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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/pinger.c
  37  *
  38  * Portal-RPC reconnection and replay operations, for use in recovery.
  39  */
  40
  41 #ifndef __KERNEL__
  42 #include <liblustre.h>
  43 #else
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #endif
  46
  47 #include <obd_support.h>
  48 #include <obd_class.h>
  49 #include "ptlrpc_internal.h"
  50
  51 struct semaphore pinger_sem;
  52 static CFS_LIST_HEAD(pinger_imports);
  53
  54 struct ptlrpc_request *
  55 ptlrpc_prep_ping(struct obd_import *imp)
  56 {
  57         struct ptlrpc_request *req;
  58
  59         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
  60                                         LUSTRE_OBD_VERSION, OBD_PING);
  61         if (req) {
  62                 ptlrpc_request_set_replen(req);
  63                 req->rq_no_resend = req->rq_no_delay = 1;
  64         }
  65         return req;
  66 }
  67
  68 int ptlrpc_obd_ping(struct obd_device *obd)
  69 {
  70         int rc;
  71         struct ptlrpc_request *req;
  72         ENTRY;
  73
  74         req = ptlrpc_prep_ping(obd->u.cli.cl_import);
  75         if (req == NULL)
  76                 RETURN(-ENOMEM);
  77
  78         req->rq_send_state = LUSTRE_IMP_FULL;
  79
  80         rc = ptlrpc_queue_wait(req);
  81
  82         ptlrpc_req_finished(req);
  83
  84         RETURN(rc);
  85 }
  86 EXPORT_SYMBOL(ptlrpc_obd_ping);
  87
  88 int ptlrpc_ping(struct obd_import *imp)
  89 {
  90         struct ptlrpc_request *req;
  91         ENTRY;
  92
  93         req = ptlrpc_prep_ping(imp);
  94         if (req == NULL) {
  95                 CERROR("OOM trying to ping %s->%s\n",
  96                        imp->imp_obd->obd_uuid.uuid,
  97                        obd2cli_tgt(imp->imp_obd));
  98                 RETURN(-ENOMEM);
  99         }
 100
 101         DEBUG_REQ(D_INFO, req, "pinging %s->%s",
 102                   imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 103         ptlrpcd_add_req(req, PSCOPE_OTHER);
 104
 105         RETURN(0);
 106 }
 107
 108 void ptlrpc_update_next_ping(struct obd_import *imp)
 109 {
 110 #ifdef ENABLE_PINGER
 111         int time = PING_INTERVAL;
 112         if (imp->imp_state == LUSTRE_IMP_DISCON) {
 113                 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
 114                                   AT_OFF ? 0 :
 115                                   at_get(&imp->imp_at.iat_net_latency));
 116                 time = min(time, dtime);
 117         }
 118         imp->imp_next_ping = cfs_time_shift(time);
 119 #endif /* ENABLE_PINGER */
 120 }
 121
 122 void ptlrpc_ping_import_soon(struct obd_import *imp)
 123 {
 124         imp->imp_next_ping = cfs_time_current();
 125 }
 126
 127 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
 128 {
 129         if (imp->imp_server_timeout)
 130                 return cfs_time_shift(obd_timeout / 2);
 131         else
 132                 return cfs_time_shift(obd_timeout);
 133 }
 134
 135 static atomic_t suspend_timeouts = ATOMIC_INIT(0);
 136 static cfs_time_t suspend_wakeup_time = 0;
 137
 138 #ifdef __KERNEL__
 139 static wait_queue_head_t suspend_timeouts_waitq;
 140 #endif
 141
 142 cfs_time_t ptlrpc_suspend_wakeup_time(void)
 143 {
 144         return suspend_wakeup_time;
 145 }
 146
 147 void ptlrpc_deactivate_timeouts(struct obd_import *imp)
 148 {
 149         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 150 #if 0
 151         if (imp->imp_no_timeout)
 152                 return;
 153         imp->imp_no_timeout = 1;
 154         atomic_inc(&suspend_timeouts);
 155         CDEBUG(D_HA|D_WARNING, "deactivate timeouts %u\n", atomic_read(&suspend_timeouts));
 156 #endif
 157 }
 158
 159 void ptlrpc_activate_timeouts(struct obd_import *imp)
 160 {
 161         /*XXX: disabled for now, will be replaced by adaptive timeouts */
 162 #if 0
 163         if (!imp->imp_no_timeout)
 164                 return;
 165         imp->imp_no_timeout = 0;
 166         LASSERT(atomic_read(&suspend_timeouts) > 0);
 167         if (atomic_dec_and_test(&suspend_timeouts)) {
 168                 suspend_wakeup_time = cfs_time_current();
 169                 wake_up(&suspend_timeouts_waitq);
 170         }
 171         CDEBUG(D_HA|D_WARNING, "activate timeouts %u\n", atomic_read(&suspend_timeouts));
 172 #endif
 173 }
 174
 175 int ptlrpc_check_suspend(void)
 176 {
 177         if (atomic_read(&suspend_timeouts))
