lustre/ptlrpc/pinger.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ptlrpc/pinger.c
  40  *
  41  * Portal-RPC reconnection and replay operations, for use in recovery.
  42  */
  43
  44 #ifndef __KERNEL__
  45 #include <liblustre.h>
  46 #else
  47 #define DEBUG_SUBSYSTEM S_RPC
  48 #endif
  49
  50 #include <obd_support.h>
  51 #include <obd_class.h>
  52 #include "ptlrpc_internal.h"
  53
  54 cfs_semaphore_t pinger_sem;
  55 static CFS_LIST_HEAD(pinger_imports);
  56 static cfs_list_t timeout_list = CFS_LIST_HEAD_INIT(timeout_list);
  57 struct ptlrpc_request *
  58 ptlrpc_prep_ping(struct obd_import *imp)
  59 {
  60         struct ptlrpc_request *req;
  61
  62         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  63                                         LUSTRE_OBD_VERSION, OBD_PING);
  64         if (req) {
  65                 ptlrpc_request_set_replen(req);
  66                 req->rq_no_resend = req->rq_no_delay = 1;
  67         }
  68         return req;
  69 }
  70
  71 int ptlrpc_obd_ping(struct obd_device *obd)
  72 {
  73         int rc;
  74         struct ptlrpc_request *req;
  75         ENTRY;
  76
  77         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  78         if (req == NULL)
  79                 RETURN(-ENOMEM);
  80
  81         req->rq_send_state = LUSTRE_IMP_FULL;
  82
  83         rc = ptlrpc_queue_wait(req);
  84
  85         ptlrpc_req_finished(req);
  86
  87         RETURN(rc);
  88 }
  89 EXPORT_SYMBOL(ptlrpc_obd_ping);
  90
  91 int ptlrpc_ping(struct obd_import *imp)
  92 {
  93         struct ptlrpc_request *req;
  94         ENTRY;
  95
  96         req = ptlrpc_prep_ping(imp);
  97         if (req == NULL) {
  98                 CERROR("OOM trying to ping %s->%s\n",
  99                        imp->imp_obd->obd_uuid.uuid,
 100                        obd2cli_tgt(imp->imp_obd));
 101                 RETURN(-ENOMEM);
 102         }
 103
 104         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 105                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 106         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 107
 108         RETURN(0);
 109 }
 110
 111 void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
 112 {
 113 #ifdef ENABLE_PINGER
 114         int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
 115         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 116                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 117                                   AT_OFF ? 0 :
 118                                   at_get(&imp->imp_at.iat_net_latency));
 119                 time = min(time, dtime);
 120         }
 121         imp->imp_next_ping = cfs_time_shift(time);
 122 #endif /* ENABLE_PINGER */
 123 }
 124
 125 void ptlrpc_ping_import_soon(struct obd_import *imp)
 126 {
 127         imp->imp_next_ping = cfs_time_current();
 128 }
 129
 130 static inline int imp_is_deactive(struct obd_import *imp)
 131 {
 132         return (imp->imp_deactive ||
 133                 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
 134 }
 135
 136 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 137 {
 138         if (imp->imp_server_timeout)
 139                 return cfs_time_shift(obd_timeout / 2);
 140         else
 141                 return cfs_time_shift(obd_timeout);
 142 }
 143
 144 static cfs_atomic_t suspend_timeouts = CFS_ATOMIC_INIT(0);
 145 static cfs_time_t suspend_wakeup_time = 0;
 146
 147 cfs_duration_t pinger_check_timeout(cfs_time_t time)
 148 {
 149         struct timeout_item *item;
 150         cfs_time_t timeout = PING_INTERVAL;
 151
 152         /* The timeout list is a increase order sorted list */
 153         cfs_mutex_down(&pinger_sem);
 154         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 155                 int ti_timeout = item->ti_timeout;
 156                 if (timeout > ti_timeout)
 157                         timeout = ti_timeout;
 158                 break;
 159         }
 160         cfs_mutex_up(&pinger_sem);
 161
 162         return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
 163                                          cfs_time_current());
 164 }
 165
 166 #ifdef __KERNEL__
 167 static cfs_waitq_t suspend_timeouts_waitq;
 168 #endif
 169
 170 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 171 {
 172         return suspend_wakeup_time;
 173 }
 174
 175 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 176 {
 177         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 178 #if 0
 179         if (imp->imp_no_timeout)
 180                 return;
 181         imp->imp_no_timeout = 1;
 182         cfs_atomic_inc(&suspend_timeouts);
 183         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n",
 184                cfs_atomic_read(&suspend_timeouts));
 185 #endif
 186 }
 187
 188 void ptlrpc_activate_timeouts(struct obd_import *imp)
 189 {
 190         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 191 #if 0
 192         if (!imp->imp_no_timeout)
