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, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 static int suppress_pings;
  52 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
  53
  54 struct mutex pinger_mutex;
  55 static CFS_LIST_HEAD(pinger_imports);
  56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  57
  58 int ptlrpc_pinger_suppress_pings()
  59 {
  60         return suppress_pings;
  61 }
  62 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  63
  64 struct ptlrpc_request *
  65 ptlrpc_prep_ping(struct obd_import *imp)
  66 {
  67         struct ptlrpc_request *req;
  68
  69         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  70                                         LUSTRE_OBD_VERSION, OBD_PING);
  71         if (req) {
  72                 ptlrpc_request_set_replen(req);
  73                 req->rq_no_resend = req->rq_no_delay = 1;
  74         }
  75         return req;
  76 }
  77
  78 int ptlrpc_obd_ping(struct obd_device *obd)
  79 {
  80         int rc;
  81         struct ptlrpc_request *req;
  82         ENTRY;
  83
  84         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  85         if (req == NULL)
  86                 RETURN(-ENOMEM);
  87
  88         req->rq_send_state = LUSTRE_IMP_FULL;
  89
  90         rc = ptlrpc_queue_wait(req);
  91
  92         ptlrpc_req_finished(req);
  93
  94         RETURN(rc);
  95 }
  96 EXPORT_SYMBOL(ptlrpc_obd_ping);
  97
  98 int ptlrpc_ping(struct obd_import *imp)
  99 {
 100         struct ptlrpc_request *req;
 101         ENTRY;
 102
 103         req = ptlrpc_prep_ping(imp);
 104         if (req == NULL) {
 105                 CERROR("OOM trying to ping %s->%s\n",
 106                        imp->imp_obd->obd_uuid.uuid,
 107                        obd2cli_tgt(imp->imp_obd));
 108                 RETURN(-ENOMEM);
 109         }
 110
 111         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 112                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 113         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 114
 115         RETURN(0);
 116 }
 117
 118 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 119 {
 120 #ifdef ENABLE_PINGER
 121         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 122         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 123                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 124                                   AT_OFF ? 0 :
 125                                   at_get(&imp->imp_at.iat_net_latency));
 126                 time = min(time, dtime);
 127         }
 128         imp->imp_next_ping = cfs_time_shift(time);
 129 #endif /* ENABLE_PINGER */
 130 }
 131
 132 void ptlrpc_ping_import_soon(struct obd_import *imp)
 133 {
 134         imp->imp_next_ping = cfs_time_current();
 135 }
 136
 137 static inline int imp_is_deactive(struct obd_import *imp)
 138 {
 139         return (imp->imp_deactive ||
 140                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 141 }
 142
 143 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 144 {
 145         if (imp->imp_server_timeout)
 146                 return cfs_time_shift(obd_timeout / 2);
 147         else
 148                 return cfs_time_shift(obd_timeout);
 149 }
 150
 151 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 152 static cfs_time_t suspend_wakeup_time = 0;
 153
 154 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 155 {
 156         struct timeout_item *item;
 157         cfs_time_t timeout = PING_INTERVAL;
 158
 159         /* The timeout list is a increase order sorted list */
 160         mutex_lock(&pinger_mutex);
 161         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 162                 int ti_timeout = item->ti_timeout;
 163                 if (timeout > ti_timeout)
 164                         timeout = ti_timeout;
 165                 break;
 166         }
 167         mutex_unlock(&pinger_mutex);
 168
 169         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 170                                          cfs_time_current());
 171 }
 172
 173 static cfs_waitq_t suspend_timeouts_waitq;
 174
 175 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 176 {
 177         return suspend_wakeup_time;
 178 }
 179
 180 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 181 {
 182         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 183 #if 0
 184         if (imp->imp_no_timeout)
 185                 return;
 186         imp->imp_no_timeout = 1;
 187         cfs_atomic_inc(&suspend_timeouts);
 188         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 189                cfs_atomic_read(&suspend_timeouts));
 190 #endif
 191 }
 192
 193 void ptlrpc_activate_timeouts(struct obd_import *imp)
 194 {
 195         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 196 #if 0
 197         if (!imp->imp_no_timeout)
 198                 return;
 199         imp->imp_no_timeout = 0;
 200         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 201         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 202                 suspend_wakeup_time = cfs_time_current();
 203                 cfs_waitq_signal(&suspend_timeouts_waitq);
 204         }
 205         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 206                cfs_atomic_read(&suspend_timeouts));
 207 #endif
 208 }
 209
 210 int ptlrpc_check_suspend(void)
 211 {