 178                 return 1;
 179         return 0;
 180 }
 181
 182 int ptlrpc_check_and_wait_suspend(struct ptlrpc_request *req)
 183 {
 184         struct l_wait_info lwi;
 185
 186         if (atomic_read(&suspend_timeouts)) {
 187                 DEBUG_REQ(D_NET, req, "-- suspend %d regular timeout",
 188                           atomic_read(&suspend_timeouts));
 189                 lwi = LWI_INTR(NULL, NULL);
 190                 l_wait_event(suspend_timeouts_waitq,
 191                              atomic_read(&suspend_timeouts) == 0, &lwi);
 192                 DEBUG_REQ(D_NET, req, "-- recharge regular timeout");
 193                 return 1;
 194         }
 195         return 0;
 196 }
 197
 198 #ifdef __KERNEL__
 199
 200 static void ptlrpc_pinger_process_import(struct obd_import *imp,
 201                                          unsigned long this_ping)
 202 {
 203         int force, level;
 204
 205         spin_lock(&imp->imp_lock);
 206         level = imp->imp_state;
 207         force = imp->imp_force_verify;
 208         if (force)
 209                 imp->imp_force_verify = 0;
 210         spin_unlock(&imp->imp_lock);
 211
 212         CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA,
 213                "level %s/%u force %u deactive %u pingable %u\n",
 214                ptlrpc_import_state_name(level), level,
 215                force, imp->imp_deactive, imp->imp_pingable);
 216
 217         if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK,
 218                              this_ping) && force == 0)
 219                 return;
 220
 221         if (level == LUSTRE_IMP_DISCON && !imp->imp_deactive) {
 222                 /* wait at least a timeout before trying recovery again */
 223                 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
 224                 ptlrpc_initiate_recovery(imp);
 225         } else if (level != LUSTRE_IMP_FULL ||
 226                    imp->imp_obd->obd_no_recov ||
 227                    imp->imp_deactive) {
 228                 CDEBUG(D_HA, "not pinging %s (in recovery "
 229                        " or recovery disabled: %s)\n",
 230                        obd2cli_tgt(imp->imp_obd),
 231                        ptlrpc_import_state_name(level));
 232         } else if (imp->imp_pingable || force) {
 233                 ptlrpc_ping(imp);
 234         }
 235 }
 236
 237 static int ptlrpc_pinger_main(void *arg)
 238 {
 239         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 240         struct ptlrpc_thread *thread = data->thread;
 241         ENTRY;
 242
 243         cfs_daemonize(data->name);
 244
 245         /* Record that the thread is running */
 246         thread->t_flags = SVC_RUNNING;
 247         cfs_waitq_signal(&thread->t_ctl_waitq);
 248
 249         /* And now, loop forever, pinging as needed. */
 250         while (1) {
 251                 cfs_time_t this_ping = cfs_time_current();
 252                 struct l_wait_info lwi;
 253                 cfs_duration_t time_to_next_ping;
 254                 struct list_head *iter;
 255
 256                 mutex_down(&pinger_sem);
 257                 list_for_each(iter, &pinger_imports) {
 258                         struct obd_import *imp =
 259                                 list_entry(iter, struct obd_import,
 260                                            imp_pinger_chain);
 261
 262                         ptlrpc_pinger_process_import(imp, this_ping);
 263                         /* obd_timeout might have changed */
 264                         if (imp->imp_pingable && imp->imp_next_ping &&
 265                             cfs_time_after(imp->imp_next_ping,
 266                                            cfs_time_add(this_ping,
 267                                                         cfs_time_seconds(PING_INTERVAL))))
 268                                 ptlrpc_update_next_ping(imp);
 269                 }
 270                 mutex_up(&pinger_sem);
 271                 /* update memory usage info */
 272                 obd_update_maxusage();
 273
 274                 /* Wait until the next ping time, or until we're stopped. */
 275                 time_to_next_ping = cfs_time_sub(cfs_time_add(this_ping,
 276                                                cfs_time_seconds(PING_INTERVAL)),
 277                                                cfs_time_current());
 278
 279                 /* The ping sent by ptlrpc_send_rpc may get sent out
 280                    say .01 second after this.