 193                 return;
 194         imp->imp_no_timeout = 0;
 195         LASSERT(cfs_atomic_read(&suspend_timeouts) > 0);
 196         if (cfs_atomic_dec_and_test(&suspend_timeouts)) {
 197                 suspend_wakeup_time = cfs_time_current();
 198                 cfs_waitq_signal(&suspend_timeouts_waitq);
 199         }
 200         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n",
 201                cfs_atomic_read(&suspend_timeouts));
 202 #endif
 203 }
 204
 205 int ptlrpc_check_suspend(void)
 206 {
 207         if (cfs_atomic_read(&suspend_timeouts))
 208                 return 1;
 209         return 0;
 210 }
 211
 212 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 213 {
 214         struct l_wait_info lwi;
 215
 216         if (cfs_atomic_read(&suspend_timeouts)) {
 217                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 218                           cfs_atomic_read(&suspend_timeouts));
 219                 lwi = LWI_INTR(NULL, NULL);
 220                 l_wait_event(suspend_timeouts_waitq,
 221                              cfs_atomic_read(&suspend_timeouts) == 0, &lwi);
 222                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 223                 return 1;
 224         }
 225         return 0;
 226 }
 227
 228 #ifdef __KERNEL__
 229
 230 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 231                                          unsigned long this_ping)
 232 {
 233         int force, level;
 234
 235         cfs_spin_lock(&imp->imp_lock);
 236         level = imp->imp_state;
 237         force = imp->imp_force_verify;
 238         if (force)
 239                 imp->imp_force_verify = 0;
 240         cfs_spin_unlock(&imp->imp_lock);
 241
 242         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 243                "level %s/%u force %u deactive %u pingable %u\n",
 244                ptlrpc_import_state_name(level), level,
 245                force, imp->imp_deactive, imp->imp_pingable);
 246
 247         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 248                              this_ping) && force == 0)
 249                 return;
 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                 if (!imp->imp_no_pinger_recover)
 255                         ptlrpc_initiate_recovery(imp);
 256         } else if (level != LUSTRE_IMP_FULL ||
 257                    imp->imp_obd->obd_no_recov ||
 258                    imp_is_deactive(imp)) {
 259                 CDEBUG(D_HA, "not pinging %s (in recovery "
 260                        " or recovery disabled: %s)\n",
 261                        obd2cli_tgt(imp->imp_obd),
 262                        ptlrpc_import_state_name(level));
 263         } else if (imp->imp_pingable || force) {
 264                 ptlrpc_ping(imp);
 265         }
 266 }
 267
 268 static int ptlrpc_pinger_main(void *arg)
 269 {
 270         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 271         struct ptlrpc_thread *thread = data->thread;
 272         ENTRY;
 273
 274         cfs_daemonize(data->name);
 275
 276         /* Record that the thread is running */
 277         thread->t_flags = SVC_RUNNING;
 278         cfs_waitq_signal(&thread->t_ctl_waitq);
 279
 280         /* And now, loop forever, pinging as needed. */
 281         while (1) {
 282                 cfs_time_t this_ping = cfs_time_current();
 283                 struct l_wait_info lwi;
 284                 cfs_duration_t time_to_next_wake;
 285                 struct timeout_item *item;
 286                 cfs_list_t *iter;
 287
 288                 cfs_mutex_down(&pinger_sem);
 289                 cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 290                         item->ti_cb(item, item->ti_cb_data);
 291                 }
 292                 cfs_list_for_each(iter, &pinger_imports) {
 293                         struct obd_import *imp =
 294                                 cfs_list_entry(iter, struct obd_import,
 295                                                imp_pinger_chain);
 296
 297                         ptlrpc_pinger_process_import(imp, this_ping);
 298                         /* obd_timeout might have changed */
 299                         if (imp->imp_pingable && imp->imp_next_ping &&
 300                             cfs_time_after(imp->imp_next_ping,
 301                                            cfs_time_add(this_ping,
 302                                                         cfs_time_seconds(PING_INTERVAL))))
 303                                 ptlrpc_update_next_ping(imp, 0);
 304                 }
 305                 cfs_mutex_up(&pinger_sem);
 306                 /* update memory usage info */
 307                 obd_update_maxusage();
 308
 309                 /* Wait until the next ping time, or until we're stopped. */
 310                 time_to_next_wake = pinger_check_timeout(this_ping);
 311                 /* The ping sent by ptlrpc_send_rpc may get sent out
 312                    say .01 second after this.
