lustre/ldlm/ldlm_lockd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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 #define DEBUG_SUBSYSTEM S_LDLM
  43
  44 #ifdef __KERNEL__
  45 # include <libcfs/libcfs.h>
  46 #else
  47 # include <liblustre.h>
  48 #endif
  49
  50 #include <lustre_dlm.h>
  51 #include <obd_class.h>
  52 #include <libcfs/list.h>
  53 #include "ldlm_internal.h"
  54
  55 static int ldlm_num_threads;
  56 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
  57                 "number of DLM service threads to start");
  58
  59 static char *ldlm_cpts;
  60 CFS_MODULE_PARM(ldlm_cpts, "s", charp, 0444,
  61                 "CPU partitions ldlm threads should run on");
  62
  63 extern cfs_mem_cache_t *ldlm_resource_slab;
  64 extern cfs_mem_cache_t *ldlm_lock_slab;
  65 static struct mutex     ldlm_ref_mutex;
  66 static int ldlm_refcount;
  67
  68 struct ldlm_cb_async_args {
  69         struct ldlm_cb_set_arg *ca_set_arg;
  70         struct ldlm_lock       *ca_lock;
  71 };
  72
  73 /* LDLM state */
  74
  75 static struct ldlm_state *ldlm_state;
  76
  77 inline cfs_time_t round_timeout(cfs_time_t timeout)
  78 {
  79         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
  80 }
  81
  82 /* timeout for initial callback (AST) reply (bz10399) */
  83 static inline unsigned int ldlm_get_rq_timeout(void)
  84 {
  85         /* Non-AT value */
  86         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
  87
  88         return timeout < 1 ? 1 : timeout;
  89 }
  90
  91 #define ELT_STOPPED   0
  92 #define ELT_READY     1
  93 #define ELT_TERMINATE 2
  94
  95 struct ldlm_bl_pool {
  96         spinlock_t              blp_lock;
  97
  98         /*
  99          * blp_prio_list is used for callbacks that should be handled
 100          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
 101          * see bug 13843
 102          */
 103         cfs_list_t              blp_prio_list;
 104
 105         /*
 106          * blp_list is used for all other callbacks which are likely
 107          * to take longer to process.
 108          */
 109         cfs_list_t              blp_list;
 110
 111         cfs_waitq_t             blp_waitq;
 112         struct completion        blp_comp;
 113         cfs_atomic_t            blp_num_threads;
 114         cfs_atomic_t            blp_busy_threads;
 115         int                     blp_min_threads;
 116         int                     blp_max_threads;
 117 };
 118
 119 struct ldlm_bl_work_item {
 120         cfs_list_t              blwi_entry;
 121         struct ldlm_namespace  *blwi_ns;
 122         struct ldlm_lock_desc   blwi_ld;
 123         struct ldlm_lock       *blwi_lock;
 124         cfs_list_t              blwi_head;
 125         int                     blwi_count;
 126         struct completion        blwi_comp;
 127         int                     blwi_mode;
 128         int                     blwi_mem_pressure;
 129 };
 130
 131 #if defined(HAVE_SERVER_SUPPORT) && defined(__KERNEL__)
 132
 133 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
 134 static spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
 135 static cfs_list_t waiting_locks_list;
 136 static cfs_timer_t waiting_locks_timer;
 137
 138 static struct expired_lock_thread {
 139         cfs_waitq_t             elt_waitq;
 140         int                     elt_state;
 141         int                     elt_dump;
 142         cfs_list_t              elt_expired_locks;
 143 } expired_lock_thread;
 144
 145 static inline int have_expired_locks(void)
 146 {
 147         int need_to_run;
 148
 149         ENTRY;
 150         spin_lock_bh(&waiting_locks_spinlock);
 151         need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
 152         spin_unlock_bh(&waiting_locks_spinlock);
 153
 154         RETURN(need_to_run);
 155 }
 156
 157 static int expired_lock_main(void *arg)
 158 {
 159         cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
 160         struct l_wait_info lwi = { 0 };
 161         int do_dump;
 162
 163         ENTRY;
 164         cfs_daemonize("ldlm_elt");
 165
 166         expired_lock_thread.elt_state = ELT_READY;
 167         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 168
 169         while (1) {
 170                 l_wait_event(expired_lock_thread.elt_waitq,
 171                              have_expired_locks() ||
 172                              expired_lock_thread.elt_state == ELT_TERMINATE,
 173                              &lwi);
 174
 175                 spin_lock_bh(&waiting_locks_spinlock);
 176                 if (expired_lock_thread.elt_dump) {
 177                         struct libcfs_debug_msg_data msgdata = {
 178                                 .msg_file = __FILE__,
 179                                 .msg_fn = "waiting_locks_callback",
 180                                 .msg_line = expired_lock_thread.elt_dump };
 181                         spin_unlock_bh(&waiting_locks_spinlock);
 182
 183                         /* from waiting_locks_callback, but not in timer */
 184                         libcfs_debug_dumplog();
 185                         libcfs_run_lbug_upcall(&msgdata);
 186
 187                         spin_lock_bh(&waiting_locks_spinlock);
 188                         expired_lock_thread.elt_dump = 0;
 189                 }
 190
 191                 do_dump = 0;
 192
 193                 while (!cfs_list_empty(expired)) {
 194                         struct obd_export *export;
 195                         struct ldlm_lock *lock;
 196
 197                         lock = cfs_list_entry(expired->next, struct ldlm_lock,
 198                                           l_pending_chain);
 199                         if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
 200                             (void *)lock >= LP_POISON) {
 201                                 spin_unlock_bh(&waiting_locks_spinlock);
 202                                 CERROR("free lock on elt list %p\n", lock);
 203                                 LBUG();
 204                         }
 205                         cfs_list_del_init(&lock->l_pending_chain);
 206                         if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
 207                             (void *)lock->l_export >= LP_POISON) {
 208                                 CERROR("lock with free export on elt list %p\n",
 209                                        lock->l_export);
 210                                 lock->l_export = NULL;
 211                                 LDLM_ERROR(lock, "free export");
 212                                 /* release extra ref grabbed by
 213                                  * ldlm_add_waiting_lock() or
 214                                  * ldlm_failed_ast() */
 215                                 LDLM_LOCK_RELEASE(lock);
 216                                 continue;
 217                         }
 218
 219                         if (lock->l_destroyed) {
 220                                 /* release the lock refcount where
 221                                  * waiting_locks_callback() founds */
 222                                 LDLM_LOCK_RELEASE(lock);
 223                                 continue;
 224                         }
 225                         export = class_export_lock_get(lock->l_export, lock);
 226                         spin_unlock_bh(&waiting_locks_spinlock);
 227
 228                         do_dump++;
 229                         class_fail_export(export);
 230                         class_export_lock_put(export, lock);
 231
 232                         /* release extra ref grabbed by ldlm_add_waiting_lock()
 233                          * or ldlm_failed_ast() */
 234                         LDLM_LOCK_RELEASE(lock);
 235
 236                         spin_lock_bh(&waiting_locks_spinlock);
 237                 }
 238                 spin_unlock_bh(&waiting_locks_spinlock);
 239
 240                 if (do_dump && obd_dump_on_eviction) {
 241                         CERROR("dump the log upon eviction\n");
 242                         libcfs_debug_dumplog();
 243                 }
 244
 245                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 246                         break;
 247         }
 248
 249         expired_lock_thread.elt_state = ELT_STOPPED;
 250         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 251         RETURN(0);
 252 }
 253
 254 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 255 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds);
 256
 257 /**
 258  * Check if there is a request in the export request list
 259  * which prevents the lock canceling.
 260  */
 261 static int ldlm_lock_busy(struct ldlm_lock *lock)
 262 {
 263         struct ptlrpc_request *req;
 264         int match = 0;
 265         ENTRY;
 266
 267         if (lock->l_export == NULL)
 268                 return 0;
 269
 270         spin_lock_bh(&lock->l_export->exp_rpc_lock);
 271         cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
 272                                 rq_exp_list) {
 273                 if (req->rq_ops->hpreq_lock_match) {
 274                         match = req->rq_ops->hpreq_lock_match(req, lock);
 275                         if (match)
 276                                 break;
 277                 }
 278         }
 279         spin_unlock_bh(&lock->l_export->exp_rpc_lock);
 280         RETURN(match);
 281 }
 282
 283 /* This is called from within a timer interrupt and cannot schedule */
 284 static void waiting_locks_callback(unsigned long unused)
 285 {
 286         struct ldlm_lock        *lock;
 287         int                     need_dump = 0;
 288
 289         spin_lock_bh(&waiting_locks_spinlock);
 290         while (!cfs_list_empty(&waiting_locks_list)) {
 291                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
 292                                       l_pending_chain);
 293                 if (cfs_time_after(lock->l_callback_timeout,
 294                                    cfs_time_current()) ||
 295                     (lock->l_req_mode == LCK_GROUP))
 296                         break;
 297
 298                 if (ptlrpc_check_suspend()) {
 299                         /* there is a case when we talk to one mds, holding
 300                          * lock from another mds. this way we easily can get
 301                          * here, if second mds is being recovered. so, we
 302                          * suspend timeouts. bug 6019 */
 303
 304                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 305                                    lock->l_export->exp_client_uuid.uuid,
 306                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 307                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 308
 309                         cfs_list_del_init(&lock->l_pending_chain);
 310                         if (lock->l_destroyed) {
 311                                 /* relay the lock refcount decrease to
 312                                  * expired lock thread */
 313                                 cfs_list_add(&lock->l_pending_chain,
 314                                         &expired_lock_thread.elt_expired_locks);
 315                         } else {
 316                                 __ldlm_add_waiting_lock(lock,
 317                                                 ldlm_get_enq_timeout(lock));
 318                         }
 319                         continue;
 320                 }
 321
 322                 /* if timeout overlaps the activation time of suspended timeouts
 323                  * then extend it to give a chance for client to reconnect */
 324                 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
 325                                                  cfs_time_seconds(obd_timeout)/2),
 326                                     ptlrpc_suspend_wakeup_time())) {
 327                         LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
 328                                    lock->l_export->exp_client_uuid.uuid,
 329                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 330                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 331
 332                         cfs_list_del_init(&lock->l_pending_chain);
 333                         if (lock->l_destroyed) {
 334                                 /* relay the lock refcount decrease to
 335                                  * expired lock thread */
 336                                 cfs_list_add(&lock->l_pending_chain,
 337                                         &expired_lock_thread.elt_expired_locks);
 338                         } else {
 339                                 __ldlm_add_waiting_lock(lock,
 340                                                 ldlm_get_enq_timeout(lock));
 341                         }
 342                         continue;
 343                 }
 344
 345                 /* Check if we need to prolong timeout */
 346                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
 347                     ldlm_lock_busy(lock)) {
 348                         int cont = 1;
 349
 350                         if (lock->l_pending_chain.next == &waiting_locks_list)
 351                                 cont = 0;
 352
 353                         LDLM_LOCK_GET(lock);
 354
 355                         spin_unlock_bh(&waiting_locks_spinlock);
 356                         LDLM_DEBUG(lock, "prolong the busy lock");
 357                         ldlm_refresh_waiting_lock(lock,
 358                                                   ldlm_get_enq_timeout(lock));
 359                         spin_lock_bh(&waiting_locks_spinlock);
 360
 361                         if (!cont) {
 362                                 LDLM_LOCK_RELEASE(lock);
 363                                 break;
 364                         }
 365
 366                         LDLM_LOCK_RELEASE(lock);
 367                         continue;
 368                 }
 369                 ldlm_lock_to_ns(lock)->ns_timeouts++;
 370                 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
 371                            "evicting client at %s ",
 372                            cfs_time_current_sec()- lock->l_last_activity,
 373                            libcfs_nid2str(
 374                                    lock->l_export->exp_connection->c_peer.nid));
 375
 376                 /* no needs to take an extra ref on the lock since it was in
 377                  * the waiting_locks_list and ldlm_add_waiting_lock()
 378                  * already grabbed a ref */
 379                 cfs_list_del(&lock->l_pending_chain);
 380                 cfs_list_add(&lock->l_pending_chain,
 381                              &expired_lock_thread.elt_expired_locks);
 382                 need_dump = 1;
 383         }
 384
 385         if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
 386                 if (obd_dump_on_timeout && need_dump)
 387                         expired_lock_thread.elt_dump = __LINE__;
 388
 389                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 390         }
 391
 392         /*
 393          * Make sure the timer will fire again if we have any locks
 394          * left.
