lustre/ldlm/ldlm_lockd.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/ldlm/ldlm_lockd.c
  37  *
  38  * Author: Peter Braam <braam@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42 #ifndef EXPORT_SYMTAB
  43 # define EXPORT_SYMTAB
  44 #endif
  45 #define DEBUG_SUBSYSTEM S_LDLM
  46
  47 #ifdef __KERNEL__
  48 # include <libcfs/libcfs.h>
  49 #else
  50 # include <liblustre.h>
  51 #endif
  52
  53 #include <lustre_dlm.h>
  54 #include <obd_class.h>
  55 #include <libcfs/list.h>
  56 #include "ldlm_internal.h"
  57
  58 #ifdef __KERNEL__
  59 static int ldlm_num_threads;
  60 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
  61                 "number of DLM service threads to start");
  62 #endif
  63
  64 extern cfs_mem_cache_t *ldlm_resource_slab;
  65 extern cfs_mem_cache_t *ldlm_lock_slab;
  66 static struct semaphore ldlm_ref_sem;
  67 static int ldlm_refcount;
  68
  69 /* LDLM state */
  70
  71 static struct ldlm_state *ldlm_state;
  72
  73 inline cfs_time_t round_timeout(cfs_time_t timeout)
  74 {
  75         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
  76 }
  77
  78 /* timeout for initial callback (AST) reply (bz10399) */
  79 static inline unsigned int ldlm_get_rq_timeout(void)
  80 {
  81         /* Non-AT value */
  82         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
  83
  84         return timeout < 1 ? 1 : timeout;
  85 }
  86
  87 #ifdef __KERNEL__
  88 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
  89 static spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
  90 static struct list_head waiting_locks_list;
  91 static cfs_timer_t waiting_locks_timer;
  92
  93 static struct expired_lock_thread {
  94         cfs_waitq_t               elt_waitq;
  95         int                       elt_state;
  96         int                       elt_dump;
  97         struct list_head          elt_expired_locks;
  98 } expired_lock_thread;
  99 #endif
 100
 101 #define ELT_STOPPED   0
 102 #define ELT_READY     1
 103 #define ELT_TERMINATE 2
 104
 105 struct ldlm_bl_pool {
 106         spinlock_t              blp_lock;
 107
 108         /*
 109          * blp_prio_list is used for callbacks that should be handled
 110          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
 111          * see bug 13843
 112          */
 113         struct list_head        blp_prio_list;
 114
 115         /*
 116          * blp_list is used for all other callbacks which are likely
 117          * to take longer to process.
 118          */
 119         struct list_head        blp_list;
 120
 121         cfs_waitq_t             blp_waitq;
 122         struct completion       blp_comp;
 123         atomic_t                blp_num_threads;
 124         atomic_t                blp_busy_threads;
 125         int                     blp_min_threads;
 126         int                     blp_max_threads;
 127 };
 128
 129 struct ldlm_bl_work_item {
 130         struct list_head        blwi_entry;
 131         struct ldlm_namespace   *blwi_ns;
 132         struct ldlm_lock_desc   blwi_ld;
 133         struct ldlm_lock        *blwi_lock;
 134         struct list_head        blwi_head;
 135         int                     blwi_count;
 136 };
 137
 138 #ifdef __KERNEL__
 139
 140 static inline int have_expired_locks(void)
 141 {
 142         int need_to_run;
 143
 144         ENTRY;
 145         spin_lock_bh(&waiting_locks_spinlock);
 146         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 147         spin_unlock_bh(&waiting_locks_spinlock);
 148
 149         RETURN(need_to_run);
 150 }
 151
 152 static int expired_lock_main(void *arg)
 153 {
 154         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 155         struct l_wait_info lwi = { 0 };
 156         int do_dump;
 157
 158         ENTRY;
 159         cfs_daemonize("ldlm_elt");
 160
 161         expired_lock_thread.elt_state = ELT_READY;
 162         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 163
 164         while (1) {
 165                 l_wait_event(expired_lock_thread.elt_waitq,
 166                              have_expired_locks() ||
 167                              expired_lock_thread.elt_state == ELT_TERMINATE,
 168                              &lwi);
 169
 170                 spin_lock_bh(&waiting_locks_spinlock);
 171                 if (expired_lock_thread.elt_dump) {
 172                         spin_unlock_bh(&waiting_locks_spinlock);
 173
 174                         /* from waiting_locks_callback, but not in timer */
 175                         libcfs_debug_dumplog();
 176                         libcfs_run_lbug_upcall(__FILE__,
 177                                                 "waiting_locks_callback",
 178                                                 expired_lock_thread.elt_dump);
 179
 180                         spin_lock_bh(&waiting_locks_spinlock);
 181                         expired_lock_thread.elt_dump = 0;
 182                 }
 183
 184                 do_dump = 0;
 185
 186                 while (!list_empty(expired)) {
 187                         struct obd_export *export;
 188                         struct ldlm_lock *lock;
 189
 190                         lock = list_entry(expired->next, struct ldlm_lock,
 191                                           l_pending_chain);
 192                         if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
 193                             (void *)lock >= LP_POISON) {
 194                                 spin_unlock_bh(&waiting_locks_spinlock);
 195                                 CERROR("free lock on elt list %p\n", lock);
 196                                 LBUG();
 197                         }
 198                         list_del_init(&lock->l_pending_chain);
 199                         if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
 200                             (void *)lock->l_export >= LP_POISON) {
 201                                 CERROR("lock with free export on elt list %p\n",
 202                                        lock->l_export);
 203                                 lock->l_export = NULL;
 204                                 LDLM_ERROR(lock, "free export");
 205                                 /* release extra ref grabbed by
 206                                  * ldlm_add_waiting_lock() or
 207                                  * ldlm_failed_ast() */
 208                                 LDLM_LOCK_RELEASE(lock);
 209                                 continue;
 210                         }
 211                         export = class_export_lock_get(lock->l_export);
 212                         spin_unlock_bh(&waiting_locks_spinlock);
 213
 214                         /* release extra ref grabbed by ldlm_add_waiting_lock()
 215                          * or ldlm_failed_ast() */
 216                         LDLM_LOCK_RELEASE(lock);
 217
 218                         do_dump++;
 219                         class_fail_export(export);
 220                         class_export_lock_put(export);
 221                         spin_lock_bh(&waiting_locks_spinlock);
 222                 }
 223                 spin_unlock_bh(&waiting_locks_spinlock);
 224
 225                 if (do_dump && obd_dump_on_eviction) {
 226                         CERROR("dump the log upon eviction\n");
 227                         libcfs_debug_dumplog();
 228                 }
 229
 230                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 231                         break;
 232         }
 233
 234         expired_lock_thread.elt_state = ELT_STOPPED;
 235         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 236         RETURN(0);
 237 }
 238
 239 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 240
 241 /**
 242  * Check if there is a request in the export request list
 243  * which prevents the lock canceling.
 244  */
 245 static int ldlm_lock_busy(struct ldlm_lock *lock)
 246 {
 247         struct ptlrpc_request *req;
 248         int match = 0;
 249         ENTRY;
 250
 251         if (lock->l_export == NULL)
 252                 return 0;
 253
 254         spin_lock(&lock->l_export->exp_lock);
 255         list_for_each_entry(req, &lock->l_export->exp_queued_rpc, rq_exp_list) {
 256                 if (req->rq_ops->hpreq_lock_match) {
 257                         match = req->rq_ops->hpreq_lock_match(req, lock);
 258                         if (match)
 259                                 break;
 260                 }
 261         }
 262         spin_unlock(&lock->l_export->exp_lock);
 263         RETURN(match);
 264 }
 265
 266 /* This is called from within a timer interrupt and cannot schedule */
 267 static void waiting_locks_callback(unsigned long unused)
 268 {
 269         struct ldlm_lock *lock, *last = NULL;
 270
 271 repeat:
 272         spin_lock_bh(&waiting_locks_spinlock);
 273         while (!list_empty(&waiting_locks_list)) {
 274                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 275                                   l_pending_chain);
 276                 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
 277                     (lock->l_req_mode == LCK_GROUP))
 278                         break;
 279
 280                 if (ptlrpc_check_suspend()) {
 281                         /* there is a case when we talk to one mds, holding
 282                          * lock from another mds. this way we easily can get
 283                          * here, if second mds is being recovered. so, we
 284                          * suspend timeouts. bug 6019 */
 285
 286                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 287                                    lock->l_export->exp_client_uuid.uuid,
 288                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 289                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 290
 291                         list_del_init(&lock->l_pending_chain);
 292                         spin_unlock_bh(&waiting_locks_spinlock);
 293                         ldlm_add_waiting_lock(lock);
 294                         goto repeat;
 295                 }
 296
 297                 /* if timeout overlaps the activation time of suspended timeouts
 298                  * then extend it to give a chance for client to reconnect */
 299                 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
 300                                                  cfs_time_seconds(obd_timeout)/2),
 301                                     ptlrpc_suspend_wakeup_time())) {
 302                         LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
 303                                    lock->l_export->exp_client_uuid.uuid,
 304                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 305                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 306
 307                         list_del_init(&lock->l_pending_chain);
 308                         spin_unlock_bh(&waiting_locks_spinlock);
 309                         ldlm_add_waiting_lock(lock);
 310                         goto repeat;
 311                 }
 312
 313                 /* Check if we need to prolong timeout */
 314                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
 315                     ldlm_lock_busy(lock)) {
 316                         int cont = 1;
 317
 318                         if (lock->l_pending_chain.next == &waiting_locks_list)
 319                                 cont = 0;
 320
 321                         LDLM_LOCK_GET(lock);
 322
 323                         spin_unlock_bh(&waiting_locks_spinlock);
 324                         LDLM_DEBUG(lock, "prolong the busy lock");
 325                         ldlm_refresh_waiting_lock(lock,
 326                                                   ldlm_get_enq_timeout(lock));
 327                         spin_lock_bh(&waiting_locks_spinlock);
 328
 329                         if (!cont) {
 330                                 LDLM_LOCK_RELEASE(lock);
 331                                 break;
 332                         }
 333
 334                         LDLM_LOCK_RELEASE(lock);
 335                         continue;
 336                 }
 337                 lock->l_resource->lr_namespace->ns_timeouts++;
 338                 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
 339                            "evicting client at %s ",
 340                            cfs_time_current_sec()- lock->l_last_activity,
 341                            libcfs_nid2str(
 342                                    lock->l_export->exp_connection->c_peer.nid));
 343
 344                 last = lock;
 345
 346                 /* no needs to take an extra ref on the lock since it was in
 347                  * the waiting_locks_list and ldlm_add_waiting_lock()
 348                  * already grabbed a ref */
 349                 list_del(&lock->l_pending_chain);
 350                 list_add(&lock->l_pending_chain,
 351                          &expired_lock_thread.elt_expired_locks);
 352         }
 353
 354         if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
 355                 if (obd_dump_on_timeout)
 356                         expired_lock_thread.elt_dump = __LINE__;
 357
 358                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 359         }
 360
 361         /*
 362          * Make sure the timer will fire again if we have any locks
 363          * left.
