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 cfs_mutex_t      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         cfs_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         cfs_completion_t        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         cfs_completion_t        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 cfs_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         cfs_spin_lock_bh(&waiting_locks_spinlock);
 151         need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
 152         cfs_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                 cfs_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                         cfs_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                         cfs_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                                 cfs_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                         cfs_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                         cfs_spin_lock_bh(&waiting_locks_spinlock);
 237                 }
 238                 cfs_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         cfs_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         cfs_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         cfs_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                         cfs_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                         cfs_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         cfs_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         cfs_spin_lock_bh(&waiting_locks_spinlock);
 458         if (lock->l_destroyed) {
 459                 static cfs_time_t next;
 460                 cfs_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         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 476
 477         if (ret) {
 478                 cfs_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                 cfs_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         cfs_spin_lock_bh(&waiting_locks_spinlock);
 537         ret = __ldlm_del_waiting_lock(lock);
 538         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 539
 540         /* remove the lock out of export blocking list */
 541         cfs_spin_lock_bh(&lock->l_export->exp_bl_list_lock);
 542         cfs_list_del_init(&lock->l_exp_list);
 543         cfs_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
 555 /*
 556  * Prolong the lock
 557  *
 558  * Called with namespace lock held.
 559  */
 560 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 561 {
 562         if (lock->l_export == NULL) {
 563                 /* We don't have a "waiting locks list" on clients. */
 564                 LDLM_DEBUG(lock, "client lock: no-op");
 565                 return 0;
 566         }
 567
 568         cfs_spin_lock_bh(&waiting_locks_spinlock);
 569
 570         if (cfs_list_empty(&lock->l_pending_chain)) {
 571                 cfs_spin_unlock_bh(&waiting_locks_spinlock);
 572                 LDLM_DEBUG(lock, "wasn't waiting");
 573                 return 0;
 574         }
 575
 576         /* we remove/add the lock to the waiting list, so no needs to
 577          * release/take a lock reference */
 578         __ldlm_del_waiting_lock(lock);
 579         __ldlm_add_waiting_lock(lock, timeout);
 580         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 581
 582         LDLM_DEBUG(lock, "refreshed");
 583         return 1;
 584 }
 585
 586 #else /* !HAVE_SERVER_SUPPORT ||  !__KERNEL__ */
 587
 588 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 589 {
 590         RETURN(0);
 591 }
 592
 593 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 594 {
 595         RETURN(0);
 596 }
 597
 598 # ifdef HAVE_SERVER_SUPPORT
 599 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 600 {
 601         LASSERT(lock->l_res_locked);
 602         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 603         RETURN(1);
 604 }
 605
 606 # endif
 607 #endif /* HAVE_SERVER_SUPPORT && __KERNEL__ */
 608
 609 #ifdef HAVE_SERVER_SUPPORT
 610
 611 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 612                             const char *ast_type)
 613 {
 614         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
 615                            "to a lock %s callback time out: rc %d\n",
 616                            lock->l_export->exp_obd->obd_name,
 617                            obd_export_nid2str(lock->l_export), ast_type, rc);
 618
 619         if (obd_dump_on_timeout)
 620                 libcfs_debug_dumplog();
 621 #ifdef __KERNEL__
 622         cfs_spin_lock_bh(&waiting_locks_spinlock);
 623         if (__ldlm_del_waiting_lock(lock) == 0)
 624                 /* the lock was not in any list, grab an extra ref before adding
 625                  * the lock to the expired list */
 626                 LDLM_LOCK_GET(lock);
 627         cfs_list_add(&lock->l_pending_chain,
 628                      &expired_lock_thread.elt_expired_locks);
 629         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 630         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 631 #else
 632         class_fail_export(lock->l_export);
 633 #endif
 634 }
 635
 636 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 637                                  struct ptlrpc_request *req, int rc,
 638                                  const char *ast_type)
 639 {
 640         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
 641
 642         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 643                 LASSERT(lock->l_export);
 644                 if (lock->l_export->exp_libclient) {
 645                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 646                                    " timeout, just cancelling lock", ast_type,
 647                                    libcfs_nid2str(peer.nid));
 648                         ldlm_lock_cancel(lock);
 649                         rc = -ERESTART;
 650                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
 651                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
 652                                    "cancel was received (AST reply lost?)",
 653                                    ast_type, libcfs_nid2str(peer.nid));
 654                         ldlm_lock_cancel(lock);
 655                         rc = -ERESTART;
 656                 } else {
 657                         ldlm_del_waiting_lock(lock);
 658                         ldlm_failed_ast(lock, rc, ast_type);
 659                 }
 660         } else if (rc) {
 661                 if (rc == -EINVAL) {
 662                         struct ldlm_resource *res = lock->l_resource;
 663                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 664                                " from %s AST - normal race",
 665                                libcfs_nid2str(peer.nid),
 666                                req->rq_repmsg ?
 667                                lustre_msg_get_status(req->rq_repmsg) : -1,
 668                                ast_type);
 669                         if (res) {
 670                                 /* update lvbo to return proper attributes.
 671                                  * see bug 23174 */
 672                                 ldlm_resource_getref(res);
 673                                 ldlm_res_lvbo_update(res, NULL, 1);
 674                                 ldlm_resource_putref(res);
 675                         }
 676
 677                 } else {
 678                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 679                                    "from %s AST", libcfs_nid2str(peer.nid),
 680                                    (req->rq_repmsg != NULL) ?
 681                                    lustre_msg_get_status(req->rq_repmsg) : 0,
 682                                    ast_type);
 683                 }
 684                 ldlm_lock_cancel(lock);
 685                 /* Server-side AST functions are called from ldlm_reprocess_all,
 686                  * which needs to be told to please restart its reprocessing. */
 687                 rc = -ERESTART;
 688         }
 689
 690         return rc;
 691 }
 692
 693 static int ldlm_cb_interpret(const struct lu_env *env,
 694                              struct ptlrpc_request *req, void *data, int rc)
 695 {
 696         struct ldlm_cb_async_args *ca   = data;
 697         struct ldlm_lock          *lock = ca->ca_lock;
 698         struct ldlm_cb_set_arg    *arg  = ca->ca_set_arg;
 699         ENTRY;
 700
 701         LASSERT(lock != NULL);
 702
 703         switch (arg->type) {
 704         case LDLM_GL_CALLBACK:
 705                 /* Update the LVB from disk if the AST failed
 706                  * (this is a legal race)
 707                  *
 708                  * - Glimpse callback of local lock just returns
 709                  *   -ELDLM_NO_LOCK_DATA.
 710                  * - Glimpse callback of remote lock might return
 711                  *   -ELDLM_NO_LOCK_DATA when inode is cleared. LU-274
 712                  */
 713                 if (rc == -ELDLM_NO_LOCK_DATA) {
 714                         LDLM_DEBUG(lock, "lost race - client has a lock but no "
 715                                    "inode");
 716                         ldlm_res_lvbo_update(lock->l_resource, NULL, 1);
 717                 } else if (rc != 0) {
 718                         rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 719                 } else {
 720                         rc = ldlm_res_lvbo_update(lock->l_resource, req, 1);
 721                 }
 722                 break;
 723         case LDLM_BL_CALLBACK:
 724                 if (rc != 0)
 725                         rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 726                 break;
 727         case LDLM_CP_CALLBACK:
 728                 if (rc != 0)
 729                         rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 730                 break;
 731         default:
 732                 LDLM_ERROR(lock, "invalid opcode for lock callback %d",
 733                            arg->type);
 734                 LBUG();
 735         }
 736
 737         /* release extra reference taken in ldlm_ast_fini() */
 738         LDLM_LOCK_RELEASE(lock);
 739
 740         if (rc == -ERESTART)
 741                 cfs_atomic_inc(&arg->restart);
 742
 743         RETURN(0);
 744 }
 745
 746 static inline int ldlm_ast_fini(struct ptlrpc_request *req,
 747                                 struct ldlm_cb_set_arg *arg,
 748                                 struct ldlm_lock *lock,
 749                                 int instant_cancel)
 750 {
 751         int rc = 0;
 752         ENTRY;
 753
 754         if (unlikely(instant_cancel)) {
 755                 rc = ptl_send_rpc(req, 1);
 756                 ptlrpc_req_finished(req);
 757                 if (rc == 0)
 758                         cfs_atomic_inc(&arg->restart);
 759         } else {
 760                 LDLM_LOCK_GET(lock);
 761                 ptlrpc_set_add_req(arg->set, req);
 762         }
 763
 764         RETURN(rc);
 765 }
 766
 767 /**
 768  * Check if there are requests in the export request list which prevent
 769  * the lock canceling and make these requests high priority ones.
 770  */
 771 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
 772 {
 773         struct ptlrpc_request *req;
 774         ENTRY;
 775
 776         if (lock->l_export == NULL) {
 777                 LDLM_DEBUG(lock, "client lock: no-op");
 778                 RETURN_EXIT;
 779         }
 780
 781         cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
 782         cfs_list_for_each_entry(req, &lock->l_export->exp_hp_rpcs,
 783                                 rq_exp_list) {
 784                 /* Do not process requests that were not yet added to there
 785                  * incoming queue or were already removed from there for
 786                  * processing */
 787                 if (!req->rq_hp && !cfs_list_empty(&req->rq_list) &&
 788                     req->rq_ops->hpreq_lock_match &&
 789                     req->rq_ops->hpreq_lock_match(req, lock))
 790                         ptlrpc_hpreq_reorder(req);
 791         }
 792         cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
 793         EXIT;
 794 }
 795
 796 /*
 797  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
 798  * enqueued server lock conflicts with given one.
 799  *
 800  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
 801  * to wait for client response.