 395          */
 396         if (!cfs_list_empty(&waiting_locks_list)) {
 397                 cfs_time_t timeout_rounded;
 398                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
 399                                       l_pending_chain);
 400                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
 401                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 402         }
 403         spin_unlock_bh(&waiting_locks_spinlock);
 404 }
 405
 406 /*
 407  * Indicate that we're waiting for a client to call us back cancelling a given
 408  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 409  * timer to fire appropriately.  (We round up to the next second, to avoid
 410  * floods of timer firings during periods of high lock contention and traffic).
 411  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
 412  * if it has been added to the waiting list (1 is returned).
 413  *
 414  * Called with the namespace lock held.
 415  */
 416 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
 417 {
 418         cfs_time_t timeout;
 419         cfs_time_t timeout_rounded;
 420
 421         if (!cfs_list_empty(&lock->l_pending_chain))
 422                 return 0;
 423
 424         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
 425             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
 426                 seconds = 1;
 427
 428         timeout = cfs_time_shift(seconds);
 429         if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
 430                 lock->l_callback_timeout = timeout;
 431
 432         timeout_rounded = round_timeout(lock->l_callback_timeout);
 433
 434         if (cfs_time_before(timeout_rounded,
 435                             cfs_timer_deadline(&waiting_locks_timer)) ||
 436             !cfs_timer_is_armed(&waiting_locks_timer)) {
 437                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 438         }
 439         /* if the new lock has a shorter timeout than something earlier on
 440            the list, we'll wait the longer amount of time; no big deal. */
 441         /* FIFO */
 442         cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
 443         return 1;
 444 }
 445
 446 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 447 {
 448         int ret;
 449         int timeout = ldlm_get_enq_timeout(lock);
 450
 451         /* NB: must be called with hold of lock_res_and_lock() */
 452         LASSERT(lock->l_res_locked);
 453         lock->l_waited = 1;
 454
 455         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 456
 457         spin_lock_bh(&waiting_locks_spinlock);
 458         if (lock->l_destroyed) {
 459                 static cfs_time_t next;
 460                 spin_unlock_bh(&waiting_locks_spinlock);
 461                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
 462                 if (cfs_time_after(cfs_time_current(), next)) {
 463                         next = cfs_time_shift(14400);
 464                         libcfs_debug_dumpstack(NULL);
 465                 }
 466                 return 0;
 467         }
 468
 469         ret = __ldlm_add_waiting_lock(lock, timeout);
 470         if (ret) {
 471                 /* grab ref on the lock if it has been added to the
 472                  * waiting list */
 473                 LDLM_LOCK_GET(lock);
 474         }
 475         spin_unlock_bh(&waiting_locks_spinlock);
 476
 477         if (ret) {
 478                 spin_lock_bh(&lock->l_export->exp_bl_list_lock);
 479                 if (cfs_list_empty(&lock->l_exp_list))
 480                         cfs_list_add(&lock->l_exp_list,
 481                                      &lock->l_export->exp_bl_list);
 482                 spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
 483         }
 484
 485         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
 486                    ret == 0 ? "not re-" : "", timeout,
 487                    AT_OFF ? "off" : "on");
 488         return ret;
 489 }
 490
 491 /*
 492  * Remove a lock from the pending list, likely because it had its cancellation
 493  * callback arrive without incident.  This adjusts the lock-timeout timer if
 494  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 495  * As done by ldlm_del_waiting_lock(), the caller must release the lock
 496  * reference when the lock is removed from any list (1 is returned).
 497  *
 498  * Called with namespace lock held.
 499  */
 500 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
 501 {
 502         cfs_list_t *list_next;
 503
 504         if (cfs_list_empty(&lock->l_pending_chain))
 505                 return 0;
 506
 507         list_next = lock->l_pending_chain.next;
 508         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 509                 /* Removing the head of the list, adjust timer. */
 510                 if (list_next == &waiting_locks_list) {
 511                         /* No more, just cancel. */
 512                         cfs_timer_disarm(&waiting_locks_timer);
 513                 } else {
 514                         struct ldlm_lock *next;
 515                         next = cfs_list_entry(list_next, struct ldlm_lock,
 516                                               l_pending_chain);
 517                         cfs_timer_arm(&waiting_locks_timer,
 518                                       round_timeout(next->l_callback_timeout));
 519                 }
 520         }
 521         cfs_list_del_init(&lock->l_pending_chain);
 522
 523         return 1;
 524 }
 525
 526 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 527 {
 528         int ret;
 529
 530         if (lock->l_export == NULL) {
 531                 /* We don't have a "waiting locks list" on clients. */
 532                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
 533                 return 0;
 534         }
 535
 536         spin_lock_bh(&waiting_locks_spinlock);
 537         ret = __ldlm_del_waiting_lock(lock);
 538         spin_unlock_bh(&waiting_locks_spinlock);
 539
 540         /* remove the lock out of export blocking list */
 541         spin_lock_bh(&lock->l_export->exp_bl_list_lock);
 542         cfs_list_del_init(&lock->l_exp_list);
 543         spin_unlock_bh(&lock->l_export->exp_bl_list_lock);
 544
 545         if (ret) {
 546                 /* release lock ref if it has indeed been removed
 547                  * from a list */
 548                 LDLM_LOCK_RELEASE(lock);
 549         }
 550
 551         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
 552         return ret;
 553 }
 554 EXPORT_SYMBOL(ldlm_del_waiting_lock);
 555
 556 /*
 557  * Prolong the lock
 558  *
 559  * Called with namespace lock held.
 560  */
 561 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 562 {
 563         if (lock->l_export == NULL) {
 564                 /* We don't have a "waiting locks list" on clients. */
 565                 LDLM_DEBUG(lock, "client lock: no-op");
 566                 return 0;
 567         }
 568
 569         spin_lock_bh(&waiting_locks_spinlock);
 570
 571         if (cfs_list_empty(&lock->l_pending_chain)) {
 572                 spin_unlock_bh(&waiting_locks_spinlock);
 573                 LDLM_DEBUG(lock, "wasn't waiting");
 574                 return 0;
 575         }
 576
 577         /* we remove/add the lock to the waiting list, so no needs to
 578          * release/take a lock reference */
 579         __ldlm_del_waiting_lock(lock);
 580         __ldlm_add_waiting_lock(lock, timeout);
 581         spin_unlock_bh(&waiting_locks_spinlock);
 582
 583         LDLM_DEBUG(lock, "refreshed");
 584         return 1;
 585 }
 586 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
 587
 588 #else /* !HAVE_SERVER_SUPPORT ||  !__KERNEL__ */
 589
 590 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 591 {
 592         RETURN(0);
 593 }
 594
 595 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 596 {
 597         RETURN(0);
 598 }
 599
 600 # ifdef HAVE_SERVER_SUPPORT
 601 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 602 {
 603         LASSERT(lock->l_res_locked);
 604         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 605         RETURN(1);
 606 }
 607
 608 # endif
 609 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
 610
 611 #ifdef HAVE_SERVER_SUPPORT
 612
 613 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 614                             const char *ast_type)
 615 {
 616         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
 617                            "to a lock %s callback time out: rc %d\n",
 618                            lock->l_export->exp_obd->obd_name,
 619                            obd_export_nid2str(lock->l_export), ast_type, rc);
 620
 621         if (obd_dump_on_timeout)
 622                 libcfs_debug_dumplog();
 623 #ifdef __KERNEL__
 624         spin_lock_bh(&waiting_locks_spinlock);
 625         if (__ldlm_del_waiting_lock(lock) == 0)
 626                 /* the lock was not in any list, grab an extra ref before adding
 627                  * the lock to the expired list */
 628                 LDLM_LOCK_GET(lock);
 629         cfs_list_add(&lock->l_pending_chain,
 630                      &expired_lock_thread.elt_expired_locks);
 631         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 632         spin_unlock_bh(&waiting_locks_spinlock);
 633 #else
 634         class_fail_export(lock->l_export);
 635 #endif
 636 }
 637
 638 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 639                                  struct ptlrpc_request *req, int rc,
 640                                  const char *ast_type)
 641 {
 642         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
 643
 644         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 645                 LASSERT(lock->l_export);
 646                 if (lock->l_export->exp_libclient) {
 647                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 648                                    " timeout, just cancelling lock", ast_type,
 649                                    libcfs_nid2str(peer.nid));
 650                         ldlm_lock_cancel(lock);
 651                         rc = -ERESTART;
 652                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
 653                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
 654                                    "cancel was received (AST reply lost?)",
 655                                    ast_type, libcfs_nid2str(peer.nid));
 656                         ldlm_lock_cancel(lock);
 657                         rc = -ERESTART;
 658                 } else {
 659                         ldlm_del_waiting_lock(lock);
 660                         ldlm_failed_ast(lock, rc, ast_type);
 661                 }
 662         } else if (rc) {
 663                 if (rc == -EINVAL) {
 664                         struct ldlm_resource *res = lock->l_resource;
 665                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 666                                " from %s AST - normal race",
 667                                libcfs_nid2str(peer.nid),
 668                                req->rq_repmsg ?
 669                                lustre_msg_get_status(req->rq_repmsg) : -1,
 670                                ast_type);
 671                         if (res) {
 672                                 /* update lvbo to return proper attributes.
 673                                  * see bug 23174 */
 674                                 ldlm_resource_getref(res);
 675                                 ldlm_res_lvbo_update(res, NULL, 1);
 676                                 ldlm_resource_putref(res);
 677                         }
 678
 679                 } else {
 680                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 681                                    "from %s AST", libcfs_nid2str(peer.nid),
 682                                    (req->rq_repmsg != NULL) ?
 683                                    lustre_msg_get_status(req->rq_repmsg) : 0,
 684                                    ast_type);
 685                 }
 686                 ldlm_lock_cancel(lock);
 687                 /* Server-side AST functions are called from ldlm_reprocess_all,
 688                  * which needs to be told to please restart its reprocessing. */
 689                 rc = -ERESTART;
 690         }
 691
 692         return rc;
 693 }
 694
 695 static int ldlm_cb_interpret(const struct lu_env *env,
 696                              struct ptlrpc_request *req, void *data, int rc)
 697 {
 698         struct ldlm_cb_async_args *ca   = data;
 699         struct ldlm_lock          *lock = ca->ca_lock;
 700         struct ldlm_cb_set_arg    *arg  = ca->ca_set_arg;
 701         ENTRY;
 702
 703         LASSERT(lock != NULL);
 704
 705         switch (arg->type) {
 706         case LDLM_GL_CALLBACK:
 707                 /* Update the LVB from disk if the AST failed
 708                  * (this is a legal race)
 709                  *
 710                  * - Glimpse callback of local lock just returns
 711                  *   -ELDLM_NO_LOCK_DATA.