 212         if (cfs_atomic_read(&suspend_timeouts))
 213                 return 1;
 214         return 0;
 215 }
 216
 217 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 218 {
 219         struct l_wait_info lwi;
 220
 221         if (cfs_atomic_read(&suspend_timeouts)) {
 222                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 223                           cfs_atomic_read(&suspend_timeouts));
 224                 lwi = LWI_INTR(NULL, NULL);
 225                 l_wait_event(suspend_timeouts_waitq,
 226                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 227                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 228                 return 1;
 229         }
 230         return 0;
 231 }
 232
 233 #ifdef __KERNEL__
 234
 235 static bool ir_up;
 236
 237 void ptlrpc_pinger_ir_up(void)
 238 {
 239         CDEBUG(D_HA, "IR up\n");
 240         ir_up = true;
 241 }
 242 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 243
 244 void ptlrpc_pinger_ir_down(void)
 245 {
 246         CDEBUG(D_HA, "IR down\n");
 247         ir_up = false;
 248 }
 249 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 250
 251 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 252                                          unsigned long this_ping)
 253 {
 254         int level;
 255         int force;
 256         int force_next;
 257         int suppress;
 258
 259         spin_lock(&imp->imp_lock);
 260
 261         level = imp->imp_state;
 262         force = imp->imp_force_verify;
 263         force_next = imp->imp_force_next_verify;
 264         /*
 265          * This will be used below only if the import is "FULL".
 266          */
 267         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 268
 269         imp->imp_force_verify = 0;
 270
 271         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 272             !force) {
 273                 spin_unlock(&imp->imp_lock);
 274                 return;
 275         }
 276
 277         imp->imp_force_next_verify = 0;
 278
 279         spin_unlock(&imp->imp_lock);
 280
 281         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
 282                "force %u force_next %u deactive %u pingable %u suppress %u\n",
 283                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 284                ptlrpc_import_state_name(level), level, force, force_next,
 285                imp->imp_deactive, imp->imp_pingable, suppress);
 286
 287         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 288                 /* wait for a while before trying recovery again */
 289                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 290                 if (!imp->imp_no_pinger_recover)
 291                         ptlrpc_initiate_recovery(imp);
 292         } else if (level != LUSTRE_IMP_FULL ||
 293                    imp->imp_obd->obd_no_recov ||
 294                    imp_is_deactive(imp)) {
 295                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
 296                        "or recovery disabled: %s)\n",
 297                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 298                        ptlrpc_import_state_name(level));
 299         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 300                 ptlrpc_ping(imp);
 301         }
 302 }
 303
 304 static int ptlrpc_pinger_main(void *arg)
 305 {
 306         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 307         ENTRY;
 308
 309         /* Record that the thread is running */
 310         thread_set_flags(thread, SVC_RUNNING);
 311         cfs_waitq_signal(&thread->t_ctl_waitq);
 312
 313         /* And now, loop forever, pinging as needed. */
 314         while (1) {
 315                 cfs_time_t this_ping = cfs_time_current();
 316                 struct l_wait_info lwi;
 317                 cfs_duration_t time_to_next_wake;
 318                 struct timeout_item *item;
 319                 cfs_list_t *iter;
 320
 321                 mutex_lock(&pinger_mutex);
 322                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 323                         item->ti_cb(item, item->ti_cb_data);
 324                 }
 325                 cfs_list_for_each(iter, &pinger_imports) {
 326                         struct obd_import *imp =
 327                                 cfs_list_entry(iter, struct obd_import,
 328                                                imp_pinger_chain);
 329
 330                         ptlrpc_pinger_process_import(imp, this_ping);
 331                         /* obd_timeout might have changed */
 332                         if (imp->imp_pingable && imp->imp_next_ping &&
 333                             cfs_time_after(imp->imp_next_ping,
 334                                            cfs_time_add(this_ping,
 335                                                         cfs_time_seconds(PING_INTERVAL))))
 336                                 ptlrpc_update_next_ping(imp, 0);
 337                 }
 338                 mutex_unlock(&pinger_mutex);
 339                 /* update memory usage info */
 340                 obd_update_maxusage();
 341
 342                 /* Wait until the next ping time, or until we're stopped. */
 343                 time_to_next_wake = pinger_check_timeout(this_ping);
 344                 /* The ping sent by ptlrpc_send_rpc may get sent out
 345                    say .01 second after this.