 802  */
 803 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 804                              struct ldlm_lock_desc *desc,
 805                              void *data, int flag)
 806 {
 807         struct ldlm_cb_async_args *ca;
 808         struct ldlm_cb_set_arg *arg = data;
 809         struct ldlm_request    *body;
 810         struct ptlrpc_request  *req;
 811         int                     instant_cancel = 0;
 812         int                     rc = 0;
 813         ENTRY;
 814
 815         if (flag == LDLM_CB_CANCELING)
 816                 /* Don't need to do anything here. */
 817                 RETURN(0);
 818
 819         LASSERT(lock);
 820         LASSERT(data != NULL);
 821         if (lock->l_export->exp_obd->obd_recovering != 0)
 822                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
 823
 824         ldlm_lock_reorder_req(lock);
 825
 826         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 827                                         &RQF_LDLM_BL_CALLBACK,
 828                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
 829         if (req == NULL)
 830                 RETURN(-ENOMEM);
 831
 832         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 833         ca = ptlrpc_req_async_args(req);
 834         ca->ca_set_arg = arg;
 835         ca->ca_lock = lock;
 836
 837         req->rq_interpret_reply = ldlm_cb_interpret;
 838         req->rq_no_resend = 1;
 839
 840         lock_res_and_lock(lock);
 841         if (lock->l_granted_mode != lock->l_req_mode) {
 842                 /* this blocking AST will be communicated as part of the
 843                  * completion AST instead */
 844                 unlock_res_and_lock(lock);
 845
 846                 ptlrpc_req_finished(req);
 847                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 848                 RETURN(0);
 849         }
 850
 851         if (lock->l_destroyed) {
 852                 /* What's the point? */
 853                 unlock_res_and_lock(lock);
 854                 ptlrpc_req_finished(req);
 855                 RETURN(0);
 856         }
 857
 858         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
 859                 instant_cancel = 1;
 860
 861         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 862         body->lock_handle[0] = lock->l_remote_handle;
 863         body->lock_desc = *desc;
 864         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 865
 866         LDLM_DEBUG(lock, "server preparing blocking AST");
 867
 868         ptlrpc_request_set_replen(req);
 869         if (instant_cancel) {
 870                 unlock_res_and_lock(lock);
 871                 ldlm_lock_cancel(lock);
 872         } else {
 873                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
 874                 ldlm_add_waiting_lock(lock);
 875                 unlock_res_and_lock(lock);
 876         }
 877
 878         req->rq_send_state = LUSTRE_IMP_FULL;
 879         /* ptlrpc_request_alloc_pack already set timeout */
 880         if (AT_OFF)
 881                 req->rq_timeout = ldlm_get_rq_timeout();
 882
 883         if (lock->l_export && lock->l_export->exp_nid_stats &&
 884             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 885                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 886                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 887
 888         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
 889
 890         RETURN(rc);
 891 }
 892
 893 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 894 {
 895         struct ldlm_cb_set_arg *arg = data;
 896         struct ldlm_request    *body;
 897         struct ptlrpc_request  *req;
 898         struct ldlm_cb_async_args *ca;
 899         long                    total_enqueue_wait;
 900         int                     instant_cancel = 0;
 901         int                     rc = 0;
 902         int                     lvb_len;
 903         ENTRY;
 904
 905         LASSERT(lock != NULL);
 906         LASSERT(data != NULL);
 907
 908         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
 909                                           lock->l_last_activity);
 910
 911         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
 912                                     &RQF_LDLM_CP_CALLBACK);
 913         if (req == NULL)
 914                 RETURN(-ENOMEM);
 915
 916         /* server namespace, doesn't need lock */
 917         lvb_len = ldlm_lvbo_size(lock);
 918         if (lvb_len > 0)
 919                  req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
 920                                       lvb_len);
 921
 922         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
 923         if (rc) {
 924                 ptlrpc_request_free(req);
 925                 RETURN(rc);
 926         }
 927
 928         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 929         ca = ptlrpc_req_async_args(req);
 930         ca->ca_set_arg = arg;
 931         ca->ca_lock = lock;
 932
 933         req->rq_interpret_reply = ldlm_cb_interpret;
 934         req->rq_no_resend = 1;
 935         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 936
 937         body->lock_handle[0] = lock->l_remote_handle;
 938         body->lock_flags = flags;
 939         ldlm_lock2desc(lock, &body->lock_desc);
 940         if (lvb_len > 0) {
 941                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
 942
 943                 lvb_len = ldlm_lvbo_fill(lock, lvb, lvb_len);
 944                 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
 945                                    lvb_len, RCL_CLIENT);
 946         }
 947
 948         LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
 949                    total_enqueue_wait);
 950
 951         /* Server-side enqueue wait time estimate, used in
 952             __ldlm_add_waiting_lock to set future enqueue timers */
 953         if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
 954                 at_measured(ldlm_lock_to_ns_at(lock),
 955                             total_enqueue_wait);
 956         else
 957                 /* bz18618. Don't add lock enqueue time we spend waiting for a
 958                    previous callback to fail. Locks waiting legitimately will
 959                    get extended by ldlm_refresh_waiting_lock regardless of the
 960                    estimate, so it's okay to underestimate here. */
 961                 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
 962                        "It is likely that a previous callback timed out.",
 963                        total_enqueue_wait,
 964                        at_get(ldlm_lock_to_ns_at(lock)));
 965
 966         ptlrpc_request_set_replen(req);
 967
 968         req->rq_send_state = LUSTRE_IMP_FULL;
 969         /* ptlrpc_request_pack already set timeout */
 970         if (AT_OFF)
 971                 req->rq_timeout = ldlm_get_rq_timeout();
 972
 973         /* We only send real blocking ASTs after the lock is granted */
 974         lock_res_and_lock(lock);
 975         if (lock->l_flags & LDLM_FL_AST_SENT) {
 976                 body->lock_flags |= LDLM_FL_AST_SENT;
 977                 /* copy ast flags like LDLM_FL_DISCARD_DATA */
 978                 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 979
 980                 /* We might get here prior to ldlm_handle_enqueue setting
 981                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
 982                  * into waiting list, but this is safe and similar code in
 983                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
 984                  * that would not only cancel the lock, but will also remove
 985                  * it from waiting list */
 986                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 987                         unlock_res_and_lock(lock);
 988                         ldlm_lock_cancel(lock);
 989                         instant_cancel = 1;
 990                         lock_res_and_lock(lock);
 991                 } else {
 992                         /* start the lock-timeout clock */
 993                         ldlm_add_waiting_lock(lock);
 994                 }
 995         }
 996         unlock_res_and_lock(lock);
 997
 998         if (lock->l_export && lock->l_export->exp_nid_stats &&
 999             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1000                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1001                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
1002
1003         rc = ldlm_ast_fini(req, arg, lock, instant_cancel);
1004
1005         RETURN(rc);
1006 }
1007
1008 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
1009 {
1010         struct ldlm_cb_set_arg          *arg = data;
1011         struct ldlm_request             *body;
1012         struct ptlrpc_request           *req;
1013         struct ldlm_cb_async_args       *ca;
1014         int                              rc;
1015         struct req_format               *req_fmt;
1016         ENTRY;
1017
1018         LASSERT(lock != NULL);
1019
1020         if (arg->gl_desc != NULL)
1021                 /* There is a glimpse descriptor to pack */
1022                 req_fmt = &RQF_LDLM_GL_DESC_CALLBACK;
1023         else
1024                 req_fmt = &RQF_LDLM_GL_CALLBACK;
1025
1026         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
1027                                         req_fmt, LUSTRE_DLM_VERSION,
1028                                         LDLM_GL_CALLBACK);
1029
1030         if (req == NULL)
1031                 RETURN(-ENOMEM);
1032
1033         if (arg->gl_desc != NULL) {
1034                 /* copy the GL descriptor */
1035                 union ldlm_gl_desc      *desc;
1036                 desc = req_capsule_client_get(&req->rq_pill, &RMF_DLM_GL_DESC);
1037                 *desc = *arg->gl_desc;
1038         }
1039
1040         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1041         body->lock_handle[0] = lock->l_remote_handle;
1042         ldlm_lock2desc(lock, &body->lock_desc);
1043
1044         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
1045         ca = ptlrpc_req_async_args(req);
1046         ca->ca_set_arg = arg;
1047         ca->ca_lock = lock;
1048
1049         /* server namespace, doesn't need lock */
1050         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
1051                              ldlm_lvbo_size(lock));
1052         ptlrpc_request_set_replen(req);
1053
1054         req->rq_send_state = LUSTRE_IMP_FULL;
1055         /* ptlrpc_request_alloc_pack already set timeout */
1056         if (AT_OFF)
1057                 req->rq_timeout = ldlm_get_rq_timeout();
1058
1059         req->rq_interpret_reply = ldlm_cb_interpret;
1060
1061         if (lock->l_export && lock->l_export->exp_nid_stats &&
1062             lock->l_export->exp_nid_stats->nid_ldlm_stats)
1063                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
1064                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
1065
1066         rc = ldlm_ast_fini(req, arg, lock, 0);
1067
1068         RETURN(rc);
1069 }
1070
1071 int ldlm_glimpse_locks(struct ldlm_resource *res, cfs_list_t *gl_work_list)
1072 {
1073         int     rc;
1074         ENTRY;
1075
1076         rc = ldlm_run_ast_work(ldlm_res_to_ns(res), gl_work_list,
1077                                LDLM_WORK_GL_AST);
1078         if (rc == -ERESTART)
1079                 ldlm_reprocess_all(res);
1080
1081         RETURN(rc);
1082 }
1083
1084 /* return ldlm lock associated with a lock callback request */
1085 struct ldlm_lock *ldlm_request_lock(struct ptlrpc_request *req)
1086 {
1087         struct ldlm_cb_async_args       *ca;
1088         struct ldlm_lock                *lock;
1089         ENTRY;
1090
1091         ca = ptlrpc_req_async_args(req);
1092         lock = ca->ca_lock;
1093         if (lock == NULL)
1094                 RETURN(ERR_PTR(-EFAULT));
1095
1096         RETURN(lock);
1097 }
1098 EXPORT_SYMBOL(ldlm_request_lock);
1099
1100 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1101                        struct lprocfs_stats *srv_stats)
1102 {
1103         int lock_type = 0, op = 0;
1104
1105         lock_type = dlm_req->lock_desc.l_resource.lr_type;
1106
1107         switch (lock_type) {
1108         case LDLM_PLAIN:
1109                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1110                 break;
1111         case LDLM_EXTENT:
1112                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1113                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1114                 else
1115                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1116                 break;
1117         case LDLM_FLOCK:
1118                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1119                 break;
1120         case LDLM_IBITS:
1121                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1122                 break;
1123         default:
1124                 op = 0;
1125                 break;
1126         }
1127
1128         if (op)
1129                 lprocfs_counter_incr(srv_stats, op);
1130
1131         return;
1132 }
1133
1134 /*
1135  * Main server-side entry point into LDLM. This is called by ptlrpc service
1136  * threads to carry out client lock enqueueing requests.