 364          */
 365         if (!list_empty(&waiting_locks_list)) {
 366                 cfs_time_t timeout_rounded;
 367                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 368                                   l_pending_chain);
 369                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
 370                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 371         }
 372         spin_unlock_bh(&waiting_locks_spinlock);
 373 }
 374
 375 /*
 376  * Indicate that we're waiting for a client to call us back cancelling a given
 377  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 378  * timer to fire appropriately.  (We round up to the next second, to avoid
 379  * floods of timer firings during periods of high lock contention and traffic).
 380  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
 381  * if it has been added to the waiting list (1 is returned).
 382  *
 383  * Called with the namespace lock held.
 384  */
 385 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
 386 {
 387         cfs_time_t timeout;
 388         cfs_time_t timeout_rounded;
 389
 390         if (!list_empty(&lock->l_pending_chain))
 391                 return 0;
 392
 393         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
 394             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
 395                 seconds = 1;
 396
 397         timeout = cfs_time_shift(seconds);
 398         if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
 399                 lock->l_callback_timeout = timeout;
 400
 401         timeout_rounded = round_timeout(lock->l_callback_timeout);
 402
 403         if (cfs_time_before(timeout_rounded,
 404                             cfs_timer_deadline(&waiting_locks_timer)) ||
 405             !cfs_timer_is_armed(&waiting_locks_timer)) {
 406                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 407         }
 408         /* if the new lock has a shorter timeout than something earlier on
 409            the list, we'll wait the longer amount of time; no big deal. */
 410         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 411         return 1;
 412 }
 413
 414 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 415 {
 416         int ret;
 417         int timeout = ldlm_get_enq_timeout(lock);
 418
 419         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 420
 421         spin_lock_bh(&waiting_locks_spinlock);
 422         if (lock->l_destroyed) {
 423                 static cfs_time_t next;
 424                 spin_unlock_bh(&waiting_locks_spinlock);
 425                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
 426                 if (cfs_time_after(cfs_time_current(), next)) {
 427                         next = cfs_time_shift(14400);
 428                         libcfs_debug_dumpstack(NULL);
 429                 }
 430                 return 0;
 431         }
 432
 433         ret = __ldlm_add_waiting_lock(lock, timeout);
 434         if (ret)
 435                 /* grab ref on the lock if it has been added to the
 436                  * waiting list */
 437                 LDLM_LOCK_GET(lock);
 438         spin_unlock_bh(&waiting_locks_spinlock);
 439
 440         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
 441                    ret == 0 ? "not re-" : "", timeout,
 442                    AT_OFF ? "off" : "on");
 443         return ret;
 444 }
 445
 446 /*
 447  * Remove a lock from the pending list, likely because it had its cancellation
 448  * callback arrive without incident.  This adjusts the lock-timeout timer if
 449  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 450  * As done by ldlm_del_waiting_lock(), the caller must release the lock
 451  * reference when the lock is removed from any list (1 is returned).
 452  *
 453  * Called with namespace lock held.
 454  */
 455 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
 456 {
 457         struct list_head *list_next;
 458
 459         if (list_empty(&lock->l_pending_chain))
 460                 return 0;
 461
 462         list_next = lock->l_pending_chain.next;
 463         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 464                 /* Removing the head of the list, adjust timer. */
 465                 if (list_next == &waiting_locks_list) {
 466                         /* No more, just cancel. */
 467                         cfs_timer_disarm(&waiting_locks_timer);
 468                 } else {
 469                         struct ldlm_lock *next;
 470                         next = list_entry(list_next, struct ldlm_lock,
 471                                           l_pending_chain);
 472                         cfs_timer_arm(&waiting_locks_timer,
 473                                       round_timeout(next->l_callback_timeout));
 474                 }
 475         }
 476         list_del_init(&lock->l_pending_chain);
 477
 478         return 1;
 479 }
 480
 481 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 482 {
 483         int ret;
 484
 485         if (lock->l_export == NULL) {
 486                 /* We don't have a "waiting locks list" on clients. */
 487                 LDLM_DEBUG(lock, "client lock: no-op");
 488                 return 0;
 489         }
 490
 491         spin_lock_bh(&waiting_locks_spinlock);
 492         ret = __ldlm_del_waiting_lock(lock);
 493         spin_unlock_bh(&waiting_locks_spinlock);
 494         if (ret)
 495                 /* release lock ref if it has indeed been removed
 496                  * from a list */
 497                 LDLM_LOCK_RELEASE(lock);
 498
 499         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
 500         return ret;
 501 }
 502
 503 /*
 504  * Prolong the lock
 505  *
 506  * Called with namespace lock held.
 507  */
 508 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 509 {
 510         if (lock->l_export == NULL) {
 511                 /* We don't have a "waiting locks list" on clients. */
 512                 LDLM_DEBUG(lock, "client lock: no-op");
 513                 return 0;
 514         }
 515
 516         spin_lock_bh(&waiting_locks_spinlock);
 517
 518         if (list_empty(&lock->l_pending_chain)) {
 519                 spin_unlock_bh(&waiting_locks_spinlock);
 520                 LDLM_DEBUG(lock, "wasn't waiting");
 521                 return 0;
 522         }
 523
 524         /* we remove/add the lock to the waiting list, so no needs to
 525          * release/take a lock reference */
 526         __ldlm_del_waiting_lock(lock);
 527         __ldlm_add_waiting_lock(lock, timeout);
 528         spin_unlock_bh(&waiting_locks_spinlock);
 529
 530         LDLM_DEBUG(lock, "refreshed");
 531         return 1;
 532 }
 533 #else /* !__KERNEL__ */
 534
 535 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 536 {
 537         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 538         RETURN(1);
 539 }
 540
 541 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 542 {
 543         RETURN(0);
 544 }
 545
 546 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 547 {
 548         RETURN(0);
 549 }
 550 #endif /* __KERNEL__ */
 551
 552 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 553                             const char *ast_type)
 554 {
 555         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
 556                            "to a lock %s callback time out: rc %d\n",
 557                            lock->l_export->exp_obd->obd_name,
 558                            obd_export_nid2str(lock->l_export), ast_type, rc);
 559
 560         if (obd_dump_on_timeout)
 561                 libcfs_debug_dumplog();
 562 #ifdef __KERNEL__
 563         spin_lock_bh(&waiting_locks_spinlock);
 564         if (__ldlm_del_waiting_lock(lock) == 0)
 565                 /* the lock was not in any list, grab an extra ref before adding
 566                  * the lock to the expired list */
 567                 LDLM_LOCK_GET(lock);
 568         list_add(&lock->l_pending_chain, &expired_lock_thread.elt_expired_locks);
 569         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 570         spin_unlock_bh(&waiting_locks_spinlock);
 571 #else
 572         class_fail_export(lock->l_export);
 573 #endif
 574 }
 575
 576 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 577                                  struct ptlrpc_request *req, int rc,
 578                                  const char *ast_type)
 579 {
 580         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
 581
 582         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 583                 LASSERT(lock->l_export);
 584                 if (lock->l_export->exp_libclient) {
 585                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 586                                    " timeout, just cancelling lock", ast_type,
 587                                    libcfs_nid2str(peer.nid));
 588                         ldlm_lock_cancel(lock);
 589                         rc = -ERESTART;
 590                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
 591                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
 592                                    "cancel was received (AST reply lost?)",
 593                                    ast_type, libcfs_nid2str(peer.nid));
 594                         ldlm_lock_cancel(lock);
 595                         rc = -ERESTART;
 596                 } else {
 597                         ldlm_del_waiting_lock(lock);
 598                         ldlm_failed_ast(lock, rc, ast_type);
 599                 }
 600         } else if (rc) {
 601                 if (rc == -EINVAL)
 602                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 603                                " from %s AST - normal race",
 604                                libcfs_nid2str(peer.nid),
 605                                req->rq_repmsg ?
 606                                lustre_msg_get_status(req->rq_repmsg) : -1,
 607                                ast_type);
 608                 else
 609                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 610                                    "from %s AST", libcfs_nid2str(peer.nid),
 611                                    (req->rq_repmsg != NULL) ?
 612                                    lustre_msg_get_status(req->rq_repmsg) : 0,
 613                                    ast_type);
 614                 ldlm_lock_cancel(lock);
 615                 /* Server-side AST functions are called from ldlm_reprocess_all,
 616                  * which needs to be told to please restart its reprocessing. */
 617                 rc = -ERESTART;
 618         }
 619
 620         return rc;
 621 }
 622
 623 static int ldlm_cb_interpret(const struct lu_env *env,
 624                              struct ptlrpc_request *req, void *data, int rc)
 625 {
 626         struct ldlm_cb_set_arg *arg;
 627         struct ldlm_lock *lock;
 628         ENTRY;
 629
 630         LASSERT(data != NULL);
 631
 632         arg = req->rq_async_args.pointer_arg[0];
 633         lock = req->rq_async_args.pointer_arg[1];
 634         LASSERT(lock != NULL);
 635         if (rc != 0) {
 636                 /* If client canceled the lock but the cancel has not
 637                  * been recieved yet, we need to update lvbo to have the
 638                  * proper attributes cached. */
 639                 if (rc == -EINVAL && arg->type == LDLM_BL_CALLBACK)
 640                         ldlm_res_lvbo_update(lock->l_resource, NULL,
 641                                              0, 1);
 642                 rc = ldlm_handle_ast_error(lock, req, rc,
 643                                            arg->type == LDLM_BL_CALLBACK
 644                                            ? "blocking" : "completion");
 645         }
 646
 647         LDLM_LOCK_RELEASE(lock);
 648
 649         if (rc == -ERESTART)
 650                 atomic_set(&arg->restart, 1);
 651
 652         RETURN(0);
 653 }
 654
 655 static inline int ldlm_bl_and_cp_ast_fini(struct ptlrpc_request *req,
 656                                           struct ldlm_cb_set_arg *arg,
 657                                           struct ldlm_lock *lock,
 658                                           int instant_cancel)
 659 {
 660         int rc = 0;
 661         ENTRY;
 662
 663         if (unlikely(instant_cancel)) {
 664                 rc = ptl_send_rpc(req, 1);
 665                 ptlrpc_req_finished(req);
 666                 if (rc == 0)
 667                         /* If we cancelled the lock, we need to restart
 668                          * ldlm_reprocess_queue */
 669                         atomic_set(&arg->restart, 1);
 670         } else {
 671                 LDLM_LOCK_GET(lock);
 672                 ptlrpc_set_add_req(arg->set, req);
 673         }
 674
 675         RETURN(rc);
 676 }
 677
 678 /**
 679  * Check if there are requests in the export request list which prevent
 680  * the lock canceling and make these requests high priority ones.