 346                    ptlrpc_pinger_sending_on_import will then set the
 347                    next ping time to next_ping + .01 sec, which means
 348                    we will SKIP the next ping at next_ping, and the
 349                    ping will get sent 2 timeouts from now!  Beware. */
 350                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
 351                        CFS_TIME_T")\n", time_to_next_wake,
 352                        cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
 353                 if (time_to_next_wake > 0) {
 354                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 355                                                 time_to_next_wake,
 356                                                 cfs_time_seconds(1)),
 357                                           NULL, NULL);
 358                         l_wait_event(thread->t_ctl_waitq,
 359                                      thread_is_stopping(thread) ||
 360                                      thread_is_event(thread),
 361                                      &lwi);
 362                         if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
 363                                 EXIT;
 364                                 break;
 365                         } else {
 366                                 /* woken after adding import to reset timer */
 367                                 thread_test_and_clear_flags(thread, SVC_EVENT);
 368                         }
 369                 }
 370         }
 371
 372         thread_set_flags(thread, SVC_STOPPED);
 373         cfs_waitq_signal(&thread->t_ctl_waitq);
 374
 375         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 376         return 0;
 377 }
 378
 379 static struct ptlrpc_thread pinger_thread;
 380
 381 int ptlrpc_start_pinger(void)
 382 {
 383         struct l_wait_info lwi = { 0 };
 384         int rc;
 385 #ifndef ENABLE_PINGER
 386         return 0;
 387 #endif
 388         ENTRY;
 389
 390         if (!thread_is_init(&pinger_thread) &&
 391             !thread_is_stopped(&pinger_thread))
 392                 RETURN(-EALREADY);
 393
 394         cfs_waitq_init(&pinger_thread.t_ctl_waitq);
 395         cfs_waitq_init(&suspend_timeouts_waitq);
 396
 397         strcpy(pinger_thread.t_name, "ll_ping");
 398
 399         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 400          * just drop the VM and FILES in kthread_run() right away. */
 401         rc = PTR_ERR(kthread_run(ptlrpc_pinger_main,
 402                                  &pinger_thread, pinger_thread.t_name));
 403         if (IS_ERR_VALUE(rc)) {
 404                 CERROR("cannot start thread: %d\n", rc);
 405                 RETURN(rc);
 406         }
 407         l_wait_event(pinger_thread.t_ctl_waitq,
 408                      thread_is_running(&pinger_thread), &lwi);
 409
 410         if (suppress_pings)
 411                 CWARN("Pings will be suppressed at the request of the "
 412                       "administrator.  The configuration shall meet the "
 413                       "additional requirements described in the manual.  "
 414                       "(Search for the \"suppress_pings\" kernel module "
 415                       "parameter.)\n");
 416
 417         RETURN(0);
 418 }
 419
 420 int ptlrpc_pinger_remove_timeouts(void);
 421
 422 int ptlrpc_stop_pinger(void)
 423 {
 424         struct l_wait_info lwi = { 0 };
 425 #ifndef ENABLE_PINGER
 426         return 0;
 427 #endif
 428         ENTRY;
 429
 430         if (thread_is_init(&pinger_thread) ||
 431             thread_is_stopped(&pinger_thread))
 432                 RETURN(-EALREADY);
 433
 434         ptlrpc_pinger_remove_timeouts();
 435
 436         thread_set_flags(&pinger_thread, SVC_STOPPING);
 437         cfs_waitq_signal(&pinger_thread.t_ctl_waitq);
 438
 439         l_wait_event(pinger_thread.t_ctl_waitq,
 440                      thread_is_stopped(&pinger_thread), &lwi);
 441         RETURN(0);
 442 }
 443
 444 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 445 {
 446         ptlrpc_update_next_ping(imp, 0);
 447 }
 448 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 449
 450 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 451 {
 452         ptlrpc_update_next_ping(imp, 1);
 453         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 454         /*
 455          * Avoid reading stale imp_connect_data.  When not sure if pings are
 456          * expected or not on next connection, we assume they are not and force
 457          * one anyway to guarantee the chance of updating
 458          * imp_peer_committed_transno.