 712                  * - Glimpse callback of remote lock might return
 713                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
 714                  */
 715                 if (rc == -ELDLM_NO_LOCK_DATA) {
 716                         LDLM_DEBUG(lock, "lost race - client has a lock but no "
 717                                    "inode");
 718                         ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
 719                 } else if (rc != 0) {
 720                         rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 721                 } else {
 722                         rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
 723                 }
 724                 break;
 725         case LDLM_BL_CALLBACK:
 726                 if (rc != 0)
 727                         rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 728                 break;
 729         case LDLM_CP_CALLBACK:
 730                 if (rc != 0)
 731                         rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 732                 break;
 733         default:
 734                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
 735                            arg->type);
 736                 LBUG();
 737         }
 738
 739         /* release extra reference taken in ldlm_ast_fini() */
 740         LDLM_LOCK_RELEASE(lock);
 741
 742         if (rc == -ERESTART)
 743                 cfs_atomic_inc(&arg->restart);
 744
 745         RETURN(0);
 746 }
 747
 748 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
 749                                 struct ldlm_cb_set_arg *arg,
 750                                 struct ldlm_lock *lock,
 751                                 int instant_cancel)
 752 {
 753         int rc = 0;
 754         ENTRY;
 755
 756         if (unlikely(instant_cancel)) {
 757                 rc = ptl_send_rpc(req, 1);
 758                 ptlrpc_req_finished(req);
 759                 if (rc == 0)
 760                         cfs_atomic_inc(&arg->restart);
 761         } else {
 762                 LDLM_LOCK_GET(lock);
 763                 ptlrpc_set_add_req(arg->set, req);
 764         }
 765
 766         RETURN(rc);
 767 }
 768
 769 /**
 770  * Check if there are requests in the export request list which prevent
 771  * the lock canceling and make these requests high priority ones.
 772  */
 773 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
 774 {
 775         struct ptlrpc_request *req;
 776         ENTRY;
 777
 778         if (lock->l_export == NULL) {
 779                 LDLM_DEBUG(lock, "client lock: no-op");
 780                 RETURN_EXIT;
 781         }
 782
 783         spin_lock_bh(&lock->l_export->exp_rpc_lock);
 784         cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
 785                                 rq_exp_list) {
 786                 /* Do not process requests that were not yet added to there
 787                  * incoming queue or were already removed from there for
 788                  * processing */
 789                 if (!req->rq_hp && !cfs_list_empty(&req->rq_list) &&
 790                     req->rq_ops->hpreq_lock_match &&
 791                     req->rq_ops->hpreq_lock_match(req, lock))
 792                         ptlrpc_hpreq_reorder(req);
 793         }
 794         spin_unlock_bh(&lock->l_export->exp_rpc_lock);
 795         EXIT;
 796 }
 797
 798 /*
 799  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
 800  * enqueued server lock conflicts with given one.
 801  *
 802  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
 803  * to wait for client response.
 804  */
 805 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 806                              struct ldlm_lock_desc *desc,
 807                              void *data, int flag)
 808 {
 809         struct ldlm_cb_async_args *ca;
 810         struct ldlm_cb_set_arg *arg = data;
 811         struct ldlm_request    *body;
 812         struct ptlrpc_request  *req;
 813         int                     instant_cancel = 0;
 814         int                     rc = 0;
 815         ENTRY;
 816
 817         if (flag == LDLM_CB_CANCELING)
 818                 /* Don't need to do anything here. */
 819                 RETURN(0);
 820
 821         LASSERT(lock);
 822         LASSERT(data != NULL);
 823         if (lock->l_export->exp_obd->obd_recovering != 0)
 824                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
 825
 826         ldlm_lock_reorder_req(lock);
 827
 828         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 829                                         &RQF_LDLM_BL_CALLBACK,
 830                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
 831         if (req == NULL)
 832                 RETURN(-ENOMEM);
 833
 834         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 835         ca = ptlrpc_req_async_args(req);
 836         ca->ca_set_arg = arg;
 837         ca->ca_lock = lock;
 838
 839         req->rq_interpret_reply = ldlm_cb_interpret;
 840         req->rq_no_resend = 1;
 841
 842         lock_res_and_lock(lock);
 843         if (lock->l_granted_mode != lock->l_req_mode) {
 844                 /* this blocking AST will be communicated as part of the
 845                  * completion AST instead */
 846                 unlock_res_and_lock(lock);
 847
 848                 ptlrpc_req_finished(req);
 849                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 850                 RETURN(0);
 851         }
 852
 853         if (lock->l_destroyed) {
 854                 /* What's the point? */
 855                 unlock_res_and_lock(lock);
 856                 ptlrpc_req_finished(req);
 857                 RETURN(0);
 858         }
 859
 860         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
 861                 instant_cancel = 1;
 862
 863         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 864         body->lock_handle[0] = lock->l_remote_handle;
 865         body->lock_desc = *desc;
 866         body->lock_flags |= ldlm_flags_to_wire(lock->l_flags & LDLM_AST_FLAGS);
 867
 868         LDLM_DEBUG(lock, "server preparing blocking AST");
 869
 870         ptlrpc_request_set_replen(req);
 871         if (instant_cancel) {
 872                 unlock_res_and_lock(lock);
 873                 ldlm_lock_cancel(lock);
 874         } else {
 875                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
 876                 ldlm_add_waiting_lock(lock);
 877                 unlock_res_and_lock(lock);
 878         }
 879
 880         req->rq_send_state = LUSTRE_IMP_FULL;
 881         /* ptlrpc_request_alloc_pack already set timeout */
 882         if (AT_OFF)
 883                 req->rq_timeout = ldlm_get_rq_timeout();
 884
 885         if (lock->l_export && lock->l_export->exp_nid_stats &&
 886             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 887                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 888                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 889
 890         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
 891
 892         RETURN(rc);
 893 }
 894 EXPORT_SYMBOL(ldlm_server_blocking_ast);
 895
 896 int ldlm_server_completion_ast(struct ldlm_lock *lock, __u64 flags, void *data)
 897 {
 898         struct ldlm_cb_set_arg *arg = data;
 899         struct ldlm_request    *body;
 900         struct ptlrpc_request  *req;
 901         struct ldlm_cb_async_args *ca;
 902         long                    total_enqueue_wait;
 903         int                     instant_cancel = 0;
 904         int                     rc = 0;
 905         int                     lvb_len;
 906         ENTRY;
 907
 908         LASSERT(lock != NULL);
 909         LASSERT(data != NULL);
 910
 911         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
 912                                           lock->l_last_activity);
 913
 914         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
 915                                     &RQF_LDLM_CP_CALLBACK);
 916         if (req == NULL)
 917                 RETURN(-ENOMEM);
 918
 919         /* server namespace, doesn't need lock */
 920         lvb_len = ldlm_lvbo_size(lock);
 921         if (lvb_len > 0)
 922                  req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
 923                                       lvb_len);
 924
 925         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
 926         if (rc) {
 927                 ptlrpc_request_free(req);
 928                 RETURN(rc);
 929         }
 930
 931         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 932         ca = ptlrpc_req_async_args(req);
 933         ca->ca_set_arg = arg;
 934         ca->ca_lock = lock;
 935
 936         req->rq_interpret_reply = ldlm_cb_interpret;
 937         req->rq_no_resend = 1;
 938         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 939
 940         body->lock_handle[0] = lock->l_remote_handle;
 941         body->lock_flags = ldlm_flags_to_wire(flags);
 942         ldlm_lock2desc(lock, &body->lock_desc);
 943         if (lvb_len > 0) {
 944                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
 945
 946                 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
 947                 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
 948                                    lvb_len, RCL_CLIENT);
 949         }
 950
 951         LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
 952                    total_enqueue_wait);
 953
 954         /* Server-side enqueue wait time estimate, used in
 955             __ldlm_add_waiting_lock to set future enqueue timers */
 956         if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
 957                 at_measured(ldlm_lock_to_ns_at(lock),
 958                             total_enqueue_wait);
 959         else
 960                 /* bz18618. Don't add lock enqueue time we spend waiting for a
 961                    previous callback to fail. Locks waiting legitimately will
 962                    get extended by ldlm_refresh_waiting_lock regardless of the
 963                    estimate, so it's okay to underestimate here. */
 964                 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
 965                        "It is likely that a previous callback timed out.",
 966                        total_enqueue_wait,
 967                        at_get(ldlm_lock_to_ns_at(lock)));
 968
 969         ptlrpc_request_set_replen(req);
 970
 971         req->rq_send_state = LUSTRE_IMP_FULL;
 972         /* ptlrpc_request_pack already set timeout */
 973         if (AT_OFF)
 974                 req->rq_timeout = ldlm_get_rq_timeout();
 975
 976         /* We only send real blocking ASTs after the lock is granted */
 977         lock_res_and_lock(lock);
 978         if (lock->l_flags & LDLM_FL_AST_SENT) {
 979                 body->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
 980                 /* copy ast flags like LDLM_FL_DISCARD_DATA */
 981                 body->lock_flags |= ldlm_flags_to_wire(lock->l_flags &
 982                                                        LDLM_AST_FLAGS);
 983
 984                 /* We might get here prior to ldlm_handle_enqueue setting
 985                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
 986                  * into waiting list, but this is safe and similar code in
 987                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
 988                  * that would not only cancel the lock, but will also remove
 989                  * it from waiting list */
 990                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 991                         unlock_res_and_lock(lock);
 992                         ldlm_lock_cancel(lock);
 993                         instant_cancel = 1;
 994                         lock_res_and_lock(lock);
 995                 } else {
 996                         /* start the lock-timeout clock */
 997                         ldlm_add_waiting_lock(lock);
 998                 }
 999         }
1000         unlock_res_and_lock(lock);
1001
1002         if (lock->l_export && lock->l_export->exp_nid_stats &&
1003             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1004                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1005                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1006
1007         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1008
1009         RETURN(rc);
1010 }
1011 EXPORT_SYMBOL(ldlm_server_completion_ast);
1012
1013 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1014 {
1015         struct ldlm_cb_set_arg          *arg = data;
1016         struct ldlm_request             *body;
1017         struct ptlrpc_request           *req;
1018         struct ldlm_cb_async_args       *ca;
1019         int                              rc;
1020         struct req_format               *req_fmt;
1021         ENTRY;
1022
1023         LASSERT(lock != NULL);
1024
1025         if (arg->gl_desc != NULL)
1026                 /* There is a glimpse descriptor to pack */
1027                 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1028         else
1029                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1030
1031         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1032                                         req_fmt, LUSTRE_DLM_VERSION,
1033                                         LDLM_GL_CALLBACK);
1034
1035         if (req == NULL)
1036                 RETURN(-ENOMEM);
1037
1038         if (arg->gl_desc != NULL) {
1039                 /* copy the GL descriptor */
1040                 union ldlm_gl_desc      *desc;
1041                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1042                 *desc = *arg->gl_desc;
1043         }
1044
1045         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1046         body->lock_handle[0] = lock->l_remote_handle;
1047         ldlm_lock2desc(lock, &body->lock_desc);
1048
1049         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1050         ca = ptlrpc_req_async_args(req);
1051         ca->ca_set_arg = arg;
1052         ca->ca_lock = lock;
1053
1054         /* server namespace, doesn't need lock */
1055         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1056                              ldlm_lvbo_size(lock));
1057         ptlrpc_request_set_replen(req);
1058
1059         req->rq_send_state = LUSTRE_IMP_FULL;
1060         /* ptlrpc_request_alloc_pack already set timeout */
1061         if (AT_OFF)
1062                 req->rq_timeout = ldlm_get_rq_timeout();
1063
1064         req->rq_interpret_reply = ldlm_cb_interpret;
1065
1066         if (lock->l_export && lock->l_export->exp_nid_stats &&
1067             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1068                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1069                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1070
1071         rc = ldlm_ast_fini(req, arg, lock, 0);
1072
1073         RETURN(rc);
1074 }
1075 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1076
1077 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1078 {
1079         int     rc;
1080         ENTRY;
1081
1082         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1083                                LDLM_WORK_GL_AST);
1084         if (rc == -ERESTART)
1085                 ldlm_reprocess_all(res);
1086
1087         RETURN(rc);
1088 }
1089 EXPORT_SYMBOL(ldlm_glimpse_locks);
1090
1091 /* return ldlm lock associated with a lock callback request */
1092 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1093 {
1094         struct ldlm_cb_async_args       *ca;
1095         struct ldlm_lock                *lock;
1096         ENTRY;
1097
1098         ca = ptlrpc_req_async_args(req);
1099         lock = ca->ca_lock;
1100         if (lock == NULL)
1101                 RETURN(ERR_PTR(-EFAULT));
1102
1103         RETURN(lock);
1104 }
1105 EXPORT_SYMBOL(ldlm_request_lock);
1106
1107 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1108                        struct lprocfs_stats *srv_stats)
1109 {
1110         int lock_type = 0, op = 0;
1111
1112         lock_type = dlm_req->lock_desc.l_resource.lr_type;
1113
1114         switch (lock_type) {
1115         case LDLM_PLAIN:
1116                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1117                 break;
1118         case LDLM_EXTENT:
1119                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1120                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1121                 else
1122                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1123                 break;
1124         case LDLM_FLOCK:
1125                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1126                 break;
1127         case LDLM_IBITS:
1128                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1129                 break;
1130         default:
1131                 op = 0;
1132                 break;
1133         }
1134
1135         if (op)
1136                 lprocfs_counter_incr(srv_stats, op);
1137
1138         return;
1139 }
1140
1141 /*
1142  * Main server-side entry point into LDLM. This is called by ptlrpc service
1143  * threads to carry out client lock enqueueing requests.
