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