 459          */
 460         if (imp->imp_state != LUSTRE_IMP_FULL ||
 461             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 462                 imp->imp_force_next_verify = 1;
 463 }
 464
 465 int ptlrpc_pinger_add_import(struct obd_import *imp)
 466 {
 467         ENTRY;
 468         if (!cfs_list_empty(&imp->imp_pinger_chain))
 469                 RETURN(-EALREADY);
 470
 471         mutex_lock(&pinger_mutex);
 472         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 473                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 474         /* if we add to pinger we want recovery on this import */
 475         imp->imp_obd->obd_no_recov = 0;
 476         ptlrpc_update_next_ping(imp, 0);
 477         /* XXX sort, blah blah */
 478         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 479         class_import_get(imp);
 480
 481         ptlrpc_pinger_wake_up();
 482         mutex_unlock(&pinger_mutex);
 483
 484         RETURN(0);
 485 }
 486 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 487
 488 int ptlrpc_pinger_del_import(struct obd_import *imp)
 489 {
 490         ENTRY;
 491         if (cfs_list_empty(&imp->imp_pinger_chain))
 492                 RETURN(-ENOENT);
 493
 494         mutex_lock(&pinger_mutex);
 495         cfs_list_del_init(&imp->imp_pinger_chain);
 496         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 497                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 498         /* if we remove from pinger we don't want recovery on this import */
 499         imp->imp_obd->obd_no_recov = 1;
 500         class_import_put(imp);
 501         mutex_unlock(&pinger_mutex);
 502         RETURN(0);
 503 }
 504 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 505
 506 /**
 507  * Register a timeout callback to the pinger list, and the callback will
 508  * be called when timeout happens.
 509  */
 510 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 511                                         timeout_cb_t cb, void *data)
 512 {
 513         struct timeout_item *ti;
 514
 515         OBD_ALLOC_PTR(ti);
 516         if (!ti)
 517                 return(NULL);
 518
 519         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 520         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 521         ti->ti_timeout = time;
 522         ti->ti_event = event;
 523         ti->ti_cb = cb;
 524         ti->ti_cb_data = data;
 525
 526         return ti;
 527 }
 528
 529 /**
 530  * Register timeout event on the the pinger thread.
 531  * Note: the timeout list is an sorted list with increased timeout value.
 532  */
 533 static struct timeout_item*
 534 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 535                                timeout_cb_t cb, void *data)
 536 {
 537         struct timeout_item *item, *tmp;
 538
 539         LASSERT_MUTEX_LOCKED(&pinger_mutex);
 540
 541         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 542                 if (item->ti_event == event)
 543                         goto out;
 544
 545         item = ptlrpc_new_timeout(time, event, cb, data);
 546         if (item) {
 547                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 548                         if (tmp->ti_timeout < time) {
 549                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 550                                 goto out;
 551                         }
 552                 }
 553                 cfs_list_add(&item->ti_chain, &timeout_list);
 554         }
 555 out:
 556         return item;
 557 }
 558
 559 /* Add a client_obd to the timeout event list, when timeout(@time)
 560  * happens, the callback(@cb) will be called.
