kernel/drivers/staging/lustre/lustre/ptlrpc/pinger.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42
  43 #include "../include/obd_support.h"
  44 #include "../include/obd_class.h"
  45 #include "ptlrpc_internal.h"
  46
  47 struct mutex pinger_mutex;
  48 static LIST_HEAD(pinger_imports);
  49 static struct list_head timeout_list = LIST_HEAD_INIT(timeout_list);
  50
  51 struct ptlrpc_request *
  52 ptlrpc_prep_ping(struct obd_import *imp)
  53 {
  54         struct ptlrpc_request *req;
  55
  56         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  57                                         LUSTRE_OBD_VERSION, OBD_PING);
  58         if (req) {
  59                 ptlrpc_request_set_replen(req);
  60                 req->rq_no_resend = req->rq_no_delay = 1;
  61         }
  62         return req;
  63 }
  64
  65 int ptlrpc_obd_ping(struct obd_device *obd)
  66 {
  67         int rc;
  68         struct ptlrpc_request *req;
  69
  70         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  71         if (req == NULL)
  72                 return -ENOMEM;
  73
  74         req->rq_send_state = LUSTRE_IMP_FULL;
  75
  76         rc = ptlrpc_queue_wait(req);
  77
  78         ptlrpc_req_finished(req);
  79
  80         return rc;
  81 }
  82 EXPORT_SYMBOL(ptlrpc_obd_ping);
  83
  84 static int ptlrpc_ping(struct obd_import *imp)
  85 {
  86         struct ptlrpc_request *req;
  87
  88         req = ptlrpc_prep_ping(imp);
  89         if (req == NULL) {
  90                 CERROR("OOM trying to ping %s->%s\n",
  91                        imp->imp_obd->obd_uuid.uuid,
  92                        obd2cli_tgt(imp->imp_obd));
  93                 return -ENOMEM;
  94         }
  95
  96         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  97                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
  98         ptlrpcd_add_req(req);
  99
 100         return 0;
 101 }
 102
 103 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 104 {
 105         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 106
 107         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 108                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 109                                   AT_OFF ? 0 :
 110                                   at_get(&imp->imp_at.iat_net_latency));
 111                 time = min(time, dtime);
 112         }
 113         imp->imp_next_ping = cfs_time_shift(time);
 114 }
 115
 116 static inline int imp_is_deactive(struct obd_import *imp)
 117 {
 118         return (imp->imp_deactive ||
 119                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 120 }
 121
 122 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 123 {
 124         if (imp->imp_server_timeout)
 125                 return cfs_time_shift(obd_timeout / 2);
 126         else
 127                 return cfs_time_shift(obd_timeout);
 128 }
 129
 130 static long pinger_check_timeout(unsigned long time)
 131 {
 132         struct timeout_item *item;
 133         unsigned long timeout = PING_INTERVAL;
 134
 135         /* The timeout list is a increase order sorted list */
 136         mutex_lock(&pinger_mutex);
 137         list_for_each_entry(item, &timeout_list, ti_chain) {
 138                 int ti_timeout = item->ti_timeout;
 139
 140                 if (timeout > ti_timeout)
 141                         timeout = ti_timeout;
 142                 break;
 143         }
 144         mutex_unlock(&pinger_mutex);
 145
 146         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 147                                          cfs_time_current());
 148 }
 149
 150 static bool ir_up;
 151
 152 void ptlrpc_pinger_ir_up(void)
 153 {
 154         CDEBUG(D_HA, "IR up\n");
 155         ir_up = true;
 156 }
 157 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 158
 159 void ptlrpc_pinger_ir_down(void)
 160 {
 161         CDEBUG(D_HA, "IR down\n");
 162         ir_up = false;
 163 }
 164 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 165
 166 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 167                                          unsigned long this_ping)
 168 {
 169         int level;
 170         int force;
 171         int force_next;
 172         int suppress;
 173
 174         spin_lock(&imp->imp_lock);
 175
 176         level = imp->imp_state;
 177         force = imp->imp_force_verify;
 178         force_next = imp->imp_force_next_verify;
 179         /*
 180          * This will be used below only if the import is "FULL".