1144  */
1145 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1146                          struct ptlrpc_request *req,
1147                          const struct ldlm_request *dlm_req,
1148                          const struct ldlm_callback_suite *cbs)
1149 {
1150         struct ldlm_reply *dlm_rep;
1151         __u64 flags;
1152         ldlm_error_t err = ELDLM_OK;
1153         struct ldlm_lock *lock = NULL;
1154         void *cookie = NULL;
1155         int rc = 0;
1156         ENTRY;
1157
1158         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1159
1160         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1161         flags = ldlm_flags_from_wire(dlm_req->lock_flags);
1162
1163         LASSERT(req->rq_export);
1164
1165         if (ptlrpc_req2svc(req)->srv_stats != NULL)
1166                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1167
1168         if (req->rq_export && req->rq_export->exp_nid_stats &&
1169             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1170                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1171                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1172
1173         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1174                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1175                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1176                           dlm_req->lock_desc.l_resource.lr_type);
1177                 GOTO(out, rc = -EFAULT);
1178         }
1179
1180         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1181                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1182                      dlm_req->lock_desc.l_req_mode &
1183                      (dlm_req->lock_desc.l_req_mode-1))) {
1184                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1185                           dlm_req->lock_desc.l_req_mode);
1186                 GOTO(out, rc = -EFAULT);
1187         }
1188
1189         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1190                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1191                              LDLM_PLAIN)) {
1192                         DEBUG_REQ(D_ERROR, req,
1193                                   "PLAIN lock request from IBITS client?");
1194                         GOTO(out, rc = -EPROTO);
1195                 }
1196         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1197                             LDLM_IBITS)) {
1198                 DEBUG_REQ(D_ERROR, req,
1199                           "IBITS lock request from unaware client?");
1200                 GOTO(out, rc = -EPROTO);
1201         }
1202
1203 #if 0
1204         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1205            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1206            ibits for mgc/mgs) */
1207
1208         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1209          * inodebits lock if client does not support them. */
1210         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1211             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1212                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1213                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1214                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1215                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1216                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1217         }
1218 #endif
1219
1220         if (unlikely(flags & LDLM_FL_REPLAY)) {
1221                 /* Find an existing lock in the per-export lock hash */
1222                 /* In the function below, .hs_keycmp resolves to
1223                  * ldlm_export_lock_keycmp() */
1224                 /* coverity[overrun-buffer-val] */
1225                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1226                                        (void *)&dlm_req->lock_handle[0]);
1227                 if (lock != NULL) {
1228                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1229                                   LPX64, lock->l_handle.h_cookie);
1230                         GOTO(existing_lock, rc = 0);
1231                 }
1232         }
1233
1234         /* The lock's callback data might be set in the policy function */
1235         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1236                                 dlm_req->lock_desc.l_resource.lr_type,
1237                                 dlm_req->lock_desc.l_req_mode,
1238                                 cbs, NULL, 0);
1239
1240         if (!lock)
1241                 GOTO(out, rc = -ENOMEM);
1242
1243         lock->l_last_activity = cfs_time_current_sec();
1244         lock->l_remote_handle = dlm_req->lock_handle[0];
1245         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1246
1247         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1248         /* Don't enqueue a lock onto the export if it is been disonnected
1249          * due to eviction (bug 3822) or server umount (bug 24324).
1250          * Cancel it now instead. */
1251         if (req->rq_export->exp_disconnected) {
1252                 LDLM_ERROR(lock, "lock on disconnected export %p",
1253                            req->rq_export);
1254                 GOTO(out, rc = -ENOTCONN);
1255         }
1256
1257         lock->l_export = class_export_lock_get(req->rq_export, lock);
1258         if (lock->l_export->exp_lock_hash)
1259                 cfs_hash_add(lock->l_export->exp_lock_hash,
1260                              &lock->l_remote_handle,
1261                              &lock->l_exp_hash);
1262
1263 existing_lock:
1264
1265         if (flags & LDLM_FL_HAS_INTENT) {
1266                 /* In this case, the reply buffer is allocated deep in
1267                  * local_lock_enqueue by the policy function. */
1268                 cookie = req;
1269         } else {
1270                 /* based on the assumption that lvb size never changes during
1271                  * resource life time otherwise it need resource->lr_lock's
1272                  * protection */
1273                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1274                                      RCL_SERVER, ldlm_lvbo_size(lock));
1275
1276                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1277                         GOTO(out, rc = -ENOMEM);
1278
1279                 rc = req_capsule_server_pack(&req->rq_pill);
1280                 if (rc)
1281                         GOTO(out, rc);
1282         }
1283
1284         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1285                 ldlm_convert_policy_to_local(req->rq_export,
1286                                           dlm_req->lock_desc.l_resource.lr_type,
1287                                           &dlm_req->lock_desc.l_policy_data,
1288                                           &lock->l_policy_data);
1289         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1290                 lock->l_req_extent = lock->l_policy_data.l_extent;
1291
1292         err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
1293         if (err)
1294                 GOTO(out, err);
1295
1296         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1297         dlm_rep->lock_flags = ldlm_flags_to_wire(flags);
1298
1299         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1300         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1301
1302         /* We never send a blocking AST until the lock is granted, but
1303          * we can tell it right now */
1304         lock_res_and_lock(lock);
1305
1306         /* Now take into account flags to be inherited from original lock
1307            request both in reply to client and in our own lock flags. */
1308         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1309         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
1310                                               LDLM_INHERIT_FLAGS);
1311
1312         /* Don't move a pending lock onto the export if it has already been
1313          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1314          * Cancel it now instead. */
1315         if (unlikely(req->rq_export->exp_disconnected ||
1316                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1317                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1318                 rc = -ENOTCONN;
1319         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1320                 dlm_rep->lock_flags |= ldlm_flags_to_wire(LDLM_FL_AST_SENT);
1321                 if (lock->l_granted_mode == lock->l_req_mode) {
1322                         /*
1323                          * Only cancel lock if it was granted, because it would
1324                          * be destroyed immediately and would never be granted
1325                          * in the future, causing timeouts on client.  Not
1326                          * granted lock will be cancelled immediately after
1327                          * sending completion AST.
1328                          */
1329                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1330                                 unlock_res_and_lock(lock);
1331                                 ldlm_lock_cancel(lock);
1332                                 lock_res_and_lock(lock);
1333                         } else
1334                                 ldlm_add_waiting_lock(lock);
1335                 }
1336         }
1337         /* Make sure we never ever grant usual metadata locks to liblustre
1338            clients */
1339         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1340             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1341              req->rq_export->exp_libclient) {
1342                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1343                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1344                         CERROR("Granting sync lock to libclient. "
1345                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1346                                dlm_req->lock_flags, dlm_rep->lock_flags,
1347                                lock->l_flags);
1348                         LDLM_ERROR(lock, "sync lock");
1349                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1350                                 struct ldlm_intent *it;
1351
1352                                 it = req_capsule_client_get(&req->rq_pill,
1353                                                             &RMF_LDLM_INTENT);
1354                                 if (it != NULL) {
1355                                         CERROR("This is intent %s ("LPU64")\n",
1356                                                ldlm_it2str(it->opc), it->opc);
1357                                 }
1358                         }
1359                 }
1360         }
1361
1362         unlock_res_and_lock(lock);
1363
1364         EXIT;
1365  out:
1366         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1367         if (!req->rq_packed_final) {
1368                 err = lustre_pack_reply(req, 1, NULL, NULL);
1369                 if (rc == 0)
1370                         rc = err;
1371         }
1372
1373         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1374          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1375         if (lock) {
1376                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1377                            "(err=%d, rc=%d)", err, rc);
1378
1379                 if (rc == 0) {
1380                         int lvb_len = ldlm_lvbo_size(lock);
1381
1382                         if (lvb_len > 0) {
1383                                 void *buf;
1384                                 int buflen;
1385
1386                                 buf = req_capsule_server_get(&req->rq_pill,
1387                                                              &RMF_DLM_LVB);
1388                                 LASSERTF(buf != NULL, "req %p, lock %p\n",
1389                                          req, lock);
1390                                 buflen = req_capsule_get_size(&req->rq_pill,
1391                                                 &RMF_DLM_LVB, RCL_SERVER);
1392                                 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1393                                 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
1394                                                    buflen, RCL_SERVER);
1395                         }
1396                 } else {
1397                         lock_res_and_lock(lock);
1398                         ldlm_resource_unlink_lock(lock);
1399                         ldlm_lock_destroy_nolock(lock);
1400                         unlock_res_and_lock(lock);
1401                 }
1402
1403                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1404                         ldlm_reprocess_all(lock->l_resource);
1405
1406                 LDLM_LOCK_RELEASE(lock);
1407         }
1408
