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 <linux/kthread.h>
  40 #include <linux/workqueue.h>
  41 #include <obd_support.h>
  42 #include <obd_class.h>
  43 #include "ptlrpc_internal.h"
  44
  45 static int suppress_pings;
  46 module_param(suppress_pings, int, 0644);
  47 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
  48
  49 struct mutex pinger_mutex;
  50 static struct list_head pinger_imports =
  51                 LIST_HEAD_INIT(pinger_imports);
  52
  53 int ptlrpc_pinger_suppress_pings(void)
  54 {
  55         return suppress_pings;
  56 }
  57 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  58
  59 struct ptlrpc_request *
  60 ptlrpc_prep_ping(struct obd_import *imp)
  61 {
  62         struct ptlrpc_request *req;
  63
  64         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  65                                         LUSTRE_OBD_VERSION, OBD_PING);
  66         if (req) {
  67                 ptlrpc_request_set_replen(req);
  68                 req->rq_no_resend = req->rq_no_delay = 1;
  69         }
  70         return req;
  71 }
  72
  73 int ptlrpc_obd_ping(struct obd_device *obd)
  74 {
  75         int rc;
  76         struct ptlrpc_request *req;
  77
  78         ENTRY;
  79
  80         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  81         if (!req)
  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 bool ptlrpc_check_import_is_idle(struct obd_import *imp)
  95 {
  96         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
  97         time64_t now;
  98
  99         if (!imp->imp_idle_timeout)
 100                 return false;
 101
 102         if (atomic_read(&imp->imp_reqs) > 0)
 103                 return false;
 104
 105         /* any lock increases ns_bref being a resource holder */
 106         if (ns && atomic_read(&ns->ns_bref) > 0)
 107                 return false;
 108
 109         now = ktime_get_real_seconds();
 110         if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
 111                 return false;
 112
 113         return true;
 114 }
 115
 116 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 117 {
 118 #ifdef CONFIG_LUSTRE_FS_PINGER
 119         time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 120
 121         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 122                 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
 123                                        AT_OFF ? 0 :
 124                                        at_get(&imp->imp_at.iat_net_latency));
 125                 time = min(time, dtime);
 126         }
 127         imp->imp_next_ping = ktime_get_seconds() + time;
 128 #endif /* CONFIG_LUSTRE_FS_PINGER */
 129 }
 130
 131 static int ptlrpc_ping(struct obd_import *imp)
 132 {
 133         struct ptlrpc_request *req;
 134
 135         ENTRY;
 136
 137         if (ptlrpc_check_import_is_idle(imp))
 138                 RETURN(ptlrpc_disconnect_and_idle_import(imp));
 139
 140         req = ptlrpc_prep_ping(imp);
 141         if (!req) {
 142                 CERROR("OOM trying to ping %s->%s\n",
 143                        imp->imp_obd->obd_uuid.uuid,
 144                        obd2cli_tgt(imp->imp_obd));
 145                 RETURN(-ENOMEM);
 146         }
 147
 148         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 149                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 150         /* Updating imp_next_ping early, it allows pinger_check_timeout to
 151          * see an actual time for next awake. request_out_callback update
 152          * happens at another thread, and ptlrpc_pinger_main may sleep
 153          * already.
