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                                        obd_at_off(imp->imp_obd) ? 0 :
 125                                        obd_at_get(imp->imp_obd,
 126                                                 &imp->imp_at.iat_net_latency));
 127                 time = min(time, dtime);
 128         }
 129         imp->imp_next_ping = ktime_get_seconds() + time;
 130 #endif /* CONFIG_LUSTRE_FS_PINGER */
 131 }
 132
 133 static int ptlrpc_ping(struct obd_import *imp)
 134 {
 135         struct ptlrpc_request *req;
 136
 137         ENTRY;
 138
 139         if (ptlrpc_check_import_is_idle(imp) &&
 140             ptlrpc_disconnect_and_idle_import(imp) == 1)
 141                         RETURN(0);
 142
 143         req = ptlrpc_prep_ping(imp);
 144         if (!req) {
 145                 CERROR("OOM trying to ping %s->%s\n",
 146                        imp->imp_obd->obd_uuid.uuid,
 147                        obd2cli_tgt(imp->imp_obd));
 148                 RETURN(-ENOMEM);
 149         }
 150
 151         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 152                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 153         /* Updating imp_next_ping early, it allows pinger_check_timeout to
 154          * see an actual time for next awake. request_out_callback update
 155          * happens at another thread, and ptlrpc_pinger_main may sleep
 156          * already.
 157          */
 158         ptlrpc_update_next_ping(imp, 0);
 159         ptlrpcd_add_req(req);
 160
 161         RETURN(0);
 162 }
 163
 164 void ptlrpc_ping_import_soon(struct obd_import *imp)
 165 {
 166         imp->imp_next_ping = ktime_get_seconds();
 167 }
 168
 169 static inline int imp_is_deactive(struct obd_import *imp)
 170 {
 171         return imp->imp_deactive ||
 172                CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE);
 173 }
 174
 175 static inline time64_t ptlrpc_next_reconnect(struct obd_import *imp)
 176 {
 177         return ktime_get_seconds() + INITIAL_CONNECT_TIMEOUT;
 178 }
 179
 180 static timeout_t pinger_check_timeout(time64_t time)
 181 {
 182         timeout_t timeout = PING_INTERVAL;
 183         timeout_t next_timeout;
 184         time64_t now;
 185         struct list_head *iter;
 186         struct obd_import *imp;
 187
 188         mutex_lock(&pinger_mutex);
 189         now = ktime_get_seconds();
 190         /* Process imports to find a nearest next ping */
 191         list_for_each(iter, &pinger_imports) {
 192                 imp = list_entry(iter, struct obd_import, imp_pinger_chain);
 193                 if (!imp->imp_pingable || imp->imp_next_ping < now)
 194                         continue;
 195                 next_timeout = imp->imp_next_ping - now;
 196                 /* make sure imp_next_ping in the future from time */
 197                 if (next_timeout > (now - time) && timeout > next_timeout)
 198                         timeout = next_timeout;
 199         }
 200         mutex_unlock(&pinger_mutex);
 201
 202         return timeout - (now - time);
 203 }
 204
 205 static bool ir_up;
 206
 207 void ptlrpc_pinger_ir_up(void)
 208 {
 209         CDEBUG(D_HA, "IR up\n");
 210         ir_up = true;
 211 }
 212 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
 213
 214 void ptlrpc_pinger_ir_down(void)
 215 {
 216         CDEBUG(D_HA, "IR down\n");
 217         ir_up = false;
 218 }
 219 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
 220
 221 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 222                                          time64_t this_ping)
 223 {
 224         int level;
 225         int force;
 226         int force_next;
 227         int suppress;
 228
 229         spin_lock(&imp->imp_lock);
 230
 231         level = imp->imp_state;
 232         force = imp->imp_force_verify;
 233         force_next = imp->imp_force_next_verify;
 234         /*
 235          * This will be used below only if the import is "FULL".