1409         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1410                           lock, rc);
1411
1412         return rc;
1413 }
1414 EXPORT_SYMBOL(ldlm_handle_enqueue0);
1415
1416 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1417                         ldlm_completion_callback completion_callback,
1418                         ldlm_blocking_callback blocking_callback,
1419                         ldlm_glimpse_callback glimpse_callback)
1420 {
1421         struct ldlm_request *dlm_req;
1422         struct ldlm_callback_suite cbs = {
1423                 .lcs_completion = completion_callback,
1424                 .lcs_blocking   = blocking_callback,
1425                 .lcs_glimpse    = glimpse_callback
1426         };
1427         int rc;
1428
1429         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1430         if (dlm_req != NULL) {
1431                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1432                                           req, dlm_req, &cbs);
1433         } else {
1434                 rc = -EFAULT;
1435         }
1436         return rc;
1437 }
1438 EXPORT_SYMBOL(ldlm_handle_enqueue);
1439
1440 int ldlm_handle_convert0(struct ptlrpc_request *req,
1441                          const struct ldlm_request *dlm_req)
1442 {
1443         struct ldlm_reply *dlm_rep;
1444         struct ldlm_lock *lock;
1445         int rc;
1446         ENTRY;
1447
1448         if (req->rq_export && req->rq_export->exp_nid_stats &&
1449             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1450                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1451                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1452
1453         rc = req_capsule_server_pack(&req->rq_pill);
1454         if (rc)
1455                 RETURN(rc);
1456
1457         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1458         dlm_rep->lock_flags = dlm_req->lock_flags;
1459
1460         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1461         if (!lock) {
1462                 req->rq_status = EINVAL;
1463         } else {
1464                 void *res = NULL;
1465
1466                 LDLM_DEBUG(lock, "server-side convert handler START");
1467
1468                 lock->l_last_activity = cfs_time_current_sec();
1469                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1470                                         &dlm_rep->lock_flags);
1471                 if (res) {
1472                         if (ldlm_del_waiting_lock(lock))
1473                                 LDLM_DEBUG(lock, "converted waiting lock");
1474                         req->rq_status = 0;
1475                 } else {
1476                         req->rq_status = EDEADLOCK;
1477                 }
1478         }
1479
1480         if (lock) {
1481                 if (!req->rq_status)
1482                         ldlm_reprocess_all(lock->l_resource);
1483                 LDLM_DEBUG(lock, "server-side convert handler END");
1484                 LDLM_LOCK_PUT(lock);
1485         } else
1486                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1487
1488         RETURN(0);
1489 }
1490 EXPORT_SYMBOL(ldlm_handle_convert0);
1491
1492 int ldlm_handle_convert(struct ptlrpc_request *req)
1493 {
1494         int rc;
1495         struct ldlm_request *dlm_req;
1496
1497         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1498         if (dlm_req != NULL) {
1499                 rc = ldlm_handle_convert0(req, dlm_req);
1500         } else {
1501                 CERROR ("Can't unpack dlm_req\n");
1502                 rc = -EFAULT;
1503         }
1504         return rc;
1505 }
1506 EXPORT_SYMBOL(ldlm_handle_convert);
1507
1508 /* Cancel all the locks whos handles are packed into ldlm_request */
1509 int ldlm_request_cancel(struct ptlrpc_request *req,
1510                         const struct ldlm_request *dlm_req, int first)
1511 {
1512         struct ldlm_resource *res, *pres = NULL;
1513         struct ldlm_lock *lock;
1514         int i, count, done = 0;
1515         ENTRY;
1516
1517         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1518         if (first >= count)
1519                 RETURN(0);
1520
1521         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1522                 RETURN(0);
1523
1524         /* There is no lock on the server at the replay time,
1525          * skip lock cancelling to make replay tests to pass. */
1526         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1527                 RETURN(0);
1528
1529         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1530                           "starting at %d", count, first);
1531
1532         for (i = first; i < count; i++) {
1533                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1534                 if (!lock) {
1535                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1536                                           "lock (cookie "LPU64")",
1537                                           dlm_req->lock_handle[i].cookie);
1538                         continue;
1539                 }
1540
1541                 res = lock->l_resource;
1542                 done++;
1543
1544                 if (res != pres) {
1545                         if (pres != NULL) {
1546                                 ldlm_reprocess_all(pres);
1547                                 LDLM_RESOURCE_DELREF(pres);
1548                                 ldlm_resource_putref(pres);
1549                         }
1550                         if (res != NULL) {
1551                                 ldlm_resource_getref(res);
1552                                 LDLM_RESOURCE_ADDREF(res);
1553                                 ldlm_res_lvbo_update(res, NULL, 1);
1554                         }
1555                         pres = res;
1556                 }
1557                 ldlm_lock_cancel(lock);
1558                 LDLM_LOCK_PUT(lock);
1559         }
1560         if (pres != NULL) {
1561                 ldlm_reprocess_all(pres);
1562                 LDLM_RESOURCE_DELREF(pres);
1563                 ldlm_resource_putref(pres);
1564         }
1565         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1566         RETURN(done);
1567 }
1568 EXPORT_SYMBOL(ldlm_request_cancel);
1569
1570 int ldlm_handle_cancel(struct ptlrpc_request *req)
1571 {
1572         struct ldlm_request *dlm_req;
1573         int rc;
1574         ENTRY;
1575
1576         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1577         if (dlm_req == NULL) {
1578                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1579                 RETURN(-EFAULT);
1580         }
1581
1582         if (req->rq_export && req->rq_export->exp_nid_stats &&
1583             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1584                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1585                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1586
1587         rc = req_capsule_server_pack(&req->rq_pill);
1588         if (rc)
1589                 RETURN(rc);
1590
1591         if (!ldlm_request_cancel(req, dlm_req, 0))
1592                 req->rq_status = ESTALE;
1593
1594         RETURN(ptlrpc_reply(req));
1595 }
1596 EXPORT_SYMBOL(ldlm_handle_cancel);
1597 #endif /* HAVE_SERVER_SUPPORT */
1598
1599 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1600                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1601 {
1602         int do_ast;
1603         ENTRY;
1604
1605         LDLM_DEBUG(lock, "client blocking AST callback handler");
1606
1607         lock_res_and_lock(lock);
1608         lock->l_flags |= LDLM_FL_CBPENDING;
1609
1610         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1611                 lock->l_flags |= LDLM_FL_CANCEL;
1612
1613         do_ast = (!lock->l_readers && !lock->l_writers);
1614         unlock_res_and_lock(lock);
1615
1616         if (do_ast) {
1617                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1618                        lock, lock->l_blocking_ast);
1619                 if (lock->l_blocking_ast != NULL)
1620                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1621                                              LDLM_CB_BLOCKING);
1622         } else {
1623                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1624                        lock);
1625         }
1626
1627         LDLM_DEBUG(lock, "client blocking callback handler END");
1628         LDLM_LOCK_RELEASE(lock);
1629         EXIT;
1630 }
1631
1632 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1633                                     struct ldlm_namespace *ns,
1634                                     struct ldlm_request *dlm_req,
1635                                     struct ldlm_lock *lock)
1636 {
1637         int lvb_len;
1638         CFS_LIST_HEAD(ast_list);
1639         ENTRY;
1640
1641         LDLM_DEBUG(lock, "client completion callback handler START");
1642
1643         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1644                 int to = cfs_time_seconds(1);
1645                 while (to > 0) {
1646                         cfs_schedule_timeout_and_set_state(
1647                                 CFS_TASK_INTERRUPTIBLE, to);
1648                         if (lock->l_granted_mode == lock->l_req_mode ||
1649                             lock->l_destroyed)
1650                                 break;
1651                 }
1652         }
1653
1654         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1655         if (lvb_len > 0) {
1656                 if (lock->l_lvb_len > 0) {
1657                         /* for extent lock, lvb contains ost_lvb{}. */
1658                         LASSERT(lock->l_lvb_data != NULL);
1659                         LASSERTF(lock->l_lvb_len == lvb_len,
1660                                 "preallocated %d, actual %d.\n",
1661                                 lock->l_lvb_len, lvb_len);
1662                 } else { /* for layout lock, lvb has variable length */
1663                         void *lvb_data;
1664
1665                         OBD_ALLOC(lvb_data, lvb_len);
1666                         if (lvb_data == NULL)
1667                                 LDLM_ERROR(lock, "no memory.\n");
1668
1669                         lock_res_and_lock(lock);
1670                         if (lvb_data == NULL) {
1671                                 lock->l_flags |= LDLM_FL_FAILED;
1672                         } else {
1673                                 LASSERT(lock->l_lvb_data == NULL);
1674                                 lock->l_lvb_data = lvb_data;
1675                                 lock->l_lvb_len = lvb_len;
1676                         }
1677                         unlock_res_and_lock(lock);
1678                 }
1679         }
1680
1681         lock_res_and_lock(lock);
1682         if (lock->l_destroyed ||
1683             lock->l_granted_mode == lock->l_req_mode) {
1684                 /* bug 11300: the lock has already been granted */
1685                 unlock_res_and_lock(lock);
1686                 LDLM_DEBUG(lock, "Double grant race happened");
1687                 LDLM_LOCK_RELEASE(lock);
1688                 EXIT;
1689                 return;
1690         }
1691
1692         /* If we receive the completion AST before the actual enqueue returned,
1693          * then we might need to switch lock modes, resources, or extents. */
1694         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1695                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1696                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1697         }
1698
1699         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1700                 ldlm_convert_policy_to_local(req->rq_export,
1701                                           dlm_req->lock_desc.l_resource.lr_type,
1702                                           &dlm_req->lock_desc.l_policy_data,
1703                                           &lock->l_policy_data);
1704                 LDLM_DEBUG(lock, "completion AST, new policy data");
1705         }
1706
1707         ldlm_resource_unlink_lock(lock);
1708         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1709                    &lock->l_resource->lr_name,
1710                    sizeof(lock->l_resource->lr_name)) != 0) {
1711                 unlock_res_and_lock(lock);
1712                 if (ldlm_lock_change_resource(ns, lock,
1713                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1714                         LDLM_ERROR(lock, "Failed to allocate resource");
1715                         LDLM_LOCK_RELEASE(lock);
1716                         EXIT;
1717                         return;
1718                 }
1719                 LDLM_DEBUG(lock, "completion AST, new resource");
1720                 CERROR("change resource!\n");
1721                 lock_res_and_lock(lock);
1722         }
1723
1724         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1725                 /* BL_AST locks are not needed in lru.