 154          */
 155         ptlrpc_update_next_ping(imp, 0);
 156         ptlrpcd_add_req(req);
 157
 158         RETURN(0);
 159 }
 160
 161 void ptlrpc_ping_import_soon(struct obd_import *imp)
 162 {
 163         imp->imp_next_ping = ktime_get_seconds();
 164 }
 165
 166 static inline int imp_is_deactive(struct obd_import *imp)
 167 {
 168         return imp->imp_deactive ||
 169                OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
 170 }
 171
 172 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
 173 {
 174         return ktime_get_seconds() + INITIAL_CONNECT_TIMEOUT;
 175 }
 176
 177 static timeout_t pinger_check_timeout(time64_t time)
 178 {
 179         timeout_t timeout = PING_INTERVAL;
 180         timeout_t next_timeout;
 181         time64_t now;
 182         struct list_head *iter;
 183         struct obd_import *imp;
 184
 185         mutex_lock(&pinger_mutex);
 186         now = ktime_get_seconds();
 187         /* Process imports to find a nearest next ping */
 188         list_for_each(iter, &pinger_imports) {
 189                 imp = list_entry(iter, struct obd_import, imp_pinger_chain);
 190                 if (!imp->imp_pingable || imp->imp_next_ping < now)
 191                         continue;
 192                 next_timeout = imp->imp_next_ping - now;
 193                 /* make sure imp_next_ping in the future from time */
 194                 if (next_timeout > (now - time) && timeout > next_timeout)
 195                         timeout = next_timeout;
 196         }
 197         mutex_unlock(&pinger_mutex);
 198
 199         return timeout - (now - time);
 200 }
 201
 202 static bool ir_up;
 203
 204 void ptlrpc_pinger_ir_up(void)
 205 {
 206         CDEBUG(D_HA, "IR up\n");
 207         ir_up = true;
 208 }
 209 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 210
 211 void ptlrpc_pinger_ir_down(void)
 212 {
 213         CDEBUG(D_HA, "IR down\n");
 214         ir_up = false;
 215 }
 216 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 217
 218 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 219                                          time64_t this_ping)
 220 {
 221         int level;
 222         int force;
 223         int force_next;
 224         int suppress;
 225
 226         spin_lock(&imp->imp_lock);
 227
 228         level = imp->imp_state;
 229         force = imp->imp_force_verify;
 230         force_next = imp->imp_force_next_verify;
 231         /*
 232          * This will be used below only if the import is "FULL".
 233          */
 234         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 235
 236         imp->imp_force_verify = 0;
 237
 238         if (imp->imp_next_ping - 5 >= this_ping && !force) {
 239                 spin_unlock(&imp->imp_lock);
 240                 return;
 241         }
 242
 243         imp->imp_force_next_verify = 0;
 244
 245         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 246                "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
 247                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 248                ptlrpc_import_state_name(level), level, force, force_next,
 249                imp->imp_deactive, imp->imp_pingable, suppress);
 250
 251         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 252                 /* wait for a while before trying recovery again */
 253                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 254                 spin_unlock(&imp->imp_lock);
 255                 if (!imp->imp_no_pinger_recover ||
 256                     imp->imp_connect_error == -EAGAIN)
 257                         ptlrpc_initiate_recovery(imp);
 258         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
 259                    imp_is_deactive(imp)) {
 260                 CDEBUG(D_HA,
 261                        "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 262                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 263                        ptlrpc_import_state_name(level));
 264                 if (force)
 265                         imp->imp_force_verify = 1;
 266                 spin_unlock(&imp->imp_lock);
 267         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 268                 spin_unlock(&imp->imp_lock);
 269                 ptlrpc_ping(imp);
 270         } else {
 271                 spin_unlock(&imp->imp_lock);
 272         }
 273 }
 274
 275 static struct workqueue_struct *pinger_wq;
 276 static void ptlrpc_pinger_main(struct work_struct *ws);
 277 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
 278
 279 static void ptlrpc_pinger_main(struct work_struct *ws)
 280 {
 281         time64_t this_ping, time_after_ping;
 282         timeout_t time_to_next_wake;
 283         struct obd_import *imp;
 284         struct list_head *iter;
 285
 286         do {
 287                 this_ping = ktime_get_seconds();
 288
 289                 mutex_lock(&pinger_mutex);
 290
 291                 list_for_each(iter, &pinger_imports) {
 292                         imp = list_entry(iter, struct obd_import,
 293                                          imp_pinger_chain);
 294
 295                         ptlrpc_pinger_process_import(imp, this_ping);
 296                         /* obd_timeout might have changed */
 297                         if (imp->imp_pingable && imp->imp_next_ping &&
 298                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 299                                 ptlrpc_update_next_ping(imp, 0);
 300                 }
 301                 mutex_unlock(&pinger_mutex);
 302
 303                 time_after_ping = ktime_get_seconds();
 304                 /* update memory usage info */
 305                 obd_update_maxusage();
 306
 307                 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
 308                         CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
 309                                this_ping, time_after_ping, ktime_get_seconds());
 310
 311                 /* Wait until the next ping time, or until we're stopped. */
 312                 time_to_next_wake = pinger_check_timeout(this_ping);
 313                 /*
 314                  * The ping sent by ptlrpc_send_rpc may get sent out
 315                  * say .01 second after this.
 316                  * ptlrpc_pinger_sending_on_import will then set the
 317                  * next ping time to next_ping + .01 sec, which means
 318                  * we will SKIP the next ping at next_ping, and the
 319                  * ping will get sent 2 timeouts from now!  Beware.
