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 static int suppress_pings;
  48 module_param(suppress_pings, int, 0644);
  49 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
  50
  51 struct mutex pinger_mutex;
  52 static LIST_HEAD(pinger_imports);
  53 static struct list_head timeout_list = LIST_HEAD_INIT(timeout_list);
  54
  55 int ptlrpc_pinger_suppress_pings(void)
  56 {
  57         return suppress_pings;
  58 }
  59 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  60
  61 struct ptlrpc_request *
  62 ptlrpc_prep_ping(struct obd_import *imp)
  63 {
  64         struct ptlrpc_request *req;
  65
  66         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  67                                         LUSTRE_OBD_VERSION, OBD_PING);
  68         if (req) {
  69                 ptlrpc_request_set_replen(req);
  70                 req->rq_no_resend = req->rq_no_delay = 1;
  71         }
  72         return req;
  73 }
  74
  75 int ptlrpc_obd_ping(struct obd_device *obd)
  76 {
  77         int rc;
  78         struct ptlrpc_request *req;
  79
  80         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  81         if (req == NULL)
  82                 return -ENOMEM;
  83
  84         req->rq_send_state = LUSTRE_IMP_FULL;
  85
  86         rc = ptlrpc_queue_wait(req);
  87
  88         ptlrpc_req_finished(req);
  89
  90         return rc;
  91 }
  92 EXPORT_SYMBOL(ptlrpc_obd_ping);
  93
  94 static int ptlrpc_ping(struct obd_import *imp)
  95 {
  96         struct ptlrpc_request *req;
  97
  98         req = ptlrpc_prep_ping(imp);
  99         if (req == NULL) {
 100                 CERROR("OOM trying to ping %s->%s\n",
 101                        imp->imp_obd->obd_uuid.uuid,
 102                        obd2cli_tgt(imp->imp_obd));
 103                 return -ENOMEM;
 104         }
 105
 106         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 107                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 108         ptlrpcd_add_req(req);
 109
 110         return 0;
 111 }
 112
 113 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 114 {
 115         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 116         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 117                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 118                                   AT_OFF ? 0 :
 119                                   at_get(&imp->imp_at.iat_net_latency));
 120                 time = min(time, dtime);
 121         }
 122         imp->imp_next_ping = cfs_time_shift(time);
 123 }
 124
 125 void ptlrpc_ping_import_soon(struct obd_import *imp)
 126 {
 127         imp->imp_next_ping = cfs_time_current();
 128 }
 129
 130 static inline int imp_is_deactive(struct obd_import *imp)
 131 {
 132         return (imp->imp_deactive ||
 133                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 134 }
 135
 136 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 137 {
 138         if (imp->imp_server_timeout)
 139                 return cfs_time_shift(obd_timeout / 2);
 140         else
 141                 return cfs_time_shift(obd_timeout);
 142 }
 143
 144 static long pinger_check_timeout(unsigned long time)
 145 {
 146         struct timeout_item *item;
 147         unsigned long timeout = PING_INTERVAL;
 148
 149         /* The timeout list is a increase order sorted list */
 150         mutex_lock(&pinger_mutex);
 151         list_for_each_entry(item, &timeout_list, ti_chain) {
 152                 int ti_timeout = item->ti_timeout;
 153                 if (timeout > ti_timeout)
 154                         timeout = ti_timeout;
 155                 break;
 156         }
 157         mutex_unlock(&pinger_mutex);
 158
 159         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 160                                          cfs_time_current());
 161 }
 162
 163 static bool ir_up;
 164
 165 void ptlrpc_pinger_ir_up(void)
 166 {
 167         CDEBUG(D_HA, "IR up\n");
 168         ir_up = true;
 169 }
 170 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 171
 172 void ptlrpc_pinger_ir_down(void)
 173 {
 174         CDEBUG(D_HA, "IR down\n");
 175         ir_up = false;
 176 }
 177 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 178
 179 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 180                                          unsigned long this_ping)
 181 {
 182         int level;
 183         int force;
 184         int force_next;
 185         int suppress;
 186
 187         spin_lock(&imp->imp_lock);
 188
 189         level = imp->imp_state;
 190         force = imp->imp_force_verify;
 191         force_next = imp->imp_force_next_verify;
 192         /*
 193          * This will be used below only if the import is "FULL".