 236          */
 237         suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
 238
 239         imp->imp_force_verify = 0;
 240
 241         if (imp->imp_next_ping - 5 >= this_ping && !force) {
 242                 spin_unlock(&imp->imp_lock);
 243                 return;
 244         }
 245
 246         imp->imp_force_next_verify = 0;
 247
 248         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 249                "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
 250                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 251                ptlrpc_import_state_name(level), level, force, force_next,
 252                imp->imp_deactive, imp->imp_pingable, suppress);
 253
 254         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 255                 /* wait for a while before trying recovery again */
 256                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 257                 if (!imp->imp_no_pinger_recover ||
 258                     imp->imp_connect_error == -EAGAIN) {
 259                         CDEBUG(D_HA, "%s: starting recovery\n",
 260                                obd2cli_tgt(imp->imp_obd));
 261                         ptlrpc_connect_import_locked(imp);
 262                 } else {
 263                         spin_unlock(&imp->imp_lock);
 264                 }
 265         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
 266                    imp_is_deactive(imp)) {
 267                 CDEBUG(D_HA,
 268                        "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 269                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 270                        ptlrpc_import_state_name(level));
 271                 if (force)
 272                         imp->imp_force_verify = 1;
 273                 spin_unlock(&imp->imp_lock);
 274         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 275                 spin_unlock(&imp->imp_lock);
 276                 ptlrpc_ping(imp);
 277         } else {
 278                 spin_unlock(&imp->imp_lock);
 279         }
 280 }
 281
 282 static struct workqueue_struct *pinger_wq;
 283 static void ptlrpc_pinger_main(struct work_struct *ws);
 284 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
 285
 286 static void ptlrpc_pinger_main(struct work_struct *ws)
 287 {
 288         time64_t this_ping, time_after_ping;
 289         timeout_t time_to_next_wake;
 290         struct obd_import *imp;
 291
 292         do {
 293                 this_ping = ktime_get_seconds();
 294
 295                 mutex_lock(&pinger_mutex);
 296
 297                 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 298                         ptlrpc_pinger_process_import(imp, this_ping);
 299                         /* obd_timeout might have changed */
 300                         if (imp->imp_pingable && imp->imp_next_ping &&
 301                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 302                                 ptlrpc_update_next_ping(imp, 0);
 303                 }
 304                 mutex_unlock(&pinger_mutex);
 305
 306                 time_after_ping = ktime_get_seconds();
 307                 /* update memory usage info */
 308                 obd_update_maxusage();
 309
 310                 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
 311                         CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
 312                                this_ping, time_after_ping, ktime_get_seconds());
 313
 314                 /* Wait until the next ping time, or until we're stopped. */
 315                 time_to_next_wake = pinger_check_timeout(this_ping);
 316                 /*
 317                  * The ping sent by ptlrpc_send_rpc may get sent out
 318                  * say .01 second after this.
 319                  * ptlrpc_pinger_sending_on_import will then set the
 320                  * next ping time to next_ping + .01 sec, which means
 321                  * we will SKIP the next ping at next_ping, and the
 322                  * ping will get sent 2 timeouts from now!  Beware.