 320                  */
 321                 CDEBUG(D_INFO, "next wakeup in %d (%lld)\n",
 322                        time_to_next_wake, this_ping + PING_INTERVAL);
 323         } while (time_to_next_wake <= 0);
 324
 325         queue_delayed_work(pinger_wq, &ping_work,
 326                            cfs_time_seconds(max(time_to_next_wake, 1)));
 327 }
 328
 329 int ptlrpc_start_pinger(void)
 330 {
 331 #ifdef CONFIG_LUSTRE_FS_PINGER
 332         if (pinger_wq)
 333                 return -EALREADY;
 334
 335         pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
 336                                            0, CFS_CPT_ANY, 1);
 337         if (IS_ERR(pinger_wq)) {
 338                 CERROR("cannot start pinger workqueue\n");
 339                 return PTR_ERR(pinger_wq);
 340         }
 341
 342         queue_delayed_work(pinger_wq, &ping_work, 0);
 343
 344         if (suppress_pings)
 345                 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");
 346 #endif
 347         return 0;
 348 }
 349
 350 int ptlrpc_stop_pinger(void)
 351 {
 352 #ifdef CONFIG_LUSTRE_FS_PINGER
 353         if (!pinger_wq)
 354                 return -EALREADY;
 355
 356         cancel_delayed_work_sync(&ping_work);
 357         destroy_workqueue(pinger_wq);
 358         pinger_wq = NULL;
 359 #endif
 360         return 0;
 361 }
 362
 363 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 364 {
 365         ptlrpc_update_next_ping(imp, 0);
 366 }
 367
 368 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 369 {
 370         ptlrpc_update_next_ping(imp, 1);
 371         assert_spin_locked(&imp->imp_lock);
 372         /*
 373          * Avoid reading stale imp_connect_data.  When not sure if pings are
 374          * expected or not on next connection, we assume they are not and force
 375          * one anyway to guarantee the chance of updating
 376          * imp_peer_committed_transno.
 377          */
 378         if (imp->imp_state != LUSTRE_IMP_FULL ||
 379             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 380                 imp->imp_force_next_verify = 1;
 381 }
 382
 383 int ptlrpc_pinger_add_import(struct obd_import *imp)
 384 {
 385         ENTRY;
 386         if (!list_empty(&imp->imp_pinger_chain))
 387                 RETURN(-EALREADY);
 388
 389         mutex_lock(&pinger_mutex);
 390         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 391                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 392         /* if we add to pinger we want recovery on this import */
 393         imp->imp_obd->obd_no_recov = 0;
 394         ptlrpc_update_next_ping(imp, 0);
 395         /* XXX sort, blah blah */
 396         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 397         class_import_get(imp);
 398
 399         ptlrpc_pinger_wake_up();
 400         mutex_unlock(&pinger_mutex);
 401
 402         RETURN(0);
 403 }
 404 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 405
 406 int ptlrpc_pinger_del_import(struct obd_import *imp)
 407 {
 408         ENTRY;
 409
 410         if (list_empty(&imp->imp_pinger_chain))
 411                 RETURN(-ENOENT);
 412
 413         mutex_lock(&pinger_mutex);
 414         list_del_init(&imp->imp_pinger_chain);
 415         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 416                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 417         /* if we remove from pinger we don't want recovery on this import */
 418         imp->imp_obd->obd_no_recov = 1;
 419         class_import_put(imp);
 420         mutex_unlock(&pinger_mutex);
 421         RETURN(0);
 422 }
 423 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 424
 425 void ptlrpc_pinger_wake_up(void)
 426 {
 427 #ifdef CONFIG_LUSTRE_FS_PINGER
 428         mod_delayed_work(pinger_wq, &ping_work, 0);
 429 #endif
 430 }
 431
 432 /* Ping evictor thread */
 433 #define PET_READY     1
 434 #define PET_TERMINATE 2
 435
 436 static int pet_refcount;
 437 static int pet_state;
 438 static wait_queue_head_t pet_waitq;
 439 static LIST_HEAD(pet_list);
 440 static DEFINE_SPINLOCK(pet_lock);
 441
 442 int ping_evictor_wake(struct obd_export *exp)
 443 {
 444         struct obd_device *obd;
 445
 446         spin_lock(&pet_lock);
 447         if (pet_state != PET_READY) {
 448                 /* eventually the new obd will call here again. */
 449                 spin_unlock(&pet_lock);
 450                 return 1;
 451         }
 452
 453         obd = class_exp2obd(exp);
 454         if (list_empty(&obd->obd_evict_list)) {
 455                 class_incref(obd, "evictor", obd);
 456                 list_add(&obd->obd_evict_list, &pet_list);
 457         }
 458         spin_unlock(&pet_lock);
 459
 460         wake_up(&pet_waitq);
 461         return 0;
 462 }
 463
 464 static int ping_evictor_main(void *arg)
 465 {
 466         struct obd_device *obd;
 467         struct obd_export *exp;
 468         time64_t expire_time;
 469
 470         ENTRY;
 471         CDEBUG(D_HA, "Starting Ping Evictor\n");
 472         pet_state = PET_READY;
 473         while (1) {
 474                 wait_event_idle(pet_waitq,
 475                                 (!list_empty(&pet_list)) ||
 476                                 (pet_state == PET_TERMINATE));
 477
 478                 /* loop until all obd's will be removed */
 479                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 480                         break;
 481
 482                 /*
 483                  * we only get here if pet_exp != NULL, and the end of this
 484                  * loop is the only place which sets it NULL again, so lock
 485                  * is not strictly necessary.