 181          */
 182         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 183
 184         imp->imp_force_verify = 0;
 185
 186         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 187             !force) {
 188                 spin_unlock(&imp->imp_lock);
 189                 return;
 190         }
 191
 192         imp->imp_force_next_verify = 0;
 193
 194         spin_unlock(&imp->imp_lock);
 195
 196         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
 197                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 198                ptlrpc_import_state_name(level), level, force, force_next,
 199                imp->imp_deactive, imp->imp_pingable, suppress);
 200
 201         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 202                 /* wait for a while before trying recovery again */
 203                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 204                 if (!imp->imp_no_pinger_recover)
 205                         ptlrpc_initiate_recovery(imp);
 206         } else if (level != LUSTRE_IMP_FULL ||
 207                    imp->imp_obd->obd_no_recov ||
 208                    imp_is_deactive(imp)) {
 209                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 210                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 211                        ptlrpc_import_state_name(level));
 212                 if (force) {
 213                         spin_lock(&imp->imp_lock);
 214                         imp->imp_force_verify = 1;
 215                         spin_unlock(&imp->imp_lock);
 216                 }
 217         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 218                 ptlrpc_ping(imp);
 219         }
 220 }
 221
 222 static int ptlrpc_pinger_main(void *arg)
 223 {
 224         struct ptlrpc_thread *thread = arg;
 225
 226         /* Record that the thread is running */
 227         thread_set_flags(thread, SVC_RUNNING);
 228         wake_up(&thread->t_ctl_waitq);
 229
 230         /* And now, loop forever, pinging as needed. */
 231         while (1) {
 232                 unsigned long this_ping = cfs_time_current();
 233                 struct l_wait_info lwi;
 234                 long time_to_next_wake;
 235                 struct timeout_item *item;
 236                 struct list_head *iter;
 237
 238                 mutex_lock(&pinger_mutex);
 239                 list_for_each_entry(item, &timeout_list, ti_chain) {
 240                         item->ti_cb(item, item->ti_cb_data);
 241                 }
 242                 list_for_each(iter, &pinger_imports) {
 243                         struct obd_import *imp =
 244                                 list_entry(iter, struct obd_import,
 245                                                imp_pinger_chain);
 246
 247                         ptlrpc_pinger_process_import(imp, this_ping);
 248                         /* obd_timeout might have changed */
 249                         if (imp->imp_pingable && imp->imp_next_ping &&
 250                             cfs_time_after(imp->imp_next_ping,
 251                                            cfs_time_add(this_ping,
 252                                                         cfs_time_seconds(PING_INTERVAL))))
 253                                 ptlrpc_update_next_ping(imp, 0);
 254                 }
 255                 mutex_unlock(&pinger_mutex);
 256
 257                 /* Wait until the next ping time, or until we're stopped. */
 258                 time_to_next_wake = pinger_check_timeout(this_ping);
 259                 /* The ping sent by ptlrpc_send_rpc may get sent out
 260                    say .01 second after this.