 323                  */
 324                 CDEBUG(D_INFO, "next wakeup in %d (%lld)\n",
 325                        time_to_next_wake, this_ping + PING_INTERVAL);
 326         } while (time_to_next_wake <= 0);
 327
 328         queue_delayed_work(pinger_wq, &ping_work,
 329                            cfs_time_seconds(max(time_to_next_wake, 1)));
 330 }
 331
 332 int ptlrpc_start_pinger(void)
 333 {
 334 #ifdef CONFIG_LUSTRE_FS_PINGER
 335         if (pinger_wq)
 336                 return -EALREADY;
 337
 338         pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
 339                                            0, CFS_CPT_ANY, 1);
 340         if (IS_ERR(pinger_wq)) {
 341                 CERROR("cannot start pinger workqueue\n");
 342                 return PTR_ERR(pinger_wq);
 343         }
 344
 345         queue_delayed_work(pinger_wq, &ping_work, 0);
 346
 347         if (suppress_pings)
 348                 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");
 349 #endif
 350         return 0;
 351 }
 352
 353 int ptlrpc_stop_pinger(void)
 354 {
 355 #ifdef CONFIG_LUSTRE_FS_PINGER
 356         if (!pinger_wq)
 357                 return -EALREADY;
 358
 359         cancel_delayed_work_sync(&ping_work);
 360         destroy_workqueue(pinger_wq);
 361         pinger_wq = NULL;
 362 #endif
 363         return 0;
 364 }
 365
 366 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 367 {
 368         ptlrpc_update_next_ping(imp, 0);
 369 }
 370
 371 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 372 {
 373         ptlrpc_update_next_ping(imp, 1);
 374         assert_spin_locked(&imp->imp_lock);
 375         /*
 376          * Avoid reading stale imp_connect_data.  When not sure if pings are
 377          * expected or not on next connection, we assume they are not and force
 378          * one anyway to guarantee the chance of updating
 379          * imp_peer_committed_transno.
 380          */
 381         if (imp->imp_state != LUSTRE_IMP_FULL ||
 382             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 383                 imp->imp_force_next_verify = 1;
 384 }
 385
 386 int ptlrpc_pinger_add_import(struct obd_import *imp)
 387 {
 388         ENTRY;
 389         if (!list_empty(&imp->imp_pinger_chain))
 390                 RETURN(-EALREADY);
 391
 392         mutex_lock(&pinger_mutex);
 393         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 394                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 395         /* if we add to pinger we want recovery on this import */
 396         imp->imp_obd->obd_no_recov = 0;
 397         ptlrpc_update_next_ping(imp, 0);
 398         /* XXX sort, blah blah */
 399         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 400         class_import_get(imp);
 401
 402         ptlrpc_pinger_wake_up();
 403         mutex_unlock(&pinger_mutex);
 404
 405         RETURN(0);
 406 }
 407 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 408
 409 int ptlrpc_pinger_del_import(struct obd_import *imp)
 410 {
 411         ENTRY;
 412
 413         if (list_empty(&imp->imp_pinger_chain))
 414                 RETURN(-ENOENT);
 415
 416         mutex_lock(&pinger_mutex);
 417         list_del_init(&imp->imp_pinger_chain);
 418         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 419                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 420         /* if we remove from pinger we don't want recovery on this import */
 421         imp->imp_obd->obd_no_recov = 1;
 422         class_import_put(imp);
 423         mutex_unlock(&pinger_mutex);
 424         RETURN(0);
 425 }
 426 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 427
 428 void ptlrpc_pinger_wake_up(void)
 429 {
 430 #ifdef CONFIG_LUSTRE_FS_PINGER
 431         mod_delayed_work(pinger_wq, &ping_work, 0);
 432 #endif
 433 }
 434
 435 /* Ping evictor thread */
 436 #define PET_READY     1
 437 #define PET_TERMINATE 2
 438
 439 static int pet_refcount;
 440 static int pet_state;
 441 static wait_queue_head_t pet_waitq;
 442 static LIST_HEAD(pet_list);
 443 static DEFINE_SPINLOCK(pet_lock);
 444
 445 int ping_evictor_wake(struct obd_export *exp)
 446 {
 447         struct obd_device *obd;
 448
 449         spin_lock(&pet_lock);
 450         if (pet_state != PET_READY) {
 451                 /* eventually the new obd will call here again. */
 452                 spin_unlock(&pet_lock);
 453                 return 1;
 454         }
 455
 456         obd = class_exp2obd(exp);
 457         if (list_empty(&obd->obd_evict_list)) {
 458                 class_incref(obd, "evictor", obd);
 459                 list_add(&obd->obd_evict_list, &pet_list);
 460         }
 461         spin_unlock(&pet_lock);
 462
 463         wake_up(&pet_waitq);
 464         return 0;
 465 }
 466
 467 static int ping_evictor_main(void *arg)
 468 {
 469         struct obd_device *obd;
 470         struct obd_export *exp;
 471         time64_t expire_time;
 472
 473         ENTRY;
 474         unshare_fs_struct();
 475         CDEBUG(D_HA, "Starting Ping Evictor\n");
 476         pet_state = PET_READY;
 477         while (1) {
 478                 wait_event_idle(pet_waitq,
 479                                 (!list_empty(&pet_list)) ||
 480                                 (pet_state == PET_TERMINATE));
 481
 482                 /* loop until all obd's will be removed */
 483                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 484                         break;
 485
 486                 /*
 487                  * we only get here if pet_exp != NULL, and the end of this
 488                  * loop is the only place which sets it NULL again, so lock
 489                  * is not strictly necessary.