 561  */
 562 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 563                               timeout_cb_t cb, void *data,
 564                               cfs_list_t *obd_list)
 565 {
 566         struct timeout_item *ti;
 567
 568         mutex_lock(&pinger_mutex);
 569         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 570         if (!ti) {
 571                 mutex_unlock(&pinger_mutex);
 572                 return (-EINVAL);
 573         }
 574         cfs_list_add(obd_list, &ti->ti_obd_list);
 575         mutex_unlock(&pinger_mutex);
 576         return 0;
 577 }
 578 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 579
 580 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 581                               enum timeout_event event)
 582 {
 583         struct timeout_item *ti = NULL, *item;
 584
 585         if (cfs_list_empty(obd_list))
 586                 return 0;
 587         mutex_lock(&pinger_mutex);
 588         cfs_list_del_init(obd_list);
 589         /**
 590          * If there are no obd attached to the timeout event
 591          * list, remove this timeout event from the pinger
 592          */
 593         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 594                 if (item->ti_event == event) {
 595                         ti = item;
 596                         break;
 597                 }
 598         }
 599         LASSERTF(ti != NULL, "ti is NULL ! \n");
 600         if (cfs_list_empty(&ti->ti_obd_list)) {
 601                 cfs_list_del(&ti->ti_chain);
 602                 OBD_FREE_PTR(ti);
 603         }
 604         mutex_unlock(&pinger_mutex);
 605         return 0;
 606 }
 607 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 608
 609 int ptlrpc_pinger_remove_timeouts(void)
 610 {
 611         struct timeout_item *item, *tmp;
 612
 613         mutex_lock(&pinger_mutex);
 614         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 615                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 616                 cfs_list_del(&item->ti_chain);
 617                 OBD_FREE_PTR(item);
 618         }
 619         mutex_unlock(&pinger_mutex);
 620         return 0;
 621 }
 622
 623 void ptlrpc_pinger_wake_up()
 624 {
 625 #ifdef ENABLE_PINGER
 626         thread_add_flags(&pinger_thread, SVC_EVENT);
 627         cfs_waitq_signal(&pinger_thread.t_ctl_waitq);
 628 #endif
 629 }
 630
 631 /* Ping evictor thread */
 632 #define PET_READY     1
 633 #define PET_TERMINATE 2
 634
 635 static int               pet_refcount = 0;
 636 static int               pet_state;
 637 static cfs_waitq_t       pet_waitq;
 638 CFS_LIST_HEAD(pet_list);
 639 static DEFINE_SPINLOCK(pet_lock);
 640
 641 int ping_evictor_wake(struct obd_export *exp)
 642 {
 643         struct obd_device *obd;
 644
 645         spin_lock(&pet_lock);
 646         if (pet_state != PET_READY) {
 647                 /* eventually the new obd will call here again. */
 648                 spin_unlock(&pet_lock);
 649                 return 1;
 650         }
 651
 652         obd = class_exp2obd(exp);
 653         if (cfs_list_empty(&obd->obd_evict_list)) {
 654                 class_incref(obd, "evictor", obd);
 655                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 656         }
 657         spin_unlock(&pet_lock);
 658
 659         cfs_waitq_signal(&pet_waitq);
 660         return 0;
 661 }
 662
 663 static int ping_evictor_main(void *arg)
 664 {
 665         struct obd_device *obd;
 666         struct obd_export *exp;
 667         struct l_wait_info lwi = { 0 };
 668         time_t expire_time;
 669         ENTRY;
 670
 671         unshare_fs_struct();
 672
 673         CDEBUG(D_HA, "Starting Ping Evictor\n");
 674         pet_state = PET_READY;
 675         while (1) {
 676                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 677                              (pet_state == PET_TERMINATE), &lwi);
 678
 679                 /* loop until all obd's will be removed */
 680                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 681                         break;
 682
 683                 /* we only get here if pet_exp != NULL, and the end of this
 684                  * loop is the only place which sets it NULL again, so lock
 685                  * is not strictly necessary. */
 686                 spin_lock(&pet_lock);
 687                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 688                                      obd_evict_list);
 689                 spin_unlock(&pet_lock);
 690
 691                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 692
 693                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 694                        obd->obd_name, expire_time);
 695
 696                 /* Exports can't be deleted out of the list while we hold
 697                  * the obd lock (class_unlink_export), which means we can't
 698                  * lose the last ref on the export.  If they've already been