 313                    ptlrpc_pinger_sending_on_import will then set the
 314                    next ping time to next_ping + .01 sec, which means
 315                    we will SKIP the next ping at next_ping, and the
 316                    ping will get sent 2 timeouts from now!  Beware. */
 317                 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 318                                 time_to_next_wake,
 319                                 cfs_time_add(this_ping, cfs_time_seconds(PING_INTERVAL)));
 320                 if (time_to_next_wake > 0) {
 321                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t, time_to_next_wake, cfs_time_seconds(1)),
 322                                             NULL, NULL);
 323                         l_wait_event(thread->t_ctl_waitq,
 324                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 325                                      &lwi);
 326                         if (thread->t_flags & SVC_STOPPING) {
 327                                 thread->t_flags &= ~SVC_STOPPING;
 328                                 EXIT;
 329                                 break;
 330                         } else if (thread->t_flags & SVC_EVENT) {
 331                                 /* woken after adding import to reset timer */
 332                                 thread->t_flags &= ~SVC_EVENT;
 333                         }
 334                 }
 335         }
 336
 337         thread->t_flags = SVC_STOPPED;
 338         cfs_waitq_signal(&thread->t_ctl_waitq);
 339
 340         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 341         return 0;
 342 }
 343
 344 static struct ptlrpc_thread *pinger_thread = NULL;
 345
 346 int ptlrpc_start_pinger(void)
 347 {
 348         struct l_wait_info lwi = { 0 };
 349         struct ptlrpc_svc_data d;
 350         int rc;
 351 #ifndef ENABLE_PINGER
 352         return 0;
 353 #endif
 354         ENTRY;
 355
 356         if (pinger_thread != NULL)
 357                 RETURN(-EALREADY);
 358
 359         OBD_ALLOC_PTR(pinger_thread);
 360         if (pinger_thread == NULL)
 361                 RETURN(-ENOMEM);
 362         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 363         cfs_waitq_init(&suspend_timeouts_waitq);
 364
 365         d.name = "ll_ping";
 366         d.thread = pinger_thread;
 367
 368         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 369          * just drop the VM and FILES in cfs_daemonize_ctxt() right away. */
 370         rc = cfs_create_thread(ptlrpc_pinger_main, &d, CFS_DAEMON_FLAGS);
 371         if (rc < 0) {
 372                 CERROR("cannot start thread: %d\n", rc);
 373                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 374                 pinger_thread = NULL;
 375                 RETURN(rc);
 376         }
 377         l_wait_event(pinger_thread->t_ctl_waitq,
 378                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 379
 380         RETURN(0);
 381 }
 382
 383 int ptlrpc_pinger_remove_timeouts(void);
 384
 385 int ptlrpc_stop_pinger(void)
 386 {
 387         struct l_wait_info lwi = { 0 };
 388         int rc = 0;
 389 #ifndef ENABLE_PINGER
 390         return 0;
 391 #endif
 392         ENTRY;
 393
 394         if (pinger_thread == NULL)
 395                 RETURN(-EALREADY);
 396
 397         ptlrpc_pinger_remove_timeouts();
 398         cfs_mutex_down(&pinger_sem);
 399         pinger_thread->t_flags = SVC_STOPPING;
 400         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 401         cfs_mutex_up(&pinger_sem);
 402
 403         l_wait_event(pinger_thread->t_ctl_waitq,
 404                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 405
 406         OBD_FREE_PTR(pinger_thread);
