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  *
  31  * lustre/ptlrpc/pinger.c
  32  *
  33  * Portal-RPC reconnection and replay operations, for use in recovery.
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_RPC
  37
  38 #include <linux/fs_struct.h>
  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         struct obd_import *imp;
  78
  79         ENTRY;
  80
  81         with_imp_locked(obd, imp, rc) {
  82                 req = ptlrpc_prep_ping(imp);
  83                 if (!req) {
  84                         rc = -ENOMEM;
  85                         continue;
  86                 }
  87                 req->rq_send_state = LUSTRE_IMP_FULL;
  88                 rc = ptlrpc_queue_wait(req);
  89                 ptlrpc_req_finished(req);
  90         }
  91         RETURN(rc);
  92 }
  93 EXPORT_SYMBOL(ptlrpc_obd_ping);
  94
  95 static bool ptlrpc_check_import_is_idle(struct obd_import *imp)
  96 {
  97         struct ldlm_namespace *ns = imp->imp_obd->obd_namespace;
  98         time64_t now;
  99
 100         if (!imp->imp_idle_timeout)
 101                 return false;
 102
 103         if (atomic_read(&imp->imp_reqs) > 0)
 104                 return false;
 105
 106         /* any lock increases ns_bref being a resource holder */
 107         if (ns && atomic_read(&ns->ns_bref) > 0)
 108                 return false;
 109
 110         now = ktime_get_real_seconds();
 111         if (now - imp->imp_last_reply_time < imp->imp_idle_timeout)
 112                 return false;
 113
 114         return true;
 115 }
 116
 117 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 118 {
 119 #ifdef CONFIG_LUSTRE_FS_PINGER
 120         time64_t time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 121
 122         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 123                 time64_t dtime = max_t(time64_t, CONNECTION_SWITCH_MIN,
 124                                        AT_OFF ? 0 :
 125                                        at_get(&imp->imp_at.iat_net_latency));
 126                 time = min(time, dtime);
 127         }
 128         imp->imp_next_ping = ktime_get_seconds() + time;
 129 #endif /* CONFIG_LUSTRE_FS_PINGER */
 130 }
 131
 132 static int ptlrpc_ping(struct obd_import *imp)
 133 {
 134         struct ptlrpc_request *req;
 135
 136         ENTRY;
 137
 138         if (ptlrpc_check_import_is_idle(imp) &&
 139             ptlrpc_disconnect_and_idle_import(imp) == 1)
 140                         RETURN(0);
 141
 142         req = ptlrpc_prep_ping(imp);
 143         if (!req) {
 144                 CERROR("OOM trying to ping %s->%s\n",
 145                        imp->imp_obd->obd_uuid.uuid,
 146                        obd2cli_tgt(imp->imp_obd));
 147                 RETURN(-ENOMEM);
 148         }
 149
 150         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 151                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 152         /* Updating imp_next_ping early, it allows pinger_check_timeout to
 153          * see an actual time for next awake. request_out_callback update
 154          * happens at another thread, and ptlrpc_pinger_main may sleep
 155          * already.
 156          */
 157         ptlrpc_update_next_ping(imp, 0);
 158         ptlrpcd_add_req(req);
 159
 160         RETURN(0);
 161 }
 162
 163 void ptlrpc_ping_import_soon(struct obd_import *imp)
 164 {
 165         imp->imp_next_ping = ktime_get_seconds();
 166 }
 167
 168 static inline int imp_is_deactive(struct obd_import *imp)
 169 {
 170         return imp->imp_deactive ||
 171                CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
 172 }
 173
 174 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
 175 {
 176         return ktime_get_seconds() + INITIAL_CONNECT_TIMEOUT;
 177 }
 178
 179 static timeout_t pinger_check_timeout(time64_t time)
 180 {
 181         timeout_t timeout = PING_INTERVAL;
 182         timeout_t next_timeout;
 183         time64_t now;
 184         struct list_head *iter;
 185         struct obd_import *imp;
 186
 187         mutex_lock(&pinger_mutex);
 188         now = ktime_get_seconds();
 189         /* Process imports to find a nearest next ping */
 190         list_for_each(iter, &pinger_imports) {
 191                 imp = list_entry(iter, struct obd_import, imp_pinger_chain);
 192                 if (!imp->imp_pingable || imp->imp_next_ping < now)
 193                         continue;
 194                 next_timeout = imp->imp_next_ping - now;
 195                 /* make sure imp_next_ping in the future from time */
 196                 if (next_timeout > (now - time) && timeout > next_timeout)
 197                         timeout = next_timeout;
 198         }
 199         mutex_unlock(&pinger_mutex);
 200
 201         return timeout - (now - time);
 202 }
 203
 204 static bool ir_up;
 205
 206 void ptlrpc_pinger_ir_up(void)
 207 {
 208         CDEBUG(D_HA, "IR up\n");
 209         ir_up = true;
 210 }
 211 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 212
 213 void ptlrpc_pinger_ir_down(void)
 214 {
 215         CDEBUG(D_HA, "IR down\n");
 216         ir_up = false;
 217 }
 218 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 219
 220 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 221                                          time64_t this_ping)
 222 {
 223         int level;
 224         int force;
 225         int force_next;
 226         int suppress;
 227
 228         spin_lock(&imp->imp_lock);
 229
 230         level = imp->imp_state;
 231         force = imp->imp_force_verify;
 232         force_next = imp->imp_force_next_verify;
 233         /*
 234          * This will be used below only if the import is "FULL".