 699                  * removed from the list, we won't find them here. */
 700                 spin_lock(&obd->obd_dev_lock);
 701                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 702                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 703                                              struct obd_export,
 704                                              exp_obd_chain_timed);
 705                         if (expire_time > exp->exp_last_request_time) {
 706                                 class_export_get(exp);
 707                                 spin_unlock(&obd->obd_dev_lock);
 708                                 LCONSOLE_WARN("%s: haven't heard from client %s"
 709                                               " (at %s) in %ld seconds. I think"
 710                                               " it's dead, and I am evicting"
 711                                               " it. exp %p, cur %ld expire %ld"
 712                                               " last %ld\n",
 713                                               obd->obd_name,
 714                                               obd_uuid2str(&exp->exp_client_uuid),
 715                                               obd_export_nid2str(exp),
 716                                               (long)(cfs_time_current_sec() -
 717                                                      exp->exp_last_request_time),
 718                                               exp, (long)cfs_time_current_sec(),
 719                                               (long)expire_time,
 720                                               (long)exp->exp_last_request_time);
 721                                 CDEBUG(D_HA, "Last request was at %ld\n",
 722                                        exp->exp_last_request_time);
 723                                 class_fail_export(exp);
 724                                 class_export_put(exp);
 725                                 spin_lock(&obd->obd_dev_lock);
 726                         } else {
 727                                 /* List is sorted, so everyone below is ok */
 728                                 break;
 729                         }
 730                 }
 731                 spin_unlock(&obd->obd_dev_lock);
 732
 733                 spin_lock(&pet_lock);
 734                 cfs_list_del_init(&obd->obd_evict_list);
 735                 spin_unlock(&pet_lock);
 736
 737                 class_decref(obd, "evictor", obd);
 738         }
 739         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 740
 741         RETURN(0);
 742 }
 743
 744 void ping_evictor_start(void)
 745 {
 746         cfs_task_t *task;
 747
 748         if (++pet_refcount > 1)
 749                 return;
 750
 751         cfs_waitq_init(&pet_waitq);
 752
 753         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 754         if (IS_ERR(task)) {
 755                 pet_refcount--;
 756                 CERROR("Cannot start ping evictor thread: %ld\n",
 757                         PTR_ERR(task));
 758         }
 759 }
 760 EXPORT_SYMBOL(ping_evictor_start);
 761
 762 void ping_evictor_stop(void)
 763 {
 764         if (--pet_refcount > 0)
 765                 return;
 766
 767         pet_state = PET_TERMINATE;
 768         cfs_waitq_signal(&pet_waitq);
 769 }
 770 EXPORT_SYMBOL(ping_evictor_stop);
 771 #else /* !__KERNEL__ */
 772
 773 /* XXX
 774  * the current implementation of pinger in liblustre is not optimized
 775  */
 776
 777 #ifdef ENABLE_PINGER
 778 static struct pinger_data {
 779         int             pd_recursion;
 780         cfs_time_t      pd_this_ping;   /* jiffies */
 781         cfs_time_t      pd_next_ping;   /* jiffies */
 782         struct ptlrpc_request_set *pd_set;
 783 } pinger_args;
 784
 785 static int pinger_check_rpcs(void *arg)
 786 {
 787         cfs_time_t curtime = cfs_time_current();
 788         struct ptlrpc_request *req;
 789         struct ptlrpc_request_set *set;
 790         cfs_list_t *iter;
 791         struct obd_import *imp;
 792         struct pinger_data *pd = &pinger_args;
 793         int rc;
 794
 795         /* prevent recursion */
 796         if (pd->pd_recursion++) {
 797                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 798                 LASSERT(pd->pd_set);
 799                 pd->pd_recursion--;
 800                 return 0;
 801         }
 802
 803         /* have we reached ping point? */
 804         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 805                 pd->pd_recursion--;
 806                 return 0;
 807         }
 808
 809         /* if we have rpc_set already, continue processing it */
 810         if (pd->pd_set) {
 811                 LASSERT(pd->pd_this_ping);
 812                 set = pd->pd_set;
 813                 goto do_check_set;
 814         }
 815
 816         pd->pd_this_ping = curtime;
 817         pd->pd_set = ptlrpc_prep_set();
 818         if (pd->pd_set == NULL)
 819                 goto out;
 820         set = pd->pd_set;
 821
 822         /* add rpcs into set */