 681  */
 682 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
 683 {
 684         struct ptlrpc_request *req;
 685         ENTRY;
 686
 687         if (lock->l_export == NULL) {
 688                 LDLM_DEBUG(lock, "client lock: no-op");
 689                 RETURN_EXIT;
 690         }
 691
 692         spin_lock(&lock->l_export->exp_lock);
 693         list_for_each_entry(req, &lock->l_export->exp_queued_rpc, rq_exp_list) {
 694                 if (!req->rq_hp && req->rq_ops->hpreq_lock_match &&
 695                     req->rq_ops->hpreq_lock_match(req, lock))
 696                         ptlrpc_hpreq_reorder(req);
 697         }
 698         spin_unlock(&lock->l_export->exp_lock);
 699         EXIT;
 700 }
 701
 702 /*
 703  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
 704  * enqueued server lock conflicts with given one.
 705  *
 706  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
 707  * to wait for client response.
 708  */
 709 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 710                              struct ldlm_lock_desc *desc,
 711                              void *data, int flag)
 712 {
 713         struct ldlm_cb_set_arg *arg = data;
 714         struct ldlm_request    *body;
 715         struct ptlrpc_request  *req;
 716         int                     instant_cancel = 0;
 717         int                     rc = 0;
 718         ENTRY;
 719
 720         if (flag == LDLM_CB_CANCELING)
 721                 /* Don't need to do anything here. */
 722                 RETURN(0);
 723
 724         LASSERT(lock);
 725         LASSERT(data != NULL);
 726         if (lock->l_export->exp_obd->obd_recovering != 0) {
 727                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
 728                 ldlm_lock_dump(D_ERROR, lock, 0);
 729         }
 730
 731         ldlm_lock_reorder_req(lock);
 732
 733         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 734                                         &RQF_LDLM_BL_CALLBACK,
 735                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
 736         if (req == NULL)
 737                 RETURN(-ENOMEM);
 738
 739         req->rq_async_args.pointer_arg[0] = arg;
 740         req->rq_async_args.pointer_arg[1] = lock;
 741         req->rq_interpret_reply = ldlm_cb_interpret;
 742         req->rq_no_resend = 1;
 743
 744         lock_res(lock->l_resource);
 745         if (lock->l_granted_mode != lock->l_req_mode) {
 746                 /* this blocking AST will be communicated as part of the
 747                  * completion AST instead */
 748                 unlock_res(lock->l_resource);
 749                 ptlrpc_req_finished(req);
 750                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 751                 RETURN(0);
 752         }
 753
 754         if (lock->l_destroyed) {
 755                 /* What's the point? */
 756                 unlock_res(lock->l_resource);
 757                 ptlrpc_req_finished(req);
 758                 RETURN(0);
 759         }
 760
 761         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
 762                 instant_cancel = 1;
 763
 764         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 765         body->lock_handle[0] = lock->l_remote_handle;
 766         body->lock_desc = *desc;
 767         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 768
 769         LDLM_DEBUG(lock, "server preparing blocking AST");
 770
 771         ptlrpc_request_set_replen(req);
 772         if (instant_cancel) {
 773                 unlock_res(lock->l_resource);
 774                 ldlm_lock_cancel(lock);
 775         } else {
 776                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
 777                 ldlm_add_waiting_lock(lock);
 778                 unlock_res(lock->l_resource);
 779         }
 780
 781         req->rq_send_state = LUSTRE_IMP_FULL;
 782         /* ptlrpc_prep_req already set timeout */
 783         if (AT_OFF)
 784                 req->rq_timeout = ldlm_get_rq_timeout();
 785
 786         if (lock->l_export && lock->l_export->exp_nid_stats &&
 787             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 788                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 789                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 790
 791         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
 792
 793         RETURN(rc);
 794 }
 795
 796 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 797 {
 798         struct ldlm_cb_set_arg *arg = data;
 799         struct ldlm_request    *body;
 800         struct ptlrpc_request  *req;
 801         long                    total_enqueue_wait;
 802         int                     instant_cancel = 0;
 803         int                     rc = 0;
 804         ENTRY;
 805
 806         LASSERT(lock != NULL);
 807         LASSERT(data != NULL);
 808
 809         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
 810                                           lock->l_last_activity);
 811
 812         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
 813                                     &RQF_LDLM_CP_CALLBACK);
 814         if (req == NULL)
 815                 RETURN(-ENOMEM);
 816
 817         lock_res_and_lock(lock);
 818         if (lock->l_resource->lr_lvb_len)
 819                  req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
 820                                       lock->l_resource->lr_lvb_len);
 821         unlock_res_and_lock(lock);
 822
 823         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
 824         if (rc) {
 825                 ptlrpc_request_free(req);
 826                 RETURN(rc);
 827         }
 828
 829         req->rq_async_args.pointer_arg[0] = arg;
 830         req->rq_async_args.pointer_arg[1] = lock;
 831         req->rq_interpret_reply = ldlm_cb_interpret;
 832         req->rq_no_resend = 1;
 833         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 834
 835         body->lock_handle[0] = lock->l_remote_handle;
 836         body->lock_flags = flags;
 837         ldlm_lock2desc(lock, &body->lock_desc);
 838         if (lock->l_resource->lr_lvb_len) {
 839                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
 840
 841                 lock_res_and_lock(lock);
 842                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 843                        lock->l_resource->lr_lvb_len);
 844                 unlock_res_and_lock(lock);
 845         }
 846
 847         LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
 848                    total_enqueue_wait);
 849
 850         /* Server-side enqueue wait time estimate, used in
 851             __ldlm_add_waiting_lock to set future enqueue timers */
 852         if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
 853                 at_add(&lock->l_resource->lr_namespace->ns_at_estimate,
 854                        total_enqueue_wait);
 855         else
 856                 /* bz18618. Don't add lock enqueue time we spend waiting for a
 857                    previous callback to fail. Locks waiting legitimately will
 858                    get extended by ldlm_refresh_waiting_lock regardless of the
 859                    estimate, so it's okay to underestimate here. */
 860                 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
 861                        "It is likely that a previous callback timed out.",
 862                        total_enqueue_wait,
 863                        at_get(&lock->l_resource->lr_namespace->ns_at_estimate));
 864
 865         ptlrpc_request_set_replen(req);
 866
 867         req->rq_send_state = LUSTRE_IMP_FULL;
 868         /* ptlrpc_prep_req already set timeout */
 869         if (AT_OFF)
 870                 req->rq_timeout = ldlm_get_rq_timeout();
 871
 872         /* We only send real blocking ASTs after the lock is granted */
 873         lock_res_and_lock(lock);
 874         if (lock->l_flags & LDLM_FL_AST_SENT) {
 875                 body->lock_flags |= LDLM_FL_AST_SENT;
 876                 /* copy ast flags like LDLM_FL_DISCARD_DATA */
 877                 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 878
 879                 /* We might get here prior to ldlm_handle_enqueue setting
 880                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
 881                  * into waiting list, but this is safe and similar code in
 882                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
 883                  * that would not only cancel the lock, but will also remove
 884                  * it from waiting list */
 885                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 886                         unlock_res_and_lock(lock);
 887                         ldlm_lock_cancel(lock);
 888                         instant_cancel = 1;
 889                         lock_res_and_lock(lock);
 890                 } else {
 891                         /* start the lock-timeout clock */
 892                         ldlm_add_waiting_lock(lock);
 893                 }
 894         }
 895         unlock_res_and_lock(lock);
 896
 897         if (lock->l_export && lock->l_export->exp_nid_stats &&
 898             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 899                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 900                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
 901
 902         rc = ldlm_bl_and_cp_ast_fini(req, arg, lock, instant_cancel);
 903
 904         RETURN(rc);
 905 }
 906
 907 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 908 {
 909         struct ldlm_resource  *res = lock->l_resource;
 910         struct ldlm_request   *body;
 911         struct ptlrpc_request *req;
 912         int                    rc;
 913         ENTRY;
 914
 915         LASSERT(lock != NULL);
 916
 917         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 918                                         &RQF_LDLM_GL_CALLBACK,
 919                                         LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
 920
 921         if (req == NULL)
 922                 RETURN(-ENOMEM);
 923
 924         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 925         body->lock_handle[0] = lock->l_remote_handle;
 926         ldlm_lock2desc(lock, &body->lock_desc);
 927
 928         lock_res_and_lock(lock);
 929         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 930                              lock->l_resource->lr_lvb_len);
 931         unlock_res_and_lock(lock);
 932         res = lock->l_resource;
 933         ptlrpc_request_set_replen(req);
 934
 935
 936         req->rq_send_state = LUSTRE_IMP_FULL;
 937         /* ptlrpc_prep_req already set timeout */
 938         if (AT_OFF)
 939                 req->rq_timeout = ldlm_get_rq_timeout();
 940
 941         if (lock->l_export && lock->l_export->exp_nid_stats &&
 942             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 943                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 944                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
 945
 946         rc = ptlrpc_queue_wait(req);
 947         if (rc == -ELDLM_NO_LOCK_DATA)
 948                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 949         else if (rc != 0)
 950                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 951         else
 952                 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
 953                                           REPLY_REC_OFF, 1);
 954         ptlrpc_req_finished(req);
 955         if (rc == -ERESTART)
 956                 ldlm_reprocess_all(res);
 957
 958         RETURN(rc);
 959 }
 960
 961 #ifdef __KERNEL__
 962 extern unsigned long long lu_time_stamp_get(void);
 963 #else
 964 #define lu_time_stamp_get() time(NULL)
 965 #endif
 966
 967 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
 968                        struct lprocfs_stats *srv_stats)
 969 {
 970         int lock_type = 0, op = 0;
 971
 972         lock_type = dlm_req->lock_desc.l_resource.lr_type;
 973
 974         switch (lock_type) {
 975         case LDLM_PLAIN:
 976                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
 977                 break;
 978         case LDLM_EXTENT:
 979                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
 980                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
 981                 else
 982                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
 983                 break;
 984         case LDLM_FLOCK:
 985                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
 986                 break;
 987         case LDLM_IBITS:
 988                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
 989                 break;
 990         default:
 991                 op = 0;
 992                 break;
 993         }
 994
 995         if (op)
 996                 lprocfs_counter_incr(srv_stats, op);
 997
 998         return;
 999 }
1000
1001 /*
1002  * Main server-side entry point into LDLM. This is called by ptlrpc service
1003  * threads to carry out client lock enqueueing requests.