 261                    ptlrpc_pinger_sending_on_import will then set the
 262                    next ping time to next_ping + .01 sec, which means
 263                    we will SKIP the next ping at next_ping, and the
 264                    ping will get sent 2 timeouts from now!  Beware. */
 265                 CDEBUG(D_INFO, "next wakeup in " CFS_DURATION_T " (%ld)\n",
 266                        time_to_next_wake,
 267                        cfs_time_add(this_ping,
 268                                     cfs_time_seconds(PING_INTERVAL)));
 269                 if (time_to_next_wake > 0) {
 270                         lwi = LWI_TIMEOUT(max_t(long, time_to_next_wake,
 271                                                 cfs_time_seconds(1)),
 272                                           NULL, NULL);
 273                         l_wait_event(thread->t_ctl_waitq,
 274                                      thread_is_stopping(thread) ||
 275                                      thread_is_event(thread),
 276                                      &lwi);
 277                         if (thread_test_and_clear_flags(thread, SVC_STOPPING))
 278                                 break;
 279                         /* woken after adding import to reset timer */
 280                         thread_test_and_clear_flags(thread, SVC_EVENT);
 281                 }
 282         }
 283
 284         thread_set_flags(thread, SVC_STOPPED);
 285         wake_up(&thread->t_ctl_waitq);
 286
 287         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
 288         return 0;
 289 }
 290
 291 static struct ptlrpc_thread pinger_thread;
 292
 293 int ptlrpc_start_pinger(void)
 294 {
 295         struct l_wait_info lwi = { 0 };
 296         int rc;
 297
 298         if (!thread_is_init(&pinger_thread) &&
 299             !thread_is_stopped(&pinger_thread))
 300                 return -EALREADY;
 301
 302         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 303
 304         strcpy(pinger_thread.t_name, "ll_ping");
 305
 306         rc = PTR_ERR(kthread_run(ptlrpc_pinger_main, &pinger_thread,
 307                                  "%s", pinger_thread.t_name));
 308         if (IS_ERR_VALUE(rc)) {
 309                 CERROR("cannot start thread: %d\n", rc);
 310                 return rc;
 311         }
 312         l_wait_event(pinger_thread.t_ctl_waitq,
 313                      thread_is_running(&pinger_thread), &lwi);
 314
 315         return 0;
 316 }
 317
 318 static int ptlrpc_pinger_remove_timeouts(void);
 319
 320 int ptlrpc_stop_pinger(void)
 321 {
 322         struct l_wait_info lwi = { 0 };
 323         int rc = 0;
 324
 325         if (thread_is_init(&pinger_thread) ||
 326             thread_is_stopped(&pinger_thread))
 327                 return -EALREADY;
 328
 329         ptlrpc_pinger_remove_timeouts();
 330         thread_set_flags(&pinger_thread, SVC_STOPPING);
 331         wake_up(&pinger_thread.t_ctl_waitq);
 332
 333         l_wait_event(pinger_thread.t_ctl_waitq,
 334                      thread_is_stopped(&pinger_thread), &lwi);
 335
 336         return rc;
 337 }
 338
 339 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 340 {
 341         ptlrpc_update_next_ping(imp, 0);
 342 }
 343 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 344
 345 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 346 {
 347         ptlrpc_update_next_ping(imp, 1);
 348         assert_spin_locked(&imp->imp_lock);
 349         /*
 350          * Avoid reading stale imp_connect_data.  When not sure if pings are
 351          * expected or not on next connection, we assume they are not and force
 352          * one anyway to guarantee the chance of updating
 353          * imp_peer_committed_transno.
 354          */
 355         if (imp->imp_state != LUSTRE_IMP_FULL ||
 356             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 357                 imp->imp_force_next_verify = 1;
 358 }
 359
 360 int ptlrpc_pinger_add_import(struct obd_import *imp)
 361 {
 362         if (!list_empty(&imp->imp_pinger_chain))
 363                 return -EALREADY;
 364
 365         mutex_lock(&pinger_mutex);
 366         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 367                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 368         /* if we add to pinger we want recovery on this import */
 369         imp->imp_obd->obd_no_recov = 0;
 370         ptlrpc_update_next_ping(imp, 0);
 371         /* XXX sort, blah blah */
 372         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 373         class_import_get(imp);
 374
 375         ptlrpc_pinger_wake_up();
 376         mutex_unlock(&pinger_mutex);
 377
 378         return 0;
 379 }
 380 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 381
 382 int ptlrpc_pinger_del_import(struct obd_import *imp)
 383 {
 384         if (list_empty(&imp->imp_pinger_chain))
 385                 return -ENOENT;
 386
 387         mutex_lock(&pinger_mutex);
 388         list_del_init(&imp->imp_pinger_chain);
 389         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 390                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 391         /* if we remove from pinger we don't want recovery on this import */
 392         imp->imp_obd->obd_no_recov = 1;
 393         class_import_put(imp);
 394         mutex_unlock(&pinger_mutex);
 395         return 0;
 396 }
 397 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 398
 399 /**
 400  * Register a timeout callback to the pinger list, and the callback will
 401  * be called when timeout happens.