 823         mutex_lock(&pinger_mutex);
 824         cfs_list_for_each(iter, &pinger_imports) {
 825                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 826                                                         imp_pinger_chain);
 827                 int generation, level;
 828
 829                 if (cfs_time_aftereq(pd->pd_this_ping,
 830                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 831                         /* Add a ping. */
 832                         spin_lock(&imp->imp_lock);
 833                         generation = imp->imp_generation;
 834                         level = imp->imp_state;
 835                         spin_unlock(&imp->imp_lock);
 836
 837                         if (level != LUSTRE_IMP_FULL) {
 838                                 CDEBUG(D_HA,
 839                                        "not pinging %s (in recovery)\n",
 840                                        obd2cli_tgt(imp->imp_obd));
 841                                 continue;
 842                         }
 843
 844                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 845                                                         LUSTRE_OBD_VERSION,
 846                                                         OBD_PING);
 847                         if (req == NULL) {
 848                                 CERROR("OOM trying to ping %s->%s\n",
 849                                        imp->imp_obd->obd_uuid.uuid,
 850                                        obd2cli_tgt(imp->imp_obd));
 851                                 break;
 852                         }
 853
 854                         req->rq_no_resend = 1;
 855                         ptlrpc_request_set_replen(req);
 856                         req->rq_send_state = LUSTRE_IMP_FULL;
 857                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 858                         req->rq_import_generation = generation;
 859                         ptlrpc_set_add_req(set, req);
 860                 } else {
 861                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 862                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 863                                imp->imp_next_ping, pd->pd_this_ping);
 864                 }
 865         }
 866         pd->pd_this_ping = curtime;
 867         mutex_unlock(&pinger_mutex);
 868
 869         /* Might be empty, that's OK. */
 870         if (cfs_atomic_read(&set->set_remaining) == 0)
 871                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 872
 873         cfs_list_for_each(iter, &set->set_requests) {
 874                 struct ptlrpc_request *req =
 875                         cfs_list_entry(iter, struct ptlrpc_request,
 876                                        rq_set_chain);
 877                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 878                           req->rq_import->imp_obd->obd_uuid.uuid,
 879                           obd2cli_tgt(req->rq_import->imp_obd));
 880                 (void)ptl_send_rpc(req, 0);
 881         }
 882
 883 do_check_set:
 884         rc = ptlrpc_check_set(NULL, set);
 885
 886         /* not finished, and we are not expired, simply return */
 887         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 888                                             cfs_time_seconds(PING_INTERVAL)))) {
 889                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 890                 pd->pd_recursion--;
 891                 return 0;
 892         }
 893
 894         /* Expire all the requests that didn't come back. */
 895         mutex_lock(&pinger_mutex);
 896         cfs_list_for_each(iter, &set->set_requests) {
 897                 req = cfs_list_entry(iter, struct ptlrpc_request,
 898                                      rq_set_chain);
 899
 900                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 901                         continue;
 902
 903                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 904                        req);
 905
 906                 /* This will also unregister reply. */
 907                 ptlrpc_expire_one_request(req, 0);
 908
 909                 /* We're done with this req, let's finally move it to complete
 910                  * phase and take care of inflights. */
 911                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 912                 imp = req->rq_import;
 913                 spin_lock(&imp->imp_lock);
 914                 if (!cfs_list_empty(&req->rq_list)) {
 915                         cfs_list_del_init(&req->rq_list);
 916                         cfs_atomic_dec(&imp->imp_inflight);
 917                 }
 918                 spin_unlock(&imp->imp_lock);
 919                 cfs_atomic_dec(&set->set_remaining);
 920         }
 921         mutex_unlock(&pinger_mutex);
 922
 923         ptlrpc_set_destroy(set);
 924         pd->pd_set = NULL;
 925
 926 out:
