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 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1085                        struct lprocfs_stats *srv_stats)
1086 {
1087         int lock_type = 0, op = 0;
1088
1089         lock_type = dlm_req->lock_desc.l_resource.lr_type;
1090
1091         switch (lock_type) {
1092         case LDLM_PLAIN:
1093                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1094                 break;
1095         case LDLM_EXTENT:
1096                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1097                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1098                 else
1099                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1100                 break;
1101         case LDLM_FLOCK:
1102                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1103                 break;
1104         case LDLM_IBITS:
1105                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1106                 break;
1107         default:
1108                 op = 0;
1109                 break;
1110         }
1111
1112         if (op)
1113                 lprocfs_counter_incr(srv_stats, op);
1114
1115         return;
1116 }
1117
1118 /*
1119  * Main server-side entry point into LDLM. This is called by ptlrpc service
1120  * threads to carry out client lock enqueueing requests.
1121  */
1122 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1123                          struct ptlrpc_request *req,
1124                          const struct ldlm_request *dlm_req,
1125                          const struct ldlm_callback_suite *cbs)
1126 {
1127         struct ldlm_reply *dlm_rep;
1128         __u32 flags;
1129         ldlm_error_t err = ELDLM_OK;
1130         struct ldlm_lock *lock = NULL;
1131         void *cookie = NULL;
1132         int rc = 0;
1133         ENTRY;
1134
1135         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1136
1137         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1138         flags = dlm_req->lock_flags;
1139
1140         LASSERT(req->rq_export);
1141
1142         if (ptlrpc_req2svc(req)->srv_stats != NULL)
1143                 ldlm_svc_get_eopc(dlm_req, ptlrpc_req2svc(req)->srv_stats);
1144
1145         if (req->rq_export && req->rq_export->exp_nid_stats &&
1146             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1147                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1148                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1149
1150         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1151                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1152                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1153                           dlm_req->lock_desc.l_resource.lr_type);
1154                 GOTO(out, rc = -EFAULT);
1155         }
1156
1157         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1158                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1159                      dlm_req->lock_desc.l_req_mode &
1160                      (dlm_req->lock_desc.l_req_mode-1))) {
1161                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1162                           dlm_req->lock_desc.l_req_mode);
1163                 GOTO(out, rc = -EFAULT);
1164         }
1165
1166         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1167                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1168                              LDLM_PLAIN)) {
1169                         DEBUG_REQ(D_ERROR, req,
1170                                   "PLAIN lock request from IBITS client?");
1171                         GOTO(out, rc = -EPROTO);
1172                 }
1173         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1174                             LDLM_IBITS)) {
1175                 DEBUG_REQ(D_ERROR, req,
1176                           "IBITS lock request from unaware client?");
1177                 GOTO(out, rc = -EPROTO);
1178         }
1179
1180 #if 0
1181         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1182            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1183            ibits for mgc/mgs) */
1184
1185         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1186          * inodebits lock if client does not support them. */
1187         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1188             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1189                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1190                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1191                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1192                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1193                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1194         }
1195 #endif
1196
1197         if (unlikely(flags & LDLM_FL_REPLAY)) {
1198                 /* Find an existing lock in the per-export lock hash */
1199                 /* In the function below, .hs_keycmp resolves to
1200                  * ldlm_export_lock_keycmp() */
1201                 /* coverity[overrun-buffer-val] */
1202                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1203                                        (void *)&dlm_req->lock_handle[0]);
1204                 if (lock != NULL) {
1205                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1206                                   LPX64, lock->l_handle.h_cookie);
1207                         GOTO(existing_lock, rc = 0);
1208                 }
1209         }
1210
1211         /* The lock's callback data might be set in the policy function */
1212         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1213                                 dlm_req->lock_desc.l_resource.lr_type,
1214                                 dlm_req->lock_desc.l_req_mode,
1215                                 cbs, NULL, 0);
1216
1217         if (!lock)
1218                 GOTO(out, rc = -ENOMEM);
1219
1220         lock->l_last_activity = cfs_time_current_sec();
1221         lock->l_remote_handle = dlm_req->lock_handle[0];
1222         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1223
1224         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1225         /* Don't enqueue a lock onto the export if it is been disonnected
1226          * due to eviction (bug 3822) or server umount (bug 24324).