1137  */
1138 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1139                          struct ptlrpc_request *req,
1140                          const struct ldlm_request *dlm_req,
1141                          const struct ldlm_callback_suite *cbs)
1142 {
1143         struct ldlm_reply *dlm_rep;
1144         __u32 flags;
1145         ldlm_error_t err = ELDLM_OK;
1146         struct ldlm_lock *lock = NULL;
1147         void *cookie = NULL;
1148         int rc = 0;
1149         ENTRY;
1150
1151         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1152
1153         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1154         flags = dlm_req->lock_flags;
1155
1156         LASSERT(req->rq_export);
1157
1158         if (ptlrpc_req2svc(req)->srv_stats != NULL)
1159                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1160
1161         if (req->rq_export && req->rq_export->exp_nid_stats &&
1162             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1163                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1164                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1165
1166         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1167                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1168                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1169                           dlm_req->lock_desc.l_resource.lr_type);
1170                 GOTO(out, rc = -EFAULT);
1171         }
1172
1173         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1174                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1175                      dlm_req->lock_desc.l_req_mode &
1176                      (dlm_req->lock_desc.l_req_mode-1))) {
1177                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1178                           dlm_req->lock_desc.l_req_mode);
1179                 GOTO(out, rc = -EFAULT);
1180         }
1181
1182         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1183                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1184                              LDLM_PLAIN)) {
1185                         DEBUG_REQ(D_ERROR, req,
1186                                   "PLAIN lock request from IBITS client?");
1187                         GOTO(out, rc = -EPROTO);
1188                 }
1189         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1190                             LDLM_IBITS)) {
1191                 DEBUG_REQ(D_ERROR, req,
1192                           "IBITS lock request from unaware client?");
1193                 GOTO(out, rc = -EPROTO);
1194         }
1195
1196 #if 0
1197         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1198            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1199            ibits for mgc/mgs) */
1200
1201         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1202          * inodebits lock if client does not support them. */
1203         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1204             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1205                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1206                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1207                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1208                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1209                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1210         }
1211 #endif
1212
1213         if (unlikely(flags & LDLM_FL_REPLAY)) {
1214                 /* Find an existing lock in the per-export lock hash */
1215                 /* In the function below, .hs_keycmp resolves to
1216                  * ldlm_export_lock_keycmp() */
1217                 /* coverity[overrun-buffer-val] */
1218                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1219                                        (void *)&dlm_req->lock_handle[0]);
1220                 if (lock != NULL) {
1221                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1222                                   LPX64, lock->l_handle.h_cookie);
1223                         GOTO(existing_lock, rc = 0);
1224                 }
1225         }
1226
1227         /* The lock's callback data might be set in the policy function */
1228         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1229                                 dlm_req->lock_desc.l_resource.lr_type,
1230                                 dlm_req->lock_desc.l_req_mode,
1231                                 cbs, NULL, 0);
1232
1233         if (!lock)
1234                 GOTO(out, rc = -ENOMEM);
1235
1236         lock->l_last_activity = cfs_time_current_sec();
1237         lock->l_remote_handle = dlm_req->lock_handle[0];
1238         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1239
1240         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1241         /* Don't enqueue a lock onto the export if it is been disonnected
1242          * due to eviction (bug 3822) or server umount (bug 24324).
1243          * Cancel it now instead. */
1244         if (req->rq_export->exp_disconnected) {
1245                 LDLM_ERROR(lock, "lock on disconnected export %p",
1246                            req->rq_export);
1247                 GOTO(out, rc = -ENOTCONN);
1248         }
1249
1250         lock->l_export = class_export_lock_get(req->rq_export, lock);
1251         if (lock->l_export->exp_lock_hash)
1252                 cfs_hash_add(lock->l_export->exp_lock_hash,
1253                              &lock->l_remote_handle,
1254                              &lock->l_exp_hash);
1255
1256 existing_lock:
1257
1258         if (flags & LDLM_FL_HAS_INTENT) {
1259                 /* In this case, the reply buffer is allocated deep in
1260                  * local_lock_enqueue by the policy function. */
1261                 cookie = req;
1262         } else {
1263                 /* based on the assumption that lvb size never changes during
1264                  * resource life time otherwise it need resource->lr_lock's
1265                  * protection */
1266                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1267                                      RCL_SERVER, ldlm_lvbo_size(lock));
1268
1269                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1270                         GOTO(out, rc = -ENOMEM);
1271
1272                 rc = req_capsule_server_pack(&req->rq_pill);
1273                 if (rc)
1274                         GOTO(out, rc);
1275         }
1276
1277         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1278                 ldlm_convert_policy_to_local(req->rq_export,
1279                                           dlm_req->lock_desc.l_resource.lr_type,
1280                                           &dlm_req->lock_desc.l_policy_data,
1281                                           &lock->l_policy_data);
1282         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1283                 lock->l_req_extent = lock->l_policy_data.l_extent;
1284
1285         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1286         if (err)
1287                 GOTO(out, err);
1288
1289         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1290         dlm_rep->lock_flags = flags;
1291
1292         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1293         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1294
1295         /* We never send a blocking AST until the lock is granted, but
1296          * we can tell it right now */
1297         lock_res_and_lock(lock);
1298
1299         /* Now take into account flags to be inherited from original lock
1300            request both in reply to client and in our own lock flags. */
1301         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1302         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1303
1304         /* Don't move a pending lock onto the export if it has already been
1305          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1306          * Cancel it now instead. */
1307         if (unlikely(req->rq_export->exp_disconnected ||
1308                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1309                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1310                 rc = -ENOTCONN;
1311         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1312                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1313                 if (lock->l_granted_mode == lock->l_req_mode) {
1314                         /*
1315                          * Only cancel lock if it was granted, because it would
1316                          * be destroyed immediately and would never be granted
1317                          * in the future, causing timeouts on client.  Not
1318                          * granted lock will be cancelled immediately after
1319                          * sending completion AST.
1320                          */
1321                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1322                                 unlock_res_and_lock(lock);
1323                                 ldlm_lock_cancel(lock);
1324                                 lock_res_and_lock(lock);
1325                         } else
1326                                 ldlm_add_waiting_lock(lock);
1327                 }
1328         }
1329         /* Make sure we never ever grant usual metadata locks to liblustre
1330            clients */
1331         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1332             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1333              req->rq_export->exp_libclient) {
1334                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1335                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1336                         CERROR("Granting sync lock to libclient. "
1337                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1338                                dlm_req->lock_flags, dlm_rep->lock_flags,
1339                                lock->l_flags);
1340                         LDLM_ERROR(lock, "sync lock");
1341                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1342                                 struct ldlm_intent *it;
1343
1344                                 it = req_capsule_client_get(&req->rq_pill,
1345                                                             &RMF_LDLM_INTENT);
1346                                 if (it != NULL) {
1347                                         CERROR("This is intent %s ("LPU64")\n",
1348                                                ldlm_it2str(it->opc), it->opc);
1349                                 }
1350                         }
1351                 }
1352         }
1353
1354         unlock_res_and_lock(lock);
1355
1356         EXIT;
1357  out:
1358         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1359         if (!req->rq_packed_final) {
1360                 err = lustre_pack_reply(req, 1, NULL, NULL);
1361                 if (rc == 0)
1362                         rc = err;
1363         }
1364
1365         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1366          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1367         if (lock) {
1368                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1369                            "(err=%d, rc=%d)", err, rc);
1370
1371                 if (rc == 0) {
1372                         int lvb_len = ldlm_lvbo_size(lock);
1373
1374                         if (lvb_len > 0) {
1375                                 void *buf;
1376                                 int buflen;
1377
1378                                 buf = req_capsule_server_get(&req->rq_pill,
1379                                                              &RMF_DLM_LVB);
1380                                 LASSERTF(buf != NULL, "req %p, lock %p\n",
1381                                          req, lock);
1382                                 buflen = req_capsule_get_size(&req->rq_pill,
1383                                                 &RMF_DLM_LVB, RCL_SERVER);
1384                                 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1385                                 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
1386                                                    buflen, RCL_SERVER);
1387                         }
1388                 } else {
1389                         lock_res_and_lock(lock);
1390                         ldlm_resource_unlink_lock(lock);
1391                         ldlm_lock_destroy_nolock(lock);
1392                         unlock_res_and_lock(lock);
1393                 }
1394
1395                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1396                         ldlm_reprocess_all(lock->l_resource);
1397