 407         pinger_thread = NULL;
 408         RETURN(rc);
 409 }
 410
 411 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 412 {
 413         ptlrpc_update_next_ping(imp, 0);
 414 }
 415
 416 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 417 {
 418         ptlrpc_update_next_ping(imp, 1);
 419 }
 420
 421 int ptlrpc_pinger_add_import(struct obd_import *imp)
 422 {
 423         ENTRY;
 424         if (!cfs_list_empty(&imp->imp_pinger_chain))
 425                 RETURN(-EALREADY);
 426
 427         cfs_mutex_down(&pinger_sem);
 428         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 429                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 430         /* if we add to pinger we want recovery on this import */
 431         imp->imp_obd->obd_no_recov = 0;
 432         ptlrpc_update_next_ping(imp, 0);
 433         /* XXX sort, blah blah */
 434         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 435         class_import_get(imp);
 436
 437         ptlrpc_pinger_wake_up();
 438         cfs_mutex_up(&pinger_sem);
 439
 440         RETURN(0);
 441 }
 442
 443 int ptlrpc_pinger_del_import(struct obd_import *imp)
 444 {
 445         ENTRY;
 446         if (cfs_list_empty(&imp->imp_pinger_chain))
 447                 RETURN(-ENOENT);
 448
 449         cfs_mutex_down(&pinger_sem);
 450         cfs_list_del_init(&imp->imp_pinger_chain);
 451         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 452                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 453         /* if we remove from pinger we don't want recovery on this import */
 454         imp->imp_obd->obd_no_recov = 1;
 455         class_import_put(imp);
 456         cfs_mutex_up(&pinger_sem);
 457         RETURN(0);
 458 }
 459
 460 /**
 461  * Register a timeout callback to the pinger list, and the callback will
 462  * be called when timeout happens.
 463  */
 464 struct timeout_item* ptlrpc_new_timeout(int time, enum timeout_event event,
 465                                         timeout_cb_t cb, void *data)
 466 {
 467         struct timeout_item *ti;
 468
 469         OBD_ALLOC_PTR(ti);
 470         if (!ti)
 471                 return(NULL);
 472
 473         CFS_INIT_LIST_HEAD(&ti->ti_obd_list);
 474         CFS_INIT_LIST_HEAD(&ti->ti_chain);
 475         ti->ti_timeout = time;
 476         ti->ti_event = event;
 477         ti->ti_cb = cb;
 478         ti->ti_cb_data = data;
 479
 480         return ti;
 481 }
 482
 483 /**
 484  * Register timeout event on the the pinger thread.
 485  * Note: the timeout list is an sorted list with increased timeout value.
 486  */
 487 static struct timeout_item*
 488 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
 489                                timeout_cb_t cb, void *data)
 490 {
 491         struct timeout_item *item, *tmp;
 492
 493         LASSERT_SEM_LOCKED(&pinger_sem);
 494
 495         cfs_list_for_each_entry(item, &timeout_list, ti_chain)
 496                 if (item->ti_event == event)
 497                         goto out;
 498
 499         item = ptlrpc_new_timeout(time, event, cb, data);
 500         if (item) {
 501                 cfs_list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
 502                         if (tmp->ti_timeout < time) {
 503                                 cfs_list_add(&item->ti_chain, &tmp->ti_chain);
 504                                 goto out;
 505                         }
 506                 }
 507                 cfs_list_add(&item->ti_chain, &timeout_list);
 508         }
 509 out:
 510         return item;
 511 }
 512
 513 /* Add a client_obd to the timeout event list, when timeout(@time)
 514  * happens, the callback(@cb) will be called.