1004  */
1005 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1006                          struct ptlrpc_request *req,
1007                          const struct ldlm_request *dlm_req,
1008                          const struct ldlm_callback_suite *cbs)
1009 {
1010         struct ldlm_reply *dlm_rep;
1011         __u32 flags;
1012         ldlm_error_t err = ELDLM_OK;
1013         struct ldlm_lock *lock = NULL;
1014         void *cookie = NULL;
1015         int rc = 0;
1016         ENTRY;
1017
1018         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1019
1020         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1021         flags = dlm_req->lock_flags;
1022
1023         LASSERT(req->rq_export);
1024
1025         if (req->rq_rqbd->rqbd_service->srv_stats)
1026                 ldlm_svc_get_eopc(dlm_req,
1027                                   req->rq_rqbd->rqbd_service->srv_stats);
1028
1029         if (req->rq_export && req->rq_export->exp_nid_stats &&
1030             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1031                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1032                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1033
1034         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1035                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1036                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1037                           dlm_req->lock_desc.l_resource.lr_type);
1038                 GOTO(out, rc = -EFAULT);
1039         }
1040
1041         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1042                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1043                      dlm_req->lock_desc.l_req_mode &
1044                      (dlm_req->lock_desc.l_req_mode-1))) {
1045                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1046                           dlm_req->lock_desc.l_req_mode);
1047                 GOTO(out, rc = -EFAULT);
1048         }
1049
1050         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1051                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1052                              LDLM_PLAIN)) {
1053                         DEBUG_REQ(D_ERROR, req,
1054                                   "PLAIN lock request from IBITS client?");
1055                         GOTO(out, rc = -EPROTO);
1056                 }
1057         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1058                             LDLM_IBITS)) {
1059                 DEBUG_REQ(D_ERROR, req,
1060                           "IBITS lock request from unaware client?");
1061                 GOTO(out, rc = -EPROTO);
1062         }
1063
1064 #if 0
1065         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1066            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1067            ibits for mgc/mgs) */
1068
1069         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1070          * inodebits lock if client does not support them. */
1071         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1072             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1073                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1074                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1075                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1076                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1077                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1078         }
1079 #endif
1080
1081         if (unlikely(flags & LDLM_FL_REPLAY)) {
1082                 /* Find an existing lock in the per-export lock hash */
1083                 lock = lustre_hash_lookup(req->rq_export->exp_lock_hash,
1084                                           (void *)&dlm_req->lock_handle[0]);
1085                 if (lock != NULL) {
1086                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1087                                   LPX64, lock->l_handle.h_cookie);
1088                         GOTO(existing_lock, rc = 0);
1089                 }
1090         }
1091
1092         /* The lock's callback data might be set in the policy function */
1093         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1094                                 dlm_req->lock_desc.l_resource.lr_type,
1095                                 dlm_req->lock_desc.l_req_mode,
1096                                 cbs, NULL, 0);
1097
1098         if (!lock)
1099                 GOTO(out, rc = -ENOMEM);
1100
1101         lock->l_last_activity = cfs_time_current_sec();
1102         lock->l_remote_handle = dlm_req->lock_handle[0];
1103         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1104
1105         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1106         /* Don't enqueue a lock onto the export if it has already
1107          * been evicted.  Cancel it now instead. (bug 3822) */
1108         if (req->rq_export->exp_failed) {
1109                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1110                 GOTO(out, rc = -ENOTCONN);
1111         }
1112         lock->l_export = class_export_lock_get(req->rq_export);
1113         if (lock->l_export->exp_lock_hash)
1114                 lustre_hash_add(lock->l_export->exp_lock_hash,
1115                                 &lock->l_remote_handle,
1116                                 &lock->l_exp_hash);
1117
1118 existing_lock:
1119
1120         if (flags & LDLM_FL_HAS_INTENT) {
1121                 /* In this case, the reply buffer is allocated deep in
1122                  * local_lock_enqueue by the policy function. */
1123                 cookie = req;
1124         } else {
1125                 lock_res_and_lock(lock);
1126                 if (lock->l_resource->lr_lvb_len) {
1127                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1128                                              RCL_SERVER,
1129                                              lock->l_resource->lr_lvb_len);
1130                 }
1131                 unlock_res_and_lock(lock);
1132
1133                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1134                         GOTO(out, rc = -ENOMEM);
1135
1136                 rc = req_capsule_server_pack(&req->rq_pill);
1137                 if (rc)
1138                         GOTO(out, rc);
1139         }
1140
1141         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1142                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1143         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1144                 lock->l_req_extent = lock->l_policy_data.l_extent;
1145
1146         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1147         if (err)
1148                 GOTO(out, err);
1149
1150         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1151         dlm_rep->lock_flags = flags;
1152
1153         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1154         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1155
1156         /* We never send a blocking AST until the lock is granted, but
1157          * we can tell it right now */
1158         lock_res_and_lock(lock);
1159
1160         /* Now take into account flags to be inherited from original lock
1161            request both in reply to client and in our own lock flags. */
1162         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1163         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1164
1165         /* Don't move a pending lock onto the export if it has already
1166          * been evicted.  Cancel it now instead. (bug 5683) */
1167         if (unlikely(req->rq_export->exp_failed ||
1168                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1169                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1170                 rc = -ENOTCONN;
1171         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1172                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1173                 if (lock->l_granted_mode == lock->l_req_mode) {
1174                         /*
1175                          * Only cancel lock if it was granted, because it would
1176                          * be destroyed immediatelly and would never be granted
1177                          * in the future, causing timeouts on client.  Not
1178                          * granted lock will be cancelled immediatelly after
1179                          * sending completion AST.
1180                          */
1181                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1182                                 unlock_res_and_lock(lock);
1183                                 ldlm_lock_cancel(lock);
1184                                 lock_res_and_lock(lock);
1185                         } else
1186                                 ldlm_add_waiting_lock(lock);
1187                 }
1188         }
1189         /* Make sure we never ever grant usual metadata locks to liblustre
1190            clients */
1191         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1192             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1193              req->rq_export->exp_libclient) {
1194                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1195                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1196                         CERROR("Granting sync lock to libclient. "
1197                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1198                                dlm_req->lock_flags, dlm_rep->lock_flags,
1199                                lock->l_flags);
1200                         LDLM_ERROR(lock, "sync lock");
1201                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1202                                 struct ldlm_intent *it;
1203
1204                                 it = req_capsule_client_get(&req->rq_pill,
1205                                                             &RMF_LDLM_INTENT);
1206                                 if (it != NULL) {
1207                                         CERROR("This is intent %s ("LPU64")\n",
1208                                                ldlm_it2str(it->opc), it->opc);
1209                                 }
1210                         }
1211                 }
1212         }
1213
1214         unlock_res_and_lock(lock);
1215
1216         EXIT;
1217  out:
1218         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1219         if (!req->rq_packed_final) {
1220                 err = lustre_pack_reply(req, 1, NULL, NULL);
1221                 if (rc == 0)
1222                         rc = err;
1223         }
1224
1225         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1226          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1227         if (lock) {
1228                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1229                            "(err=%d, rc=%d)", err, rc);
1230
1231                 lock_res_and_lock(lock);
1232                 if (rc == 0) {
1233                         if (lock->l_resource->lr_lvb_len > 0) {
1234                                 void *lvb;
1235
1236                                 lvb = req_capsule_server_get(&req->rq_pill,
1237                                                              &RMF_DLM_LVB);
1238                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1239                                          req, lock);
1240
1241                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
1242                                        lock->l_resource->lr_lvb_len);