1398                 LDLM_LOCK_RELEASE(lock);
1399         }
1400
1401         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1402                           lock, rc);
1403
1404         return rc;
1405 }
1406
1407 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1408                         ldlm_completion_callback completion_callback,
1409                         ldlm_blocking_callback blocking_callback,
1410                         ldlm_glimpse_callback glimpse_callback)
1411 {
1412         struct ldlm_request *dlm_req;
1413         struct ldlm_callback_suite cbs = {
1414                 .lcs_completion = completion_callback,
1415                 .lcs_blocking   = blocking_callback,
1416                 .lcs_glimpse    = glimpse_callback
1417         };
1418         int rc;
1419
1420         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1421         if (dlm_req != NULL) {
1422                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1423                                           req, dlm_req, &cbs);
1424         } else {
1425                 rc = -EFAULT;
1426         }
1427         return rc;
1428 }
1429
1430 int ldlm_handle_convert0(struct ptlrpc_request *req,
1431                          const struct ldlm_request *dlm_req)
1432 {
1433         struct ldlm_reply *dlm_rep;
1434         struct ldlm_lock *lock;
1435         int rc;
1436         ENTRY;
1437
1438         if (req->rq_export && req->rq_export->exp_nid_stats &&
1439             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1440                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1441                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1442
1443         rc = req_capsule_server_pack(&req->rq_pill);
1444         if (rc)
1445                 RETURN(rc);
1446
1447         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1448         dlm_rep->lock_flags = dlm_req->lock_flags;
1449
1450         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1451         if (!lock) {
1452                 req->rq_status = EINVAL;
1453         } else {
1454                 void *res = NULL;
1455
1456                 LDLM_DEBUG(lock, "server-side convert handler START");
1457
1458                 lock->l_last_activity = cfs_time_current_sec();
1459                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1460                                         &dlm_rep->lock_flags);
1461                 if (res) {
1462                         if (ldlm_del_waiting_lock(lock))
1463                                 LDLM_DEBUG(lock, "converted waiting lock");
1464                         req->rq_status = 0;
1465                 } else {
1466                         req->rq_status = EDEADLOCK;
1467                 }
1468         }
1469
1470         if (lock) {
1471                 if (!req->rq_status)
1472                         ldlm_reprocess_all(lock->l_resource);
1473                 LDLM_DEBUG(lock, "server-side convert handler END");
1474                 LDLM_LOCK_PUT(lock);
1475         } else
1476                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1477
1478         RETURN(0);
1479 }
1480
1481 int ldlm_handle_convert(struct ptlrpc_request *req)
1482 {
1483         int rc;
1484         struct ldlm_request *dlm_req;
1485
1486         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1487         if (dlm_req != NULL) {
1488                 rc = ldlm_handle_convert0(req, dlm_req);
1489         } else {
1490                 CERROR ("Can't unpack dlm_req\n");
1491                 rc = -EFAULT;
1492         }
1493         return rc;
1494 }
1495
1496 /* Cancel all the locks whos handles are packed into ldlm_request */
1497 int ldlm_request_cancel(struct ptlrpc_request *req,
1498                         const struct ldlm_request *dlm_req, int first)
1499 {
1500         struct ldlm_resource *res, *pres = NULL;
1501         struct ldlm_lock *lock;
1502         int i, count, done = 0;
1503         ENTRY;
1504
1505         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1506         if (first >= count)
1507                 RETURN(0);
1508
1509         if (count == 1 && dlm_req->lock_handle[0].cookie == 0)
1510                 RETURN(0);
1511
1512         /* There is no lock on the server at the replay time,
1513          * skip lock cancelling to make replay tests to pass. */
1514         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1515                 RETURN(0);
1516
1517         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1518                           "starting at %d", count, first);
1519
1520         for (i = first; i < count; i++) {
1521                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1522                 if (!lock) {
1523                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1524                                           "lock (cookie "LPU64")",
1525                                           dlm_req->lock_handle[i].cookie);
1526                         continue;
1527                 }
1528
1529                 res = lock->l_resource;
1530                 done++;
1531
1532                 if (res != pres) {
1533                         if (pres != NULL) {
1534                                 ldlm_reprocess_all(pres);
1535                                 LDLM_RESOURCE_DELREF(pres);
1536                                 ldlm_resource_putref(pres);
1537                         }
1538                         if (res != NULL) {
1539                                 ldlm_resource_getref(res);
1540                                 LDLM_RESOURCE_ADDREF(res);
1541                                 ldlm_res_lvbo_update(res, NULL, 1);
1542                         }
1543                         pres = res;
1544                 }
1545                 ldlm_lock_cancel(lock);
1546                 LDLM_LOCK_PUT(lock);
1547         }
1548         if (pres != NULL) {
1549                 ldlm_reprocess_all(pres);
1550                 LDLM_RESOURCE_DELREF(pres);
1551                 ldlm_resource_putref(pres);
1552         }
1553         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1554         RETURN(done);
1555 }
1556
1557 int ldlm_handle_cancel(struct ptlrpc_request *req)
1558 {
1559         struct ldlm_request *dlm_req;
1560         int rc;
1561         ENTRY;
1562
1563         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1564         if (dlm_req == NULL) {
1565                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1566                 RETURN(-EFAULT);
1567         }
1568
1569         if (req->rq_export && req->rq_export->exp_nid_stats &&
1570             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1571                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1572                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1573
1574         rc = req_capsule_server_pack(&req->rq_pill);
1575         if (rc)
1576                 RETURN(rc);
1577
1578         if (!ldlm_request_cancel(req, dlm_req, 0))
1579                 req->rq_status = ESTALE;
1580
1581         RETURN(ptlrpc_reply(req));
1582 }
1583 #endif /* HAVE_SERVER_SUPPORT */
1584
1585 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1586                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1587 {
1588         int do_ast;
1589         ENTRY;
1590
1591         LDLM_DEBUG(lock, "client blocking AST callback handler");
1592
1593         lock_res_and_lock(lock);
1594         lock->l_flags |= LDLM_FL_CBPENDING;
1595
1596         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1597                 lock->l_flags |= LDLM_FL_CANCEL;
1598
1599         do_ast = (!lock->l_readers && !lock->l_writers);
1600         unlock_res_and_lock(lock);
1601
1602         if (do_ast) {
1603                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1604                        lock, lock->l_blocking_ast);
1605                 if (lock->l_blocking_ast != NULL)
1606                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1607                                              LDLM_CB_BLOCKING);
1608         } else {
1609                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1610                        lock);
1611         }
1612
1613         LDLM_DEBUG(lock, "client blocking callback handler END");
1614         LDLM_LOCK_RELEASE(lock);
1615         EXIT;
1616 }
1617
1618 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1619                                     struct ldlm_namespace *ns,
1620                                     struct ldlm_request *dlm_req,
1621                                     struct ldlm_lock *lock)
1622 {
1623         int lvb_len;
1624         CFS_LIST_HEAD(ast_list);
1625         ENTRY;
1626
1627         LDLM_DEBUG(lock, "client completion callback handler START");
1628
1629         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1630                 int to = cfs_time_seconds(1);
1631                 while (to > 0) {
1632                         cfs_schedule_timeout_and_set_state(
1633                                 CFS_TASK_INTERRUPTIBLE, to);
1634                         if (lock->l_granted_mode == lock->l_req_mode ||
1635                             lock->l_destroyed)
1636                                 break;
1637                 }
1638         }
1639
1640         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1641         if (lvb_len > 0) {
1642                 if (lock->l_lvb_len > 0) {
1643                         /* for extent lock, lvb contains ost_lvb{}. */
1644                         LASSERT(lock->l_lvb_data != NULL);
1645                         LASSERTF(lock->l_lvb_len == lvb_len,
1646                                 "preallocated %d, actual %d.\n",
1647                                 lock->l_lvb_len, lvb_len);
1648                 } else { /* for layout lock, lvb has variable length */
1649                         void *lvb_data;
1650
1651                         OBD_ALLOC(lvb_data, lvb_len);
1652                         if (lvb_data == NULL)
1653                                 LDLM_ERROR(lock, "no memory.\n");
1654
1655                         lock_res_and_lock(lock);
1656                         if (lvb_data == NULL) {
1657                                 lock->l_flags |= LDLM_FL_FAILED;
1658                         } else {
1659                                 LASSERT(lock->l_lvb_data == NULL);
1660                                 lock->l_lvb_data = lvb_data;
1661                                 lock->l_lvb_len = lvb_len;
1662                         }
1663                         unlock_res_and_lock(lock);
1664                 }
1665         }
1666
1667         lock_res_and_lock(lock);
1668         if (lock->l_destroyed ||
1669             lock->l_granted_mode == lock->l_req_mode) {
1670                 /* bug 11300: the lock has already been granted */
1671                 unlock_res_and_lock(lock);
1672                 LDLM_DEBUG(lock, "Double grant race happened");
1673                 LDLM_LOCK_RELEASE(lock);
1674                 EXIT;
1675                 return;
1676         }
1677
1678         /* If we receive the completion AST before the actual enqueue returned,
1679          * then we might need to switch lock modes, resources, or extents. */
1680         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1681                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1682                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1683         }
1684
1685         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1686                 ldlm_convert_policy_to_local(req->rq_export,
1687                                           dlm_req->lock_desc.l_resource.lr_type,
1688                                           &dlm_req->lock_desc.l_policy_data,
1689                                           &lock->l_policy_data);
1690                 LDLM_DEBUG(lock, "completion AST, new policy data");
1691         }
1692
1693         ldlm_resource_unlink_lock(lock);
1694         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1695                    &lock->l_resource->lr_name,
1696                    sizeof(lock->l_resource->lr_name)) != 0) {
1697                 unlock_res_and_lock(lock);
1698                 if (ldlm_lock_change_resource(ns, lock,
1699                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1700                         LDLM_ERROR(lock, "Failed to allocate resource");
1701                         LDLM_LOCK_RELEASE(lock);
1702                         EXIT;
1703                         return;
1704                 }
1705                 LDLM_DEBUG(lock, "completion AST, new resource");
1706                 CERROR("change resource!\n");
1707                 lock_res_and_lock(lock);
1708         }
1709
1710         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1711                 /* BL_AST locks are not needed in lru.