 515  */
 516 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 517                               timeout_cb_t cb, void *data,
 518                               cfs_list_t *obd_list)
 519 {
 520         struct timeout_item *ti;
 521
 522         cfs_mutex_down(&pinger_sem);
 523         ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
 524         if (!ti) {
 525                 cfs_mutex_up(&pinger_sem);
 526                 return (-EINVAL);
 527         }
 528         cfs_list_add(obd_list, &ti->ti_obd_list);
 529         cfs_mutex_up(&pinger_sem);
 530         return 0;
 531 }
 532
 533 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 534                               enum timeout_event event)
 535 {
 536         struct timeout_item *ti = NULL, *item;
 537
 538         if (cfs_list_empty(obd_list))
 539                 return 0;
 540         cfs_mutex_down(&pinger_sem);
 541         cfs_list_del_init(obd_list);
 542         /**
 543          * If there are no obd attached to the timeout event
 544          * list, remove this timeout event from the pinger
 545          */
 546         cfs_list_for_each_entry(item, &timeout_list, ti_chain) {
 547                 if (item->ti_event == event) {
 548                         ti = item;
 549                         break;
 550                 }
 551         }
 552         LASSERTF(ti != NULL, "ti is NULL ! \n");
 553         if (cfs_list_empty(&ti->ti_obd_list)) {
 554                 cfs_list_del(&ti->ti_chain);
 555                 OBD_FREE_PTR(ti);
 556         }
 557         cfs_mutex_up(&pinger_sem);
 558         return 0;
 559 }
 560
 561 int ptlrpc_pinger_remove_timeouts(void)
 562 {
 563         struct timeout_item *item, *tmp;
 564
 565         cfs_mutex_down(&pinger_sem);
 566         cfs_list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
 567                 LASSERT(cfs_list_empty(&item->ti_obd_list));
 568                 cfs_list_del(&item->ti_chain);
 569                 OBD_FREE_PTR(item);
 570         }
 571         cfs_mutex_up(&pinger_sem);
 572         return 0;
 573 }
 574
 575 void ptlrpc_pinger_wake_up()
 576 {
 577 #ifdef ENABLE_PINGER
 578         pinger_thread->t_flags |= SVC_EVENT;
 579         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 580 #endif
 581 }
 582
 583 /* Ping evictor thread */
 584 #define PET_READY     1
 585 #define PET_TERMINATE 2
 586
 587 static int               pet_refcount = 0;
 588 static int               pet_state;
 589 static cfs_waitq_t       pet_waitq;
 590 CFS_LIST_HEAD(pet_list);
 591 static cfs_spinlock_t    pet_lock = CFS_SPIN_LOCK_UNLOCKED;
 592
 593 int ping_evictor_wake(struct obd_export *exp)
 594 {
 595         struct obd_device *obd;
 596
 597         cfs_spin_lock(&pet_lock);
 598         if (pet_state != PET_READY) {
 599                 /* eventually the new obd will call here again. */
 600                 cfs_spin_unlock(&pet_lock);
 601                 return 1;
 602         }
 603
 604         obd = class_exp2obd(exp);
 605         if (cfs_list_empty(&obd->obd_evict_list)) {
 606                 class_incref(obd, "evictor", obd);
 607                 cfs_list_add(&obd->obd_evict_list, &pet_list);
 608         }
 609         cfs_spin_unlock(&pet_lock);
 610
 611         cfs_waitq_signal(&pet_waitq);
 612         return 0;
 613 }
 614
 615 static int ping_evictor_main(void *arg)
 616 {
 617         struct obd_device *obd;
 618         struct obd_export *exp;
 619         struct l_wait_info lwi = { 0 };
 620         time_t expire_time;
 621         ENTRY;
 622
 623         cfs_daemonize_ctxt("ll_evictor");
 624
 625         CDEBUG(D_HA, "Starting Ping Evictor\n");
 626         pet_state = PET_READY;
 627         while (1) {
 628                 l_wait_event(pet_waitq, (!cfs_list_empty(&pet_list)) ||
 629                              (pet_state == PET_TERMINATE), &lwi);
 630
 631                 /* loop until all obd's will be removed */
 632                 if ((pet_state == PET_TERMINATE) && cfs_list_empty(&pet_list))
 633                         break;
 634
 635                 /* we only get here if pet_exp != NULL, and the end of this
 636                  * loop is the only place which sets it NULL again, so lock
 637                  * is not strictly necessary. */
 638                 cfs_spin_lock(&pet_lock);
 639                 obd = cfs_list_entry(pet_list.next, struct obd_device,
 640                                      obd_evict_list);
 641                 cfs_spin_unlock(&pet_lock);
 642
 643                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 644
 645                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 646                        obd->obd_name, expire_time);
 647
 648                 /* Exports can't be deleted out of the list while we hold
 649                  * the obd lock (class_unlink_export), which means we can't
 650                  * lose the last ref on the export.  If they've already been
 651                  * removed from the list, we won't find them here. */