1726                  * let ldlm_cancel_lru() be fast. */
1727                 ldlm_lock_remove_from_lru(lock);
1728                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1729                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1730         }
1731
1732         if (lock->l_lvb_len) {
1733                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1734                                          RCL_CLIENT) < lock->l_lvb_len) {
1735                         LDLM_ERROR(lock, "completion AST did not contain "
1736                                    "expected LVB!");
1737                 } else {
1738                         void *lvb = req_capsule_client_get(&req->rq_pill,
1739                                                            &RMF_DLM_LVB);
1740                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1741                 }
1742         }
1743
1744         ldlm_grant_lock(lock, &ast_list);
1745         unlock_res_and_lock(lock);
1746
1747         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1748
1749         /* Let Enqueue to call osc_lock_upcall() and initialize
1750          * l_ast_data */
1751         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1752
1753         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1754
1755         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1756                           lock);
1757         LDLM_LOCK_RELEASE(lock);
1758         EXIT;
1759 }
1760
1761 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1762                                     struct ldlm_namespace *ns,
1763                                     struct ldlm_request *dlm_req,
1764                                     struct ldlm_lock *lock)
1765 {
1766         int rc = -ENOSYS;
1767         ENTRY;
1768
1769         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1770
1771         if (lock->l_glimpse_ast != NULL)
1772                 rc = lock->l_glimpse_ast(lock, req);
1773
1774         if (req->rq_repmsg != NULL) {
1775                 ptlrpc_reply(req);
1776         } else {
1777                 req->rq_status = rc;
1778                 ptlrpc_error(req);
1779         }
1780
1781         lock_res_and_lock(lock);
1782         if (lock->l_granted_mode == LCK_PW &&
1783             !lock->l_readers && !lock->l_writers &&
1784             cfs_time_after(cfs_time_current(),
1785                            cfs_time_add(lock->l_last_used,
1786                                         cfs_time_seconds(10)))) {
1787                 unlock_res_and_lock(lock);
1788                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1789                         ldlm_handle_bl_callback(ns, NULL, lock);
1790
1791                 EXIT;
1792                 return;
1793         }
1794         unlock_res_and_lock(lock);
1795         LDLM_LOCK_RELEASE(lock);
1796         EXIT;
1797 }
1798
1799 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1800 {
1801         if (req->rq_no_reply)
1802                 return 0;
1803
1804         req->rq_status = rc;
1805         if (!req->rq_packed_final) {
1806                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1807                 if (rc)
1808                         return rc;
1809         }
1810         return ptlrpc_reply(req);
1811 }
1812
1813 #ifdef __KERNEL__
1814 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1815 {
1816         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1817         ENTRY;
1818
1819         spin_lock(&blp->blp_lock);
1820         if (blwi->blwi_lock &&
1821             blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1822                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1823                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1824         } else {
1825                 /* other blocking callbacks are added to the regular list */
1826                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1827         }
1828         spin_unlock(&blp->blp_lock);
1829
1830         cfs_waitq_signal(&blp->blp_waitq);
1831
1832         /* can not use blwi->blwi_mode as blwi could be already freed in
1833            LDLM_ASYNC mode */
1834         if (mode == LDLM_SYNC)
1835                 wait_for_completion(&blwi->blwi_comp);
1836
1837         RETURN(0);
1838 }
1839
1840 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1841                              struct ldlm_namespace *ns,
1842                              struct ldlm_lock_desc *ld,
1843                              cfs_list_t *cancels, int count,
1844                              struct ldlm_lock *lock,
1845                              int mode)
1846 {
1847         init_completion(&blwi->blwi_comp);
1848         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1849
1850         if (cfs_memory_pressure_get())
1851                 blwi->blwi_mem_pressure = 1;
1852
1853         blwi->blwi_ns = ns;
1854         blwi->blwi_mode = mode;
1855         if (ld != NULL)
1856                 blwi->blwi_ld = *ld;
1857         if (count) {
1858                 cfs_list_add(&blwi->blwi_head, cancels);
1859                 cfs_list_del_init(cancels);
1860                 blwi->blwi_count = count;
1861         } else {
1862                 blwi->blwi_lock = lock;
1863         }
1864 }
1865
1866 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1867                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1868                              cfs_list_t *cancels, int count, int mode)
1869 {
1870         ENTRY;
1871
1872         if (cancels && count == 0)
1873                 RETURN(0);
1874
1875         if (mode == LDLM_SYNC) {
1876                 /* if it is synchronous call do minimum mem alloc, as it could
1877                  * be triggered from kernel shrinker
1878                  */
1879                 struct ldlm_bl_work_item blwi;
1880                 memset(&blwi, 0, sizeof(blwi));
1881                 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1882                 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1883         } else {
1884                 struct ldlm_bl_work_item *blwi;
1885                 OBD_ALLOC(blwi, sizeof(*blwi));
1886                 if (blwi == NULL)
1887                         RETURN(-ENOMEM);
1888                 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1889
1890                 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1891         }
1892 }
1893
1894 #endif
1895
1896 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1897                            struct ldlm_lock *lock)
1898 {
1899 #ifdef __KERNEL__
1900         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1901 #else
1902         RETURN(-ENOSYS);
1903 #endif
1904 }
1905
1906 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1907                            cfs_list_t *cancels, int count, int mode)
1908 {
1909 #ifdef __KERNEL__
1910         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1911 #else
1912         RETURN(-ENOSYS);
1913 #endif
1914 }
1915
1916 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1917 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1918 {
1919         struct obd_device *obd = req->rq_export->exp_obd;
1920         char *key;
1921         void *val;
1922         int keylen, vallen;
1923         int rc = -ENOSYS;
1924         ENTRY;
1925
1926         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1927
1928         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1929
1930         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1931         if (key == NULL) {
1932                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1933                 RETURN(-EFAULT);
1934         }
1935         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1936                                       RCL_CLIENT);
1937         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1938         if (val == NULL) {
1939                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1940                 RETURN(-EFAULT);
1941         }
1942         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1943                                       RCL_CLIENT);
1944
1945         /* We are responsible for swabbing contents of val */
1946
1947         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1948                 /* Pass it on to mdc (the "export" in this case) */
1949                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
1950                                         req->rq_export,
1951                                         sizeof(KEY_HSM_COPYTOOL_SEND),
1952                                         KEY_HSM_COPYTOOL_SEND,
1953                                         vallen, val, NULL);
1954         else
1955                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1956
1957         return rc;
1958 }
1959
1960 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1961                                         const char *msg, int rc,
1962                                         struct lustre_handle *handle)
1963 {
1964         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1965                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
1966                   msg, libcfs_id2str(req->rq_peer), rc,
1967                   handle ? handle->cookie : 0);
1968         if (req->rq_no_reply)
1969                 CWARN("No reply was sent, maybe cause bug 21636.\n");
1970         else if (rc)
1971                 CWARN("Send reply failed, maybe cause bug 21636.\n");
1972 }
1973
1974 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1975 static int ldlm_callback_handler(struct ptlrpc_request *req)
1976 {
1977         struct ldlm_namespace *ns;
1978         struct ldlm_request *dlm_req;
1979         struct ldlm_lock *lock;
1980         int rc;
1981         ENTRY;
1982
1983         /* Requests arrive in sender's byte order.  The ptlrpc service
1984          * handler has already checked and, if necessary, byte-swapped the
1985          * incoming request message body, but I am responsible for the
1986          * message buffers. */
1987
1988         /* do nothing for sec context finalize */
1989         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1990                 RETURN(0);
1991
1992         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1993
1994         if (req->rq_export == NULL) {
1995                 rc = ldlm_callback_reply(req, -ENOTCONN);
1996                 ldlm_callback_errmsg(req, "Operate on unconnected server",
1997                                      rc, NULL);
1998                 RETURN(0);
1999         }
2000
2001         LASSERT(req->rq_export != NULL);
2002         LASSERT(req->rq_export->exp_obd != NULL);
2003
2004         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2005         case LDLM_BL_CALLBACK:
2006                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
2007                         RETURN(0);
2008                 break;
2009         case LDLM_CP_CALLBACK:
2010                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
2011                         RETURN(0);
2012                 break;
2013         case LDLM_GL_CALLBACK:
2014                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2015                         RETURN(0);
2016                 break;
2017         case LDLM_SET_INFO:
2018                 rc = ldlm_handle_setinfo(req);
2019                 ldlm_callback_reply(req, rc);
2020                 RETURN(0);
2021         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
2022                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
2023                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2024                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2025                         RETURN(0);
2026                 rc = llog_origin_handle_cancel(req);
2027                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2028                         RETURN(0);
2029                 ldlm_callback_reply(req, rc);
2030                 RETURN(0);
2031         case LLOG_ORIGIN_HANDLE_CREATE:
2032                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2033                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2034                         RETURN(0);
2035                 rc = llog_origin_handle_open(req);
2036                 ldlm_callback_reply(req, rc);
2037                 RETURN(0);
2038         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2039                 req_capsule_set(&req->rq_pill,
2040                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2041                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2042                         RETURN(0);
2043                 rc = llog_origin_handle_next_block(req);
2044                 ldlm_callback_reply(req, rc);
2045                 RETURN(0);
2046         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2047                 req_capsule_set(&req->rq_pill,
2048                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2049                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2050                         RETURN(0);
2051                 rc = llog_origin_handle_read_header(req);
2052                 ldlm_callback_reply(req, rc);
2053                 RETURN(0);
2054         case LLOG_ORIGIN_HANDLE_CLOSE:
2055                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2056                         RETURN(0);
2057                 rc = llog_origin_handle_close(req);
2058                 ldlm_callback_reply(req, rc);
2059                 RETURN(0);
2060         default:
2061                 CERROR("unknown opcode %u\n",
2062                        lustre_msg_get_opc(req->rq_reqmsg));
2063                 ldlm_callback_reply(req, -EPROTO);
2064                 RETURN(0);
2065         }
2066
2067         ns = req->rq_export->exp_obd->obd_namespace;
2068         LASSERT(ns != NULL);
2069
2070         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2071
2072         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2073         if (dlm_req == NULL) {
2074                 rc = ldlm_callback_reply(req, -EPROTO);
2075                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2076                                      NULL);
2077                 RETURN(0);
2078         }
2079
2080         /* Force a known safe race, send a cancel to the server for a lock
2081          * which the server has already started a blocking callback on. */
2082         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2083             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2084                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
2085                 if (rc < 0)
2086                         CERROR("ldlm_cli_cancel: %d\n", rc);
2087         }
2088
2089         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2090         if (!lock) {
2091                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2092                        "disappeared\n", dlm_req->lock_handle[0].cookie);
2093                 rc = ldlm_callback_reply(req, -EINVAL);
2094                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2095                                      &dlm_req->lock_handle[0]);
2096                 RETURN(0);
2097         }
2098
2099         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2100             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2101                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2102
2103         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2104         lock_res_and_lock(lock);
2105         lock->l_flags |= ldlm_flags_from_wire(dlm_req->lock_flags &
2106                                               LDLM_AST_FLAGS);
2107         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2108                 /* If somebody cancels lock and cache is already dropped,
2109                  * or lock is failed before cp_ast received on client,
2110                  * we can tell the server we have no lock. Otherwise, we
2111                  * should send cancel after dropping the cache. */
2112                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2113                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
2114                     (lock->l_flags & LDLM_FL_FAILED)) {
2115                         LDLM_DEBUG(lock, "callback on lock "
2116                                    LPX64" - lock disappeared\n",
2117                                    dlm_req->lock_handle[0].cookie);
2118                         unlock_res_and_lock(lock);
2119                         LDLM_LOCK_RELEASE(lock);
2120                         rc = ldlm_callback_reply(req, -EINVAL);
2121                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2122                                              &dlm_req->lock_handle[0]);
2123                         RETURN(0);
2124                 }
2125                 /* BL_AST locks are not needed in lru.
2126                  * let ldlm_cancel_lru() be fast. */
2127                 ldlm_lock_remove_from_lru(lock);
2128                 lock->l_flags |= LDLM_FL_BL_AST;
2129         }
2130         unlock_res_and_lock(lock);
2131
2132         /* We want the ost thread to get this reply so that it can respond
2133          * to ost requests (write cache writeback) that might be triggered
2134          * in the callback.
