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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/ptlrpc/pinger.c
  33  *
  34  * Portal-RPC reconnection and replay operations, for use in recovery.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  39 #include "../include/obd_support.h"
  40 #include "../include/obd_class.h"
  41 #include "ptlrpc_internal.h"
  42
  43 struct mutex pinger_mutex;
  44 static LIST_HEAD(pinger_imports);
  45 static struct list_head timeout_list = LIST_HEAD_INIT(timeout_list);
  46
  47 struct ptlrpc_request *
  48 ptlrpc_prep_ping(struct obd_import *imp)
  49 {
  50         struct ptlrpc_request *req;
  51
  52         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  53                                         LUSTRE_OBD_VERSION, OBD_PING);
  54         if (req) {
  55                 ptlrpc_request_set_replen(req);
  56                 req->rq_no_resend = 1;
  57                 req->rq_no_delay = 1;
  58         }
  59         return req;
  60 }
  61
  62 int ptlrpc_obd_ping(struct obd_device *obd)
  63 {
  64         int rc;
  65         struct ptlrpc_request *req;
  66
  67         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  68         if (!req)
  69                 return -ENOMEM;
  70
  71         req->rq_send_state = LUSTRE_IMP_FULL;
  72
  73         rc = ptlrpc_queue_wait(req);
  74
  75         ptlrpc_req_finished(req);
  76
  77         return rc;
  78 }
  79 EXPORT_SYMBOL(ptlrpc_obd_ping);
  80
  81 static int ptlrpc_ping(struct obd_import *imp)
  82 {
  83         struct ptlrpc_request *req;
  84
  85         req = ptlrpc_prep_ping(imp);
  86         if (!req) {
  87                 CERROR("OOM trying to ping %s->%s\n",
  88                        imp->imp_obd->obd_uuid.uuid,
  89                        obd2cli_tgt(imp->imp_obd));
  90                 return -ENOMEM;
  91         }
  92
  93         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
  94                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
  95         ptlrpcd_add_req(req);
  96
  97         return 0;
  98 }
  99
 100 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 101 {
 102         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 103
 104         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 105                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 106                                   AT_OFF ? 0 :
 107                                   at_get(&imp->imp_at.iat_net_latency));
 108                 time = min(time, dtime);
 109         }
 110         imp->imp_next_ping = cfs_time_shift(time);
 111 }
 112
 113 static inline int imp_is_deactive(struct obd_import *imp)
 114 {
 115         return (imp->imp_deactive ||
 116                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 117 }
 118
 119 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 120 {
 121         if (imp->imp_server_timeout)
 122                 return cfs_time_shift(obd_timeout / 2);
 123         else
 124                 return cfs_time_shift(obd_timeout);
 125 }
 126
 127 static long pinger_check_timeout(unsigned long time)
 128 {
 129         struct timeout_item *item;
 130         unsigned long timeout = PING_INTERVAL;
 131
 132         /* The timeout list is a increase order sorted list */
 133         mutex_lock(&pinger_mutex);
 134         list_for_each_entry(item, &timeout_list, ti_chain) {
 135                 int ti_timeout = item->ti_timeout;
 136
 137                 if (timeout > ti_timeout)
 138                         timeout = ti_timeout;
 139                 break;
 140         }
 141         mutex_unlock(&pinger_mutex);
 142
 143         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 144                                          cfs_time_current());
 145 }
 146
 147 static bool ir_up;
 148
 149 void ptlrpc_pinger_ir_up(void)
 150 {
 151         CDEBUG(D_HA, "IR up\n");
 152         ir_up = true;
 153 }
 154 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 155
 156 void ptlrpc_pinger_ir_down(void)
 157 {
 158         CDEBUG(D_HA, "IR down\n");
 159         ir_up = false;
 160 }
 161 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 162
 163 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 164                                          unsigned long this_ping)
 165 {
 166         int level;
 167         int force;
 168         int force_next;
 169         int suppress;
 170
 171         spin_lock(&imp->imp_lock);
 172
 173         level = imp->imp_state;
 174         force = imp->imp_force_verify;
 175         force_next = imp->imp_force_next_verify;
 176         /*
 177          * This will be used below only if the import is "FULL".
