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_put(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         enum lustre_imp_state 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 || level == LUSTRE_IMP_IDLE) ?
 249                 D_INFO : D_HA,
 250                "%s->%s: level %s/%u force %u force_next %u deactive %u pingable %u suppress %u\n",
 251                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 252                ptlrpc_import_state_name(level), level, force, force_next,
 253                imp->imp_deactive, imp->imp_pingable, suppress);
 254
 255         if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
 256                 /* wait for a while before trying recovery again */
 257                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 258                 if (!imp->imp_no_pinger_recover ||
 259                     imp->imp_connect_error == -EAGAIN) {
 260                         CDEBUG(D_HA, "%s: starting recovery\n",
 261                                obd2cli_tgt(imp->imp_obd));
 262                         ptlrpc_connect_import_locked(imp);
 263                 } else {
 264                         spin_unlock(&imp->imp_lock);
 265                 }
 266         } else if (level != LUSTRE_IMP_FULL || imp->imp_obd->obd_no_recov ||
 267                    imp_is_deactive(imp)) {
 268                 CDEBUG(level == LUSTRE_IMP_IDLE ? D_INFO : D_HA,
 269                        "%s->%s: not pinging (in recovery or recovery disabled: %s)\n",
 270                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
 271                        ptlrpc_import_state_name(level));
 272                 if (force)
 273                         imp->imp_force_verify = 1;
 274                 spin_unlock(&imp->imp_lock);
 275         } else if ((imp->imp_pingable && !suppress) || force_next || force) {
 276                 spin_unlock(&imp->imp_lock);
 277                 ptlrpc_ping(imp);
 278         } else {
 279                 spin_unlock(&imp->imp_lock);
 280         }
 281 }
 282
 283 static struct workqueue_struct *pinger_wq;
 284 static void ptlrpc_pinger_main(struct work_struct *ws);
 285 static DECLARE_DELAYED_WORK(ping_work, ptlrpc_pinger_main);
 286
 287 static void ptlrpc_pinger_main(struct work_struct *ws)
 288 {
 289         time64_t this_ping, time_after_ping;
 290         timeout_t time_to_next_wake;
 291         struct obd_import *imp;
 292
 293         do {
 294                 this_ping = ktime_get_seconds();
 295
 296                 mutex_lock(&pinger_mutex);
 297
 298                 list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 299                         ptlrpc_pinger_process_import(imp, this_ping);
 300                         /* obd_timeout might have changed */
 301                         if (imp->imp_pingable && imp->imp_next_ping &&
 302                             imp->imp_next_ping > this_ping + PING_INTERVAL)
 303                                 ptlrpc_update_next_ping(imp, 0);
 304                 }
 305                 mutex_unlock(&pinger_mutex);
 306
 307                 time_after_ping = ktime_get_seconds();
 308                 /* update memory usage info */
 309                 obd_update_maxusage();
 310
 311                 if ((ktime_get_seconds() - this_ping - 3) > PING_INTERVAL)
 312                         CDEBUG(D_HA, "long time to ping: %lld, %lld, %lld\n",
 313                                this_ping, time_after_ping, ktime_get_seconds());
 314
 315                 /* Wait until the next ping time, or until we're stopped. */
 316                 time_to_next_wake = pinger_check_timeout(this_ping);
 317                 /*
 318                  * The ping sent by ptlrpc_send_rpc may get sent out
 319                  * say .01 second after this.
 320                  * ptlrpc_pinger_sending_on_import will then set the
 321                  * next ping time to next_ping + .01 sec, which means
 322                  * we will SKIP the next ping at next_ping, and the
 323                  * ping will get sent 2 timeouts from now!  Beware.