2135          *
2136          * But we'd also like to be able to indicate in the reply that we're
2137          * cancelling right now, because it's unused, or have an intent result
2138          * in the reply, so we might have to push the responsibility for sending
2139          * the reply down into the AST handlers, alas. */
2140
2141         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2142         case LDLM_BL_CALLBACK:
2143                 CDEBUG(D_INODE, "blocking ast\n");
2144                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2145                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2146                         rc = ldlm_callback_reply(req, 0);
2147                         if (req->rq_no_reply || rc)
2148                                 ldlm_callback_errmsg(req, "Normal process", rc,
2149                                                      &dlm_req->lock_handle[0]);
2150                 }
2151                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2152                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2153                 break;
2154         case LDLM_CP_CALLBACK:
2155                 CDEBUG(D_INODE, "completion ast\n");
2156                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2157                 ldlm_callback_reply(req, 0);
2158                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2159                 break;
2160         case LDLM_GL_CALLBACK:
2161                 CDEBUG(D_INODE, "glimpse ast\n");
2162                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2163                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2164                 break;
2165         default:
2166                 LBUG();                         /* checked above */
2167         }
2168
2169         RETURN(0);
2170 }
2171
2172 #ifdef HAVE_SERVER_SUPPORT
2173 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2174 {
2175         int rc;
2176         ENTRY;
2177
2178         /* Requests arrive in sender's byte order.  The ptlrpc service
2179          * handler has already checked and, if necessary, byte-swapped the
2180          * incoming request message body, but I am responsible for the
2181          * message buffers. */
2182
2183         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2184
2185         if (req->rq_export == NULL) {
2186                 struct ldlm_request *dlm_req;
2187
2188                 CERROR("%s from %s arrived at %lu with bad export cookie "
2189                        LPU64"\n",
2190                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2191                        libcfs_nid2str(req->rq_peer.nid),
2192                        req->rq_arrival_time.tv_sec,
2193                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2194
2195                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2196                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2197                         dlm_req = req_capsule_client_get(&req->rq_pill,
2198                                                          &RMF_DLM_REQ);
2199                         if (dlm_req != NULL)
2200                                 ldlm_lock_dump_handle(D_ERROR,
2201                                                       &dlm_req->lock_handle[0]);
2202                 }
2203                 ldlm_callback_reply(req, -ENOTCONN);
2204                 RETURN(0);
2205         }
2206
2207         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2208
2209         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2210         case LDLM_CANCEL:
2211                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2212                 CDEBUG(D_INODE, "cancel\n");
2213                 if (CFS_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET) ||
2214                     CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_CANCEL_RESEND))
2215                         RETURN(0);
2216                 rc = ldlm_handle_cancel(req);
2217                 if (rc)
2218                         break;
2219                 RETURN(0);
2220         case OBD_LOG_CANCEL:
2221                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2222                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2223                         RETURN(0);
2224                 rc = llog_origin_handle_cancel(req);
2225                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2226                         RETURN(0);
2227                 ldlm_callback_reply(req, rc);
2228                 RETURN(0);
2229         default:
2230                 CERROR("invalid opcode %d\n",
2231                        lustre_msg_get_opc(req->rq_reqmsg));
2232                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2233                 ldlm_callback_reply(req, -EINVAL);
2234         }
2235
2236         RETURN(0);
2237 }
2238
2239 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2240                                         struct ldlm_lock *lock)
2241 {
2242         struct ldlm_request *dlm_req;
2243         struct lustre_handle lockh;
2244         int rc = 0;
2245         int i;
2246         ENTRY;
2247
2248         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2249         if (dlm_req == NULL)
2250                 RETURN(0);
2251
2252         ldlm_lock2handle(lock, &lockh);
2253         for (i = 0; i < dlm_req->lock_count; i++) {
2254                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2255                                         &lockh)) {
2256                         DEBUG_REQ(D_RPCTRACE, req,
2257                                   "Prio raised by lock "LPX64".", lockh.cookie);
2258
2259                         rc = 1;
2260                         break;
2261                 }
2262         }
2263
2264         RETURN(rc);
2265
2266 }
2267
2268 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2269 {
2270         struct ldlm_request *dlm_req;
2271         int rc = 0;
2272         int i;
2273         ENTRY;
2274
2275         /* no prolong in recovery */
2276         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2277                 RETURN(0);
2278
2279         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2280         if (dlm_req == NULL)
2281                 RETURN(-EFAULT);
2282
2283         for (i = 0; i < dlm_req->lock_count; i++) {
2284                 struct ldlm_lock *lock;
2285
2286                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2287                 if (lock == NULL)
2288                         continue;
2289
2290                 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2291                 if (rc)
2292                         LDLM_DEBUG(lock, "hpreq cancel lock");
2293                 LDLM_LOCK_PUT(lock);
2294
2295                 if (rc)
2296                         break;
2297         }
2298
2299         RETURN(rc);
2300 }
2301
2302 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2303         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2304         .hpreq_check      = ldlm_cancel_hpreq_check,
2305         .hpreq_fini       = NULL,
2306 };
2307
2308 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2309 {
2310         ENTRY;
2311
2312         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2313
2314         if (req->rq_export == NULL)
2315                 RETURN(0);
2316
2317         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2318                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2319                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2320         }
2321         RETURN(0);
2322 }
2323
2324 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2325                         cfs_hlist_node_t *hnode, void *data)
2326
2327 {
2328         cfs_list_t         *rpc_list = data;
2329         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2330
2331         lock_res_and_lock(lock);
2332
2333         if (lock->l_req_mode != lock->l_granted_mode) {
2334                 unlock_res_and_lock(lock);
2335                 return 0;
2336         }
2337
2338         LASSERT(lock->l_resource);
2339         if (lock->l_resource->lr_type != LDLM_IBITS &&
2340             lock->l_resource->lr_type != LDLM_PLAIN) {
2341                 unlock_res_and_lock(lock);
2342                 return 0;
2343         }
2344
2345         if (lock->l_flags & LDLM_FL_AST_SENT) {
2346                 unlock_res_and_lock(lock);
2347                 return 0;
2348         }
2349
2350         LASSERT(lock->l_blocking_ast);
2351         LASSERT(!lock->l_blocking_lock);
2352
2353         lock->l_flags |= LDLM_FL_AST_SENT;
2354         if (lock->l_export && lock->l_export->exp_lock_hash) {
2355                 /* NB: it's safe to call cfs_hash_del() even lock isn't
2356                  * in exp_lock_hash. */
2357                 /* In the function below, .hs_keycmp resolves to
2358                  * ldlm_export_lock_keycmp() */
2359                 /* coverity[overrun-buffer-val] */
2360                 cfs_hash_del(lock->l_export->exp_lock_hash,
2361                              &lock->l_remote_handle, &lock->l_exp_hash);
2362         }
2363
2364         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2365         LDLM_LOCK_GET(lock);
2366
2367         unlock_res_and_lock(lock);
2368         return 0;
2369 }
2370
2371 void ldlm_revoke_export_locks(struct obd_export *exp)
2372 {
2373         cfs_list_t  rpc_list;
2374         ENTRY;
2375
2376         CFS_INIT_LIST_HEAD(&rpc_list);
2377         cfs_hash_for_each_empty(exp->exp_lock_hash,
2378                                 ldlm_revoke_lock_cb, &rpc_list);
2379         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2380                           LDLM_WORK_REVOKE_AST);
2381
2382         EXIT;
2383 }
2384 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2385 #endif /* HAVE_SERVER_SUPPORT */
2386
2387 #ifdef __KERNEL__
2388 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2389 {
2390         struct ldlm_bl_work_item *blwi = NULL;
2391         static unsigned int num_bl = 0;
2392
2393         spin_lock(&blp->blp_lock);
2394         /* process a request from the blp_list at least every blp_num_threads */
2395         if (!cfs_list_empty(&blp->blp_list) &&
2396             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2397                 blwi = cfs_list_entry(blp->blp_list.next,
2398                                       struct ldlm_bl_work_item, blwi_entry);
2399         else
2400                 if (!cfs_list_empty(&blp->blp_prio_list))
2401                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2402                                               struct ldlm_bl_work_item,
2403                                               blwi_entry);
2404
2405         if (blwi) {
2406                 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2407                         num_bl = 0;
2408                 cfs_list_del(&blwi->blwi_entry);
2409         }
2410         spin_unlock(&blp->blp_lock);
2411
2412         return blwi;
2413 }
2414
2415 /* This only contains temporary data until the thread starts */
2416 struct ldlm_bl_thread_data {
2417         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2418         struct ldlm_bl_pool     *bltd_blp;
2419         struct completion       bltd_comp;
2420         int                     bltd_num;
2421 };
2422
2423 static int ldlm_bl_thread_main(void *arg);
2424
2425 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2426 {
2427         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2428         int rc;
2429
2430         init_completion(&bltd.bltd_comp);
2431         rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2432         if (rc < 0) {
2433                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2434                        cfs_atomic_read(&blp->blp_num_threads), rc);
2435                 return rc;
2436         }
2437         wait_for_completion(&bltd.bltd_comp);
2438
2439         return 0;
2440 }
2441
2442 static int ldlm_bl_thread_main(void *arg)
2443 {
2444         struct ldlm_bl_pool *blp;
2445         ENTRY;
2446
2447         {
2448                 struct ldlm_bl_thread_data *bltd = arg;
2449
2450                 blp = bltd->bltd_blp;
2451
2452                 bltd->bltd_num =
2453                         cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2454                 cfs_atomic_inc(&blp->blp_busy_threads);
2455
2456                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2457                         "ldlm_bl_%02d", bltd->bltd_num);
2458                 cfs_daemonize(bltd->bltd_name);
2459
2460                 complete(&bltd->bltd_comp);
2461                 /* cannot use bltd after this, it is only on caller's stack */
2462         }
2463
2464         while (1) {
2465                 struct l_wait_info lwi = { 0 };
2466                 struct ldlm_bl_work_item *blwi = NULL;
2467                 int busy;
2468
2469                 blwi = ldlm_bl_get_work(blp);
2470
2471                 if (blwi == NULL) {
2472                         cfs_atomic_dec(&blp->blp_busy_threads);
2473                         l_wait_event_exclusive(blp->blp_waitq,
2474                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2475                                          &lwi);
2476                         busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2477                 } else {
2478                         busy = cfs_atomic_read(&blp->blp_busy_threads);
2479                 }
2480
2481                 if (blwi->blwi_ns == NULL)
2482                         /* added by ldlm_cleanup() */
2483                         break;
2484
2485                 /* Not fatal if racy and have a few too many threads */
2486                 if (unlikely(busy < blp->blp_max_threads &&
2487                              busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2488                              !blwi->blwi_mem_pressure))
2489                         /* discard the return value, we tried */
2490                         ldlm_bl_thread_start(blp);
2491
2492                 if (blwi->blwi_mem_pressure)
2493                         cfs_memory_pressure_set();
2494
2495                 if (blwi->blwi_count) {
2496                         int count;
2497                         /* The special case when we cancel locks in lru
2498                          * asynchronously, we pass the list of locks here.
2499                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2500                          * canceled locally yet. */
2501                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2502                                                            blwi->blwi_count,
2503                                                            LCF_BL_AST);
2504                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2505                 } else {
2506                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2507                                                 blwi->blwi_lock);
2508                 }
2509                 if (blwi->blwi_mem_pressure)
2510                         cfs_memory_pressure_clr();
2511
2512                 if (blwi->blwi_mode == LDLM_ASYNC)
2513                         OBD_FREE(blwi, sizeof(*blwi));
2514                 else
2515                         complete(&blwi->blwi_comp);
2516         }
2517
2518         cfs_atomic_dec(&blp->blp_busy_threads);
2519         cfs_atomic_dec(&blp->blp_num_threads);
2520         complete(&blp->blp_comp);
2521         RETURN(0);
2522 }
2523
2524 #endif
2525
2526 static int ldlm_setup(void);
2527 static int ldlm_cleanup(void);
2528
2529 int ldlm_get_ref(void)
2530 {
2531         int rc = 0;
2532         ENTRY;
2533         mutex_lock(&ldlm_ref_mutex);
2534         if (++ldlm_refcount == 1) {
2535                 rc = ldlm_setup();
2536                 if (rc)
2537                         ldlm_refcount--;
2538         }
2539         mutex_unlock(&ldlm_ref_mutex);
2540
2541         RETURN(rc);
2542 }
2543 EXPORT_SYMBOL(ldlm_get_ref);
2544
2545 void ldlm_put_ref(void)
2546 {
2547         ENTRY;
2548         mutex_lock(&ldlm_ref_mutex);
2549         if (ldlm_refcount == 1) {
2550                 int rc = ldlm_cleanup();
2551                 if (rc)
2552                         CERROR("ldlm_cleanup failed: %d\n", rc);
2553                 else
2554                         ldlm_refcount--;
2555         } else {
2556                 ldlm_refcount--;
2557         }
2558         mutex_unlock(&ldlm_ref_mutex);
2559
2560         EXIT;
2561 }
2562 EXPORT_SYMBOL(ldlm_put_ref);
2563
2564 /*
2565  * Export handle<->lock hash operations.