 194          */
 195         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 196
 197         imp->imp_force_verify = 0;
 198
 199         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 200             !force) {
 201                 spin_unlock(&imp->imp_lock);
 202                 return;
 203         }
 204
 205         imp->imp_force_next_verify = 0;
 206
 207         spin_unlock(&imp->imp_lock);
 208
 209         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",
 210                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 211                ptlrpc_import_state_name(level), level, force, force_next,
 212                imp->imp_deactive, imp->imp_pingable, suppress);
 213
 214         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 215                 /* wait for a while before trying recovery again */
 216                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 217                 if (!imp->imp_no_pinger_recover)
 218                         ptlrpc_initiate_recovery(imp);
 219         } else if (level != LUSTRE_IMP_FULL ||
 220                    imp->imp_obd->obd_no_recov ||
 221                    imp_is_deactive(imp)) {
 222                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 223                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 224                        ptlrpc_import_state_name(level));
 225                 if (force) {
 226                         spin_lock(&imp->imp_lock);
 227                         imp->imp_force_verify = 1;
 228                         spin_unlock(&imp->imp_lock);
 229                 }
 230         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 231                 ptlrpc_ping(imp);
 232         }
 233 }
 234
 235 static int ptlrpc_pinger_main(void *arg)
 236 {
 237         struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
 238
 239         /* Record that the thread is running */
 240         thread_set_flags(thread, SVC_RUNNING);
 241         wake_up(&thread->t_ctl_waitq);
 242
 243         /* And now, loop forever, pinging as needed. */
 244         while (1) {
 245                 unsigned long this_ping = cfs_time_current();
 246                 struct l_wait_info lwi;
 247                 long time_to_next_wake;
 248                 struct timeout_item *item;
 249                 struct list_head *iter;
 250
 251                 mutex_lock(&pinger_mutex);
 252                 list_for_each_entry(item, &timeout_list, ti_chain) {
 253                         item->ti_cb(item, item->ti_cb_data);
 254                 }
 255                 list_for_each(iter, &pinger_imports) {
 256                         struct obd_import *imp =
 257                                 list_entry(iter, struct obd_import,
 258                                                imp_pinger_chain);
 259
 260                         ptlrpc_pinger_process_import(imp, this_ping);
 261                         /* obd_timeout might have changed */
 262                         if (imp->imp_pingable && imp->imp_next_ping &&
 263                             cfs_time_after(imp->imp_next_ping,
 264                                            cfs_time_add(this_ping,
 265                                                         cfs_time_seconds(PING_INTERVAL))))
 266                                 ptlrpc_update_next_ping(imp, 0);
 267                 }
 268                 mutex_unlock(&pinger_mutex);
 269
 270                 /* Wait until the next ping time, or until we're stopped. */
 271                 time_to_next_wake = pinger_check_timeout(this_ping);
 272                 /* The ping sent by ptlrpc_send_rpc may get sent out
 273                    say .01 second after this.