 652                 cfs_spin_lock(&obd->obd_dev_lock);
 653                 while (!cfs_list_empty(&obd->obd_exports_timed)) {
 654                         exp = cfs_list_entry(obd->obd_exports_timed.next,
 655                                              struct obd_export,
 656                                              exp_obd_chain_timed);
 657                         if (expire_time > exp->exp_last_request_time) {
 658                                 class_export_get(exp);
 659                                 cfs_spin_unlock(&obd->obd_dev_lock);
 660                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 661                                               " (at %s) in %ld seconds. I think"
 662                                               " it's dead, and I am evicting"
 663                                               " it. exp %p, cur %ld expire %ld"
 664                                               " last %ld\n",
 665                                               obd->obd_name,
 666                                               obd_uuid2str(&exp->exp_client_uuid),
 667                                               obd_export_nid2str(exp),
 668                                               (long)(cfs_time_current_sec() -
 669                                                      exp->exp_last_request_time),
 670                                               exp, (long)cfs_time_current_sec(),
 671                                               (long)expire_time,
 672                                               (long)exp->exp_last_request_time);
 673                                 CDEBUG(D_HA, "Last request was at %ld\n",
 674                                        exp->exp_last_request_time);
 675                                 class_fail_export(exp);
 676                                 class_export_put(exp);
 677                                 cfs_spin_lock(&obd->obd_dev_lock);
 678                         } else {
 679                                 /* List is sorted, so everyone below is ok */
 680                                 break;
 681                         }
 682                 }
 683                 cfs_spin_unlock(&obd->obd_dev_lock);
 684
 685                 cfs_spin_lock(&pet_lock);
 686                 cfs_list_del_init(&obd->obd_evict_list);
 687                 cfs_spin_unlock(&pet_lock);
 688
 689                 class_decref(obd, "evictor", obd);
 690         }
 691         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 692
 693         RETURN(0);
 694 }
 695
 696 void ping_evictor_start(void)
 697 {
 698         int rc;
 699
 700         if (++pet_refcount > 1)
 701                 return;
 702
 703         cfs_waitq_init(&pet_waitq);
 704
 705         rc = cfs_create_thread(ping_evictor_main, NULL, CFS_DAEMON_FLAGS);
 706         if (rc < 0) {
 707                 pet_refcount--;
 708                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 709         }
 710 }
 711 EXPORT_SYMBOL(ping_evictor_start);
 712
 713 void ping_evictor_stop(void)
 714 {
 715         if (--pet_refcount > 0)
 716                 return;
 717
 718         pet_state = PET_TERMINATE;
 719         cfs_waitq_signal(&pet_waitq);
 720 }
 721 EXPORT_SYMBOL(ping_evictor_stop);
 722 #else /* !__KERNEL__ */
 723
 724 /* XXX
 725  * the current implementation of pinger in liblustre is not optimized
 726  */
 727
 728 #ifdef ENABLE_PINGER
 729 static struct pinger_data {
 730         int             pd_recursion;
 731         cfs_time_t      pd_this_ping;   /* jiffies */
 732         cfs_time_t      pd_next_ping;   /* jiffies */
 733         struct ptlrpc_request_set *pd_set;
 734 } pinger_args;
 735
 736 static int pinger_check_rpcs(void *arg)
 737 {
 738         cfs_time_t curtime = cfs_time_current();
 739         struct ptlrpc_request *req;
 740         struct ptlrpc_request_set *set;
 741         cfs_list_t *iter;
 742         struct obd_import *imp;
 743         struct pinger_data *pd = &pinger_args;
 744         int rc;
 745
 746         /* prevent recursion */
 747         if (pd->pd_recursion++) {
 748                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 749                 LASSERT(pd->pd_set);
 750                 pd->pd_recursion--;
 751                 return 0;
 752         }
 753
 754         /* have we reached ping point? */
 755         if (!pd->pd_set && cfs_time_before(curtime, pd->pd_next_ping)) {
 756                 pd->pd_recursion--;
 757                 return 0;
 758         }
 759
 760         /* if we have rpc_set already, continue processing it */
 761         if (pd->pd_set) {
 762                 LASSERT(pd->pd_this_ping);
 763                 set = pd->pd_set;
 764                 goto do_check_set;
 765         }
 766
 767         pd->pd_this_ping = curtime;
 768         pd->pd_set = ptlrpc_prep_set();
 769         if (pd->pd_set == NULL)
 770                 goto out;
 771         set = pd->pd_set;
 772
 773         /* add rpcs into set */
 774         cfs_mutex_down(&pinger_sem);