1712                  * let ldlm_cancel_lru() be fast. */
1713                 ldlm_lock_remove_from_lru(lock);
1714                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1715                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1716         }
1717
1718         if (lock->l_lvb_len) {
1719                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1720                                          RCL_CLIENT) < lock->l_lvb_len) {
1721                         LDLM_ERROR(lock, "completion AST did not contain "
1722                                    "expected LVB!");
1723                 } else {
1724                         void *lvb = req_capsule_client_get(&req->rq_pill,
1725                                                            &RMF_DLM_LVB);
1726                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1727                 }
1728         }
1729
1730         ldlm_grant_lock(lock, &ast_list);
1731         unlock_res_and_lock(lock);
1732
1733         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1734
1735         /* Let Enqueue to call osc_lock_upcall() and initialize
1736          * l_ast_data */
1737         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1738
1739         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1740
1741         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1742                           lock);
1743         LDLM_LOCK_RELEASE(lock);
1744         EXIT;
1745 }
1746
1747 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1748                                     struct ldlm_namespace *ns,
1749                                     struct ldlm_request *dlm_req,
1750                                     struct ldlm_lock *lock)
1751 {
1752         int rc = -ENOSYS;
1753         ENTRY;
1754
1755         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1756
1757         if (lock->l_glimpse_ast != NULL)
1758                 rc = lock->l_glimpse_ast(lock, req);
1759
1760         if (req->rq_repmsg != NULL) {
1761                 ptlrpc_reply(req);
1762         } else {
1763                 req->rq_status = rc;
1764                 ptlrpc_error(req);
1765         }
1766
1767         lock_res_and_lock(lock);
1768         if (lock->l_granted_mode == LCK_PW &&
1769             !lock->l_readers && !lock->l_writers &&
1770             cfs_time_after(cfs_time_current(),
1771                            cfs_time_add(lock->l_last_used,
1772                                         cfs_time_seconds(10)))) {
1773                 unlock_res_and_lock(lock);
1774                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1775                         ldlm_handle_bl_callback(ns, NULL, lock);
1776
1777                 EXIT;
1778                 return;
1779         }
1780         unlock_res_and_lock(lock);
1781         LDLM_LOCK_RELEASE(lock);
1782         EXIT;
1783 }
1784
1785 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1786 {
1787         if (req->rq_no_reply)
1788                 return 0;
1789
1790         req->rq_status = rc;
1791         if (!req->rq_packed_final) {
1792                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1793                 if (rc)
1794                         return rc;
1795         }
1796         return ptlrpc_reply(req);
1797 }
1798
1799 #ifdef __KERNEL__
1800 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1801 {
1802         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1803         ENTRY;
1804
1805         cfs_spin_lock(&blp->blp_lock);
1806         if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1807                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1808                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1809         } else {
1810                 /* other blocking callbacks are added to the regular list */
1811                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1812         }
1813         cfs_spin_unlock(&blp->blp_lock);
1814
1815         cfs_waitq_signal(&blp->blp_waitq);
1816
1817         /* can not use blwi->blwi_mode as blwi could be already freed in
1818            LDLM_ASYNC mode */
1819         if (mode == LDLM_SYNC)
1820                 cfs_wait_for_completion(&blwi->blwi_comp);
1821
1822         RETURN(0);
1823 }
1824
1825 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1826                              struct ldlm_namespace *ns,
1827                              struct ldlm_lock_desc *ld,
1828                              cfs_list_t *cancels, int count,
1829                              struct ldlm_lock *lock,
1830                              int mode)
1831 {
1832         cfs_init_completion(&blwi->blwi_comp);
1833         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1834
1835         if (cfs_memory_pressure_get())
1836                 blwi->blwi_mem_pressure = 1;
1837
1838         blwi->blwi_ns = ns;
1839         blwi->blwi_mode = mode;
1840         if (ld != NULL)
1841                 blwi->blwi_ld = *ld;
1842         if (count) {
1843                 cfs_list_add(&blwi->blwi_head, cancels);
1844                 cfs_list_del_init(cancels);
1845                 blwi->blwi_count = count;
1846         } else {
1847                 blwi->blwi_lock = lock;
1848         }
1849 }
1850
1851 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1852                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1853                              cfs_list_t *cancels, int count, int mode)
1854 {
1855         ENTRY;
1856
1857         if (cancels && count == 0)
1858                 RETURN(0);
1859
1860         if (mode == LDLM_SYNC) {
1861                 /* if it is synchronous call do minimum mem alloc, as it could
1862                  * be triggered from kernel shrinker
1863                  */
1864                 struct ldlm_bl_work_item blwi;
1865                 memset(&blwi, 0, sizeof(blwi));
1866                 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1867                 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1868         } else {
1869                 struct ldlm_bl_work_item *blwi;
1870                 OBD_ALLOC(blwi, sizeof(*blwi));
1871                 if (blwi == NULL)
1872                         RETURN(-ENOMEM);
1873                 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1874
1875                 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1876         }
1877 }
1878
1879 #endif
1880
1881 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1882                            struct ldlm_lock *lock)
1883 {
1884 #ifdef __KERNEL__
1885         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1886 #else
1887         RETURN(-ENOSYS);
1888 #endif
1889 }
1890
1891 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1892                            cfs_list_t *cancels, int count, int mode)
1893 {
1894 #ifdef __KERNEL__
1895         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1896 #else
1897         RETURN(-ENOSYS);
1898 #endif
1899 }
1900
1901 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1902 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1903 {
1904         struct obd_device *obd = req->rq_export->exp_obd;
1905         char *key;
1906         void *val;
1907         int keylen, vallen;
1908         int rc = -ENOSYS;
1909         ENTRY;
1910
1911         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1912
1913         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1914
1915         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1916         if (key == NULL) {
1917                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1918                 RETURN(-EFAULT);
1919         }
1920         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1921                                       RCL_CLIENT);
1922         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1923         if (val == NULL) {
1924                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1925                 RETURN(-EFAULT);
1926         }
1927         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1928                                       RCL_CLIENT);
1929
1930         /* We are responsible for swabbing contents of val */
1931
1932         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1933                 /* Pass it on to mdc (the "export" in this case) */
1934                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
1935                                         req->rq_export,
1936                                         sizeof(KEY_HSM_COPYTOOL_SEND),
1937                                         KEY_HSM_COPYTOOL_SEND,
1938                                         vallen, val, NULL);
1939         else
1940                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1941
1942         return rc;
1943 }
1944
1945 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1946                                         const char *msg, int rc,
1947                                         struct lustre_handle *handle)
1948 {
1949         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1950                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
1951                   msg, libcfs_id2str(req->rq_peer), rc,
1952                   handle ? handle->cookie : 0);
1953         if (req->rq_no_reply)
1954                 CWARN("No reply was sent, maybe cause bug 21636.\n");
1955         else if (rc)
1956                 CWARN("Send reply failed, maybe cause bug 21636.\n");
1957 }
1958
1959 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1960 static int ldlm_callback_handler(struct ptlrpc_request *req)
1961 {
1962         struct ldlm_namespace *ns;
1963         struct ldlm_request *dlm_req;
1964         struct ldlm_lock *lock;
1965         int rc;
1966         ENTRY;
1967
1968         /* Requests arrive in sender's byte order.  The ptlrpc service
1969          * handler has already checked and, if necessary, byte-swapped the
1970          * incoming request message body, but I am responsible for the
1971          * message buffers. */
1972
1973         /* do nothing for sec context finalize */
1974         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1975                 RETURN(0);
1976
1977         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1978
1979         if (req->rq_export == NULL) {
1980                 rc = ldlm_callback_reply(req, -ENOTCONN);
1981                 ldlm_callback_errmsg(req, "Operate on unconnected server",
1982                                      rc, NULL);
1983                 RETURN(0);
1984         }
1985
1986         LASSERT(req->rq_export != NULL);
1987         LASSERT(req->rq_export->exp_obd != NULL);
1988
1989         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1990         case LDLM_BL_CALLBACK:
1991                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK_NET))
1992                         RETURN(0);
1993                 break;
1994         case LDLM_CP_CALLBACK:
1995                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK_NET))
1996                         RETURN(0);
1997                 break;
1998         case LDLM_GL_CALLBACK:
1999                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK_NET))
2000                         RETURN(0);
2001                 break;
2002         case LDLM_SET_INFO:
2003                 rc = ldlm_handle_setinfo(req);
2004                 ldlm_callback_reply(req, rc);
2005                 RETURN(0);
2006         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
2007                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
2008                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2009                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2010                         RETURN(0);
2011                 rc = llog_origin_handle_cancel(req);
2012                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2013                         RETURN(0);
2014                 ldlm_callback_reply(req, rc);
2015                 RETURN(0);
2016         case LLOG_ORIGIN_HANDLE_CREATE:
2017                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
2018                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2019                         RETURN(0);
2020                 rc = llog_origin_handle_open(req);
2021                 ldlm_callback_reply(req, rc);
2022                 RETURN(0);
2023         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2024                 req_capsule_set(&req->rq_pill,
2025                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2026                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2027                         RETURN(0);
2028                 rc = llog_origin_handle_next_block(req);
2029                 ldlm_callback_reply(req, rc);
2030                 RETURN(0);
2031         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2032                 req_capsule_set(&req->rq_pill,
2033                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2034                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2035                         RETURN(0);
2036                 rc = llog_origin_handle_read_header(req);
2037                 ldlm_callback_reply(req, rc);
2038                 RETURN(0);
2039         case LLOG_ORIGIN_HANDLE_CLOSE:
2040                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2041                         RETURN(0);
2042                 rc = llog_origin_handle_close(req);
2043                 ldlm_callback_reply(req, rc);
2044                 RETURN(0);
2045         default:
2046                 CERROR("unknown opcode %u\n",
2047                        lustre_msg_get_opc(req->rq_reqmsg));
2048                 ldlm_callback_reply(req, -EPROTO);
2049                 RETURN(0);
2050         }
2051
2052         ns = req->rq_export->exp_obd->obd_namespace;
2053         LASSERT(ns != NULL);
2054
2055         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2056
2057         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2058         if (dlm_req == NULL) {
2059                 rc = ldlm_callback_reply(req, -EPROTO);
2060                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2061                                      NULL);
2062                 RETURN(0);
2063         }
2064
2065         /* Force a known safe race, send a cancel to the server for a lock
2066          * which the server has already started a blocking callback on. */
2067         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2068             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2069                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
2070                 if (rc < 0)
2071                         CERROR("ldlm_cli_cancel: %d\n", rc);
2072         }
2073
2074         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2075         if (!lock) {
2076                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2077                        "disappeared\n", dlm_req->lock_handle[0].cookie);
2078                 rc = ldlm_callback_reply(req, -EINVAL);
2079                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2080                                      &dlm_req->lock_handle[0]);
2081                 RETURN(0);
2082         }
2083
2084         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2085             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2086                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2087
2088         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2089         lock_res_and_lock(lock);
2090         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
2091         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2092                 /* If somebody cancels lock and cache is already dropped,
2093                  * or lock is failed before cp_ast received on client,
2094                  * we can tell the server we have no lock. Otherwise, we
2095                  * should send cancel after dropping the cache. */
2096                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2097                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
2098                     (lock->l_flags & LDLM_FL_FAILED)) {
2099                         LDLM_DEBUG(lock, "callback on lock "
2100                                    LPX64" - lock disappeared\n",
2101                                    dlm_req->lock_handle[0].cookie);
2102                         unlock_res_and_lock(lock);
2103                         LDLM_LOCK_RELEASE(lock);
2104                         rc = ldlm_callback_reply(req, -EINVAL);
2105                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2106                                              &dlm_req->lock_handle[0]);
2107                         RETURN(0);
2108                 }
2109                 /* BL_AST locks are not needed in lru.