 274                    ptlrpc_pinger_sending_on_import will then set the
 275                    next ping time to next_ping + .01 sec, which means
 276                    we will SKIP the next ping at next_ping, and the
 277                    ping will get sent 2 timeouts from now!  Beware. */
 278                 CDEBUG(D_INFO, "next wakeup in " CFS_DURATION_T " (%ld)\n",
 279                        time_to_next_wake,
 280                        cfs_time_add(this_ping,
 281                                     cfs_time_seconds(PING_INTERVAL)));
 282                 if (time_to_next_wake > 0) {
 283                         lwi = LWI_TIMEOUT(max_t(long, time_to_next_wake,
 284                                                 cfs_time_seconds(1)),
 285                                           NULL, NULL);
 286                         l_wait_event(thread->t_ctl_waitq,
 287                                      thread_is_stopping(thread) ||
 288                                      thread_is_event(thread),
 289                                      &lwi);
 290                         if (thread_test_and_clear_flags(thread, SVC_STOPPING))
 291                                 break;
 292                         /* woken after adding import to reset timer */
 293                         thread_test_and_clear_flags(thread, SVC_EVENT);
 294                 }
 295         }
 296
 297         thread_set_flags(thread, SVC_STOPPED);
 298         wake_up(&thread->t_ctl_waitq);
 299
 300         CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
 301         return 0;
 302 }
 303
 304 static struct ptlrpc_thread pinger_thread;
 305
 306 int ptlrpc_start_pinger(void)
 307 {
 308         struct l_wait_info lwi = { 0 };
 309         int rc;
 310
 311         if (!thread_is_init(&pinger_thread) &&
 312             !thread_is_stopped(&pinger_thread))
 313                 return -EALREADY;
 314
 315         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 316
 317         strcpy(pinger_thread.t_name, "ll_ping");
 318
 319         rc = PTR_ERR(kthread_run(ptlrpc_pinger_main, &pinger_thread,
 320                                  "%s", pinger_thread.t_name));
 321         if (IS_ERR_VALUE(rc)) {
 322                 CERROR("cannot start thread: %d\n", rc);
 323                 return rc;
 324         }
 325         l_wait_event(pinger_thread.t_ctl_waitq,
 326                      thread_is_running(&pinger_thread), &lwi);
 327
 328         if (suppress_pings)
 329                 CWARN("Pings will be suppressed at the request of the administrator.  The configuration shall meet the additional requirements described in the manual.  (Search for the \"suppress_pings\" kernel module parameter.)\n");
 330
 331         return 0;
 332 }
 333
 334 int ptlrpc_pinger_remove_timeouts(void);
 335
 336 int ptlrpc_stop_pinger(void)
 337 {
 338         struct l_wait_info lwi = { 0 };
 339         int rc = 0;
 340
 341         if (thread_is_init(&pinger_thread) ||
 342             thread_is_stopped(&pinger_thread))
 343                 return -EALREADY;
 344
 345         ptlrpc_pinger_remove_timeouts();
 346         thread_set_flags(&pinger_thread, SVC_STOPPING);
 347         wake_up(&pinger_thread.t_ctl_waitq);
 348
 349         l_wait_event(pinger_thread.t_ctl_waitq,
 350                      thread_is_stopped(&pinger_thread), &lwi);
 351
 352         return rc;
 353 }
 354
 355 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 356 {
 357         ptlrpc_update_next_ping(imp, 0);
 358 }
 359 EXPORT_SYMBOL(ptlrpc_pinger_sending_on_import);
 360
 361 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 362 {
 363         ptlrpc_update_next_ping(imp, 1);
 364         assert_spin_locked(&imp->imp_lock);
 365         /*
 366          * Avoid reading stale imp_connect_data.  When not sure if pings are
 367          * expected or not on next connection, we assume they are not and force
 368          * one anyway to guarantee the chance of updating
 369          * imp_peer_committed_transno.
 370          */
 371         if (imp->imp_state != LUSTRE_IMP_FULL ||
 372             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 373                 imp->imp_force_next_verify = 1;
 374 }
 375
 376 int ptlrpc_pinger_add_import(struct obd_import *imp)
 377 {
 378         if (!list_empty(&imp->imp_pinger_chain))
 379                 return -EALREADY;
 380
 381         mutex_lock(&pinger_mutex);
 382         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 383                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 384         /* if we add to pinger we want recovery on this import */
 385         imp->imp_obd->obd_no_recov = 0;
 386         ptlrpc_update_next_ping(imp, 0);
 387         /* XXX sort, blah blah */
 388         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 389         class_import_get(imp);
 390
 391         ptlrpc_pinger_wake_up();
 392         mutex_unlock(&pinger_mutex);
 393
 394         return 0;
 395 }
 396 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 397
 398 int ptlrpc_pinger_del_import(struct obd_import *imp)
 399 {
 400         if (list_empty(&imp->imp_pinger_chain))
 401                 return -ENOENT;
 402
 403         mutex_lock(&pinger_mutex);
 404         list_del_init(&imp->imp_pinger_chain);
 405         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 406                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 407         /* if we remove from pinger we don't want recovery on this import */
 408         imp->imp_obd->obd_no_recov = 1;
 409         class_import_put(imp);
 410         mutex_unlock(&pinger_mutex);
 411         return 0;
 412 }
 413 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 414
 415 /**
 416  * Register a timeout callback to the pinger list, and the callback will
 417  * be called when timeout happens.