1243                         }
1244                 } else {
1245                         ldlm_resource_unlink_lock(lock);
1246                         ldlm_lock_destroy_nolock(lock);
1247                 }
1248                 unlock_res_and_lock(lock);
1249
1250                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1251                         ldlm_reprocess_all(lock->l_resource);
1252
1253                 LDLM_LOCK_RELEASE(lock);
1254         }
1255
1256         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1257                           lock, rc);
1258
1259         return rc;
1260 }
1261
1262 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1263                         ldlm_completion_callback completion_callback,
1264                         ldlm_blocking_callback blocking_callback,
1265                         ldlm_glimpse_callback glimpse_callback)
1266 {
1267         struct ldlm_request *dlm_req;
1268         struct ldlm_callback_suite cbs = {
1269                 .lcs_completion = completion_callback,
1270                 .lcs_blocking   = blocking_callback,
1271                 .lcs_glimpse    = glimpse_callback
1272         };
1273         int rc;
1274
1275         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1276         if (dlm_req != NULL) {
1277                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1278                                           req, dlm_req, &cbs);
1279         } else {
1280                 rc = -EFAULT;
1281         }
1282         return rc;
1283 }
1284
1285 int ldlm_handle_convert0(struct ptlrpc_request *req,
1286                          const struct ldlm_request *dlm_req)
1287 {
1288         struct ldlm_reply *dlm_rep;
1289         struct ldlm_lock *lock;
1290         int rc;
1291         ENTRY;
1292
1293         if (req->rq_export && req->rq_export->exp_nid_stats &&
1294             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1295                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1296                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1297
1298         rc = req_capsule_server_pack(&req->rq_pill);
1299         if (rc)
1300                 RETURN(rc);
1301
1302         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1303         dlm_rep->lock_flags = dlm_req->lock_flags;
1304
1305         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1306         if (!lock) {
1307                 req->rq_status = EINVAL;
1308         } else {
1309                 void *res = NULL;
1310
1311                 LDLM_DEBUG(lock, "server-side convert handler START");
1312
1313                 lock->l_last_activity = cfs_time_current_sec();
1314                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1315                                         &dlm_rep->lock_flags);
1316                 if (res) {
1317                         if (ldlm_del_waiting_lock(lock))
1318                                 LDLM_DEBUG(lock, "converted waiting lock");
1319                         req->rq_status = 0;
1320                 } else {
1321                         req->rq_status = EDEADLOCK;
1322                 }
1323         }
1324
1325         if (lock) {
1326                 if (!req->rq_status)
1327                         ldlm_reprocess_all(lock->l_resource);
1328                 LDLM_DEBUG(lock, "server-side convert handler END");
1329                 LDLM_LOCK_PUT(lock);
1330         } else
1331                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1332
1333         RETURN(0);
1334 }
1335
1336 int ldlm_handle_convert(struct ptlrpc_request *req)
1337 {
1338         int rc;
1339         struct ldlm_request *dlm_req;
1340
1341         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1342         if (dlm_req != NULL) {
1343                 rc = ldlm_handle_convert0(req, dlm_req);
1344         } else {
1345                 CERROR ("Can't unpack dlm_req\n");
1346                 rc = -EFAULT;
1347         }
1348         return rc;
1349 }
1350
1351 /* Cancel all the locks whos handles are packed into ldlm_request */
1352 int ldlm_request_cancel(struct ptlrpc_request *req,
1353                         const struct ldlm_request *dlm_req, int first)
1354 {
1355         struct ldlm_resource *res, *pres = NULL;
1356         struct ldlm_lock *lock;
1357         int i, count, done = 0;
1358         ENTRY;
1359
1360         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1361         if (first >= count)
1362                 RETURN(0);
1363
1364         /* There is no lock on the server at the replay time,
1365          * skip lock cancelling to make replay tests to pass. */
1366         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1367                 RETURN(0);
1368
1369         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1370                           "starting at %d", count, first);
1371
1372         for (i = first; i < count; i++) {
1373                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1374                 if (!lock) {
1375                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1376                                           "lock (cookie "LPU64")",
1377                                           dlm_req->lock_handle[i].cookie);
1378                         continue;
1379                 }
1380
1381                 res = lock->l_resource;
1382                 done++;
1383
1384                 if (res != pres) {
1385                         if (pres != NULL) {
1386                                 ldlm_reprocess_all(pres);
1387                                 LDLM_RESOURCE_DELREF(pres);
1388                                 ldlm_resource_putref(pres);
1389                         }
1390                         if (res != NULL) {
1391                                 ldlm_resource_getref(res);
1392                                 LDLM_RESOURCE_ADDREF(res);
1393                                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1394                         }
1395                         pres = res;
1396                 }
1397                 ldlm_lock_cancel(lock);
1398                 LDLM_LOCK_PUT(lock);
1399         }
1400         if (pres != NULL) {
1401                 ldlm_reprocess_all(pres);
1402                 LDLM_RESOURCE_DELREF(pres);
1403                 ldlm_resource_putref(pres);
1404         }
1405         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1406         RETURN(done);
1407 }
1408
1409 int ldlm_handle_cancel(struct ptlrpc_request *req)
1410 {
1411         struct ldlm_request *dlm_req;
1412         int rc;
1413         ENTRY;
1414
1415         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1416         if (dlm_req == NULL) {
1417                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1418                 RETURN(-EFAULT);
1419         }
1420
1421         if (req->rq_export && req->rq_export->exp_nid_stats &&
1422             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1423                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1424                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1425
1426         rc = req_capsule_server_pack(&req->rq_pill);
1427         if (rc)
1428                 RETURN(rc);
1429
1430         if (!ldlm_request_cancel(req, dlm_req, 0))
1431                 req->rq_status = ESTALE;
1432
1433         if (ptlrpc_reply(req) != 0)
1434                 LBUG();
1435
1436         RETURN(0);
1437 }
1438
1439 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1440                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1441 {
1442         int do_ast;
1443         ENTRY;
1444
1445         LDLM_DEBUG(lock, "client blocking AST callback handler START");
1446
1447         lock_res_and_lock(lock);
1448         lock->l_flags |= LDLM_FL_CBPENDING;
1449
1450         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1451                 lock->l_flags |= LDLM_FL_CANCEL;
1452
1453         do_ast = (!lock->l_readers && !lock->l_writers);
1454         unlock_res_and_lock(lock);
1455
1456         if (do_ast) {
1457                 LDLM_DEBUG(lock, "already unused, calling "
1458                            "callback (%p)", lock->l_blocking_ast);
1459                 if (lock->l_blocking_ast != NULL)
1460                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1461                                              LDLM_CB_BLOCKING);
1462         } else {
1463                 LDLM_DEBUG(lock, "Lock still has references, will be"
1464                            " cancelled later");
1465         }
1466
1467         LDLM_DEBUG(lock, "client blocking callback handler END");
1468         LDLM_LOCK_RELEASE(lock);
1469         EXIT;
1470 }
1471
1472 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1473                                     struct ldlm_namespace *ns,
1474                                     struct ldlm_request *dlm_req,
1475                                     struct ldlm_lock *lock)
1476 {
1477         CFS_LIST_HEAD(ast_list);
1478         ENTRY;
1479
1480         LDLM_DEBUG(lock, "client completion callback handler START");
1481
1482         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1483                 int to = cfs_time_seconds(1);
1484                 while (to > 0) {
1485                         cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE, to);
1486                         if (lock->l_granted_mode == lock->l_req_mode ||
1487                             lock->l_destroyed)
1488                                 break;
1489                 }
1490         }
1491
1492         lock_res_and_lock(lock);
1493         if (lock->l_destroyed ||
1494             lock->l_granted_mode == lock->l_req_mode) {
1495                 /* bug 11300: the lock has already been granted */
1496                 unlock_res_and_lock(lock);
1497                 LDLM_DEBUG(lock, "Double grant race happened");
1498                 LDLM_LOCK_RELEASE(lock);
1499                 EXIT;
1500                 return;
1501         }
1502
1503         /* If we receive the completion AST before the actual enqueue returned,
1504          * then we might need to switch lock modes, resources, or extents. */
1505         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1506                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1507                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1508         }
1509
1510         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1511                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1512                 LDLM_DEBUG(lock, "completion AST, new policy data");
1513         }
1514
1515         ldlm_resource_unlink_lock(lock);
1516         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1517                    &lock->l_resource->lr_name,
1518                    sizeof(lock->l_resource->lr_name)) != 0) {
1519                 unlock_res_and_lock(lock);
1520                 if (ldlm_lock_change_resource(ns, lock,
1521                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1522                         LDLM_ERROR(lock, "Failed to allocate resource");
1523                         LDLM_LOCK_RELEASE(lock);
1524                         EXIT;
1525                         return;
1526                 }
1527                 LDLM_DEBUG(lock, "completion AST, new resource");
1528                 CERROR("change resource!\n");
1529                 lock_res_and_lock(lock);
1530         }
1531
1532         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1533                 /* BL_AST locks are not needed in lru.