2110                  * let ldlm_cancel_lru() be fast. */
2111                 ldlm_lock_remove_from_lru(lock);
2112                 lock->l_flags |= LDLM_FL_BL_AST;
2113         }
2114         unlock_res_and_lock(lock);
2115
2116         /* We want the ost thread to get this reply so that it can respond
2117          * to ost requests (write cache writeback) that might be triggered
2118          * in the callback.
2119          *
2120          * But we'd also like to be able to indicate in the reply that we're
2121          * cancelling right now, because it's unused, or have an intent result
2122          * in the reply, so we might have to push the responsibility for sending
2123          * the reply down into the AST handlers, alas. */
2124
2125         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2126         case LDLM_BL_CALLBACK:
2127                 CDEBUG(D_INODE, "blocking ast\n");
2128                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2129                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2130                         rc = ldlm_callback_reply(req, 0);
2131                         if (req->rq_no_reply || rc)
2132                                 ldlm_callback_errmsg(req, "Normal process", rc,
2133                                                      &dlm_req->lock_handle[0]);
2134                 }
2135                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2136                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2137                 break;
2138         case LDLM_CP_CALLBACK:
2139                 CDEBUG(D_INODE, "completion ast\n");
2140                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2141                 ldlm_callback_reply(req, 0);
2142                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2143                 break;
2144         case LDLM_GL_CALLBACK:
2145                 CDEBUG(D_INODE, "glimpse ast\n");
2146                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2147                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2148                 break;
2149         default:
2150                 LBUG();                         /* checked above */
2151         }
2152
2153         RETURN(0);
2154 }
2155
2156 #ifdef HAVE_SERVER_SUPPORT
2157 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2158 {
2159         int rc;
2160         ENTRY;
2161
2162         /* Requests arrive in sender's byte order.  The ptlrpc service
2163          * handler has already checked and, if necessary, byte-swapped the
2164          * incoming request message body, but I am responsible for the
2165          * message buffers. */
2166
2167         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2168
2169         if (req->rq_export == NULL) {
2170                 struct ldlm_request *dlm_req;
2171
2172                 CERROR("%s from %s arrived at %lu with bad export cookie "
2173                        LPU64"\n",
2174                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2175                        libcfs_nid2str(req->rq_peer.nid),
2176                        req->rq_arrival_time.tv_sec,
2177                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2178
2179                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2180                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2181                         dlm_req = req_capsule_client_get(&req->rq_pill,
2182                                                          &RMF_DLM_REQ);
2183                         if (dlm_req != NULL)
2184                                 ldlm_lock_dump_handle(D_ERROR,
2185                                                       &dlm_req->lock_handle[0]);
2186                 }
2187                 ldlm_callback_reply(req, -ENOTCONN);
2188                 RETURN(0);
2189         }
2190
2191         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2192
2193         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2194         case LDLM_CANCEL:
2195                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2196                 CDEBUG(D_INODE, "cancel\n");
2197                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET))
2198                         RETURN(0);
2199                 rc = ldlm_handle_cancel(req);
2200                 if (rc)
2201                         break;
2202                 RETURN(0);
2203         case OBD_LOG_CANCEL:
2204                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2205                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2206                         RETURN(0);
2207                 rc = llog_origin_handle_cancel(req);
2208                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2209                         RETURN(0);
2210                 ldlm_callback_reply(req, rc);
2211                 RETURN(0);
2212         default:
2213                 CERROR("invalid opcode %d\n",
2214                        lustre_msg_get_opc(req->rq_reqmsg));
2215                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2216                 ldlm_callback_reply(req, -EINVAL);
2217         }
2218
2219         RETURN(0);
2220 }
2221
2222 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2223                                         struct ldlm_lock *lock)
2224 {
2225         struct ldlm_request *dlm_req;
2226         struct lustre_handle lockh;
2227         int rc = 0;
2228         int i;
2229         ENTRY;
2230
2231         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2232         if (dlm_req == NULL)
2233                 RETURN(0);
2234
2235         ldlm_lock2handle(lock, &lockh);
2236         for (i = 0; i < dlm_req->lock_count; i++) {
2237                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2238                                         &lockh)) {
2239                         DEBUG_REQ(D_RPCTRACE, req,
2240                                   "Prio raised by lock "LPX64".", lockh.cookie);
2241
2242                         rc = 1;
2243                         break;
2244                 }
2245         }
2246
2247         RETURN(rc);
2248
2249 }
2250
2251 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2252 {
2253         struct ldlm_request *dlm_req;
2254         int rc = 0;
2255         int i;
2256         ENTRY;
2257
2258         /* no prolong in recovery */
2259         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2260                 RETURN(0);
2261
2262         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2263         if (dlm_req == NULL)
2264                 RETURN(-EFAULT);
2265
2266         for (i = 0; i < dlm_req->lock_count; i++) {
2267                 struct ldlm_lock *lock;
2268
2269                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2270                 if (lock == NULL)
2271                         continue;
2272
2273                 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2274                 if (rc)
2275                         LDLM_DEBUG(lock, "hpreq cancel lock");
2276                 LDLM_LOCK_PUT(lock);
2277
2278                 if (rc)
2279                         break;
2280         }
2281
2282         RETURN(rc);
2283 }
2284
2285 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2286         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2287         .hpreq_check      = ldlm_cancel_hpreq_check,
2288         .hpreq_fini       = NULL,
2289 };
2290
2291 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2292 {
2293         ENTRY;
2294
2295         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2296
2297         if (req->rq_export == NULL)
2298                 RETURN(0);
2299
2300         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2301                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2302                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2303         }
2304         RETURN(0);
2305 }
2306
2307 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2308                         cfs_hlist_node_t *hnode, void *data)
2309
2310 {
2311         cfs_list_t         *rpc_list = data;
2312         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2313
2314         lock_res_and_lock(lock);
2315
2316         if (lock->l_req_mode != lock->l_granted_mode) {
2317                 unlock_res_and_lock(lock);
2318                 return 0;
2319         }
2320
2321         LASSERT(lock->l_resource);
2322         if (lock->l_resource->lr_type != LDLM_IBITS &&
2323             lock->l_resource->lr_type != LDLM_PLAIN) {
2324                 unlock_res_and_lock(lock);
2325                 return 0;
2326         }
2327
2328         if (lock->l_flags & LDLM_FL_AST_SENT) {
2329                 unlock_res_and_lock(lock);
2330                 return 0;
2331         }
2332
2333         LASSERT(lock->l_blocking_ast);
2334         LASSERT(!lock->l_blocking_lock);
2335
2336         lock->l_flags |= LDLM_FL_AST_SENT;
2337         if (lock->l_export && lock->l_export->exp_lock_hash) {
2338                 /* NB: it's safe to call cfs_hash_del() even lock isn't
2339                  * in exp_lock_hash. */
2340                 /* In the function below, .hs_keycmp resolves to
2341                  * ldlm_export_lock_keycmp() */
2342                 /* coverity[overrun-buffer-val] */
2343                 cfs_hash_del(lock->l_export->exp_lock_hash,
2344                              &lock->l_remote_handle, &lock->l_exp_hash);
2345         }
2346
2347         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2348         LDLM_LOCK_GET(lock);
2349
2350         unlock_res_and_lock(lock);
2351         return 0;
2352 }
2353
2354 void ldlm_revoke_export_locks(struct obd_export *exp)
2355 {
2356         cfs_list_t  rpc_list;
2357         ENTRY;
2358
2359         CFS_INIT_LIST_HEAD(&rpc_list);
2360         cfs_hash_for_each_empty(exp->exp_lock_hash,
2361                                 ldlm_revoke_lock_cb, &rpc_list);
2362         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2363                           LDLM_WORK_REVOKE_AST);
2364
2365         EXIT;
2366 }
2367 #endif /* HAVE_SERVER_SUPPORT */
2368
2369 #ifdef __KERNEL__
2370 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2371 {
2372         struct ldlm_bl_work_item *blwi = NULL;
2373         static unsigned int num_bl = 0;
2374
2375         cfs_spin_lock(&blp->blp_lock);
2376         /* process a request from the blp_list at least every blp_num_threads */
2377         if (!cfs_list_empty(&blp->blp_list) &&
2378             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2379                 blwi = cfs_list_entry(blp->blp_list.next,
2380                                       struct ldlm_bl_work_item, blwi_entry);
2381         else
2382                 if (!cfs_list_empty(&blp->blp_prio_list))
2383                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2384                                               struct ldlm_bl_work_item,
2385                                               blwi_entry);
2386
2387         if (blwi) {
2388                 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2389                         num_bl = 0;
2390                 cfs_list_del(&blwi->blwi_entry);
2391         }
2392         cfs_spin_unlock(&blp->blp_lock);
2393
2394         return blwi;
2395 }
2396
2397 /* This only contains temporary data until the thread starts */
2398 struct ldlm_bl_thread_data {
2399         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2400         struct ldlm_bl_pool     *bltd_blp;
2401         cfs_completion_t        bltd_comp;
2402         int                     bltd_num;
2403 };
2404
2405 static int ldlm_bl_thread_main(void *arg);
2406
2407 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2408 {
2409         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2410         int rc;
2411
2412         cfs_init_completion(&bltd.bltd_comp);
2413         rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2414         if (rc < 0) {
2415                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2416                        cfs_atomic_read(&blp->blp_num_threads), rc);
2417                 return rc;
2418         }
2419         cfs_wait_for_completion(&bltd.bltd_comp);
2420
2421         return 0;
2422 }
2423
2424 static int ldlm_bl_thread_main(void *arg)
2425 {
2426         struct ldlm_bl_pool *blp;
2427         ENTRY;
2428
2429         {
2430                 struct ldlm_bl_thread_data *bltd = arg;
2431
2432                 blp = bltd->bltd_blp;
2433
2434                 bltd->bltd_num =
2435                         cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2436                 cfs_atomic_inc(&blp->blp_busy_threads);
2437
2438                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2439                         "ldlm_bl_%02d", bltd->bltd_num);
2440                 cfs_daemonize(bltd->bltd_name);
2441
2442                 cfs_complete(&bltd->bltd_comp);
2443                 /* cannot use bltd after this, it is only on caller's stack */
2444         }
2445
2446         while (1) {
2447                 struct l_wait_info lwi = { 0 };
2448                 struct ldlm_bl_work_item *blwi = NULL;
2449                 int busy;
2450
2451                 blwi = ldlm_bl_get_work(blp);
2452
2453                 if (blwi == NULL) {
2454                         cfs_atomic_dec(&blp->blp_busy_threads);
2455                         l_wait_event_exclusive(blp->blp_waitq,
2456                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2457                                          &lwi);
2458                         busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2459                 } else {
2460                         busy = cfs_atomic_read(&blp->blp_busy_threads);
2461                 }
2462
2463                 if (blwi->blwi_ns == NULL)
2464                         /* added by ldlm_cleanup() */
2465                         break;
2466
2467                 /* Not fatal if racy and have a few too many threads */
2468                 if (unlikely(busy < blp->blp_max_threads &&
2469                              busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2470                              !blwi->blwi_mem_pressure))
2471                         /* discard the return value, we tried */
2472                         ldlm_bl_thread_start(blp);
2473
2474                 if (blwi->blwi_mem_pressure)
2475                         cfs_memory_pressure_set();
2476
2477                 if (blwi->blwi_count) {
2478                         int count;
2479                         /* The special case when we cancel locks in lru
2480                          * asynchronously, we pass the list of locks here.