 927         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 928                                         cfs_time_seconds(PING_INTERVAL));
 929         pd->pd_this_ping = 0; /* XXX for debug */
 930
 931         CDEBUG(D_INFO, "finished a round ping\n");
 932         pd->pd_recursion--;
 933         return 0;
 934 }
 935
 936 static void *pinger_callback = NULL;
 937 #endif /* ENABLE_PINGER */
 938
 939 int ptlrpc_start_pinger(void)
 940 {
 941 #ifdef ENABLE_PINGER
 942         memset(&pinger_args, 0, sizeof(pinger_args));
 943         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 944                                                            &pinger_check_rpcs,
 945                                                            &pinger_args);
 946 #endif
 947         return 0;
 948 }
 949
 950 int ptlrpc_stop_pinger(void)
 951 {
 952 #ifdef ENABLE_PINGER
 953         if (pinger_callback)
 954                 liblustre_deregister_wait_callback(pinger_callback);
 955 #endif
 956         return 0;
 957 }
 958
 959 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 960 {
 961 #ifdef ENABLE_PINGER
 962         mutex_lock(&pinger_mutex);
 963         ptlrpc_update_next_ping(imp, 0);
 964         if (pinger_args.pd_set == NULL &&
 965             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 966                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 967                         imp->imp_next_ping, cfs_time_current());
 968                 pinger_args.pd_next_ping = imp->imp_next_ping;
 969         }
 970         mutex_unlock(&pinger_mutex);
 971 #endif
 972 }
 973
 974 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 975 {
 976 #ifdef ENABLE_PINGER
 977         mutex_lock(&pinger_mutex);
 978         ptlrpc_update_next_ping(imp, 1);
 979         if (pinger_args.pd_set == NULL &&
 980             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 981                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 982                         imp->imp_next_ping, cfs_time_current());
 983                 pinger_args.pd_next_ping = imp->imp_next_ping;
 984         }
 985         mutex_unlock(&pinger_mutex);
 986 #endif
 987 }
 988
 989 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 990                               timeout_cb_t cb, void *data,
 991                               cfs_list_t *obd_list)
 992 {
 993         return 0;
 994 }
 995
 996 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 997                               enum timeout_event event)
 998 {
 999         return 0;
1000 }
1001
1002 int ptlrpc_pinger_add_import(struct obd_import *imp)
1003 {
1004         ENTRY;
1005         if (!cfs_list_empty(&imp->imp_pinger_chain))
1006                 RETURN(-EALREADY);
1007
1008         CDEBUG(D_HA, "adding pingable import %s->%s\n",
1009                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1010         ptlrpc_pinger_sending_on_import(imp);
1011
1012         mutex_lock(&pinger_mutex);
1013         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
1014         class_import_get(imp);
1015         mutex_unlock(&pinger_mutex);
1016
1017         RETURN(0);
1018 }
1019
1020 int ptlrpc_pinger_del_import(struct obd_import *imp)
1021 {
1022         ENTRY;
1023         if (cfs_list_empty(&imp->imp_pinger_chain))
1024                 RETURN(-ENOENT);
1025
1026         mutex_lock(&pinger_mutex);
1027         cfs_list_del_init(&imp->imp_pinger_chain);
1028         CDEBUG(D_HA, "removing pingable import %s->%s\n",
1029                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1030         class_import_put(imp);
1031         mutex_unlock(&pinger_mutex);
1032         RETURN(0);
1033 }
1034
1035 void ptlrpc_pinger_wake_up()
1036 {
1037 #ifdef ENABLE_PINGER
1038         /* XXX force pinger to run, if needed */
1039         struct obd_import *imp;
1040         ENTRY;
1041         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
1042                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
1043                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
1044 #ifdef ENABLE_LIBLUSTRE_RECOVERY
1045                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
1046                     !imp_is_deactive(imp))
1047 #else
1048                 /*XXX only recover for the initial connection */
1049                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1050                     imp->imp_state == LUSTRE_IMP_DISCON &&
1051                     !imp_is_deactive(imp))
1052 #endif
1053                         ptlrpc_initiate_recovery(imp);
1054                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1055                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1056                                      "state %d, deactive %d\n",
1057                                      imp->imp_obd->obd_uuid.uuid,
1058                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1059                                      imp_is_deactive(imp));
1060         }
1061         EXIT;
1062 #endif
1063 }
1064 #endif /* !__KERNEL__ */