1534                  * let ldlm_cancel_lru() be fast. */
1535                 ldlm_lock_remove_from_lru(lock);
1536                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1537                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1538         }
1539
1540         if (lock->l_lvb_len) {
1541                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1542                                          RCL_CLIENT) < lock->l_lvb_len) {
1543                         LDLM_ERROR(lock, "completion AST did not contain "
1544                                    "expected LVB!");
1545                 } else {
1546                         void *lvb = req_capsule_client_swab_get(&req->rq_pill,
1547                                                                 &RMF_DLM_LVB,
1548                                                   (void *)lock->l_lvb_swabber);
1549                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1550                 }
1551         }
1552
1553         ldlm_grant_lock(lock, &ast_list);
1554         unlock_res_and_lock(lock);
1555
1556         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1557
1558         ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1559
1560         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1561                           lock);
1562         LDLM_LOCK_RELEASE(lock);
1563         EXIT;
1564 }
1565
1566 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1567                                     struct ldlm_namespace *ns,
1568                                     struct ldlm_request *dlm_req,
1569                                     struct ldlm_lock *lock)
1570 {
1571         int rc = -ENOSYS;
1572         ENTRY;
1573
1574         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1575
1576         if (lock->l_glimpse_ast != NULL)
1577                 rc = lock->l_glimpse_ast(lock, req);
1578
1579         if (req->rq_repmsg != NULL) {
1580                 ptlrpc_reply(req);
1581         } else {
1582                 req->rq_status = rc;
1583                 ptlrpc_error(req);
1584         }
1585
1586         lock_res_and_lock(lock);
1587         if (lock->l_granted_mode == LCK_PW &&
1588             !lock->l_readers && !lock->l_writers &&
1589             cfs_time_after(cfs_time_current(),
1590                            cfs_time_add(lock->l_last_used,
1591                                         cfs_time_seconds(10)))) {
1592                 unlock_res_and_lock(lock);
1593                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1594                         ldlm_handle_bl_callback(ns, NULL, lock);
1595
1596                 EXIT;
1597                 return;
1598         }
1599         unlock_res_and_lock(lock);
1600         LDLM_LOCK_RELEASE(lock);
1601         EXIT;
1602 }
1603
1604 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1605 {
1606         if (req->rq_no_reply)
1607                 return 0;
1608
1609         req->rq_status = rc;
1610         if (!req->rq_packed_final) {
1611                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1612                 if (rc)
1613                         return rc;
1614         }
1615         return ptlrpc_reply(req);
1616 }
1617
1618 #ifdef __KERNEL__
1619 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1620                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1621                              struct list_head *cancels, int count)
1622 {
1623         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1624         struct ldlm_bl_work_item *blwi;
1625         ENTRY;
1626
1627         if (cancels && count == 0)
1628                 RETURN(0);
1629
1630         OBD_ALLOC(blwi, sizeof(*blwi));
1631         if (blwi == NULL)
1632                 RETURN(-ENOMEM);
1633
1634         blwi->blwi_ns = ns;
1635         if (ld != NULL)
1636                 blwi->blwi_ld = *ld;
1637         if (count) {
1638                 list_add(&blwi->blwi_head, cancels);
1639                 list_del_init(cancels);
1640                 blwi->blwi_count = count;
1641         } else {
1642                 blwi->blwi_lock = lock;
1643         }
1644         spin_lock(&blp->blp_lock);
1645         if (lock && lock->l_flags & LDLM_FL_DISCARD_DATA) {
1646                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1647                 list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1648         } else {
1649                 /* other blocking callbacks are added to the regular list */
1650                 list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1651         }
1652         cfs_waitq_signal(&blp->blp_waitq);
1653         spin_unlock(&blp->blp_lock);
1654
1655         RETURN(0);
1656 }
1657 #endif
1658
1659 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1660                            struct ldlm_lock *lock)
1661 {
1662 #ifdef __KERNEL__
1663         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0));
1664 #else
1665         RETURN(-ENOSYS);
1666 #endif
1667 }
1668
1669 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1670                            struct list_head *cancels, int count)
1671 {
1672 #ifdef __KERNEL__
1673         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count));
1674 #else
1675         RETURN(-ENOSYS);
1676 #endif
1677 }
1678
1679 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1680 static int ldlm_callback_handler(struct ptlrpc_request *req)
1681 {
1682         struct ldlm_namespace *ns;
1683         struct ldlm_request *dlm_req;
1684         struct ldlm_lock *lock;
1685         int rc;
1686         ENTRY;
1687
1688         /* Requests arrive in sender's byte order.  The ptlrpc service
1689          * handler has already checked and, if necessary, byte-swapped the
1690          * incoming request message body, but I am responsible for the
1691          * message buffers. */
1692
1693         /* do nothing for sec context finalize */
1694         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1695                 RETURN(0);
1696
1697         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1698
1699         if (req->rq_export == NULL) {
1700                 ldlm_callback_reply(req, -ENOTCONN);
1701                 RETURN(0);
1702         }
1703
1704         LASSERT(req->rq_export != NULL);
1705         LASSERT(req->rq_export->exp_obd != NULL);
1706
1707         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1708         case LDLM_BL_CALLBACK:
1709                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1710                         RETURN(0);
1711                 break;
1712         case LDLM_CP_CALLBACK:
1713                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1714                         RETURN(0);
1715                 break;
1716         case LDLM_GL_CALLBACK:
1717                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1718                         RETURN(0);
1719                 break;
1720         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1721                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1722                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1723                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1724                         RETURN(0);
1725                 rc = llog_origin_handle_cancel(req);
1726                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1727                         RETURN(0);
1728                 ldlm_callback_reply(req, rc);
1729                 RETURN(0);
1730         case OBD_QC_CALLBACK:
1731                 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1732                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1733                         RETURN(0);
1734                 rc = target_handle_qc_callback(req);
1735                 ldlm_callback_reply(req, rc);
1736                 RETURN(0);
1737         case QUOTA_DQACQ:
1738         case QUOTA_DQREL:
1739                 /* reply in handler */
1740                 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1741                 rc = target_handle_dqacq_callback(req);
1742                 RETURN(0);
1743         case LLOG_ORIGIN_HANDLE_CREATE:
1744                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1745                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1746                         RETURN(0);
1747                 rc = llog_origin_handle_create(req);
1748                 ldlm_callback_reply(req, rc);
1749                 RETURN(0);
1750         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1751                 req_capsule_set(&req->rq_pill,
1752                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1753                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1754                         RETURN(0);
1755                 rc = llog_origin_handle_next_block(req);
1756                 ldlm_callback_reply(req, rc);
1757                 RETURN(0);
1758         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1759                 req_capsule_set(&req->rq_pill,
1760                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1761                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1762                         RETURN(0);
1763                 rc = llog_origin_handle_read_header(req);
1764                 ldlm_callback_reply(req, rc);
1765                 RETURN(0);
1766         case LLOG_ORIGIN_HANDLE_CLOSE:
1767                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1768                         RETURN(0);
1769                 rc = llog_origin_handle_close(req);
1770                 ldlm_callback_reply(req, rc);
1771                 RETURN(0);
1772         default:
1773                 CERROR("unknown opcode %u\n",
1774                        lustre_msg_get_opc(req->rq_reqmsg));
1775                 ldlm_callback_reply(req, -EPROTO);
1776                 RETURN(0);
1777         }
1778
1779         ns = req->rq_export->exp_obd->obd_namespace;
1780         LASSERT(ns != NULL);
1781
1782         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1783
1784         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1785         if (dlm_req == NULL) {
1786                 ldlm_callback_reply(req, -EPROTO);
1787                 RETURN(0);
1788         }
1789
1790         /* Force a known safe race, send a cancel to the server for a lock
1791          * which the server has already started a blocking callback on. */
1792         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1793             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1794                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1795                 if (rc < 0)
1796                         CERROR("ldlm_cli_cancel: %d\n", rc);
1797         }
1798
1799         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
1800         if (!lock) {
1801                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1802                        "disappeared\n", dlm_req->lock_handle[0].cookie);
1803                 ldlm_callback_reply(req, -EINVAL);
1804                 RETURN(0);
1805         }
1806
1807         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
1808             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
1809                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
1810
1811         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1812         lock_res_and_lock(lock);
1813         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1814         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1815                 /* If somebody cancels lock and cache is already droped,
1816                  * or lock is failed before cp_ast received on client,
1817                  * we can tell the server we have no lock. Otherwise, we
1818                  * should send cancel after dropping the cache. */
1819                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
1820                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
1821                     (lock->l_flags & LDLM_FL_FAILED)) {
1822                         LDLM_DEBUG(lock, "callback on lock "
1823                                    LPX64" - lock disappeared\n",
1824                                    dlm_req->lock_handle[0].cookie);
1825                         unlock_res_and_lock(lock);
1826                         LDLM_LOCK_RELEASE(lock);
1827                         ldlm_callback_reply(req, -EINVAL);
1828                         RETURN(0);
1829                 }
1830                 /* BL_AST locks are not needed in lru.
1831                  * let ldlm_cancel_lru() be fast. */
1832                 ldlm_lock_remove_from_lru(lock);
1833                 lock->l_flags |= LDLM_FL_BL_AST;
1834         }
1835         unlock_res_and_lock(lock);
1836
1837         /* We want the ost thread to get this reply so that it can respond
1838          * to ost requests (write cache writeback) that might be triggered
1839          * in the callback.
1840          *
1841          * But we'd also like to be able to indicate in the reply that we're
1842          * cancelling right now, because it's unused, or have an intent result
1843          * in the reply, so we might have to push the responsibility for sending
1844          * the reply down into the AST handlers, alas. */
1845
1846         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1847         case LDLM_BL_CALLBACK:
1848                 CDEBUG(D_INODE, "blocking ast\n");
1849                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1850                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1851                         ldlm_callback_reply(req, 0);