 235          */
 236         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 237
 238         imp->imp_force_verify = 0;
 239
 240         if (imp->imp_next_ping - 5 >= this_ping && !force) {
 241                 spin_unlock(&imp->imp_lock);
 242                 return;
 243         }
 244
 245         imp->imp_force_next_verify = 0;
 246
 247         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 248                "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
 249                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 250                ptlrpc_import_state_name(level), level, force, force_next,
 251                imp->imp_deactive, imp->imp_pingable, suppress);
 252
 253         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 254                 /* wait for a while before trying recovery again */
 255                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 256                 spin_unlock(&imp->imp_lock);
 257                 if (!imp->imp_no_pinger_recover ||
 258                     imp->imp_connect_error == -EAGAIN)
 259                         ptlrpc_initiate_recovery(imp);
 260         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
 261                    imp_is_deactive(imp)) {
 262                 CDEBUG(D_HA,
 263                        "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 264                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 265                        ptlrpc_import_state_name(level));
 266                 if (force)
 267                         imp->imp_force_verify = 1;
 268                 spin_unlock(&imp->imp_lock);
 269         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 270                 spin_unlock(&imp->imp_lock);
 271                 ptlrpc_ping(imp);
 272         } else {
 273                 spin_unlock(&imp->imp_lock);
 274         }
 275 }
 276
 277 static struct workqueue_struct *pinger_wq;
 278 static void ptlrpc_pinger_main(struct work_struct *ws);
 279 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
 280
 281 static void ptlrpc_pinger_main(struct work_struct *ws)
 282 {
 283         time64_t this_ping, time_after_ping;
 284         timeout_t time_to_next_wake;
 285         struct obd_import *imp;
 286
 287         do {
 288                 this_ping = ktime_get_seconds();
 289
 290                 mutex_lock(&pinger_mutex);
 291
 292                 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 293                         ptlrpc_pinger_process_import(imp, this_ping);
 294                         /* obd_timeout might have changed */
 295                         if (imp->imp_pingable && imp->imp_next_ping &&
 296                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 297                                 ptlrpc_update_next_ping(imp, 0);
 298                 }
 299                 mutex_unlock(&pinger_mutex);
 300
 301                 time_after_ping = ktime_get_seconds();
 302                 /* update memory usage info */
 303                 obd_update_maxusage();
 304
 305                 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
 306                         CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
 307                                this_ping, time_after_ping, ktime_get_seconds());
 308
 309                 /* Wait until the next ping time, or until we're stopped. */
 310                 time_to_next_wake = pinger_check_timeout(this_ping);
 311                 /*
 312                  * The ping sent by ptlrpc_send_rpc may get sent out
 313                  * say .01 second after this.
 314                  * ptlrpc_pinger_sending_on_import will then set the
 315                  * next ping time to next_ping + .01 sec, which means
 316                  * we will SKIP the next ping at next_ping, and the
 317                  * ping will get sent 2 timeouts from now!  Beware.