 775         cfs_list_for_each(iter, &pinger_imports) {
 776                 struct obd_import *imp = cfs_list_entry(iter, struct obd_import,
 777                                                         imp_pinger_chain);
 778                 int generation, level;
 779
 780                 if (cfs_time_aftereq(pd->pd_this_ping,
 781                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 782                         /* Add a ping. */
 783                         cfs_spin_lock(&imp->imp_lock);
 784                         generation = imp->imp_generation;
 785                         level = imp->imp_state;
 786                         cfs_spin_unlock(&imp->imp_lock);
 787
 788                         if (level != LUSTRE_IMP_FULL) {
 789                                 CDEBUG(D_HA,
 790                                        "not pinging %s (in recovery)\n",
 791                                        obd2cli_tgt(imp->imp_obd));
 792                                 continue;
 793                         }
 794
 795                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 796                                                         LUSTRE_OBD_VERSION,
 797                                                         OBD_PING);
 798                         if (req == NULL) {
 799                                 CERROR("OOM trying to ping %s->%s\n",
 800                                        imp->imp_obd->obd_uuid.uuid,
 801                                        obd2cli_tgt(imp->imp_obd));
 802                                 break;
 803                         }
 804
 805                         req->rq_no_resend = 1;
 806                         ptlrpc_request_set_replen(req);
 807                         req->rq_send_state = LUSTRE_IMP_FULL;
 808                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 809                         req->rq_import_generation = generation;
 810                         ptlrpc_set_add_req(set, req);
 811                 } else {
 812                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 813                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 814                                imp->imp_next_ping, pd->pd_this_ping);
 815                 }
 816         }
 817         pd->pd_this_ping = curtime;
 818         cfs_mutex_up(&pinger_sem);
 819
 820         /* Might be empty, that's OK. */
 821         if (cfs_atomic_read(&set->set_remaining) == 0)
 822                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 823
 824         cfs_list_for_each(iter, &set->set_requests) {
 825                 struct ptlrpc_request *req =
 826                         cfs_list_entry(iter, struct ptlrpc_request,
 827                                        rq_set_chain);
 828                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 829                           req->rq_import->imp_obd->obd_uuid.uuid,
 830                           obd2cli_tgt(req->rq_import->imp_obd));
 831                 (void)ptl_send_rpc(req, 0);
 832         }
 833
 834 do_check_set:
 835         rc = ptlrpc_check_set(NULL, set);
 836
 837         /* not finished, and we are not expired, simply return */
 838         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 839                                             cfs_time_seconds(PING_INTERVAL)))) {
 840                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 841                 pd->pd_recursion--;
 842                 return 0;
 843         }
 844
 845         /* Expire all the requests that didn't come back. */
 846         cfs_mutex_down(&pinger_sem);
 847         cfs_list_for_each(iter, &set->set_requests) {
 848                 req = cfs_list_entry(iter, struct ptlrpc_request,
 849                                      rq_set_chain);
 850
 851                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 852                         continue;
 853
 854                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 855                        req);
 856
 857                 /* This will also unregister reply. */
 858                 ptlrpc_expire_one_request(req, 0);
 859
 860                 /* We're done with this req, let's finally move it to complete
 861                  * phase and take care of inflights. */
 862                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 863                 imp = req->rq_import;
 864                 cfs_spin_lock(&imp->imp_lock);
 865                 if (!cfs_list_empty(&req->rq_list)) {
 866                         cfs_list_del_init(&req->rq_list);
 867                         cfs_atomic_dec(&imp->imp_inflight);
 868                 }
 869                 cfs_spin_unlock(&imp->imp_lock);
 870                 cfs_atomic_dec(&set->set_remaining);
 871         }
 872         cfs_mutex_up(&pinger_sem);
 873
 874         ptlrpc_set_destroy(set);
 875         pd->pd_set = NULL;
 876
 877 out:
 878         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 879                                         cfs_time_seconds(PING_INTERVAL));