 486                  */
 487                 spin_lock(&pet_lock);
 488                 obd = list_entry(pet_list.next, struct obd_device,
 489                                  obd_evict_list);
 490                 spin_unlock(&pet_lock);
 491
 492                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 493
 494                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 495                        obd->obd_name, expire_time);
 496
 497                 /*
 498                  * Exports can't be deleted out of the list while we hold
 499                  * the obd lock (class_unlink_export), which means we can't
 500                  * lose the last ref on the export.  If they've already been
 501                  * removed from the list, we won't find them here.
 502                  */
 503                 spin_lock(&obd->obd_dev_lock);
 504                 while (!list_empty(&obd->obd_exports_timed)) {
 505                         exp = list_entry(obd->obd_exports_timed.next,
 506                                          struct obd_export,
 507                                          exp_obd_chain_timed);
 508                         if (expire_time > exp->exp_last_request_time) {
 509                                 struct obd_uuid *client_uuid;
 510
 511                                 class_export_get(exp);
 512                                 client_uuid = &exp->exp_client_uuid;
 513                                 spin_unlock(&obd->obd_dev_lock);
 514                                 LCONSOLE_WARN("%s: haven't heard from client %s (at %s) in %lld seconds. I think it's dead, and I am evicting it. exp %p, cur %lld expire %lld last %lld\n",
 515                                               obd->obd_name,
 516                                               obd_uuid2str(client_uuid),
 517                                               obd_export_nid2str(exp),
 518                                               ktime_get_real_seconds() -
 519                                               exp->exp_last_request_time,
 520                                               exp, ktime_get_real_seconds(),
 521                                               expire_time,
 522                                               exp->exp_last_request_time);
 523                                 CDEBUG(D_HA, "Last request was at %lld\n",
 524                                        exp->exp_last_request_time);
 525                                 class_fail_export(exp);
 526                                 class_export_put(exp);
 527                                 spin_lock(&obd->obd_dev_lock);
 528                         } else {
 529                                 /* List is sorted, so everyone below is ok */
 530                                 break;
 531                         }
 532                 }
 533                 spin_unlock(&obd->obd_dev_lock);
 534
 535                 spin_lock(&pet_lock);
 536                 list_del_init(&obd->obd_evict_list);
 537                 spin_unlock(&pet_lock);
 538
 539                 class_decref(obd, "evictor", obd);
 540         }
 541         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 542
 543         RETURN(0);
 544 }
 545
 546 void ping_evictor_start(void)
 547 {
 548         struct task_struct *task;
 549
 550         if (++pet_refcount > 1)
 551                 return;
 552
 553         init_waitqueue_head(&pet_waitq);
 554
 555         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 556         if (IS_ERR(task)) {
 557                 pet_refcount--;
 558                 CERROR("Cannot start ping evictor thread: %ld\n",
 559                         PTR_ERR(task));
 560         }
 561 }
 562 EXPORT_SYMBOL(ping_evictor_start);
 563
 564 void ping_evictor_stop(void)
 565 {
 566         if (--pet_refcount > 0)
 567                 return;
 568
 569         pet_state = PET_TERMINATE;
 570         wake_up(&pet_waitq);
 571 }
 572 EXPORT_SYMBOL(ping_evictor_stop);