                         goto out;
 495
 496         item = ptlrpc_new_timeout(time, event, cb, data);
 497         if (item) {
 498                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 499                         if (tmp->ti_timeout < time) {
 500                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 501                                 goto out;
 502                         }
 503                 }
 504                 cfs_list_add(&item->ti_chain, &timeout_list);
 505         }
 506 out:
 507         return item;
 508 }
 509
 510 /* Add a client_obd to the timeout event list, when timeout(@time)
 511  * happens, the callback(@cb) will be called.
 512  */
 513 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 514                               timeout_cb_t cb, void *data,
 515                               cfs_list_t *obd_list)
 516 {
 517         struct timeout_item *ti;
 518
 519         cfs_mutex_lock(&pinger_mutex);
 520         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 521         if (!ti) {
 522                 cfs_mutex_unlock(&pinger_mutex);
 523                 return (-EINVAL);
 524         }
 525         cfs_list_add(obd_list, &ti->ti_obd_list);
 526         cfs_mutex_unlock(&pinger_mutex);
 527         return 0;
 528 }
 529
 530 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 531                               enum timeout_event event)
 532 {
 533         struct timeout_item *ti = NULL, *item;
 534
 535         if (cfs_list_empty(obd_list))
 536                 return 0;
 537         cfs_mutex_lock(&pinger_mutex);
 538         cfs_list_del_init(obd_list);
 539         /**
 540          * If there are no obd attached to the timeout event
 541          * list, remove this timeout event from the pinger
 542          */
 543         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 544                 if (item->ti_event == event) {
 545                         ti = item;
 546                         break;
 547                 }
 548         }
 549         LASSERTF(ti != NULL, "ti is NULL ! \n");
 550         if (cfs_list_empty(&ti->ti_obd_list)) {
 551                 cfs_list_del(&ti->ti_chain);
 552                 OBD_FREE_PTR(ti);
 553         }
 554         cfs_mutex_unlock(&pinger_mutex);
 555         return 0;
 556 }
 557
 558 int ptlrpc_pinger_remove_timeouts(void)
 559 {
 560         struct timeout_item *item, *tmp;
 561
 562         cfs_mutex_lock(&pinger_mutex);
 563         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 564                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 565                 cfs_list_del(&item->ti_chain);
 566                 OBD_FREE_PTR(item);
 567         }
 568         cfs_mutex_unlock(&pinger_mutex);
 569         return 0;
 570 }
 571
 572 void ptlrpc_pinger_wake_up()
 573 {
 574 #ifdef ENABLE_PINGER
 575         thread_add_flags(pinger_thread, SVC_EVENT);
 576         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 577 #endif
 578 }
 579
 580 /* Ping evictor thread */
 581 #define PET_READY     1
 582 #define PET_TERMINATE 2
 583
 584 static int               pet_refcount = 0;
 585 static int               pet_state;
 586 static cfs_waitq_t       pet_waitq;
 587 CFS_LIST_HEAD(pet_list);
 588 static cfs_spinlock_t    pet_lock = CFS_SPIN_LOCK_UNLOCKED;
 589
 590 int ping_evictor_wake(struct obd_export *exp)
 591 {
 592         struct obd_device *obd;
 593
 594         cfs_spin_lock(&pet_lock);
 595         if (pet_state != PET_READY) {
 596                 /* eventually the new obd will call here again. */
 597                 cfs_spin_unlock(&pet_lock);
 598                 return 1;
 599         }
 600
 601         obd = class_exp2obd(exp);
 602         if (cfs_list_empty(&obd->obd_evict_list)) {
 603                 class_incref(obd, "evictor", obd);
 604                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 605         }
 606         cfs_spin_unlock(&pet_lock);
 607
 608         cfs_waitq_signal(&pet_waitq);
 609         return 0;
 610 }
 611
 612 static int ping_evictor_main(void *arg)
 613 {
 614         struct obd_device *obd;
 615         struct obd_export *exp;
 616         struct l_wait_info lwi = { 0 };
 617         time_t expire_time;
 618         ENTRY;
 619
 620         cfs_daemonize_ctxt("ll_evictor");
 621
 622         CDEBUG(D_HA, "Starting Ping Evictor\n");
 623         pet_state = PET_READY;
 624         while (1) {
 625                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 626                              (pet_state == PET_TERMINATE), &lwi);
 627
 628                 /* loop until all obd's will be removed */
 629                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 630                         break;
 631
 632                 /* we only get here if pet_exp != NULL, and the end of this
 633                  * loop is the only place which sets it NULL again, so lock
 634                  * is not strictly necessary. */
 635                 cfs_spin_lock(&pet_lock);
 636                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 637                                      obd_evict_list);
 638                 cfs_spin_unlock(&pet_lock);