 324                  */
 325                 CDEBUG(D_INFO, "next wakeup in %d (%lld)\n",
 326                        time_to_next_wake, this_ping + PING_INTERVAL);
 327         } while (time_to_next_wake <= 0);
 328
 329         queue_delayed_work(pinger_wq, &ping_work,
 330                            cfs_time_seconds(max(time_to_next_wake, 1)));
 331 }
 332
 333 int ptlrpc_start_pinger(void)
 334 {
 335 #ifdef CONFIG_LUSTRE_FS_PINGER
 336         if (pinger_wq)
 337                 return -EALREADY;
 338
 339         pinger_wq = cfs_cpt_bind_workqueue("ptlrpc_pinger", cfs_cpt_tab,
 340                                            0, CFS_CPT_ANY, 1);
 341         if (IS_ERR(pinger_wq)) {
 342                 CERROR("cannot start pinger workqueue\n");
 343                 return PTR_ERR(pinger_wq);
 344         }
 345
 346         queue_delayed_work(pinger_wq, &ping_work, 0);
 347
 348         if (suppress_pings)
 349                 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");
 350 #endif
 351         return 0;
 352 }
 353
 354 int ptlrpc_stop_pinger(void)
 355 {
 356 #ifdef CONFIG_LUSTRE_FS_PINGER
 357         if (!pinger_wq)
 358                 return -EALREADY;
 359
 360         cancel_delayed_work_sync(&ping_work);
 361         destroy_workqueue(pinger_wq);
 362         pinger_wq = NULL;
 363 #endif
 364         return 0;
 365 }
 366
 367 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 368 {
 369         ptlrpc_update_next_ping(imp, 0);
 370 }
 371
 372 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 373 {
 374         ptlrpc_update_next_ping(imp, 1);
 375         assert_spin_locked(&imp->imp_lock);
 376         /*
 377          * Avoid reading stale imp_connect_data.  When not sure if pings are
 378          * expected or not on next connection, we assume they are not and force
 379          * one anyway to guarantee the chance of updating
 380          * imp_peer_committed_transno.
 381          */
 382         if (imp->imp_state != LUSTRE_IMP_FULL ||
 383             OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
 384                 imp->imp_force_next_verify = 1;
 385 }
 386
 387 int ptlrpc_pinger_add_import(struct obd_import *imp)
 388 {
 389         ENTRY;
 390         if (!list_empty(&imp->imp_pinger_chain))
 391                 RETURN(-EALREADY);
 392
 393         mutex_lock(&pinger_mutex);
 394         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 395                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 396         /* if we add to pinger we want recovery on this import */
 397         imp->imp_obd->obd_no_recov = 0;
 398         ptlrpc_update_next_ping(imp, 0);
 399         /* XXX sort, blah blah */
 400         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 401         class_import_get(imp);
 402
 403         ptlrpc_pinger_wake_up();
 404         mutex_unlock(&pinger_mutex);
 405
 406         RETURN(0);
 407 }
 408 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
 409
 410 int ptlrpc_pinger_del_import(struct obd_import *imp)
 411 {
 412         ENTRY;
 413
 414         if (list_empty(&imp->imp_pinger_chain))
 415                 RETURN(-ENOENT);
 416
 417         mutex_lock(&pinger_mutex);
 418         list_del_init(&imp->imp_pinger_chain);
 419         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 420                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 421         /* if we remove from pinger we don't want recovery on this import */
 422         imp->imp_obd->obd_no_recov = 1;
 423         class_import_put(imp);
 424         mutex_unlock(&pinger_mutex);
 425         RETURN(0);
 426 }
 427 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
 428
 429 void ptlrpc_pinger_wake_up(void)
 430 {
 431 #ifdef CONFIG_LUSTRE_FS_PINGER
 432         mod_delayed_work(pinger_wq, &ping_work, 0);
 433 #endif
 434 }
 435
 436 /* Ping evictor thread */
 437 #define PET_READY     1
 438 #define PET_TERMINATE 2
 439
 440 static int pet_refcount;
 441 static int pet_state;
 442 static wait_queue_head_t pet_waitq;
 443 static LIST_HEAD(pet_list);
 444 static DEFINE_SPINLOCK(pet_lock);
 445
 446 int ping_evictor_wake(struct obd_export *exp)
 447 {
 448         struct obd_device *obd;
 449
 450         spin_lock(&pet_lock);
 451         if (pet_state != PET_READY) {
 452                 /* eventually the new obd will call here again. */
 453                 spin_unlock(&pet_lock);
 454                 return 1;
 455         }
 456
 457         obd = class_exp2obd(exp);
 458         if (list_empty(&obd->obd_evict_list)) {
 459                 class_incref(obd, "evictor", obd);
 460                 list_add(&obd->obd_evict_list, &pet_list);
 461         }
 462         spin_unlock(&pet_lock);
 463
 464         wake_up(&pet_waitq);
 465         return 0;
 466 }
 467
 468 static int ping_evictor_main(void *arg)
 469 {
 470         struct obd_device *obd;
 471         struct obd_export *exp;
 472         time64_t expire_time;
 473
 474         ENTRY;
 475         unshare_fs_struct();
 476         CDEBUG(D_HA, "Starting Ping Evictor\n");
 477         pet_state = PET_READY;
 478         while (1) {
 479                 wait_event_idle(pet_waitq,
 480                                 (!list_empty(&pet_list)) ||
 481                                 (pet_state == PET_TERMINATE));
 482
 483                 /* loop until all obd's will be removed */
 484                 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
 485                         break;
 486
 487                 /*
 488                  * we only get here if pet_exp != NULL, and the end of this
 489                  * loop is the only place which sets it NULL again, so lock
 490                  * is not strictly necessary.