 490                  */
 491                 spin_lock(&pet_lock);
 492                 obd = list_first_entry(&pet_list, struct obd_device,
 493                                        obd_evict_list);
 494                 spin_unlock(&pet_lock);
 495
 496                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 497
 498                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 499                        obd->obd_name, expire_time);
 500
 501                 /*
 502                  * Exports can't be deleted out of the list while we hold
 503                  * the obd lock (class_unlink_export), which means we can't
 504                  * lose the last ref on the export.  If they've already been
 505                  * removed from the list, we won't find them here.
 506                  */
 507                 spin_lock(&obd->obd_dev_lock);
 508                 while (!list_empty(&obd->obd_exports_timed)) {
 509                         exp = list_first_entry(&obd->obd_exports_timed,
 510                                                struct obd_export,
 511                                                exp_obd_chain_timed);
 512                         if (expire_time > exp->exp_last_request_time) {
 513                                 struct obd_uuid *client_uuid;
 514
 515                                 class_export_get(exp);
 516                                 client_uuid = &exp->exp_client_uuid;
 517                                 spin_unlock(&obd->obd_dev_lock);
 518                                 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",
 519                                               obd->obd_name,
 520                                               obd_uuid2str(client_uuid),
 521                                               obd_export_nid2str(exp),
 522                                               ktime_get_real_seconds() -
 523                                               exp->exp_last_request_time,
 524                                               exp, ktime_get_real_seconds(),
 525                                               expire_time,
 526                                               exp->exp_last_request_time);
 527                                 CDEBUG(D_HA, "Last request was at %lld\n",
 528                                        exp->exp_last_request_time);
 529                                 class_fail_export(exp);
 530                                 class_export_put(exp);
 531                                 spin_lock(&obd->obd_dev_lock);
 532                         } else {
 533                                 /* List is sorted, so everyone below is ok */
 534                                 break;
 535                         }
 536                 }
 537                 spin_unlock(&obd->obd_dev_lock);
 538
 539                 spin_lock(&pet_lock);
 540                 list_del_init(&obd->obd_evict_list);
 541                 spin_unlock(&pet_lock);
 542
 543                 class_decref(obd, "evictor", obd);
 544         }
 545         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 546
 547         RETURN(0);
 548 }
 549
 550 void ping_evictor_start(void)
 551 {
 552         struct task_struct *task;
 553
 554         if (++pet_refcount > 1)
 555                 return;
 556
 557         init_waitqueue_head(&pet_waitq);
 558
 559         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 560         if (IS_ERR(task)) {
 561                 pet_refcount--;
 562                 CERROR("Cannot start ping evictor thread: %ld\n",
 563                         PTR_ERR(task));
 564         }
 565 }
 566 EXPORT_SYMBOL(ping_evictor_start);
 567
 568 void ping_evictor_stop(void)
 569 {
 570         if (--pet_refcount > 0)
 571                 return;
 572
 573         pet_state = PET_TERMINATE;
 574         wake_up(&pet_waitq);
 575 }
 576 EXPORT_SYMBOL(ping_evictor_stop);