2566  */
2567 static unsigned
2568 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2569 {
2570         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2571 }
2572
2573 static void *
2574 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2575 {
2576         struct ldlm_lock *lock;
2577
2578         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2579         return &lock->l_remote_handle;
2580 }
2581
2582 static void
2583 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2584 {
2585         struct ldlm_lock     *lock;
2586
2587         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2588         lock->l_remote_handle = *(struct lustre_handle *)key;
2589 }
2590
2591 static int
2592 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2593 {
2594         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2595 }
2596
2597 static void *
2598 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2599 {
2600         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2601 }
2602
2603 static void
2604 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2605 {
2606         struct ldlm_lock *lock;
2607
2608         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2609         LDLM_LOCK_GET(lock);
2610 }
2611
2612 static void
2613 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2614 {
2615         struct ldlm_lock *lock;
2616
2617         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2618         LDLM_LOCK_RELEASE(lock);
2619 }
2620
2621 static cfs_hash_ops_t ldlm_export_lock_ops = {
2622         .hs_hash        = ldlm_export_lock_hash,
2623         .hs_key         = ldlm_export_lock_key,
2624         .hs_keycmp      = ldlm_export_lock_keycmp,
2625         .hs_keycpy      = ldlm_export_lock_keycpy,
2626         .hs_object      = ldlm_export_lock_object,
2627         .hs_get         = ldlm_export_lock_get,
2628         .hs_put         = ldlm_export_lock_put,
2629         .hs_put_locked  = ldlm_export_lock_put,
2630 };
2631
2632 int ldlm_init_export(struct obd_export *exp)
2633 {
2634         ENTRY;
2635
2636         exp->exp_lock_hash =
2637                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2638                                 HASH_EXP_LOCK_CUR_BITS,
2639                                 HASH_EXP_LOCK_MAX_BITS,
2640                                 HASH_EXP_LOCK_BKT_BITS, 0,
2641                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2642                                 &ldlm_export_lock_ops,
2643                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2644                                 CFS_HASH_NBLK_CHANGE);
2645
2646         if (!exp->exp_lock_hash)
2647                 RETURN(-ENOMEM);
2648
2649         RETURN(0);
2650 }
2651 EXPORT_SYMBOL(ldlm_init_export);
2652
2653 void ldlm_destroy_export(struct obd_export *exp)
2654 {
2655         ENTRY;
2656         cfs_hash_putref(exp->exp_lock_hash);
2657         exp->exp_lock_hash = NULL;
2658
2659         ldlm_destroy_flock_export(exp);
2660         EXIT;
2661 }
2662 EXPORT_SYMBOL(ldlm_destroy_export);
2663
2664 static int ldlm_setup(void)
2665 {
2666         static struct ptlrpc_service_conf       conf;
2667         struct ldlm_bl_pool                     *blp = NULL;
2668         int rc = 0;
2669 #ifdef __KERNEL__
2670         int i;
2671 #endif
2672         ENTRY;
2673
2674         if (ldlm_state != NULL)
2675                 RETURN(-EALREADY);
2676
2677         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2678         if (ldlm_state == NULL)
2679                 RETURN(-ENOMEM);
2680
2681 #ifdef LPROCFS
2682         rc = ldlm_proc_setup();
2683         if (rc != 0)
2684                 GOTO(out, rc);
2685 #endif
2686
2687         memset(&conf, 0, sizeof(conf));
2688         conf = (typeof(conf)) {
2689                 .psc_name               = "ldlm_cbd",
2690                 .psc_watchdog_factor    = 2,
2691                 .psc_buf                = {
2692                         .bc_nbufs               = LDLM_NBUFS,
2693                         .bc_buf_size            = LDLM_BUFSIZE,
2694                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2695                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2696                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
2697                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
2698                 },
2699                 .psc_thr                = {
2700                         .tc_thr_name            = "ldlm_cb",
2701                         .tc_thr_factor          = LDLM_THR_FACTOR,
2702                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2703                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2704                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2705                         .tc_nthrs_user          = ldlm_num_threads,
2706                         .tc_cpu_affinity        = 1,
2707                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
2708                 },
2709                 .psc_cpt                = {
2710                         .cc_pattern             = ldlm_cpts,
2711                 },
2712                 .psc_ops                = {
2713                         .so_req_handler         = ldlm_callback_handler,
2714                 },
2715         };
2716         ldlm_state->ldlm_cb_service = \
2717                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2718         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2719                 CERROR("failed to start service\n");
2720                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2721                 ldlm_state->ldlm_cb_service = NULL;
2722                 GOTO(out, rc);
2723         }
2724
2725 #ifdef HAVE_SERVER_SUPPORT
2726         memset(&conf, 0, sizeof(conf));
2727         conf = (typeof(conf)) {
2728                 .psc_name               = "ldlm_canceld",
2729                 .psc_watchdog_factor    = 6,
2730                 .psc_buf                = {
2731                         .bc_nbufs               = LDLM_NBUFS,
2732                         .bc_buf_size            = LDLM_BUFSIZE,
2733                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2734                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2735                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
2736                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
2737
2738                 },
2739                 .psc_thr                = {
2740                         .tc_thr_name            = "ldlm_cn",
2741                         .tc_thr_factor          = LDLM_THR_FACTOR,
2742                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2743                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2744                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2745                         .tc_nthrs_user          = ldlm_num_threads,
2746                         .tc_cpu_affinity        = 1,
2747                         .tc_ctx_tags            = LCT_MD_THREAD | \
2748                                                   LCT_DT_THREAD | \
2749                                                   LCT_CL_THREAD,
2750                 },
2751                 .psc_cpt                = {
2752                         .cc_pattern             = ldlm_cpts,
2753                 },
2754                 .psc_ops                = {
2755                         .so_req_handler         = ldlm_cancel_handler,
2756                         .so_hpreq_handler       = ldlm_hpreq_handler,
2757                 },
2758         };
2759         ldlm_state->ldlm_cancel_service = \
2760                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2761         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2762                 CERROR("failed to start service\n");
2763                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2764                 ldlm_state->ldlm_cancel_service = NULL;
2765                 GOTO(out, rc);
2766         }
2767 #endif
2768
2769         OBD_ALLOC(blp, sizeof(*blp));
2770         if (blp == NULL)
2771                 GOTO(out, rc = -ENOMEM);
2772         ldlm_state->ldlm_bl_pool = blp;
2773
2774         spin_lock_init(&blp->blp_lock);
2775         CFS_INIT_LIST_HEAD(&blp->blp_list);
2776         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2777         cfs_waitq_init(&blp->blp_waitq);
2778         cfs_atomic_set(&blp->blp_num_threads, 0);
2779         cfs_atomic_set(&blp->blp_busy_threads, 0);
2780
2781 #ifdef __KERNEL__
2782         if (ldlm_num_threads == 0) {
2783                 blp->blp_min_threads = LDLM_NTHRS_INIT;
2784                 blp->blp_max_threads = LDLM_NTHRS_MAX;
2785         } else {
2786                 blp->blp_min_threads = blp->blp_max_threads = \
2787                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2788                                                          ldlm_num_threads));
2789         }
2790
2791         for (i = 0; i < blp->blp_min_threads; i++) {
2792                 rc = ldlm_bl_thread_start(blp);
2793                 if (rc < 0)
2794                         GOTO(out, rc);
2795         }
2796
2797 # ifdef HAVE_SERVER_SUPPORT
2798         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2799         expired_lock_thread.elt_state = ELT_STOPPED;
2800         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2801
2802         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2803         spin_lock_init(&waiting_locks_spinlock);
2804         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2805
2806         rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2807         if (rc < 0) {
2808                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2809                 GOTO(out, rc);
2810         }
2811
2812         cfs_wait_event(expired_lock_thread.elt_waitq,
2813                        expired_lock_thread.elt_state == ELT_READY);
2814 # endif /* HAVE_SERVER_SUPPORT */
2815
2816         rc = ldlm_pools_init();
2817         if (rc) {
2818                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2819                 GOTO(out, rc);
2820         }
2821 #endif
2822         RETURN(0);
2823
2824  out:
2825         ldlm_cleanup();
2826         RETURN(rc);
2827 }
2828
2829 static int ldlm_cleanup(void)
2830 {
2831         ENTRY;
2832
2833         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2834             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2835                 CERROR("ldlm still has namespaces; clean these up first.\n");
2836                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2837                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2838                 RETURN(-EBUSY);
2839         }
2840
2841 #ifdef __KERNEL__
2842         ldlm_pools_fini();
2843
2844         if (ldlm_state->ldlm_bl_pool != NULL) {
2845                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2846
2847                 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2848                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2849
2850                         init_completion(&blp->blp_comp);
2851
2852                         spin_lock(&blp->blp_lock);
2853                         cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2854                         cfs_waitq_signal(&blp->blp_waitq);
2855                         spin_unlock(&blp->blp_lock);
2856
2857                         wait_for_completion(&blp->blp_comp);
2858                 }
2859
2860                 OBD_FREE(blp, sizeof(*blp));
2861         }
2862 #endif /* __KERNEL__ */
2863
2864         if (ldlm_state->ldlm_cb_service != NULL)
2865                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2866 # ifdef HAVE_SERVER_SUPPORT
2867         if (ldlm_state->ldlm_cancel_service != NULL)
2868                 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2869 # endif
2870
2871 #ifdef __KERNEL__
2872         ldlm_proc_cleanup();
2873
2874 # ifdef HAVE_SERVER_SUPPORT
2875         if (expired_lock_thread.elt_state != ELT_STOPPED) {
2876                 expired_lock_thread.elt_state = ELT_TERMINATE;
2877                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2878                 cfs_wait_event(expired_lock_thread.elt_waitq,
2879                                expired_lock_thread.elt_state == ELT_STOPPED);
2880         }
2881 # endif
2882 #endif /* __KERNEL__ */
2883
2884         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2885         ldlm_state = NULL;
2886
2887         RETURN(0);
2888 }
2889
2890 int ldlm_init(void)
2891 {
2892         mutex_init(&ldlm_ref_mutex);
2893         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2894         mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2895         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2896                                                sizeof(struct ldlm_resource), 0,
2897                                                CFS_SLAB_HWCACHE_ALIGN);
2898         if (ldlm_resource_slab == NULL)
2899                 return -ENOMEM;
2900
2901         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2902                               sizeof(struct ldlm_lock), 0,
2903                               CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
2904         if (ldlm_lock_slab == NULL) {
2905                 cfs_mem_cache_destroy(ldlm_resource_slab);
2906                 return -ENOMEM;
2907         }
2908
2909         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2910                                         sizeof(struct ldlm_interval),
2911                                         0, CFS_SLAB_HWCACHE_ALIGN);
2912         if (ldlm_interval_slab == NULL) {
2913                 cfs_mem_cache_destroy(ldlm_resource_slab);
2914                 cfs_mem_cache_destroy(ldlm_lock_slab);
2915                 return -ENOMEM;
2916         }
2917 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2918         class_export_dump_hook = ldlm_dump_export_locks;
2919 #endif
2920         return 0;
2921 }
2922
2923 void ldlm_exit(void)
2924 {
2925         int rc;
2926         if (ldlm_refcount)
2927                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2928         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2929         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2930 #ifdef __KERNEL__
2931         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2932          * synchronize_rcu() to wait a grace period elapsed, so that
2933          * ldlm_lock_free() get a chance to be called. */
2934         synchronize_rcu();
2935 #endif
2936         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2937         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2938         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2939         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2940 }