 491                  */
 492                 spin_lock(&pet_lock);
 493                 obd = list_first_entry(&pet_list, struct obd_device,
 494                                        obd_evict_list);
 495                 spin_unlock(&pet_lock);
 496
 497                 expire_time = ktime_get_real_seconds() - PING_EVICT_TIMEOUT;
 498
 499                 CDEBUG(D_HA, "evicting all exports of obd %s older than %lld\n",
 500                        obd->obd_name, expire_time);
 501
 502                 /*
 503                  * Exports can't be deleted out of the list while we hold
 504                  * the obd lock (class_unlink_export), which means we can't
 505                  * lose the last ref on the export.  If they've already been
 506                  * removed from the list, we won't find them here.
 507                  */
 508                 spin_lock(&obd->obd_dev_lock);
 509                 while (!list_empty(&obd->obd_exports_timed)) {
 510                         exp = list_first_entry(&obd->obd_exports_timed,
 511                                                struct obd_export,
 512                                                exp_obd_chain_timed);
 513                         if (expire_time > exp->exp_last_request_time) {
 514                                 struct obd_uuid *client_uuid;
 515
 516                                 class_export_get(exp);
 517                                 client_uuid = &exp->exp_client_uuid;
 518                                 spin_unlock(&obd->obd_dev_lock);
 519                                 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",
 520                                               obd->obd_name,
 521                                               obd_uuid2str(client_uuid),
 522                                               obd_export_nid2str(exp),
 523                                               ktime_get_real_seconds() -
 524                                               exp->exp_last_request_time,
 525                                               exp, ktime_get_real_seconds(),
 526                                               expire_time,
 527                                               exp->exp_last_request_time);
 528                                 CDEBUG(D_HA, "Last request was at %lld\n",
 529                                        exp->exp_last_request_time);
 530                                 class_fail_export(exp);
 531                                 class_export_put(exp);
 532                                 spin_lock(&obd->obd_dev_lock);
 533                         } else {
 534                                 /* List is sorted, so everyone below is ok */
 535                                 break;
 536                         }
 537                 }
 538                 spin_unlock(&obd->obd_dev_lock);
 539
 540                 spin_lock(&pet_lock);
 541                 list_del_init(&obd->obd_evict_list);
 542                 spin_unlock(&pet_lock);
 543
 544                 class_decref(obd, "evictor", obd);
 545         }
 546         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 547
 548         RETURN(0);
 549 }
 550
 551 void ping_evictor_start(void)
 552 {
 553         struct task_struct *task;
 554
 555         if (++pet_refcount > 1)
 556                 return;
 557
 558         init_waitqueue_head(&pet_waitq);
 559
 560         task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
 561         if (IS_ERR(task)) {
 562                 pet_refcount--;
 563                 CERROR("Cannot start ping evictor thread: %ld\n",
 564                         PTR_ERR(task));
 565         }
 566 }
 567 EXPORT_SYMBOL(ping_evictor_start);
 568
 569 void ping_evictor_stop(void)
 570 {
 571         if (--pet_refcount > 0)
 572                 return;
 573
 574         pet_state = PET_TERMINATE;
 575         wake_up(&pet_waitq);
 576 }
 577 EXPORT_SYMBOL(ping_evictor_stop);