2481                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2482                          * canceled locally yet. */
2483                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2484                                                            blwi->blwi_count,
2485                                                            LCF_BL_AST);
2486                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2487                 } else {
2488                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2489                                                 blwi->blwi_lock);
2490                 }
2491                 if (blwi->blwi_mem_pressure)
2492                         cfs_memory_pressure_clr();
2493
2494                 if (blwi->blwi_mode == LDLM_ASYNC)
2495                         OBD_FREE(blwi, sizeof(*blwi));
2496                 else
2497                         cfs_complete(&blwi->blwi_comp);
2498         }
2499
2500         cfs_atomic_dec(&blp->blp_busy_threads);
2501         cfs_atomic_dec(&blp->blp_num_threads);
2502         cfs_complete(&blp->blp_comp);
2503         RETURN(0);
2504 }
2505
2506 #endif
2507
2508 static int ldlm_setup(void);
2509 static int ldlm_cleanup(void);
2510
2511 int ldlm_get_ref(void)
2512 {
2513         int rc = 0;
2514         ENTRY;
2515         cfs_mutex_lock(&ldlm_ref_mutex);
2516         if (++ldlm_refcount == 1) {
2517                 rc = ldlm_setup();
2518                 if (rc)
2519                         ldlm_refcount--;
2520         }
2521         cfs_mutex_unlock(&ldlm_ref_mutex);
2522
2523         RETURN(rc);
2524 }
2525
2526 void ldlm_put_ref(void)
2527 {
2528         ENTRY;
2529         cfs_mutex_lock(&ldlm_ref_mutex);
2530         if (ldlm_refcount == 1) {
2531                 int rc = ldlm_cleanup();
2532                 if (rc)
2533                         CERROR("ldlm_cleanup failed: %d\n", rc);
2534                 else
2535                         ldlm_refcount--;
2536         } else {
2537                 ldlm_refcount--;
2538         }
2539         cfs_mutex_unlock(&ldlm_ref_mutex);
2540
2541         EXIT;
2542 }
2543
2544 /*
2545  * Export handle<->lock hash operations.
2546  */
2547 static unsigned
2548 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2549 {
2550         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2551 }
2552
2553 static void *
2554 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2555 {
2556         struct ldlm_lock *lock;
2557
2558         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2559         return &lock->l_remote_handle;
2560 }
2561
2562 static void
2563 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2564 {
2565         struct ldlm_lock     *lock;
2566
2567         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2568         lock->l_remote_handle = *(struct lustre_handle *)key;
2569 }
2570
2571 static int
2572 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2573 {
2574         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2575 }
2576
2577 static void *
2578 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2579 {
2580         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2581 }
2582
2583 static void
2584 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2585 {
2586         struct ldlm_lock *lock;
2587
2588         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2589         LDLM_LOCK_GET(lock);
2590 }
2591
2592 static void
2593 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2594 {
2595         struct ldlm_lock *lock;
2596
2597         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2598         LDLM_LOCK_RELEASE(lock);
2599 }
2600
2601 static cfs_hash_ops_t ldlm_export_lock_ops = {
2602         .hs_hash        = ldlm_export_lock_hash,
2603         .hs_key         = ldlm_export_lock_key,
2604         .hs_keycmp      = ldlm_export_lock_keycmp,
2605         .hs_keycpy      = ldlm_export_lock_keycpy,
2606         .hs_object      = ldlm_export_lock_object,
2607         .hs_get         = ldlm_export_lock_get,
2608         .hs_put         = ldlm_export_lock_put,
2609         .hs_put_locked  = ldlm_export_lock_put,
2610 };
2611
2612 int ldlm_init_export(struct obd_export *exp)
2613 {
2614         ENTRY;
2615
2616         exp->exp_lock_hash =
2617                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2618                                 HASH_EXP_LOCK_CUR_BITS,
2619                                 HASH_EXP_LOCK_MAX_BITS,
2620                                 HASH_EXP_LOCK_BKT_BITS, 0,
2621                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2622                                 &ldlm_export_lock_ops,
2623                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2624                                 CFS_HASH_NBLK_CHANGE);
2625
2626         if (!exp->exp_lock_hash)
2627                 RETURN(-ENOMEM);
2628
2629         RETURN(0);
2630 }
2631 EXPORT_SYMBOL(ldlm_init_export);
2632
2633 void ldlm_destroy_export(struct obd_export *exp)
2634 {
2635         ENTRY;
2636         cfs_hash_putref(exp->exp_lock_hash);
2637         exp->exp_lock_hash = NULL;
2638
2639         ldlm_destroy_flock_export(exp);
2640         EXIT;
2641 }
2642 EXPORT_SYMBOL(ldlm_destroy_export);
2643
2644 static int ldlm_setup(void)
2645 {
2646         static struct ptlrpc_service_conf       conf;
2647         struct ldlm_bl_pool                     *blp = NULL;
2648         int rc = 0;
2649 #ifdef __KERNEL__
2650         int i;
2651 #endif
2652         ENTRY;
2653
2654         if (ldlm_state != NULL)
2655                 RETURN(-EALREADY);
2656
2657         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2658         if (ldlm_state == NULL)
2659                 RETURN(-ENOMEM);
2660
2661 #ifdef LPROCFS
2662         rc = ldlm_proc_setup();
2663         if (rc != 0)
2664                 GOTO(out, rc);
2665 #endif
2666
2667         memset(&conf, 0, sizeof(conf));
2668         conf = (typeof(conf)) {
2669                 .psc_name               = "ldlm_cbd",
2670                 .psc_watchdog_factor    = 2,
2671                 .psc_buf                = {
2672                         .bc_nbufs               = LDLM_NBUFS,
2673                         .bc_buf_size            = LDLM_BUFSIZE,
2674                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2675                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2676                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
2677                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
2678                 },
2679                 .psc_thr                = {
2680                         .tc_thr_name            = "ldlm_cb",
2681                         .tc_thr_factor          = LDLM_THR_FACTOR,
2682                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2683                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2684                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2685                         .tc_nthrs_user          = ldlm_num_threads,
2686                         .tc_cpu_affinity        = 1,
2687                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
2688                 },
2689                 .psc_cpt                = {
2690                         .cc_pattern             = ldlm_cpts,
2691                 },
2692                 .psc_ops                = {
2693                         .so_req_handler         = ldlm_callback_handler,
2694                 },
2695         };
2696         ldlm_state->ldlm_cb_service = \
2697                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2698         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2699                 CERROR("failed to start service\n");
2700                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2701                 ldlm_state->ldlm_cb_service = NULL;
2702                 GOTO(out, rc);
2703         }
2704
2705 #ifdef HAVE_SERVER_SUPPORT
2706         memset(&conf, 0, sizeof(conf));
2707         conf = (typeof(conf)) {
2708                 .psc_name               = "ldlm_canceld",
2709                 .psc_watchdog_factor    = 6,
2710                 .psc_buf                = {
2711                         .bc_nbufs               = LDLM_NBUFS,
2712                         .bc_buf_size            = LDLM_BUFSIZE,
2713                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2714                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2715                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
2716                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
2717
2718                 },
2719                 .psc_thr                = {
2720                         .tc_thr_name            = "ldlm_cn",
2721                         .tc_thr_factor          = LDLM_THR_FACTOR,
2722                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2723                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2724                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2725                         .tc_nthrs_user          = ldlm_num_threads,
2726                         .tc_cpu_affinity        = 1,
2727                         .tc_ctx_tags            = LCT_MD_THREAD | \
2728                                                   LCT_DT_THREAD | \
2729                                                   LCT_CL_THREAD,
2730                 },
2731                 .psc_cpt                = {
2732                         .cc_pattern             = ldlm_cpts,
2733                 },
2734                 .psc_ops                = {
2735                         .so_req_handler         = ldlm_cancel_handler,
2736                         .so_hpreq_handler       = ldlm_hpreq_handler,
2737                 },
2738         };
2739         ldlm_state->ldlm_cancel_service = \
2740                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2741         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2742                 CERROR("failed to start service\n");
2743                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2744                 ldlm_state->ldlm_cancel_service = NULL;
2745                 GOTO(out, rc);
2746         }
2747 #endif
2748
2749         OBD_ALLOC(blp, sizeof(*blp));
2750         if (blp == NULL)
2751                 GOTO(out, rc = -ENOMEM);
2752         ldlm_state->ldlm_bl_pool = blp;
2753
2754         cfs_spin_lock_init(&blp->blp_lock);
2755         CFS_INIT_LIST_HEAD(&blp->blp_list);
2756         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2757         cfs_waitq_init(&blp->blp_waitq);
2758         cfs_atomic_set(&blp->blp_num_threads, 0);
2759         cfs_atomic_set(&blp->blp_busy_threads, 0);
2760
2761 #ifdef __KERNEL__