 418  */
 419 static struct timeout_item *ptlrpc_new_timeout(int time,
 420         enum timeout_event event, timeout_cb_t cb, void *data)
 421 {
 422         struct timeout_item *ti;
 423
 424         ti = kzalloc(sizeof(*ti), GFP_NOFS);
 425         if (!ti)
 426                 return NULL;
 427
 428         INIT_LIST_HEAD(&ti->ti_obd_list);
 429         INIT_LIST_HEAD(&ti->ti_chain);
 430         ti->ti_timeout = time;
 431         ti->ti_event = event;
 432         ti->ti_cb = cb;
 433         ti->ti_cb_data = data;
 434
 435         return ti;
 436 }
 437
 438 /**
 439  * Register timeout event on the pinger thread.
 440  * Note: the timeout list is an sorted list with increased timeout value.
 441  */
 442 static struct timeout_item*
 443 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 444                                timeout_cb_t cb, void *data)
 445 {
 446         struct timeout_item *item, *tmp;
 447
 448         LASSERT(mutex_is_locked(&pinger_mutex));
 449
 450         list_for_each_entry(item, &timeout_list, ti_chain)
 451                 if (item->ti_event == event)
 452                         goto out;
 453
 454         item = ptlrpc_new_timeout(time, event, cb, data);
 455         if (item) {
 456                 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 457                         if (tmp->ti_timeout < time) {
 458                                 list_add(&item->ti_chain, &tmp->ti_chain);
 459                                 goto out;
 460                         }
 461                 }
 462                 list_add(&item->ti_chain, &timeout_list);
 463         }
 464 out:
 465         return item;
 466 }
 467
 468 /* Add a client_obd to the timeout event list, when timeout(@time)
 469  * happens, the callback(@cb) will be called.
 470  */
 471 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 472                               timeout_cb_t cb, void *data,
 473                               struct list_head *obd_list)
 474 {
 475         struct timeout_item *ti;
 476
 477         mutex_lock(&pinger_mutex);
 478         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 479         if (!ti) {
 480                 mutex_unlock(&pinger_mutex);
 481                 return -EINVAL;
 482         }
 483         list_add(obd_list, &ti->ti_obd_list);
 484         mutex_unlock(&pinger_mutex);
 485         return 0;
 486 }
 487 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
 488
 489 int ptlrpc_del_timeout_client(struct list_head *obd_list,
 490                               enum timeout_event event)
 491 {
 492         struct timeout_item *ti = NULL, *item;
 493
 494         if (list_empty(obd_list))
 495                 return 0;
 496         mutex_lock(&pinger_mutex);
 497         list_del_init(obd_list);
 498         /**
 499          * If there are no obd attached to the timeout event
 500          * list, remove this timeout event from the pinger
 501          */
 502         list_for_each_entry(item, &timeout_list, ti_chain) {
 503                 if (item->ti_event == event) {
 504                         ti = item;
 505                         break;
 506                 }
 507         }
 508         LASSERTF(ti != NULL, "ti is NULL !\n");
 509         if (list_empty(&ti->ti_obd_list)) {
 510                 list_del(&ti->ti_chain);
 511                 kfree(ti);
 512         }
 513         mutex_unlock(&pinger_mutex);
 514         return 0;
 515 }
 516 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 517
 518 int ptlrpc_pinger_remove_timeouts(void)
 519 {
 520         struct timeout_item *item, *tmp;
 521
 522         mutex_lock(&pinger_mutex);
 523         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 524                 LASSERT(list_empty(&item->ti_obd_list));
 525                 list_del(&item->ti_chain);
 526                 kfree(item);
 527         }
 528         mutex_unlock(&pinger_mutex);
 529         return 0;
 530 }
 531
 532 void ptlrpc_pinger_wake_up(void)
 533 {
 534         thread_add_flags(&pinger_thread, SVC_EVENT);
 535         wake_up(&pinger_thread.t_ctl_waitq);
 536 }