 880         pd->pd_this_ping = 0; /* XXX for debug */
 881
 882         CDEBUG(D_INFO, "finished a round ping\n");
 883         pd->pd_recursion--;
 884         return 0;
 885 }
 886
 887 static void *pinger_callback = NULL;
 888 #endif /* ENABLE_PINGER */
 889
 890 int ptlrpc_start_pinger(void)
 891 {
 892 #ifdef ENABLE_PINGER
 893         memset(&pinger_args, 0, sizeof(pinger_args));
 894         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 895                                                            &pinger_check_rpcs,
 896                                                            &pinger_args);
 897 #endif
 898         return 0;
 899 }
 900
 901 int ptlrpc_stop_pinger(void)
 902 {
 903 #ifdef ENABLE_PINGER
 904         if (pinger_callback)
 905                 liblustre_deregister_wait_callback(pinger_callback);
 906 #endif
 907         return 0;
 908 }
 909
 910 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 911 {
 912 #ifdef ENABLE_PINGER
 913         cfs_mutex_down(&pinger_sem);
 914         ptlrpc_update_next_ping(imp, 0);
 915         if (pinger_args.pd_set == NULL &&
 916             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 917                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 918                         imp->imp_next_ping, cfs_time_current());
 919                 pinger_args.pd_next_ping = imp->imp_next_ping;
 920         }
 921         cfs_mutex_up(&pinger_sem);
 922 #endif
 923 }
 924
 925 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
 926 {
 927 #ifdef ENABLE_PINGER
 928         cfs_mutex_down(&pinger_sem);
 929         ptlrpc_update_next_ping(imp, 1);
 930         if (pinger_args.pd_set == NULL &&
 931             cfs_time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 932                 CDEBUG(D_HA,"set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 933                         imp->imp_next_ping, cfs_time_current());
 934                 pinger_args.pd_next_ping = imp->imp_next_ping;
 935         }
 936         cfs_mutex_up(&pinger_sem);
 937 #endif
 938 }
 939
 940 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
 941                               timeout_cb_t cb, void *data,
 942                               cfs_list_t *obd_list)
 943 {
 944         return 0;
 945 }
 946
 947 int ptlrpc_del_timeout_client(cfs_list_t *obd_list,
 948                               enum timeout_event event)
 949 {
 950         return 0;
 951 }
 952
 953 int ptlrpc_pinger_add_import(struct obd_import *imp)
 954 {
 955         ENTRY;
 956         if (!cfs_list_empty(&imp->imp_pinger_chain))
 957                 RETURN(-EALREADY);
 958
 959         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 960                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 961         ptlrpc_pinger_sending_on_import(imp);
 962
 963         cfs_mutex_down(&pinger_sem);
 964         cfs_list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 965         class_import_get(imp);
 966         cfs_mutex_up(&pinger_sem);
 967
 968         RETURN(0);
 969 }
 970
 971 int ptlrpc_pinger_del_import(struct obd_import *imp)
 972 {
 973         ENTRY;
 974         if (cfs_list_empty(&imp->imp_pinger_chain))
 975                 RETURN(-ENOENT);
 976
 977         cfs_mutex_down(&pinger_sem);
 978         cfs_list_del_init(&imp->imp_pinger_chain);
 979         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 980                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 981         class_import_put(imp);
 982         cfs_mutex_up(&pinger_sem);
 983         RETURN(0);
 984 }
 985
 986 void ptlrpc_pinger_wake_up()
 987 {
 988 #ifdef ENABLE_PINGER
 989         /* XXX force pinger to run, if needed */
 990         struct obd_import *imp;
 991         ENTRY;
 992         cfs_list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 993                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 994                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 995 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 996                 if (imp->imp_state == LUSTRE_IMP_DISCON &&
 997                     !imp_is_deactive(imp))
 998 #else
 999                 /*XXX only recover for the initial connection */
1000                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
1001                     imp->imp_state == LUSTRE_IMP_DISCON &&
1002                     !imp_is_deactive(imp))
1003 #endif
1004                         ptlrpc_initiate_recovery(imp);
1005                 else if (imp->imp_state != LUSTRE_IMP_FULL)
1006                         CDEBUG(D_HA, "Refused to recover import %s->%s "
1007                                      "state %d, deactive %d\n",
1008                                      imp->imp_obd->obd_uuid.uuid,
1009                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
1010                                      imp_is_deactive(imp));
1011         }
1012         EXIT;
1013 #endif
1014 }
1015 #endif /* !__KERNEL__ */