 318                  */
 319                 CDEBUG(D_INFO, "next wakeup in %d (%lld)\n",
 320                        time_to_next_wake, this_ping + PING_INTERVAL);
 321         } while (time_to_next_wake <= 0);
 322
 323         queue_delayed_work(pinger_wq, &ping_work,
 324                            cfs_time_seconds(max(time_to_next_wake, 1)));
 325 }
 326
 327 int ptlrpc_start_pinger(void)
 328 {
 329 #ifdef CONFIG_LUSTRE_FS_PINGER
 330         if (pinger_wq)
 331                 return -EALREADY;
 332
 333         pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
 334                                            0, CFS_CPT_ANY, 1);
 335         if (IS_ERR(pinger_wq)) {
 336                 CERROR("cannot start pinger workqueue\n");
 337                 return PTR_ERR(pinger_wq);
 338         }
 339
 340         queue_delayed_work(pinger_wq, &ping_work, 0);
 341
 342         if (suppress_pings)
 343                 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");
 344 #endif
 345         return 0;
 346 }
 347
 348 int ptlrpc_stop_pinger(void)
 349 {
 350 #ifdef CONFIG_LUSTRE_FS_PINGER
 351         if (!pinger_wq)
 352                 return -EALREADY;
 353
 354         cancel_delayed_work_sync(&ping_work);
 355         destroy_workqueue(pinger_wq);
 356         pinger_wq = NULL;
 357 #endif
 358         return 0;
 359 }
 360
 361 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 362 {
 363         ptlrpc_update_next_ping(imp, 0);
 364 }
 365
 366 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 367 {
 368         ptlrpc_update_next_ping(imp, 1);
 369         assert_spin_locked(&imp->imp_lock);
 370         /*
 371          * Avoid reading stale imp_connect_data.  When not sure if pings are
 372          * expected or not on next connection, we assume they are not and force
 373          * one anyway to guarantee the chance of updating
 374          * imp_peer_committed_transno.
 375          */
 376         if (imp->imp_state != LUSTRE_IMP_FULL ||
 377             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 378                 imp->imp_force_next_verify = 1;
 379 }
 380
 381 int ptlrpc_pinger_add_import(struct obd_import *imp)
 382 {
 383         ENTRY;
 384         if (!list_empty(&imp->imp_pinger_chain))
 385                 RETURN(-EALREADY);
 386
 387         mutex_lock(&pinger_mutex);
 388         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 389                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 390         /* if we add to pinger we want recovery on this import */
 391         imp->imp_obd->obd_no_recov = 0;
 392         ptlrpc_update_next_ping(imp, 0);
 393         /* XXX sort, blah blah */
 394         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 395         class_import_get(imp);
 396
 397         ptlrpc_pinger_wake_up();
 398         mutex_unlock(&pinger_mutex);
 399
 400         RETURN(0);
 401 }
 402 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 403
 404 int ptlrpc_pinger_del_import(struct obd_import *imp)
 405 {
 406         ENTRY;
 407
 408         if (list_empty(&imp->imp_pinger_chain))
 409                 RETURN(-ENOENT);
 410
 411         mutex_lock(&pinger_mutex);
 412         list_del_init(&imp->imp_pinger_chain);
 413         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 414                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 415         /* if we remove from pinger we don't want recovery on this import */
 416         imp->imp_obd->obd_no_recov = 1;
 417         class_import_put(imp);
 418         mutex_unlock(&pinger_mutex);
 419         RETURN(0);
 420 }
 421 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 422
 423 void ptlrpc_pinger_wake_up(void)
 424 {
 425 #ifdef CONFIG_LUSTRE_FS_PINGER
 426         mod_delayed_work(pinger_wq, &ping_work, 0);
 427 #endif
 428 }
 429
 430 /* Ping evictor thread */
 431 #define PET_READY     1
 432 #define PET_TERMINATE 2
 433
 434 static int pet_refcount;
 435 static int pet_state;
 436 static wait_queue_head_t pet_waitq;
 437 static LIST_HEAD(pet_list);
 438 static DEFINE_SPINLOCK(pet_lock);
 439
 440 int ping_evictor_wake(struct obd_export *exp)
 441 {
 442         struct obd_device *obd;
 443
 444         spin_lock(&pet_lock);
 445         if (pet_state != PET_READY) {
 446                 /* eventually the new obd will call here again. */
 447                 spin_unlock(&pet_lock);
 448                 return 1;
 449         }
 450
 451         obd = class_exp2obd(exp);
 452         if (list_empty(&obd->obd_evict_list)) {
 453                 class_incref(obd, "evictor", obd);
 454                 list_add(&obd->obd_evict_list, &pet_list);
 455         }
 456         spin_unlock(&pet_lock);
 457
 458         wake_up(&pet_waitq);
 459         return 0;
 460 }
 461
 462 static int ping_evictor_main(void *arg)
 463 {
 464         struct obd_device *obd;
 465         struct obd_export *exp;
 466         time64_t expire_time;
 467
 468         ENTRY;
 469         unshare_fs_struct();
 470         CDEBUG(D_HA, "Starting Ping Evictor\n");
 471         pet_state = PET_READY;
 472         while (1) {
 473                 wait_event_idle(pet_waitq,
 474                                 (!list_empty(&pet_list)) ||
 475                                 (pet_state == PET_TERMINATE));
 476
 477                 /* loop until all obd's will be removed */
 478                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 479                         break;
 480
 481                 /*
 482                  * we only get here if pet_exp != NULL, and the end of this
 483                  * loop is the only place which sets it NULL again, so lock
 484                  * is not strictly necessary.