1852                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1853                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1854                 break;
1855         case LDLM_CP_CALLBACK:
1856                 CDEBUG(D_INODE, "completion ast\n");
1857                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1858                 ldlm_callback_reply(req, 0);
1859                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1860                 break;
1861         case LDLM_GL_CALLBACK:
1862                 CDEBUG(D_INODE, "glimpse ast\n");
1863                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1864                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1865                 break;
1866         default:
1867                 LBUG();                         /* checked above */
1868         }
1869
1870         RETURN(0);
1871 }
1872
1873 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1874 {
1875         int rc;
1876         ENTRY;
1877
1878         /* Requests arrive in sender's byte order.  The ptlrpc service
1879          * handler has already checked and, if necessary, byte-swapped the
1880          * incoming request message body, but I am responsible for the
1881          * message buffers. */
1882
1883         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1884
1885         if (req->rq_export == NULL) {
1886                 struct ldlm_request *dlm_req;
1887
1888                 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1889                        lustre_msg_get_opc(req->rq_reqmsg),
1890                        libcfs_id2str(req->rq_peer),
1891                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1892
1893                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1894                 dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1895                 if (dlm_req != NULL)
1896                         ldlm_lock_dump_handle(D_ERROR,
1897                                               &dlm_req->lock_handle[0]);
1898                 ldlm_callback_reply(req, -ENOTCONN);
1899                 RETURN(0);
1900         }
1901
1902         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1903
1904         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1905         case LDLM_CANCEL:
1906                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1907                 CDEBUG(D_INODE, "cancel\n");
1908                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1909                         RETURN(0);
1910                 rc = ldlm_handle_cancel(req);
1911                 if (rc)
1912                         break;
1913                 RETURN(0);
1914         case OBD_LOG_CANCEL:
1915                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1916                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1917                         RETURN(0);
1918                 rc = llog_origin_handle_cancel(req);
1919                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1920                         RETURN(0);
1921                 ldlm_callback_reply(req, rc);
1922                 RETURN(0);
1923         default:
1924                 CERROR("invalid opcode %d\n",
1925                        lustre_msg_get_opc(req->rq_reqmsg));
1926                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1927                 ldlm_callback_reply(req, -EINVAL);
1928         }
1929
1930         RETURN(0);
1931 }
1932
1933 void ldlm_revoke_lock_cb(void *obj, void *data)
1934 {
1935         struct list_head   *rpc_list = data;
1936         struct ldlm_lock   *lock = obj;
1937
1938         lock_res_and_lock(lock);
1939
1940         if (lock->l_req_mode != lock->l_granted_mode) {
1941                 unlock_res_and_lock(lock);
1942                 return;
1943         }
1944
1945         LASSERT(lock->l_resource);
1946         if (lock->l_resource->lr_type != LDLM_IBITS &&
1947             lock->l_resource->lr_type != LDLM_PLAIN) {
1948                 unlock_res_and_lock(lock);
1949                 return;
1950         }
1951
1952         if (lock->l_flags & LDLM_FL_AST_SENT) {
1953                 unlock_res_and_lock(lock);
1954                 return;
1955         }
1956
1957         LASSERT(lock->l_blocking_ast);
1958         LASSERT(!lock->l_blocking_lock);
1959
1960         lock->l_flags |= LDLM_FL_AST_SENT;
1961         if (lock->l_export && lock->l_export->exp_lock_hash &&
1962             !hlist_unhashed(&lock->l_exp_hash))
1963                 lustre_hash_del(lock->l_export->exp_lock_hash,
1964                                 &lock->l_remote_handle, &lock->l_exp_hash);
1965         list_add_tail(&lock->l_rk_ast, rpc_list);
1966         LDLM_LOCK_GET(lock);
1967
1968         unlock_res_and_lock(lock);
1969 }
1970
1971 void ldlm_revoke_export_locks(struct obd_export *exp)
1972 {
1973         struct list_head  rpc_list;
1974         ENTRY;
1975
1976         CFS_INIT_LIST_HEAD(&rpc_list);
1977         lustre_hash_for_each_empty(exp->exp_lock_hash,
1978                                    ldlm_revoke_lock_cb, &rpc_list);
1979         ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
1980
1981         EXIT;
1982 }
1983
1984 #ifdef __KERNEL__
1985 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1986 {
1987         struct ldlm_bl_work_item *blwi = NULL;
1988         static unsigned int num_bl = 0;
1989
1990         spin_lock(&blp->blp_lock);
1991         /* process a request from the blp_list at least every blp_num_threads */
1992         if (!list_empty(&blp->blp_list) &&
1993             (list_empty(&blp->blp_prio_list) || num_bl == 0))
1994                 blwi = list_entry(blp->blp_list.next,
1995                                   struct ldlm_bl_work_item, blwi_entry);
1996         else
1997                 if (!list_empty(&blp->blp_prio_list))
1998                         blwi = list_entry(blp->blp_prio_list.next,
1999                                           struct ldlm_bl_work_item, blwi_entry);
2000
2001         if (blwi) {
2002                 if (++num_bl >= atomic_read(&blp->blp_num_threads))
2003                         num_bl = 0;
2004                 list_del(&blwi->blwi_entry);
2005         }
2006         spin_unlock(&blp->blp_lock);
2007
2008         return blwi;
2009 }
2010
2011 /* This only contains temporary data until the thread starts */
2012 struct ldlm_bl_thread_data {
2013         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2014         struct ldlm_bl_pool     *bltd_blp;
2015         struct completion       bltd_comp;
2016         int                     bltd_num;
2017 };
2018
2019 static int ldlm_bl_thread_main(void *arg);
2020
2021 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2022 {
2023         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2024         int rc;
2025
2026         init_completion(&bltd.bltd_comp);
2027         rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
2028         if (rc < 0) {
2029                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2030                        atomic_read(&blp->blp_num_threads), rc);
2031                 return rc;
2032         }
2033         wait_for_completion(&bltd.bltd_comp);
2034
2035         return 0;
2036 }
2037
2038 static int ldlm_bl_thread_main(void *arg)
2039 {
2040         struct ldlm_bl_pool *blp;
2041         ENTRY;
2042
2043         {
2044                 struct ldlm_bl_thread_data *bltd = arg;
2045
2046                 blp = bltd->bltd_blp;
2047
2048                 bltd->bltd_num = atomic_inc_return(&blp->blp_num_threads) - 1;
2049                 atomic_inc(&blp->blp_busy_threads);
2050
2051                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2052                         "ldlm_bl_%02d", bltd->bltd_num);
2053                 cfs_daemonize(bltd->bltd_name);
2054
2055                 complete(&bltd->bltd_comp);
2056                 /* cannot use bltd after this, it is only on caller's stack */
2057         }
2058
2059         while (1) {
2060                 struct l_wait_info lwi = { 0 };
2061                 struct ldlm_bl_work_item *blwi = NULL;
2062
2063                 blwi = ldlm_bl_get_work(blp);
2064
2065                 if (blwi == NULL) {
2066                         int busy;
2067
2068                         atomic_dec(&blp->blp_busy_threads);
2069                         l_wait_event_exclusive(blp->blp_waitq,
2070                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2071                                          &lwi);
2072                         busy = atomic_inc_return(&blp->blp_busy_threads);
2073
2074                         if (blwi->blwi_ns == NULL)
2075                                 /* added by ldlm_cleanup() */
2076                                 break;
2077
2078                         /* Not fatal if racy and have a few too many threads */
2079                         if (unlikely(busy < blp->blp_max_threads &&
2080                                     busy >= atomic_read(&blp->blp_num_threads)))
2081                                 /* discard the return value, we tried */
2082                                 ldlm_bl_thread_start(blp);
2083                 } else {
2084                         if (blwi->blwi_ns == NULL)
2085                                 /* added by ldlm_cleanup() */
2086                                 break;
2087                 }
2088
2089                 if (blwi->blwi_count) {
2090                         /* The special case when we cancel locks in lru
2091                          * asynchronously, we pass the list of locks here.
2092                          * Thus lock is marked LDLM_FL_CANCELING, and already
2093                          * canceled locally. */
2094                         ldlm_cli_cancel_list(&blwi->blwi_head,
2095                                              blwi->blwi_count, NULL, 0);
2096                 } else {
2097                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2098                                                 blwi->blwi_lock);
2099                 }
2100                 OBD_FREE(blwi, sizeof(*blwi));
2101         }
2102
2103         atomic_dec(&blp->blp_busy_threads);
2104         atomic_dec(&blp->blp_num_threads);
2105         complete(&blp->blp_comp);
2106         RETURN(0);
2107 }
2108
2109 #endif
2110
2111 static int ldlm_setup(void);
2112 static int ldlm_cleanup(void);
2113
2114 int ldlm_get_ref(void)
2115 {
2116         int rc = 0;
2117         ENTRY;
2118         mutex_down(&ldlm_ref_sem);
2119         if (++ldlm_refcount == 1) {
2120                 rc = ldlm_setup();
2121                 if (rc)
2122                         ldlm_refcount--;
2123         }
2124         mutex_up(&ldlm_ref_sem);
2125
2126         RETURN(rc);
2127 }
2128
2129 void ldlm_put_ref(void)
2130 {
2131         ENTRY;
2132         mutex_down(&ldlm_ref_sem);
2133         if (ldlm_refcount == 1) {
2134                 int rc = ldlm_cleanup();
2135                 if (rc)
2136                         CERROR("ldlm_cleanup failed: %d\n", rc);
2137                 else
2138                         ldlm_refcount--;
2139         } else {
2140                 ldlm_refcount--;
2141         }
2142         mutex_up(&ldlm_ref_sem);
2143
2144         EXIT;
2145 }
2146
2147 /*
2148  * Export handle<->lock hash operations.
2149  */
2150 static unsigned
2151 ldlm_export_lock_hash(lustre_hash_t *lh, void *key, unsigned mask)
2152 {
2153         return lh_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2154 }
2155
2156 static void *
2157 ldlm_export_lock_key(struct hlist_node *hnode)
2158 {
2159         struct ldlm_lock *lock;
2160         ENTRY;
2161
2162         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2163         RETURN(&lock->l_remote_handle);
2164 }
2165
2166 static int
2167 ldlm_export_lock_compare(void *key, struct hlist_node *hnode)
2168 {
2169         ENTRY;
2170         RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
2171 }
2172
2173 static void *
2174 ldlm_export_lock_get(struct hlist_node *hnode)
2175 {
2176         struct ldlm_lock *lock;
2177         ENTRY;
2178
2179         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2180         LDLM_LOCK_GET(lock);
2181
2182         RETURN(lock);
2183 }
2184
2185 static void *
2186 ldlm_export_lock_put(struct hlist_node *hnode)
2187 {
2188         struct ldlm_lock *lock;
2189         ENTRY;
2190
2191         lock = hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2192         LDLM_LOCK_RELEASE(lock);
2193
2194         RETURN(lock);
2195 }
2196
2197 static lustre_hash_ops_t ldlm_export_lock_ops = {
2198         .lh_hash    = ldlm_export_lock_hash,
2199         .lh_key     = ldlm_export_lock_key,
2200         .lh_compare = ldlm_export_lock_compare,
2201         .lh_get     = ldlm_export_lock_get,
2202         .lh_put     = ldlm_export_lock_put
2203 };
2204
2205 int ldlm_init_export(struct obd_export *exp)
2206 {
2207         ENTRY;
2208
2209         exp->exp_lock_hash =
2210                 lustre_hash_init(obd_uuid2str(&exp->exp_client_uuid),
2211                                  7, 16, &ldlm_export_lock_ops, LH_REHASH);
2212
2213         if (!exp->exp_lock_hash)
2214                 RETURN(-ENOMEM);
2215
2216         RETURN(0);
2217 }
2218 EXPORT_SYMBOL(ldlm_init_export);
2219
2220 void ldlm_destroy_export(struct obd_export *exp)
2221 {
2222         ENTRY;
2223         lustre_hash_exit(exp->exp_lock_hash);
2224         exp->exp_lock_hash = NULL;
2225         EXIT;
2226 }
2227 EXPORT_SYMBOL(ldlm_destroy_export);
2228
2229 static int ldlm_setup(void)
2230 {