 402  */
 403 static struct timeout_item *ptlrpc_new_timeout(int time,
 404         enum timeout_event event, timeout_cb_t cb, void *data)
 405 {
 406         struct timeout_item *ti;
 407
 408         ti = kzalloc(sizeof(*ti), GFP_NOFS);
 409         if (!ti)
 410                 return NULL;
 411
 412         INIT_LIST_HEAD(&ti->ti_obd_list);
 413         INIT_LIST_HEAD(&ti->ti_chain);
 414         ti->ti_timeout = time;
 415         ti->ti_event = event;
 416         ti->ti_cb = cb;
 417         ti->ti_cb_data = data;
 418
 419         return ti;
 420 }
 421
 422 /**
 423  * Register timeout event on the pinger thread.
 424  * Note: the timeout list is an sorted list with increased timeout value.
 425  */
 426 static struct timeout_item*
 427 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 428                                timeout_cb_t cb, void *data)
 429 {
 430         struct timeout_item *item, *tmp;
 431
 432         LASSERT(mutex_is_locked(&pinger_mutex));
 433
 434         list_for_each_entry(item, &timeout_list, ti_chain)
 435                 if (item->ti_event == event)
 436                         goto out;
 437
 438         item = ptlrpc_new_timeout(time, event, cb, data);
 439         if (item) {
 440                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 441                         if (tmp->ti_timeout < time) {
 442                                 list_add(&item->ti_chain, &tmp->ti_chain);
 443                                 goto out;
 444                         }
 445                 }
 446                 list_add(&item->ti_chain, &timeout_list);
 447         }
 448 out:
 449         return item;
 450 }
 451
 452 /* Add a client_obd to the timeout event list, when timeout(@time)
 453  * happens, the callback(@cb) will be called.
 454  */
 455 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 456                               timeout_cb_t cb, void *data,
 457                               struct list_head *obd_list)
 458 {
 459         struct timeout_item *ti;
 460
 461         mutex_lock(&pinger_mutex);
 462         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 463         if (!ti) {
 464                 mutex_unlock(&pinger_mutex);
 465                 return -EINVAL;
 466         }
 467         list_add(obd_list, &ti->ti_obd_list);
 468         mutex_unlock(&pinger_mutex);
 469         return 0;
 470 }
 471 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 472
 473 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 474                               enum timeout_event event)
 475 {
 476         struct timeout_item *ti = NULL, *item;
 477
 478         if (list_empty(obd_list))
 479                 return 0;
 480         mutex_lock(&pinger_mutex);
 481         list_del_init(obd_list);
 482         /**
 483          * If there are no obd attached to the timeout event
 484          * list, remove this timeout event from the pinger
 485          */
 486         list_for_each_entry(item, &timeout_list, ti_chain) {
 487                 if (item->ti_event == event) {
 488                         ti = item;
 489                         break;
 490                 }
 491         }
 492         LASSERTF(ti != NULL, "ti is NULL !\n");
 493         if (list_empty(&ti->ti_obd_list)) {
 494                 list_del(&ti->ti_chain);
 495                 kfree(ti);
 496         }
 497         mutex_unlock(&pinger_mutex);
 498         return 0;
 499 }
 500 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 501
 502 static int ptlrpc_pinger_remove_timeouts(void)
 503 {
 504         struct timeout_item *item, *tmp;
 505
 506         mutex_lock(&pinger_mutex);
 507         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 508                 LASSERT(list_empty(&item->ti_obd_list));
 509                 list_del(&item->ti_chain);
 510                 kfree(item);
 511         }
 512         mutex_unlock(&pinger_mutex);
 513         return 0;
 514 }
 515
 516 void ptlrpc_pinger_wake_up(void)
 517 {
 518         thread_add_flags(&pinger_thread, SVC_EVENT);
 519         wake_up(&pinger_thread.t_ctl_waitq);
 520 }