 281                    ptlrpc_pinger_sending_on_import will then set the
 282                    next ping time to next_ping + .01 sec, which means
 283                    we will SKIP the next ping at next_ping, and the
 284                    ping will get sent 2 timeouts from now!  Beware. */
 285                 CDEBUG(D_INFO, "next ping in "CFS_DURATION_T" ("CFS_TIME_T")\n",
 286                                time_to_next_ping,
 287                                cfs_time_add(this_ping,
 288                                             cfs_time_seconds(PING_INTERVAL)));
 289                 if (time_to_next_ping > 0) {
 290                         lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
 291                                                 time_to_next_ping,
 292                                                 cfs_time_seconds(1)),
 293                                           NULL, NULL);
 294                         l_wait_event(thread->t_ctl_waitq,
 295                                      thread->t_flags & (SVC_STOPPING|SVC_EVENT),
 296                                      &lwi);
 297                         if (thread->t_flags & SVC_STOPPING) {
 298                                 thread->t_flags &= ~SVC_STOPPING;
 299                                 EXIT;
 300                                 break;
 301                         } else if (thread->t_flags & SVC_EVENT) {
 302                                 /* woken after adding import to reset timer */
 303                                 thread->t_flags &= ~SVC_EVENT;
 304                         }
 305                 }
 306         }
 307
 308         thread->t_flags = SVC_STOPPED;
 309         cfs_waitq_signal(&thread->t_ctl_waitq);
 310
 311         CDEBUG(D_NET, "pinger thread exiting, process %d\n", cfs_curproc_pid());
 312         return 0;
 313 }
 314
 315 static struct ptlrpc_thread *pinger_thread = NULL;
 316
 317 int ptlrpc_start_pinger(void)
 318 {
 319         struct l_wait_info lwi = { 0 };
 320         struct ptlrpc_svc_data d;
 321         int rc;
 322 #ifndef ENABLE_PINGER
 323         return 0;
 324 #endif
 325         ENTRY;
 326
 327         if (pinger_thread != NULL)
 328                 RETURN(-EALREADY);
 329
 330         OBD_ALLOC_PTR(pinger_thread);
 331         if (pinger_thread == NULL)
 332                 RETURN(-ENOMEM);
 333         cfs_waitq_init(&pinger_thread->t_ctl_waitq);
 334         cfs_waitq_init(&suspend_timeouts_waitq);
 335
 336         d.name = "ll_ping";
 337         d.thread = pinger_thread;
 338
 339         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
 340          * just drop the VM and FILES in ptlrpc_daemonize() right away. */
 341         rc = cfs_kernel_thread(ptlrpc_pinger_main, &d, CLONE_VM | CLONE_FILES);
 342         if (rc < 0) {
 343                 CERROR("cannot start thread: %d\n", rc);
 344                 OBD_FREE(pinger_thread, sizeof(*pinger_thread));
 345                 pinger_thread = NULL;
 346                 RETURN(rc);
 347         }
 348         l_wait_event(pinger_thread->t_ctl_waitq,
 349                      pinger_thread->t_flags & SVC_RUNNING, &lwi);
 350
 351         RETURN(0);
 352 }
 353
 354 int ptlrpc_stop_pinger(void)
 355 {
 356         struct l_wait_info lwi = { 0 };
 357         int rc = 0;
 358 #ifndef ENABLE_PINGER
 359         return 0;
 360 #endif
 361         ENTRY;
 362
 363         if (pinger_thread == NULL)
 364                 RETURN(-EALREADY);
 365         mutex_down(&pinger_sem);
 366         pinger_thread->t_flags = SVC_STOPPING;
 367         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 368         mutex_up(&pinger_sem);
 369
 370         l_wait_event(pinger_thread->t_ctl_waitq,
 371                      (pinger_thread->t_flags & SVC_STOPPED), &lwi);
 372
 373         OBD_FREE_PTR(pinger_thread);
 374         pinger_thread = NULL;
 375         RETURN(rc);
 376 }
 377
 378 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 379 {
 380         ptlrpc_update_next_ping(imp);
 381 }
 382
 383 int ptlrpc_pinger_add_import(struct obd_import *imp)
 384 {
 385         ENTRY;
 386         if (!list_empty(&imp->imp_pinger_chain))
 387                 RETURN(-EALREADY);
 388
 389         mutex_down(&pinger_sem);
 390         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 391                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 392         /* if we add to pinger we want recovery on this import */