2231         struct ldlm_bl_pool *blp;
2232         int rc = 0;
2233         int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2234         int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2235 #ifdef __KERNEL__
2236         int i;
2237 #endif
2238         ENTRY;
2239
2240         if (ldlm_state != NULL)
2241                 RETURN(-EALREADY);
2242
2243         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2244         if (ldlm_state == NULL)
2245                 RETURN(-ENOMEM);
2246
2247 #ifdef LPROCFS
2248         rc = ldlm_proc_setup();
2249         if (rc != 0)
2250                 GOTO(out_free, rc);
2251 #endif
2252
2253 #ifdef __KERNEL__
2254         if (ldlm_num_threads) {
2255                 /* If ldlm_num_threads is set, it is the min and the max. */
2256                 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2257                         ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2258                 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2259                         ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2260                 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2261         }
2262 #endif
2263
2264         ldlm_state->ldlm_cb_service =
2265                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2266                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2267                                 LDLM_CB_REPLY_PORTAL, 2,
2268                                 ldlm_callback_handler, "ldlm_cbd",
2269                                 ldlm_svc_proc_dir, NULL,
2270                                 ldlm_min_threads, ldlm_max_threads,
2271                                 "ldlm_cb",
2272                                 LCT_MD_THREAD|LCT_DT_THREAD, NULL);
2273
2274         if (!ldlm_state->ldlm_cb_service) {
2275                 CERROR("failed to start service\n");
2276                 GOTO(out_proc, rc = -ENOMEM);
2277         }
2278
2279         ldlm_state->ldlm_cancel_service =
2280                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2281                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2282                                 LDLM_CANCEL_REPLY_PORTAL, 6,
2283                                 ldlm_cancel_handler, "ldlm_canceld",
2284                                 ldlm_svc_proc_dir, NULL,
2285                                 ldlm_min_threads, ldlm_max_threads,
2286                                 "ldlm_cn",
2287                                 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
2288                                 NULL);
2289
2290         if (!ldlm_state->ldlm_cancel_service) {
2291                 CERROR("failed to start service\n");
2292                 GOTO(out_proc, rc = -ENOMEM);
2293         }
2294
2295         OBD_ALLOC(blp, sizeof(*blp));
2296         if (blp == NULL)
2297                 GOTO(out_proc, rc = -ENOMEM);
2298         ldlm_state->ldlm_bl_pool = blp;
2299
2300         spin_lock_init(&blp->blp_lock);
2301         CFS_INIT_LIST_HEAD(&blp->blp_list);
2302         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2303         cfs_waitq_init(&blp->blp_waitq);
2304         atomic_set(&blp->blp_num_threads, 0);
2305         atomic_set(&blp->blp_busy_threads, 0);
2306         blp->blp_min_threads = ldlm_min_threads;
2307         blp->blp_max_threads = ldlm_max_threads;
2308
2309 #ifdef __KERNEL__
2310         for (i = 0; i < blp->blp_min_threads; i++) {
2311                 rc = ldlm_bl_thread_start(blp);
2312                 if (rc < 0)
2313                         GOTO(out_thread, rc);
2314         }
2315
2316         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
2317         if (rc)
2318                 GOTO(out_thread, rc);
2319
2320         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
2321         if (rc)
2322                 GOTO(out_thread, rc);
2323
2324         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2325         expired_lock_thread.elt_state = ELT_STOPPED;
2326         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2327
2328         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2329         spin_lock_init(&waiting_locks_spinlock);
2330         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2331
2332         rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2333         if (rc < 0) {
2334                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2335                 GOTO(out_thread, rc);
2336         }
2337
2338         wait_event(expired_lock_thread.elt_waitq,
2339                    expired_lock_thread.elt_state == ELT_READY);
2340 #endif
2341
2342 #ifdef __KERNEL__
2343         rc = ldlm_pools_init();
2344         if (rc)
2345                 GOTO(out_thread, rc);
2346 #endif
2347         RETURN(0);
2348
2349 #ifdef __KERNEL__
2350  out_thread:
2351         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2352         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2353 #endif
2354
2355  out_proc:
2356 #ifdef LPROCFS
2357         ldlm_proc_cleanup();
2358  out_free:
2359 #endif
2360         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2361         ldlm_state = NULL;
2362         return rc;
2363 }
2364
2365 static int ldlm_cleanup(void)
2366 {
2367 #ifdef __KERNEL__
2368         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2369 #endif
2370         ENTRY;
2371
2372         if (!list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2373             !list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2374                 CERROR("ldlm still has namespaces; clean these up first.\n");
2375                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2376                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2377                 RETURN(-EBUSY);
2378         }
2379
2380 #ifdef __KERNEL__
2381         ldlm_pools_fini();
2382 #endif
2383
2384 #ifdef __KERNEL__
2385         while (atomic_read(&blp->blp_num_threads) > 0) {
2386                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2387
2388                 init_completion(&blp->blp_comp);
2389
2390                 spin_lock(&blp->blp_lock);
2391                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2392                 cfs_waitq_signal(&blp->blp_waitq);
2393                 spin_unlock(&blp->blp_lock);
2394
2395                 wait_for_completion(&blp->blp_comp);
2396         }
2397         OBD_FREE(blp, sizeof(*blp));
2398
2399         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2400         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2401         ldlm_proc_cleanup();
2402
2403         expired_lock_thread.elt_state = ELT_TERMINATE;
2404         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2405         wait_event(expired_lock_thread.elt_waitq,
2406                    expired_lock_thread.elt_state == ELT_STOPPED);
2407 #else
2408         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2409         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2410 #endif
2411
2412         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2413         ldlm_state = NULL;
2414
2415         RETURN(0);
2416 }
2417
2418 int __init ldlm_init(void)
2419 {
2420         init_mutex(&ldlm_ref_sem);
2421         init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2422         init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2423         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2424                                                sizeof(struct ldlm_resource), 0,
2425                                                SLAB_HWCACHE_ALIGN);
2426         if (ldlm_resource_slab == NULL)
2427                 return -ENOMEM;
2428
2429         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2430                                       sizeof(struct ldlm_lock), 0,
2431                                       SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
2432         if (ldlm_lock_slab == NULL) {
2433                 cfs_mem_cache_destroy(ldlm_resource_slab);
2434                 return -ENOMEM;
2435         }
2436
2437         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2438                                         sizeof(struct ldlm_interval),
2439                                         0, SLAB_HWCACHE_ALIGN);
2440         if (ldlm_interval_slab == NULL) {
2441                 cfs_mem_cache_destroy(ldlm_resource_slab);
2442                 cfs_mem_cache_destroy(ldlm_lock_slab);
2443                 return -ENOMEM;
2444         }
2445
2446         return 0;
2447 }
2448
2449 void __exit ldlm_exit(void)
2450 {
2451         int rc;
2452         if (ldlm_refcount)
2453                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2454         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2455         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2456 #ifdef __KERNEL__
2457         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2458          * synchronize_rcu() to wait a grace period elapsed, so that
2459          * ldlm_lock_free() get a chance to be called. */
2460         synchronize_rcu();
2461 #endif
2462         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2463         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2464         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2465         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2466 }
2467
2468 /* ldlm_extent.c */
2469 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2470
2471 /* ldlm_lock.c */
2472 EXPORT_SYMBOL(ldlm_get_processing_policy);
2473 EXPORT_SYMBOL(ldlm_lock2desc);
2474 EXPORT_SYMBOL(ldlm_register_intent);
2475 EXPORT_SYMBOL(ldlm_lockname);
2476 EXPORT_SYMBOL(ldlm_typename);
2477 EXPORT_SYMBOL(ldlm_lock2handle);
2478 EXPORT_SYMBOL(__ldlm_handle2lock);
2479 EXPORT_SYMBOL(ldlm_lock_get);
2480 EXPORT_SYMBOL(ldlm_lock_put);
2481 EXPORT_SYMBOL(ldlm_lock_match);
2482 EXPORT_SYMBOL(ldlm_lock_cancel);
2483 EXPORT_SYMBOL(ldlm_lock_addref);
2484 EXPORT_SYMBOL(ldlm_lock_addref_try);
2485 EXPORT_SYMBOL(ldlm_lock_decref);
2486 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2487 EXPORT_SYMBOL(ldlm_lock_change_resource);
2488 EXPORT_SYMBOL(ldlm_it2str);
2489 EXPORT_SYMBOL(ldlm_lock_dump);
2490 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2491 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2492 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2493 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2494 EXPORT_SYMBOL(ldlm_lock_allow_match);
2495 EXPORT_SYMBOL(ldlm_lock_downgrade);
2496 EXPORT_SYMBOL(ldlm_lock_convert);
2497
2498 /* ldlm_request.c */
2499 EXPORT_SYMBOL(ldlm_completion_ast_async);
2500 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2501 EXPORT_SYMBOL(ldlm_completion_ast);
2502 EXPORT_SYMBOL(ldlm_blocking_ast);
2503 EXPORT_SYMBOL(ldlm_glimpse_ast);
2504 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2505 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2506 EXPORT_SYMBOL(ldlm_prep_elc_req);
2507 EXPORT_SYMBOL(ldlm_cli_convert);
2508 EXPORT_SYMBOL(ldlm_cli_enqueue);
2509 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2510 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2511 EXPORT_SYMBOL(ldlm_cli_cancel);
2512 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2513 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2514 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2515 EXPORT_SYMBOL(ldlm_replay_locks);
2516 EXPORT_SYMBOL(ldlm_resource_foreach);
2517 EXPORT_SYMBOL(ldlm_namespace_foreach);
2518 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2519 EXPORT_SYMBOL(ldlm_resource_iterate);
2520 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2521 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2522
2523 /* ldlm_lockd.c */
2524 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2525 EXPORT_SYMBOL(ldlm_server_completion_ast);
2526 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2527 EXPORT_SYMBOL(ldlm_handle_enqueue);
2528 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2529 EXPORT_SYMBOL(ldlm_handle_cancel);
2530 EXPORT_SYMBOL(ldlm_request_cancel);
2531 EXPORT_SYMBOL(ldlm_handle_convert);
2532 EXPORT_SYMBOL(ldlm_handle_convert0);
2533 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2534 EXPORT_SYMBOL(ldlm_get_ref);
2535 EXPORT_SYMBOL(ldlm_put_ref);
2536 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2537 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2538
2539 /* ldlm_resource.c */
2540 EXPORT_SYMBOL(ldlm_namespace_new);
2541 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2542 EXPORT_SYMBOL(ldlm_namespace_free);
2543 EXPORT_SYMBOL(ldlm_namespace_dump);
2544 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2545 EXPORT_SYMBOL(ldlm_resource_get);
2546 EXPORT_SYMBOL(ldlm_resource_putref);
2547 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2548
2549 /* ldlm_lib.c */
2550 EXPORT_SYMBOL(client_import_add_conn);
2551 EXPORT_SYMBOL(client_import_del_conn);
2552 EXPORT_SYMBOL(client_obd_setup);
2553 EXPORT_SYMBOL(client_obd_cleanup);
2554 EXPORT_SYMBOL(client_connect_import);
2555 EXPORT_SYMBOL(client_disconnect_export);
2556 EXPORT_SYMBOL(target_stop_recovery_thread);
2557 EXPORT_SYMBOL(target_handle_connect);
2558 EXPORT_SYMBOL(target_cleanup_recovery);
2559 EXPORT_SYMBOL(target_destroy_export);
2560 EXPORT_SYMBOL(target_cancel_recovery_timer);
2561 EXPORT_SYMBOL(target_send_reply);
2562 EXPORT_SYMBOL(target_queue_recovery_request);
2563 EXPORT_SYMBOL(target_handle_ping);
2564 EXPORT_SYMBOL(target_pack_pool_reply);
2565 EXPORT_SYMBOL(target_handle_disconnect);
2566
2567 /* l_lock.c */
2568 EXPORT_SYMBOL(lock_res_and_lock);
2569 EXPORT_SYMBOL(unlock_res_and_lock);