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/kthread.h>
  39 #include <linux/workqueue.h>
  40 #include <obd_support.h>
  41 #include <obd_class.h>
  42 #include "ptlrpc_internal.h"
  43
  44 static int suppress_pings;
  45 module_param(suppress_pings, int, 0644);
  46 MODULE_PARM_DESC(suppress_pings, "Suppress pings");
  47
  48 struct mutex pinger_mutex;
  49 static struct list_head pinger_imports =
  50                 LIST_HEAD_INIT(pinger_imports);
  51
  52 int ptlrpc_pinger_suppress_pings(void)
  53 {
  54         return suppress_pings;
  55 }
  56 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
  57
  58 struct ptlrpc_request *
  59 ptlrpc_prep_ping(struct obd_import *imp)
  60 {
  61         struct ptlrpc_request *req;
  62
  63         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  64                                         LUSTRE_OBD_VERSION, OBD_PING);
  65         if (req) {
  66                 ptlrpc_request_set_replen(req);
  67                 req->rq_no_resend = req->rq_no_delay = 1;
  68         }
  69         return req;
  70 }
  71
  72 int ptlrpc_obd_ping(struct obd_device *obd)
  73 {
  74         int rc;
  75         struct ptlrpc_request *req;
  76         struct obd_import *imp;
  77
  78         ENTRY;
  79
  80         with_imp_locked(obd, imp, rc) {
  81                 req = ptlrpc_prep_ping(imp);
  82                 if (!req) {
  83                         rc = -ENOMEM;
  84                         continue;
  85                 }
  86                 req->rq_send_state = LUSTRE_IMP_FULL;
  87                 rc = ptlrpc_queue_wait(req);
  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
 285         do {
 286                 this_ping = ktime_get_seconds();
 287
 288                 mutex_lock(&pinger_mutex);
 289
 290                 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 291                         ptlrpc_pinger_process_import(imp, this_ping);
 292                         /* obd_timeout might have changed */
 293                         if (imp->imp_pingable && imp->imp_next_ping &&
 294                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 295                                 ptlrpc_update_next_ping(imp, 0);
 296                 }
 297                 mutex_unlock(&pinger_mutex);
 298
 299                 time_after_ping = ktime_get_seconds();
 300                 /* update memory usage info */
 301                 obd_update_maxusage();
 302
 303                 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
 304                         CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
 305                                this_ping, time_after_ping, ktime_get_seconds());
 306
 307                 /* Wait until the next ping time, or until we're stopped. */
 308                 time_to_next_wake = pinger_check_timeout(this_ping);
 309                 /*
 310                  * The ping sent by ptlrpc_send_rpc may get sent out
 311                  * say .01 second after this.
 312                  * ptlrpc_pinger_sending_on_import will then set the
 313                  * next ping time to next_ping + .01 sec, which means
 314                  * we will SKIP the next ping at next_ping, and the
 315                  * ping will get sent 2 timeouts from now!  Beware.
 316                  */
 317                 CDEBUG(D_INFO, "next wakeup in %d (%lld)\n",
 318                        time_to_next_wake, this_ping + PING_INTERVAL);
 319         } while (time_to_next_wake <= 0);
 320
 321         queue_delayed_work(pinger_wq, &ping_work,
 322                            cfs_time_seconds(max(time_to_next_wake, 1)));
 323 }
 324
 325 int ptlrpc_start_pinger(void)
 326 {
 327 #ifdef CONFIG_LUSTRE_FS_PINGER
 328         if (pinger_wq)
 329                 return -EALREADY;
 330
 331         pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
 332                                            0, CFS_CPT_ANY, 1);
 333         if (IS_ERR(pinger_wq)) {
 334                 CERROR("cannot start pinger workqueue\n");
 335                 return PTR_ERR(pinger_wq);
 336         }
 337
 338         queue_delayed_work(pinger_wq, &ping_work, 0);
 339
 340         if (suppress_pings)
 341                 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");
 342 #endif
 343         return 0;
 344 }
 345
 346 int ptlrpc_stop_pinger(void)
 347 {
 348 #ifdef CONFIG_LUSTRE_FS_PINGER
 349         if (!pinger_wq)
 350                 return -EALREADY;
 351
 352         cancel_delayed_work_sync(&ping_work);
 353         destroy_workqueue(pinger_wq);
 354         pinger_wq = NULL;
 355 #endif
 356         return 0;
 357 }
 358
 359 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 360 {
 361         ptlrpc_update_next_ping(imp, 0);
 362 }
 363
 364 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 365 {
 366         ptlrpc_update_next_ping(imp, 1);
 367         assert_spin_locked(&imp->imp_lock);
 368         /*
 369          * Avoid reading stale imp_connect_data.  When not sure if pings are
 370          * expected or not on next connection, we assume they are not and force
 371          * one anyway to guarantee the chance of updating
 372          * imp_peer_committed_transno.