 393         imp->imp_obd->obd_no_recov = 0;
 394         ptlrpc_update_next_ping(imp);
 395         /* XXX sort, blah blah */
 396         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 397         class_import_get(imp);
 398
 399         ptlrpc_pinger_wake_up();
 400         mutex_up(&pinger_sem);
 401
 402         RETURN(0);
 403 }
 404
 405 int ptlrpc_pinger_del_import(struct obd_import *imp)
 406 {
 407         ENTRY;
 408         if (list_empty(&imp->imp_pinger_chain))
 409                 RETURN(-ENOENT);
 410
 411         mutex_down(&pinger_sem);
 412         list_del_init(&imp->imp_pinger_chain);
 413         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 414                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 415         /* if we remove from pinger we don't want recovery on this import */
 416         imp->imp_obd->obd_no_recov = 1;
 417         class_import_put(imp);
 418         mutex_up(&pinger_sem);
 419         RETURN(0);
 420 }
 421
 422 void ptlrpc_pinger_wake_up()
 423 {
 424 #ifdef ENABLE_PINGER
 425         pinger_thread->t_flags |= SVC_EVENT;
 426         cfs_waitq_signal(&pinger_thread->t_ctl_waitq);
 427 #endif
 428 }
 429
 430 /* Ping evictor thread */
 431 #define PET_READY     1
 432 #define PET_TERMINATE 2
 433
 434 static int               pet_refcount = 0;
 435 static int               pet_state;
 436 static wait_queue_head_t pet_waitq;
 437 static struct obd_export *pet_exp = NULL;
 438 static spinlock_t        pet_lock = SPIN_LOCK_UNLOCKED;
 439
 440 int ping_evictor_wake(struct obd_export *exp)
 441 {
 442         spin_lock(&pet_lock);
 443         if (pet_exp || (pet_state != PET_READY)) {
 444                 /* eventually the new obd will call here again. */
 445                 spin_unlock(&pet_lock);
 446                 return 1;
 447         }
 448
 449         /* We have to make sure the obd isn't destroyed between now and when
 450          * the ping evictor runs.  We'll take a reference here, and drop it
 451          * when we finish in the evictor.  We don't really care about this
 452          * export in particular; we just need one to keep the obd alive. */
 453         pet_exp = class_export_get(exp);
 454         spin_unlock(&pet_lock);
 455
 456         wake_up(&pet_waitq);
 457         return 0;
 458 }
 459
 460 static int ping_evictor_main(void *arg)
 461 {
 462         struct obd_device *obd;
 463         struct obd_export *exp;
 464         struct l_wait_info lwi = { 0 };
 465         time_t expire_time;
 466         ENTRY;
 467
 468         ptlrpc_daemonize("ll_evictor");
 469
 470         CDEBUG(D_HA, "Starting Ping Evictor\n");
 471         pet_exp = NULL;
 472         pet_state = PET_READY;
 473         while (1) {
 474                 l_wait_event(pet_waitq, pet_exp ||
 475                              (pet_state == PET_TERMINATE), &lwi);
 476                 if (pet_state == PET_TERMINATE)
 477                         break;
 478
 479                 /* we only get here if pet_exp != NULL, and the end of this
 480                  * loop is the only place which sets it NULL again, so lock
 481                  * is not strictly necessary. */
 482                 spin_lock(&pet_lock);
 483                 obd = pet_exp->exp_obd;
 484                 spin_unlock(&pet_lock);
 485
 486                 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
 487
 488                 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
 489                        obd->obd_name, expire_time);
 490
 491                 /* Exports can't be deleted out of the list while we hold
 492                  * the obd lock (class_unlink_export), which means we can't
 493                  * lose the last ref on the export.  If they've already been
 494                  * removed from the list, we won't find them here. */
 495                 spin_lock(&obd->obd_dev_lock);
 496                 while (!list_empty(&obd->obd_exports_timed)) {
 497                         exp = list_entry(obd->obd_exports_timed.next,
 498                                          struct obd_export,exp_obd_chain_timed);
 499                         if (expire_time > exp->exp_last_request_time) {
 500                                 class_export_get(exp);
 501                                 spin_unlock(&obd->obd_dev_lock);
 502                                  LCONSOLE_WARN("%s: haven't heard from client %s"