2762         if (ldlm_num_threads == 0) {
2763                 blp->blp_min_threads = LDLM_NTHRS_INIT;
2764                 blp->blp_max_threads = LDLM_NTHRS_MAX;
2765         } else {
2766                 blp->blp_min_threads = blp->blp_max_threads = \
2767                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2768                                                          ldlm_num_threads));
2769         }
2770
2771         for (i = 0; i < blp->blp_min_threads; i++) {
2772                 rc = ldlm_bl_thread_start(blp);
2773                 if (rc < 0)
2774                         GOTO(out, rc);
2775         }
2776
2777 # ifdef HAVE_SERVER_SUPPORT
2778         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2779         expired_lock_thread.elt_state = ELT_STOPPED;
2780         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2781
2782         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2783         cfs_spin_lock_init(&waiting_locks_spinlock);
2784         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2785
2786         rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2787         if (rc < 0) {
2788                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2789                 GOTO(out, rc);
2790         }
2791
2792         cfs_wait_event(expired_lock_thread.elt_waitq,
2793                        expired_lock_thread.elt_state == ELT_READY);
2794 # endif /* HAVE_SERVER_SUPPORT */
2795
2796         rc = ldlm_pools_init();
2797         if (rc) {
2798                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2799                 GOTO(out, rc);
2800         }
2801 #endif
2802         RETURN(0);
2803
2804  out:
2805         ldlm_cleanup();
2806         RETURN(rc);
2807 }
2808
2809 static int ldlm_cleanup(void)
2810 {
2811         ENTRY;
2812
2813         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2814             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2815                 CERROR("ldlm still has namespaces; clean these up first.\n");
2816                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2817                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2818                 RETURN(-EBUSY);
2819         }
2820
2821 #ifdef __KERNEL__
2822         ldlm_pools_fini();
2823
2824         if (ldlm_state->ldlm_bl_pool != NULL) {
2825                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2826
2827                 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2828                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2829
2830                         cfs_init_completion(&blp->blp_comp);
2831
2832                         cfs_spin_lock(&blp->blp_lock);
2833                         cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2834                         cfs_waitq_signal(&blp->blp_waitq);
2835                         cfs_spin_unlock(&blp->blp_lock);
2836
2837                         cfs_wait_for_completion(&blp->blp_comp);
2838                 }
2839
2840                 OBD_FREE(blp, sizeof(*blp));
2841         }
2842 #endif /* __KERNEL__ */
2843
2844         if (ldlm_state->ldlm_cb_service != NULL)
2845                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2846 # ifdef HAVE_SERVER_SUPPORT
2847         if (ldlm_state->ldlm_cancel_service != NULL)
2848                 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2849 # endif
2850
2851 #ifdef __KERNEL__
2852         ldlm_proc_cleanup();
2853
2854 # ifdef HAVE_SERVER_SUPPORT
2855         if (expired_lock_thread.elt_state != ELT_STOPPED) {
2856                 expired_lock_thread.elt_state = ELT_TERMINATE;
2857                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2858                 cfs_wait_event(expired_lock_thread.elt_waitq,
2859                                expired_lock_thread.elt_state == ELT_STOPPED);
2860         }
2861 # endif
2862 #endif /* __KERNEL__ */
2863
2864         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2865         ldlm_state = NULL;
2866
2867         RETURN(0);
2868 }
2869
2870 int ldlm_init(void)
2871 {
2872         cfs_mutex_init(&ldlm_ref_mutex);
2873         cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2874         cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2875         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2876                                                sizeof(struct ldlm_resource), 0,
2877                                                CFS_SLAB_HWCACHE_ALIGN);
2878         if (ldlm_resource_slab == NULL)
2879                 return -ENOMEM;
2880
2881         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2882                               sizeof(struct ldlm_lock), 0,
2883                               CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
2884         if (ldlm_lock_slab == NULL) {
2885                 cfs_mem_cache_destroy(ldlm_resource_slab);
2886                 return -ENOMEM;
2887         }
2888
2889         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2890                                         sizeof(struct ldlm_interval),
2891                                         0, CFS_SLAB_HWCACHE_ALIGN);
2892         if (ldlm_interval_slab == NULL) {
2893                 cfs_mem_cache_destroy(ldlm_resource_slab);
2894                 cfs_mem_cache_destroy(ldlm_lock_slab);
2895                 return -ENOMEM;
2896         }
2897 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2898         class_export_dump_hook = ldlm_dump_export_locks;
2899 #endif
2900         return 0;
2901 }
2902
2903 void ldlm_exit(void)
2904 {
2905         int rc;
2906         if (ldlm_refcount)
2907                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2908         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2909         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2910 #ifdef __KERNEL__
2911         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2912          * synchronize_rcu() to wait a grace period elapsed, so that
2913          * ldlm_lock_free() get a chance to be called. */
2914         synchronize_rcu();
2915 #endif
2916         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2917         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2918         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2919         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2920 }
2921
2922 /* ldlm_extent.c */
2923 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2924
2925 /* ldlm_lock.c */
2926 #ifdef HAVE_SERVER_SUPPORT
2927 EXPORT_SYMBOL(ldlm_get_processing_policy);
2928 #endif
2929 EXPORT_SYMBOL(ldlm_lock2desc);
2930 EXPORT_SYMBOL(ldlm_register_intent);
2931 EXPORT_SYMBOL(ldlm_lockname);
2932 EXPORT_SYMBOL(ldlm_typename);
2933 EXPORT_SYMBOL(ldlm_lock2handle);
2934 EXPORT_SYMBOL(__ldlm_handle2lock);
2935 EXPORT_SYMBOL(ldlm_lock_get);
2936 EXPORT_SYMBOL(ldlm_lock_put);
2937 EXPORT_SYMBOL(ldlm_lock_match);
2938 EXPORT_SYMBOL(ldlm_lock_cancel);
2939 EXPORT_SYMBOL(ldlm_lock_addref);
2940 EXPORT_SYMBOL(ldlm_lock_addref_try);
2941 EXPORT_SYMBOL(ldlm_lock_decref);
2942 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2943 EXPORT_SYMBOL(ldlm_lock_change_resource);
2944 EXPORT_SYMBOL(ldlm_it2str);
2945 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2946 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2947 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2948 EXPORT_SYMBOL(ldlm_lock_allow_match);
2949 EXPORT_SYMBOL(ldlm_lock_downgrade);
2950 EXPORT_SYMBOL(ldlm_lock_convert);
2951
2952 /* ldlm_request.c */
2953 EXPORT_SYMBOL(ldlm_completion_ast_async);
2954 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2955 EXPORT_SYMBOL(ldlm_completion_ast);
2956 EXPORT_SYMBOL(ldlm_blocking_ast);
2957 EXPORT_SYMBOL(ldlm_glimpse_ast);
2958 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2959 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2960 EXPORT_SYMBOL(ldlm_prep_elc_req);
2961 EXPORT_SYMBOL(ldlm_cli_convert);
2962 EXPORT_SYMBOL(ldlm_cli_enqueue);
2963 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2964 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2965 EXPORT_SYMBOL(ldlm_cli_cancel);
2966 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2967 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2968 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2969 EXPORT_SYMBOL(ldlm_replay_locks);
2970 EXPORT_SYMBOL(ldlm_resource_foreach);
2971 EXPORT_SYMBOL(ldlm_namespace_foreach);
2972 EXPORT_SYMBOL(ldlm_resource_iterate);
2973 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2974 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2975 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2976
2977 /* ldlm_lockd.c */
2978 #ifdef HAVE_SERVER_SUPPORT
2979 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2980 EXPORT_SYMBOL(ldlm_server_completion_ast);
2981 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2982 EXPORT_SYMBOL(ldlm_glimpse_locks);
2983 EXPORT_SYMBOL(ldlm_handle_enqueue);
2984 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2985 EXPORT_SYMBOL(ldlm_handle_cancel);
2986 EXPORT_SYMBOL(ldlm_request_cancel);
2987 EXPORT_SYMBOL(ldlm_handle_convert);
2988 EXPORT_SYMBOL(ldlm_handle_convert0);
2989 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2990 #endif
2991 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2992 EXPORT_SYMBOL(ldlm_get_ref);
2993 EXPORT_SYMBOL(ldlm_put_ref);
2994 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2995
2996 /* ldlm_resource.c */
2997 EXPORT_SYMBOL(ldlm_namespace_new);
2998 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2999 EXPORT_SYMBOL(ldlm_namespace_free);
3000 EXPORT_SYMBOL(ldlm_namespace_dump);
3001 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
3002 EXPORT_SYMBOL(ldlm_resource_get);
3003 EXPORT_SYMBOL(ldlm_resource_putref);
3004 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
3005
3006 /* ldlm_lib.c */
3007 EXPORT_SYMBOL(client_import_add_conn);
3008 EXPORT_SYMBOL(client_import_del_conn);
3009 EXPORT_SYMBOL(client_obd_setup);
3010 EXPORT_SYMBOL(client_obd_cleanup);
3011 EXPORT_SYMBOL(client_connect_import);
3012 EXPORT_SYMBOL(client_disconnect_export);
3013 EXPORT_SYMBOL(target_send_reply);
3014 EXPORT_SYMBOL(target_pack_pool_reply);
3015
3016 #ifdef HAVE_SERVER_SUPPORT
3017 EXPORT_SYMBOL(server_disconnect_export);
3018 EXPORT_SYMBOL(target_stop_recovery_thread);
3019 EXPORT_SYMBOL(target_handle_connect);
3020 EXPORT_SYMBOL(target_cleanup_recovery);
3021 EXPORT_SYMBOL(target_destroy_export);
3022 EXPORT_SYMBOL(target_cancel_recovery_timer);
3023 EXPORT_SYMBOL(target_queue_recovery_request);
3024 EXPORT_SYMBOL(target_handle_ping);
3025 EXPORT_SYMBOL(target_handle_disconnect);
3026 #endif
3027
3028 /* l_lock.c */
3029 EXPORT_SYMBOL(lock_res_and_lock);
3030 EXPORT_SYMBOL(unlock_res_and_lock);