 639
 640                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 641
 642                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 643                        obd->obd_name, expire_time);
 644
 645                 /* Exports can't be deleted out of the list while we hold
 646                  * the obd lock (class_unlink_export), which means we can't
 647                  * lose the last ref on the export.  If they've already been
 648                  * removed from the list, we won't find them here. */
 649                 cfs_spin_lock(&obd->obd_dev_lock);
 650                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 651                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 652                                              struct obd_export,
 653                                              exp_obd_chain_timed);
 654                         if (expire_time > exp->exp_last_request_time) {
 655                                 class_export_get(exp);
 656                                 cfs_spin_unlock(&obd->obd_dev_lock);
 657                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 658                                               " (at %s) in %ld seconds. I think"
 659                                               " it's dead, and I am evicting"
 660                                               " it. exp %p, cur %ld expire %ld"
 661                                               " last %ld\n",
 662                                               obd->obd_name,
 663                                               obd_uuid2str(&exp->exp_client_uuid),
 664                                               obd_export_nid2str(exp),
 665                                               (long)(cfs_time_current_sec() -
 666                                                      exp->exp_last_request_time),
 667                                               exp, (long)cfs_time_current_sec(),
 668                                               (long)expire_time,
 669                                               (long)exp->exp_last_request_time);
 670                                 CDEBUG(D_HA, "Last request was at %ld\n",
 671                                        exp->exp_last_request_time);
 672                                 class_fail_export(exp);
 673                                 class_export_put(exp);
 674                                 cfs_spin_lock(&obd->obd_dev_lock);
 675                         } else {
 676                                 /* List is sorted, so everyone below is ok */
 677                                 break;
 678                         }
 679                 }
 680                 cfs_spin_unlock(&obd->obd_dev_lock);
 681
 682                 cfs_spin_lock(&pet_lock);
 683                 cfs_list_del_init(&obd->obd_evict_list);
 684                 cfs_spin_unlock(&pet_lock);
 685
 686                 class_decref(obd, "evictor", obd);
 687         }
 688         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 689
 690         RETURN(0);
 691 }
 692
 693 void ping_evictor_start(void)
 694 {
 695         int rc;
 696
 697         if (++pet_refcount > 1)
 698                 return;
 699
 700         cfs_waitq_init(&pet_waitq);
 701
 702         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 703         if (rc < 0) {
 704                 pet_refcount--;
 705                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 706         }
 707 }
 708 EXPORT_SYMBOL(ping_evictor_start);
 709
 710 void ping_evictor_stop(void)
 711 {
 712         if (--pet_refcount > 0)
 713                 return;
 714
 715         pet_state = PET_TERMINATE;
 716         cfs_waitq_signal(&pet_waitq);
 717 }
 718 EXPORT_SYMBOL(ping_evictor_stop);
 719 #else /* !__KERNEL__ */
 720
 721 /* XXX
 722  * the current implementation of pinger in liblustre is not optimized
 723  */
 724
 725 #ifdef ENABLE_PINGER
 726 static struct pinger_data {
 727         int             pd_recursion;
 728         cfs_time_t      pd_this_ping;   /* jiffies */
 729         cfs_time_t      pd_next_ping;   /* jiffies */
 730         struct ptlrpc_request_set *pd_set;
 731 } pinger_args;
 732
 733 static int pinger_check_rpcs(void *arg)
 734 {
 735         cfs_time_t curtime = cfs_time_current();
 736         struct ptlrpc_request *req;
 737         struct ptlrpc_request_set *set;
 738         cfs_list_t *iter;
 739         struct obd_import *imp;
 740         struct pinger_data *pd = &pinger_args;
 741         int rc;
 742
 743         /* prevent recursion */
 744         if (pd->pd_recursion++) {
 745                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 746                 LASSERT(pd->pd_set);
 747                 pd->pd_recursion--;
 748                 return 0;
 749         }
 750
 751         /* have we reached ping point? */
 752         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 753                 pd->pd_recursion--;
 754                 return 0;
 755         }
 756
 757         /* if we have rpc_set already, continue processing it */
 758         if (pd->pd_set) {
 759                 LASSERT(pd->pd_this_ping);
 760                 set = pd->pd_set;