 503                                               " (at %s) in %ld seconds. I think"
 504                                               " it's dead, and I am evicting"
 505                                               " it. exp %p, cur %ld expire %ld"
 506                                               " last %ld\n",
 507                                               obd->obd_name,
 508                                               obd_uuid2str(&exp->exp_client_uuid),
 509                                               obd_export_nid2str(exp),
 510                                               (long)(cfs_time_current_sec() -
 511                                                      exp->exp_last_request_time),
 512                                               exp, (long)cfs_time_current_sec(),
 513                                               (long)expire_time,
 514                                               (long)exp->exp_last_request_time);
 515                                 CDEBUG(D_HA, "Last request was at %ld\n",
 516                                        exp->exp_last_request_time);
 517                                 class_fail_export(exp);
 518                                 class_export_put(exp);
 519                                 spin_lock(&obd->obd_dev_lock);
 520                         } else {
 521                                 /* List is sorted, so everyone below is ok */
 522                                 break;
 523                         }
 524                 }
 525                 spin_unlock(&obd->obd_dev_lock);
 526
 527                 class_export_put(pet_exp);
 528
 529                 spin_lock(&pet_lock);
 530                 pet_exp = NULL;
 531                 spin_unlock(&pet_lock);
 532         }
 533         CDEBUG(D_HA, "Exiting Ping Evictor\n");
 534
 535         RETURN(0);
 536 }
 537
 538 void ping_evictor_start(void)
 539 {
 540         int rc;
 541
 542         if (++pet_refcount > 1)
 543                 return;
 544
 545         init_waitqueue_head(&pet_waitq);
 546
 547         rc = cfs_kernel_thread(ping_evictor_main, NULL, CLONE_VM | CLONE_FILES);
 548         if (rc < 0) {
 549                 pet_refcount--;
 550                 CERROR("Cannot start ping evictor thread: %d\n", rc);
 551         }
 552 }
 553 EXPORT_SYMBOL(ping_evictor_start);
 554
 555 void ping_evictor_stop(void)
 556 {
 557         if (--pet_refcount > 0)
 558                 return;
 559
 560         pet_state = PET_TERMINATE;
 561         wake_up(&pet_waitq);
 562 }
 563 EXPORT_SYMBOL(ping_evictor_stop);
 564 #else /* !__KERNEL__ */
 565
 566 /* XXX
 567  * the current implementation of pinger in liblustre is not optimized
 568  */
 569
 570 #ifdef ENABLE_PINGER
 571 static struct pinger_data {
 572         int             pd_recursion;
 573         cfs_time_t      pd_this_ping;   /* jiffies */
 574         cfs_time_t      pd_next_ping;   /* jiffies */
 575         struct ptlrpc_request_set *pd_set;
 576 } pinger_args;
 577
 578 static int pinger_check_rpcs(void *arg)
 579 {
 580         cfs_time_t curtime = cfs_time_current();
 581         struct ptlrpc_request *req;
 582         struct ptlrpc_request_set *set;
 583         struct list_head *iter;
 584         struct obd_import *imp;
 585         struct pinger_data *pd = &pinger_args;
 586         int rc;
 587
 588         /* prevent recursion */
 589         if (pd->pd_recursion++) {
 590                 CDEBUG(D_HA, "pinger: recursion! quit\n");
 591                 LASSERT(pd->pd_set);
 592                 pd->pd_recursion--;
 593                 return 0;
 594         }
 595
 596         /* have we reached ping point? */
 597         if (!pd->pd_set && time_before(curtime, pd->pd_next_ping)) {
 598                 pd->pd_recursion--;
 599                 return 0;
 600         }
 601
 602         /* if we have rpc_set already, continue processing it */
 603         if (pd->pd_set) {
 604                 LASSERT(pd->pd_this_ping);
 605                 set = pd->pd_set;
 606                 goto do_check_set;
 607         }
 608
 609         pd->pd_this_ping = curtime;
 610         pd->pd_set = ptlrpc_prep_set();
 611         if (pd->pd_set == NULL)
 612                 goto out;
 613         set = pd->pd_set;
 614
 615         /* add rpcs into set */
 616         mutex_down(&pinger_sem);
 617         list_for_each(iter, &pinger_imports) {
 618                 struct obd_import *imp =