 178          */
 179         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 180
 181         imp->imp_force_verify = 0;
 182
 183         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
 184             !force) {
 185                 spin_unlock(&imp->imp_lock);
 186                 return;
 187         }
 188
 189         imp->imp_force_next_verify = 0;
 190
 191         spin_unlock(&imp->imp_lock);
 192
 193         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",
 194                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 195                ptlrpc_import_state_name(level), level, force, force_next,
 196                imp->imp_deactive, imp->imp_pingable, suppress);
 197
 198         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 199                 /* wait for a while before trying recovery again */
 200                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 201                 if (!imp->imp_no_pinger_recover)
 202                         ptlrpc_initiate_recovery(imp);
 203         } else if (level != LUSTRE_IMP_FULL ||
 204                    imp->imp_obd->obd_no_recov ||
 205                    imp_is_deactive(imp)) {
 206                 CDEBUG(D_HA, "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 207                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 208                        ptlrpc_import_state_name(level));
 209                 if (force) {
 210                         spin_lock(&imp->imp_lock);
 211                         imp->imp_force_verify = 1;
 212                         spin_unlock(&imp->imp_lock);
 213                 }
 214         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 215                 ptlrpc_ping(imp);
 216         }
 217 }
 218
 219 static int ptlrpc_pinger_main(void *arg)
 220 {
 221         struct ptlrpc_thread *thread = arg;
 222
 223         /* Record that the thread is running */
 224         thread_set_flags(thread, SVC_RUNNING);
 225         wake_up(&thread->t_ctl_waitq);
 226
 227         /* And now, loop forever, pinging as needed. */
 228         while (1) {
 229                 unsigned long this_ping = cfs_time_current();
 230                 struct l_wait_info lwi;
 231                 long time_to_next_wake;
 232                 struct timeout_item *item;
 233                 struct list_head *iter;
 234
 235                 mutex_lock(&pinger_mutex);
 236                 list_for_each_entry(item, &timeout_list, ti_chain) {
 237                         item->ti_cb(item, item->ti_cb_data);
 238                 }
 239                 list_for_each(iter, &pinger_imports) {
 240                         struct obd_import *imp =
 241                                 list_entry(iter, struct obd_import,
 242                                            imp_pinger_chain);
 243
 244                         ptlrpc_pinger_process_import(imp, this_ping);
 245                         /* obd_timeout might have changed */
 246                         if (imp->imp_pingable && imp->imp_next_ping &&
 247                             cfs_time_after(imp->imp_next_ping,
 248                                            cfs_time_add(this_ping,
 249                                                         cfs_time_seconds(PING_INTERVAL))))
 250                                 ptlrpc_update_next_ping(imp, 0);
 251                 }
 252                 mutex_unlock(&pinger_mutex);
 253
 254                 /* Wait until the next ping time, or until we're stopped. */
 255                 time_to_next_wake = pinger_check_timeout(this_ping);
 256                 /* The ping sent by ptlrpc_send_rpc may get sent out
 257                  * say .01 second after this.
 258                  * ptlrpc_pinger_sending_on_import will then set the
 259                  * next ping time to next_ping + .01 sec, which means
 260                  * we will SKIP the next ping at next_ping, and the
 261                  * ping will get sent 2 timeouts from now!  Beware.
 262                  */
 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         struct task_struct *task;
 295         int rc;
 296
 297         if (!thread_is_init(&pinger_thread) &&
 298             !thread_is_stopped(&pinger_thread))
 299                 return -EALREADY;
 300
 301         init_waitqueue_head(&pinger_thread.t_ctl_waitq);
 302
 303         strcpy(pinger_thread.t_name, "ll_ping");
 304
 305         task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
 306                            pinger_thread.t_name);
 307         if (IS_ERR(task)) {
 308                 rc = PTR_ERR(task);
 309                 CERROR("cannot start pinger thread: rc = %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
 344 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 345 {
 346         ptlrpc_update_next_ping(imp, 1);
 347         assert_spin_locked(&imp->imp_lock);
 348         /*
 349          * Avoid reading stale imp_connect_data.  When not sure if pings are
 350          * expected or not on next connection, we assume they are not and force
 351          * one anyway to guarantee the chance of updating
 352          * imp_peer_committed_transno.
 353          */
 354         if (imp->imp_state != LUSTRE_IMP_FULL ||
 355             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 356                 imp->imp_force_next_verify = 1;
 357 }
 358
 359 int ptlrpc_pinger_add_import(struct obd_import *imp)
 360 {
 361         if (!list_empty(&imp->imp_pinger_chain))
 362                 return -EALREADY;
 363
 364         mutex_lock(&pinger_mutex);
 365         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 366                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 367         /* if we add to pinger we want recovery on this import */
 368         imp->imp_obd->obd_no_recov = 0;
 369         ptlrpc_update_next_ping(imp, 0);
 370         /* XXX sort, blah blah */
 371         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 372         class_import_get(imp);
 373
 374         ptlrpc_pinger_wake_up();
 375         mutex_unlock(&pinger_mutex);
 376
 377         return 0;
 378 }
 379 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 380
 381 int ptlrpc_pinger_del_import(struct obd_import *imp)
 382 {
 383         if (list_empty(&imp->imp_pinger_chain))
 384                 return -ENOENT;
 385
 386         mutex_lock(&pinger_mutex);
 387         list_del_init(&imp->imp_pinger_chain);
 388         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 389                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 390         /* if we remove from pinger we don't want recovery on this import */
 391         imp->imp_obd->obd_no_recov = 1;
 392         class_import_put(imp);
 393         mutex_unlock(&pinger_mutex);
 394         return 0;
 395 }
 396 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 397
 398 /**
 399  * Register a timeout callback to the pinger list, and the callback will
 400  * be called when timeout happens.
 401  */
 402 static struct timeout_item *ptlrpc_new_timeout(int time,
 403                                                enum timeout_event event,
 404                                                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         if (list_empty(&ti->ti_obd_list)) {
 493                 list_del(&ti->ti_chain);
 494                 kfree(ti);
 495         }
 496         mutex_unlock(&pinger_mutex);
 497         return 0;
 498 }
 499 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
 500
 501 static int ptlrpc_pinger_remove_timeouts(void)
 502 {
 503         struct timeout_item *item, *tmp;
 504
 505         mutex_lock(&pinger_mutex);
 506         list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 507                 LASSERT(list_empty(&item->ti_obd_list));
 508                 list_del(&item->ti_chain);
 509                 kfree(item);
 510         }
 511         mutex_unlock(&pinger_mutex);
 512         return 0;
 513 }
 514
 515 void ptlrpc_pinger_wake_up(void)
 516 {
 517         thread_add_flags(&pinger_thread, SVC_EVENT);
 518         wake_up(&pinger_thread.t_ctl_waitq);
 519 }