 761                 goto do_check_set;
 762         }
 763
 764         pd->pd_this_ping = curtime;
 765         pd->pd_set = ptlrpc_prep_set();
 766         if (pd->pd_set == NULL)
 767                 goto out;
 768         set = pd->pd_set;
 769
 770         /* add rpcs into set */
 771         cfs_mutex_lock(&pinger_mutex);
 772         cfs_list_for_each(iter, &pinger_imports) {
 773                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 774                                                         imp_pinger_chain);
 775                 int generation, level;
 776
 777                 if (cfs_time_aftereq(pd->pd_this_ping,
 778                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 779                         /* Add a ping. */
 780                         cfs_spin_lock(&imp->imp_lock);
 781                         generation = imp->imp_generation;
 782                         level = imp->imp_state;
 783                         cfs_spin_unlock(&imp->imp_lock);
 784
 785                         if (level != LUSTRE_IMP_FULL) {
 786                                 CDEBUG(D_HA,
 787                                        "not pinging %s (in recovery)\n",
 788                                        obd2cli_tgt(imp->imp_obd));
 789                                 continue;
 790                         }
 791
 792                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 793                                                         LUSTRE_OBD_VERSION,
 794                                                         OBD_PING);
 795                         if (req == NULL) {
 796                                 CERROR("OOM trying to ping %s->%s\n",
 797                                        imp->imp_obd->obd_uuid.uuid,
 798                                        obd2cli_tgt(imp->imp_obd));
 799                                 break;
 800                         }
 801
 802                         req->rq_no_resend = 1;
 803                         ptlrpc_request_set_replen(req);
 804                         req->rq_send_state = LUSTRE_IMP_FULL;
 805                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 806                         req->rq_import_generation = generation;
 807                         ptlrpc_set_add_req(set, req);
 808                 } else {
 809                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 810                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 811                                imp->imp_next_ping, pd->pd_this_ping);
 812                 }
 813         }
 814         pd->pd_this_ping = curtime;
 815         cfs_mutex_unlock(&pinger_mutex);
 816
 817         /* Might be empty, that's OK. */
 818         if (cfs_atomic_read(&set->set_remaining) == 0)
 819                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 820
 821         cfs_list_for_each(iter, &set->set_requests) {
 822                 struct ptlrpc_request *req =
 823                         cfs_list_entry(iter, struct ptlrpc_request,
 824                                        rq_set_chain);
 825                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 826                           req->rq_import->imp_obd->obd_uuid.uuid,
 827                           obd2cli_tgt(req->rq_import->imp_obd));
 828                 (void)ptl_send_rpc(req, 0);
 829         }
 830
 831 do_check_set:
 832         rc = ptlrpc_check_set(NULL, set);
 833
 834         /* not finished, and we are not expired, simply return */
 835         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 836                                             cfs_time_seconds(PING_INTERVAL)))) {
 837                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 838                 pd->pd_recursion--;
 839                 return 0;
 840         }
 841
 842         /* Expire all the requests that didn't come back. */
 843         cfs_mutex_lock(&pinger_mutex);
 844         cfs_list_for_each(iter, &set->set_requests) {
 845                 req = cfs_list_entry(iter, struct ptlrpc_request,
 846                                      rq_set_chain);
 847
 848                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 849                         continue;
 850
 851                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 852                        req);
 853
 854                 /* This will also unregister reply. */
 855                 ptlrpc_expire_one_request(req, 0);
 856
 857                 /* We're done with this req, let's finally move it to complete
 858                  * phase and take care of inflights. */
 859                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 860                 imp = req->rq_import;
 861                 cfs_spin_lock(&imp->imp_lock);
 862                 if (!cfs_list_empty(&req->rq_list)) {
 863                         cfs_list_del_init(&req->rq_list);
 864                         cfs_atomic_dec(&imp->imp_inflight);
 865                 }
 866                 cfs_spin_unlock(&imp->imp_lock);
 867                 cfs_atomic_dec(&set->set_remaining);
 868         }
 869         cfs_mutex_unlock(&pinger_mutex);
 870
 871         ptlrpc_set_destroy(set);