 619                         list_entry(iter, struct obd_import, imp_pinger_chain);
 620                 int generation, level;
 621
 622                 if (cfs_time_aftereq(pd->pd_this_ping,
 623                                      imp->imp_next_ping - 5 * CFS_TICK)) {
 624                         /* Add a ping. */
 625                         spin_lock(&imp->imp_lock);
 626                         generation = imp->imp_generation;
 627                         level = imp->imp_state;
 628                         spin_unlock(&imp->imp_lock);
 629
 630                         if (level != LUSTRE_IMP_FULL) {
 631                                 CDEBUG(D_HA,
 632                                        "not pinging %s (in recovery)\n",
 633                                        obd2cli_tgt(imp->imp_obd));
 634                                 continue;
 635                         }
 636
 637                         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
 638                                                         LUSTRE_OBD_VERSION,
 639                                                         OBD_PING);
 640                         if (req == NULL) {
 641                                 CERROR("OOM trying to ping %s->%s\n",
 642                                        imp->imp_obd->obd_uuid.uuid,
 643                                        obd2cli_tgt(imp->imp_obd));
 644                                 break;
 645                         }
 646
 647                         req->rq_no_resend = 1;
 648                         ptlrpc_request_set_replen(req);
 649                         req->rq_send_state = LUSTRE_IMP_FULL;
 650                         ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
 651                         req->rq_import_generation = generation;
 652                         ptlrpc_set_add_req(set, req);
 653                 } else {
 654                         CDEBUG(D_INFO, "don't need to ping %s ("CFS_TIME_T
 655                                " > "CFS_TIME_T")\n", obd2cli_tgt(imp->imp_obd),
 656                                imp->imp_next_ping, pd->pd_this_ping);
 657                 }
 658         }
 659         pd->pd_this_ping = curtime;
 660         mutex_up(&pinger_sem);
 661
 662         /* Might be empty, that's OK. */
 663         if (set->set_remaining == 0)
 664                 CDEBUG(D_RPCTRACE, "nothing to ping\n");
 665
 666         list_for_each(iter, &set->set_requests) {
 667                 struct ptlrpc_request *req =
 668                         list_entry(iter, struct ptlrpc_request,
 669                                    rq_set_chain);
 670                 DEBUG_REQ(D_RPCTRACE, req, "pinging %s->%s",
 671                           req->rq_import->imp_obd->obd_uuid.uuid,
 672                           obd2cli_tgt(req->rq_import->imp_obd));
 673                 (void)ptl_send_rpc(req, 0);
 674         }
 675
 676 do_check_set:
 677         rc = ptlrpc_check_set(NULL, set);
 678
 679         /* not finished, and we are not expired, simply return */
 680         if (!rc && cfs_time_before(curtime, cfs_time_add(pd->pd_this_ping,
 681                                             cfs_time_seconds(PING_INTERVAL)))) {
 682                 CDEBUG(D_RPCTRACE, "not finished, but also not expired\n");
 683                 pd->pd_recursion--;
 684                 return 0;
 685         }
 686
 687         /* Expire all the requests that didn't come back. */
 688         mutex_down(&pinger_sem);
 689         list_for_each(iter, &set->set_requests) {
 690                 req = list_entry(iter, struct ptlrpc_request,
 691                                  rq_set_chain);
 692
 693                 if (req->rq_phase == RQ_PHASE_COMPLETE)
 694                         continue;
 695
 696                 CDEBUG(D_RPCTRACE, "Pinger initiate expire request(%p)\n",
 697                        req);
 698
 699                 /* This will also unregister reply. */
 700                 ptlrpc_expire_one_request(req, 0);
 701
 702                 /* We're done with this req, let's finally move it to complete
 703                  * phase and take care of inflights. */
 704                 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
 705                 imp = req->rq_import;
 706                 spin_lock(&imp->imp_lock);
 707                 if (!list_empty(&req->rq_list)) {
 708                         list_del_init(&req->rq_list);