 485                  */
 486                 spin_lock(&pet_lock);
 487                 obd = list_first_entry(&pet_list, struct obd_device,
 488                                        obd_evict_list);
 489                 spin_unlock(&pet_lock);
 490
 491                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 492
 493                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 494                        obd->obd_name, expire_time);
 495
 496                 /*
 497                  * Exports can't be deleted out of the list while we hold
 498                  * the obd lock (class_unlink_export), which means we can't
 499                  * lose the last ref on the export.  If they've already been
 500                  * removed from the list, we won't find them here.
 501                  */
 502                 spin_lock(&obd->obd_dev_lock);
 503                 while (!list_empty(&obd->obd_exports_timed)) {
 504                         exp = list_first_entry(&obd->obd_exports_timed,
 505                                                struct obd_export,
 506                                                exp_obd_chain_timed);
 507                         if (expire_time > exp->exp_last_request_time) {
 508                                 struct obd_uuid *client_uuid;
 509
 510                                 class_export_get(exp);
 511                                 client_uuid = &exp->exp_client_uuid;
 512                                 spin_unlock(&obd->obd_dev_lock);
 513                                 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",
 514                                               obd->obd_name,
 515                                               obd_uuid2str(client_uuid),
 516                                               obd_export_nid2str(exp),
 517                                               ktime_get_real_seconds() -
 518                                               exp->exp_last_request_time,
 519                                               exp, ktime_get_real_seconds(),
 520                                               expire_time,
 521                                               exp->exp_last_request_time);
 522                                 CDEBUG(D_HA, "Last request was at %lld\n",
 523                                        exp->exp_last_request_time);
 524                                 class_fail_export(exp);
 525                                 class_export_put(exp);
 526                                 spin_lock(&obd->obd_dev_lock);
 527                         } else {
 528                                 /* List is sorted, so everyone below is ok */
 529                                 break;
 530                         }
 531                 }
 532                 spin_unlock(&obd->obd_dev_lock);
 533
 534                 spin_lock(&pet_lock);
 535                 list_del_init(&obd->obd_evict_list);
 536                 spin_unlock(&pet_lock);
 537
 538                 class_decref(obd, "evictor", obd);
 539         }
 540         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 541
 542         RETURN(0);
 543 }
 544
 545 void ping_evictor_start(void)
 546 {
 547         struct task_struct *task;
 548
 549         if (++pet_refcount > 1)
 550                 return;
 551
 552         init_waitqueue_head(&pet_waitq);
 553
 554         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 555         if (IS_ERR(task)) {
 556                 pet_refcount--;
 557                 CERROR("Cannot start ping evictor thread: %ld\n",
 558                         PTR_ERR(task));
 559         }
 560 }
 561 EXPORT_SYMBOL(ping_evictor_start);
 562
 563 void ping_evictor_stop(void)
 564 {
 565         if (--pet_refcount > 0)
 566                 return;
 567
 568         pet_state = PET_TERMINATE;
 569         wake_up(&pet_waitq);
 570 }
 571 EXPORT_SYMBOL(ping_evictor_stop);