 373          */
 374         if (imp->imp_state != LUSTRE_IMP_FULL ||
 375             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 376                 imp->imp_force_next_verify = 1;
 377 }
 378
 379 int ptlrpc_pinger_add_import(struct obd_import *imp)
 380 {
 381         ENTRY;
 382         if (!list_empty(&imp->imp_pinger_chain))
 383                 RETURN(-EALREADY);
 384
 385         mutex_lock(&pinger_mutex);
 386         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 387                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 388         /* if we add to pinger we want recovery on this import */
 389         imp->imp_obd->obd_no_recov = 0;
 390         ptlrpc_update_next_ping(imp, 0);
 391         /* XXX sort, blah blah */
 392         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 393         class_import_get(imp);
 394
 395         ptlrpc_pinger_wake_up();
 396         mutex_unlock(&pinger_mutex);
 397
 398         RETURN(0);
 399 }
 400 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 401
 402 int ptlrpc_pinger_del_import(struct obd_import *imp)
 403 {
 404         ENTRY;
 405
 406         if (list_empty(&imp->imp_pinger_chain))
 407                 RETURN(-ENOENT);
 408
 409         mutex_lock(&pinger_mutex);
 410         list_del_init(&imp->imp_pinger_chain);
 411         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 412                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 413         /* if we remove from pinger we don't want recovery on this import */
 414         imp->imp_obd->obd_no_recov = 1;
 415         class_import_put(imp);
 416         mutex_unlock(&pinger_mutex);
 417         RETURN(0);
 418 }
 419 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 420
 421 void ptlrpc_pinger_wake_up(void)
 422 {
 423 #ifdef CONFIG_LUSTRE_FS_PINGER
 424         mod_delayed_work(pinger_wq, &ping_work, 0);
 425 #endif
 426 }
 427
 428 /* Ping evictor thread */
 429 #define PET_READY     1
 430 #define PET_TERMINATE 2
 431
 432 static int pet_refcount;
 433 static int pet_state;
 434 static wait_queue_head_t pet_waitq;
 435 static LIST_HEAD(pet_list);
 436 static DEFINE_SPINLOCK(pet_lock);
 437
 438 int ping_evictor_wake(struct obd_export *exp)
 439 {
 440         struct obd_device *obd;
 441
 442         spin_lock(&pet_lock);
 443         if (pet_state != PET_READY) {
 444                 /* eventually the new obd will call here again. */
 445                 spin_unlock(&pet_lock);
 446                 return 1;
 447         }
 448
 449         obd = class_exp2obd(exp);
 450         if (list_empty(&obd->obd_evict_list)) {
 451                 class_incref(obd, "evictor", obd);
 452                 list_add(&obd->obd_evict_list, &pet_list);
 453         }
 454         spin_unlock(&pet_lock);
 455
 456         wake_up(&pet_waitq);
 457         return 0;
 458 }
 459
 460 static int ping_evictor_main(void *arg)
 461 {
 462         struct obd_device *obd;
 463         struct obd_export *exp;
 464         time64_t expire_time;
 465
 466         ENTRY;
 467         CDEBUG(D_HA, "Starting Ping Evictor\n");
 468         pet_state = PET_READY;
 469         while (1) {
 470                 wait_event_idle(pet_waitq,
 471                                 (!list_empty(&pet_list)) ||
 472                                 (pet_state == PET_TERMINATE));
 473
 474                 /* loop until all obd's will be removed */
 475                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 476                         break;
 477
 478                 /*
 479                  * we only get here if pet_exp != NULL, and the end of this
 480                  * loop is the only place which sets it NULL again, so lock
 481                  * is not strictly necessary.
 482                  */
 483                 spin_lock(&pet_lock);
 484                 obd = list_entry(pet_list.next, struct obd_device,
 485                                  obd_evict_list);
 486                 spin_unlock(&pet_lock);
 487
 488                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 489
 490                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 491                        obd->obd_name, expire_time);
 492
 493                 /*
 494                  * Exports can't be deleted out of the list while we hold
 495                  * the obd lock (class_unlink_export), which means we can't
 496                  * lose the last ref on the export.  If they've already been
 497                  * removed from the list, we won't find them here.
 498                  */
 499                 spin_lock(&obd->obd_dev_lock);
 500                 while (!list_empty(&obd->obd_exports_timed)) {
 501                         exp = list_entry(obd->obd_exports_timed.next,
 502                                          struct obd_export,
 503                                          exp_obd_chain_timed);
 504                         if (expire_time > exp->exp_last_request_time) {
 505                                 struct obd_uuid *client_uuid;
 506
 507                                 class_export_get(exp);
 508                                 client_uuid = &exp->exp_client_uuid;
 509                                 spin_unlock(&obd->obd_dev_lock);
 510                                 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",
 511                                               obd->obd_name,
 512                                               obd_uuid2str(client_uuid),
 513                                               obd_export_nid2str(exp),
 514                                               ktime_get_real_seconds() -
 515                                               exp->exp_last_request_time,
 516                                               exp, ktime_get_real_seconds(),
 517                                               expire_time,
 518                                               exp->exp_last_request_time);
 519                                 CDEBUG(D_HA, "Last request was at %lld\n",
 520                                        exp->exp_last_request_time);
 521                                 class_fail_export(exp);
 522                                 class_export_put(exp);
 523                                 spin_lock(&obd->obd_dev_lock);
 524                         } else {
 525                                 /* List is sorted, so everyone below is ok */
 526                                 break;
 527                         }
 528                 }
 529                 spin_unlock(&obd->obd_dev_lock);
 530
 531                 spin_lock(&pet_lock);
 532                 list_del_init(&obd->obd_evict_list);
 533                 spin_unlock(&pet_lock);
 534
 535                 class_decref(obd, "evictor", obd);
 536         }
 537         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 538
 539         RETURN(0);
 540 }
 541
 542 void ping_evictor_start(void)
 543 {
 544         struct task_struct *task;
 545
 546         if (++pet_refcount > 1)
 547                 return;
 548
 549         init_waitqueue_head(&pet_waitq);
 550
 551         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 552         if (IS_ERR(task)) {
 553                 pet_refcount--;
 554                 CERROR("Cannot start ping evictor thread: %ld\n",
 555                         PTR_ERR(task));
 556         }
 557 }
 558 EXPORT_SYMBOL(ping_evictor_start);
 559
 560 void ping_evictor_stop(void)
 561 {
 562         if (--pet_refcount > 0)
 563                 return;
 564
 565         pet_state = PET_TERMINATE;
 566         wake_up(&pet_waitq);
 567 }
 568 EXPORT_SYMBOL(ping_evictor_stop);