 709                         atomic_dec(&imp->imp_inflight);
 710                 }
 711                 spin_unlock(&imp->imp_lock);
 712                 set->set_remaining--;
 713         }
 714         mutex_up(&pinger_sem);
 715
 716         ptlrpc_set_destroy(set);
 717         pd->pd_set = NULL;
 718
 719 out:
 720         pd->pd_next_ping = cfs_time_add(pd->pd_this_ping,
 721                                         cfs_time_seconds(PING_INTERVAL));
 722         pd->pd_this_ping = 0; /* XXX for debug */
 723
 724         CDEBUG(D_INFO, "finished a round ping\n");
 725         pd->pd_recursion--;
 726         return 0;
 727 }
 728
 729 static void *pinger_callback = NULL;
 730 #endif /* ENABLE_PINGER */
 731
 732 int ptlrpc_start_pinger(void)
 733 {
 734 #ifdef ENABLE_PINGER
 735         memset(&pinger_args, 0, sizeof(pinger_args));
 736         pinger_callback = liblustre_register_wait_callback("pinger_check_rpcs",
 737                                                            &pinger_check_rpcs,
 738                                                            &pinger_args);
 739 #endif
 740         return 0;
 741 }
 742
 743 int ptlrpc_stop_pinger(void)
 744 {
 745 #ifdef ENABLE_PINGER
 746         if (pinger_callback)
 747                 liblustre_deregister_wait_callback(pinger_callback);
 748 #endif
 749         return 0;
 750 }
 751
 752 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
 753 {
 754 #ifdef ENABLE_PINGER
 755         mutex_down(&pinger_sem);
 756         ptlrpc_update_next_ping(imp);
 757         if (pinger_args.pd_set == NULL &&
 758             time_before(imp->imp_next_ping, pinger_args.pd_next_ping)) {
 759                 CDEBUG(D_HA, "set next ping to "CFS_TIME_T"(cur "CFS_TIME_T")\n",
 760                         imp->imp_next_ping, cfs_time_current());
 761                 pinger_args.pd_next_ping = imp->imp_next_ping;
 762         }
 763         mutex_up(&pinger_sem);
 764 #endif
 765 }
 766
 767 int ptlrpc_pinger_add_import(struct obd_import *imp)
 768 {
 769         ENTRY;
 770         if (!list_empty(&imp->imp_pinger_chain))
 771                 RETURN(-EALREADY);
 772
 773         CDEBUG(D_HA, "adding pingable import %s->%s\n",
 774                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 775         ptlrpc_pinger_sending_on_import(imp);
 776
 777         mutex_down(&pinger_sem);
 778         list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
 779         class_import_get(imp);
 780         mutex_up(&pinger_sem);
 781
 782         RETURN(0);
 783 }
 784
 785 int ptlrpc_pinger_del_import(struct obd_import *imp)
 786 {
 787         ENTRY;
 788         if (list_empty(&imp->imp_pinger_chain))
 789                 RETURN(-ENOENT);
 790
 791         mutex_down(&pinger_sem);
 792         list_del_init(&imp->imp_pinger_chain);
 793         CDEBUG(D_HA, "removing pingable import %s->%s\n",
 794                imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 795         class_import_put(imp);
 796         mutex_up(&pinger_sem);
 797         RETURN(0);
 798 }
 799
 800 void ptlrpc_pinger_wake_up()
 801 {
 802 #ifdef ENABLE_PINGER
 803         /* XXX force pinger to run, if needed */
 804         struct obd_import *imp;
 805         ENTRY;
 806         list_for_each_entry(imp, &pinger_imports, imp_pinger_chain) {
 807                 CDEBUG(D_RPCTRACE, "checking import %s->%s\n",
 808                        imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
 809 #ifdef ENABLE_LIBLUSTRE_RECOVERY
 810                 if (imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 811 #else
 812                 /*XXX only recover for the initial connection */
 813                 if (!lustre_handle_is_used(&imp->imp_remote_handle) &&
 814                     imp->imp_state == LUSTRE_IMP_DISCON && !imp->imp_deactive)
 815 #endif
 816                         ptlrpc_initiate_recovery(imp);
 817                 else if (imp->imp_state != LUSTRE_IMP_FULL)
 818                         CDEBUG(D_HA, "Refused to recover import %s->%s "
 819                                      "state %d, deactive %d\n",
 820                                      imp->imp_obd->obd_uuid.uuid,
 821                                      obd2cli_tgt(imp->imp_obd), imp->imp_state,
 822                                      imp->imp_deactive);
 823         }
 824         EXIT;
 825 #endif
 826 }
 827 #endif /* !__KERNEL__ */