 872         pd->pd_set = NULL;
 873
 874 out:
 875         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 876                                         cfs_time_seconds(PING_INTERVAL));
 877         pd->pd_this_ping = 0; /* XXX for debug */
 878
 879         CDEBUG(D_INFO, "finished a round ping\n");
 880         pd->pd_recursion--;
 881         return 0;
 882 }
 883
 884 static void *pinger_callback = NULL;
 885 #endif /* ENABLE_PINGER */
 886
 887 int ptlrpc_start_pinger(void)
 888 {
 889 #ifdef ENABLE_PINGER
 890         memset(&pinger_args, 0, sizeof(pinger_args));
 891         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 892                                                            &pinger_check_rpcs,
 893                                                            &pinger_args);
 894 #endif
 895         return 0;
 896 }
 897
 898 int ptlrpc_stop_pinger(void)
 899 {
 900 #ifdef ENABLE_PINGER
 901         if (pinger_callback)
 902                 liblustre_deregister_wait_callback(pinger_callback);
 903 #endif
 904         return 0;
 905 }
 906
 907 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 908 {
 909 #ifdef ENABLE_PINGER
 910         cfs_mutex_lock(&pinger_mutex);
 911         ptlrpc_update_next_ping(imp, 0);
 912         if (pinger_args.pd_set == NULL &&
 913             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 914                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 915                         imp->imp_next_ping, cfs_time_current());
 916                 pinger_args.pd_next_ping = imp->imp_next_ping;
 917         }
 918         cfs_mutex_unlock(&pinger_mutex);
 919 #endif
 920 }
 921
 922 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 923 {
 924 #ifdef ENABLE_PINGER
 925         cfs_mutex_lock(&pinger_mutex);
 926         ptlrpc_update_next_ping(imp, 1);
 927         if (pinger_args.pd_set == NULL &&
 928             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 929                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 930                         imp->imp_next_ping, cfs_time_current());
 931                 pinger_args.pd_next_ping = imp->imp_next_ping;
 932         }
 933         cfs_mutex_unlock(&pinger_mutex);
 934 #endif
 935 }
 936
 937 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 938                               timeout_cb_t cb, void *data,
 939                               cfs_list_t *obd_list)
 940 {
 941         return 0;
 942 }
 943
 944 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 945                               enum timeout_event event)
 946 {
 947         return 0;
 948 }
 949
 950 int ptlrpc_pinger_add_import(struct obd_import *imp)
 951 {
 952         ENTRY;
 953         if (!cfs_list_empty(&imp->imp_pinger_chain))
 954                 RETURN(-EALREADY);
 955
 956         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 957                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 958         ptlrpc_pinger_sending_on_import(imp);
 959
 960         cfs_mutex_lock(&pinger_mutex);
 961         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 962         class_import_get(imp);
 963         cfs_mutex_unlock(&pinger_mutex);
 964
 965         RETURN(0);
 966 }
 967
 968 int ptlrpc_pinger_del_import(struct obd_import *imp)
 969 {
 970         ENTRY;
 971         if (cfs_list_empty(&imp->imp_pinger_chain))
 972                 RETURN(-ENOENT);
 973
 974         cfs_mutex_lock(&pinger_mutex);
 975         cfs_list_del_init(&imp->imp_pinger_chain);
 976         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 977                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 978         class_import_put(imp);
 979         cfs_mutex_unlock(&pinger_mutex);
 980         RETURN(0);
 981 }
 982
 983 void ptlrpc_pinger_wake_up()
 984 {
 985 #ifdef ENABLE_PINGER
 986         /* XXX force pinger to run, if needed */
 987         struct obd_import *imp;
 988         ENTRY;
 989         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 990                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 991                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 992 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 993                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 994                     !imp_is_deactive(imp))
 995 #else
 996                 /*XXX only recover for the initial connection */
 997                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 998                     imp->imp_state == LUSTRE_IMP_DISCON &&
 999                     !imp_is_deactive(imp))
1000 #endif
1001                         ptlrpc_initiate_recovery(imp);
1002                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1003                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1004                                      "state %d, deactive %d\n",
1005                                      imp->imp_obd->obd_uuid.uuid,
1006                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1007                                      imp_is_deactive(imp));
1008         }
1009         EXIT;
1010 #endif
1011 }
1012 #endif /* !__KERNEL__ */