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