1227          * Cancel it now instead. */
1228         if (req->rq_export->exp_disconnected) {
1229                 LDLM_ERROR(lock, "lock on disconnected export %p",
1230                            req->rq_export);
1231                 GOTO(out, rc = -ENOTCONN);
1232         }
1233
1234         lock->l_export = class_export_lock_get(req->rq_export, lock);
1235         if (lock->l_export->exp_lock_hash)
1236                 cfs_hash_add(lock->l_export->exp_lock_hash,
1237                              &lock->l_remote_handle,
1238                              &lock->l_exp_hash);
1239
1240 existing_lock:
1241
1242         if (flags & LDLM_FL_HAS_INTENT) {
1243                 /* In this case, the reply buffer is allocated deep in
1244                  * local_lock_enqueue by the policy function. */
1245                 cookie = req;
1246         } else {
1247                 /* based on the assumption that lvb size never changes during
1248                  * resource life time otherwise it need resource->lr_lock's
1249                  * protection */
1250                 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1251                                      RCL_SERVER, ldlm_lvbo_size(lock));
1252
1253                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1254                         GOTO(out, rc = -ENOMEM);
1255
1256                 rc = req_capsule_server_pack(&req->rq_pill);
1257                 if (rc)
1258                         GOTO(out, rc);
1259         }
1260
1261         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1262                 ldlm_convert_policy_to_local(req->rq_export,
1263                                           dlm_req->lock_desc.l_resource.lr_type,
1264                                           &dlm_req->lock_desc.l_policy_data,
1265                                           &lock->l_policy_data);
1266         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1267                 lock->l_req_extent = lock->l_policy_data.l_extent;
1268
1269         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1270         if (err)
1271                 GOTO(out, err);
1272
1273         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1274         dlm_rep->lock_flags = flags;
1275
1276         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1277         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1278
1279         /* We never send a blocking AST until the lock is granted, but
1280          * we can tell it right now */
1281         lock_res_and_lock(lock);
1282
1283         /* Now take into account flags to be inherited from original lock
1284            request both in reply to client and in our own lock flags. */
1285         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1286         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1287
1288         /* Don't move a pending lock onto the export if it has already been
1289          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1290          * Cancel it now instead. */
1291         if (unlikely(req->rq_export->exp_disconnected ||
1292                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1293                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1294                 rc = -ENOTCONN;
1295         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1296                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1297                 if (lock->l_granted_mode == lock->l_req_mode) {
1298                         /*
1299                          * Only cancel lock if it was granted, because it would
1300                          * be destroyed immediately and would never be granted
1301                          * in the future, causing timeouts on client.  Not
1302                          * granted lock will be cancelled immediately after
1303                          * sending completion AST.
1304                          */
1305                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1306                                 unlock_res_and_lock(lock);
1307                                 ldlm_lock_cancel(lock);
1308                                 lock_res_and_lock(lock);
1309                         } else
1310                                 ldlm_add_waiting_lock(lock);
1311                 }
1312         }
1313         /* Make sure we never ever grant usual metadata locks to liblustre
1314            clients */
1315         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1316             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1317              req->rq_export->exp_libclient) {
1318                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1319                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1320                         CERROR("Granting sync lock to libclient. "
1321                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1322                                dlm_req->lock_flags, dlm_rep->lock_flags,
1323                                lock->l_flags);
1324                         LDLM_ERROR(lock, "sync lock");
1325                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1326                                 struct ldlm_intent *it;
1327
1328                                 it = req_capsule_client_get(&req->rq_pill,
1329                                                             &RMF_LDLM_INTENT);
1330                                 if (it != NULL) {
1331                                         CERROR("This is intent %s ("LPU64")\n",
1332                                                ldlm_it2str(it->opc), it->opc);
1333                                 }
1334                         }
1335                 }
1336         }
1337
1338         unlock_res_and_lock(lock);
1339
1340         EXIT;
1341  out:
1342         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1343         if (!req->rq_packed_final) {
1344                 err = lustre_pack_reply(req, 1, NULL, NULL);
1345                 if (rc == 0)
1346                         rc = err;
1347         }
1348
1349         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1350          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1351         if (lock) {
1352                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1353                            "(err=%d, rc=%d)", err, rc);
1354
1355                 if (rc == 0) {
1356                         int lvb_len = ldlm_lvbo_size(lock);
1357
1358                         if (lvb_len > 0) {
1359                                 void *buf;
1360                                 int buflen;
1361
1362                                 buf = req_capsule_server_get(&req->rq_pill,
1363                                                              &RMF_DLM_LVB);
1364                                 LASSERTF(buf != NULL, "req %p, lock %p\n",
1365                                          req, lock);
1366                                 buflen = req_capsule_get_size(&req->rq_pill,
1367                                                 &RMF_DLM_LVB, RCL_SERVER);
1368                                 buflen = ldlm_lvbo_fill(lock, buf, buflen);
1369                                 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
1370                                                    buflen, RCL_SERVER);
1371                         }
1372                 } else {
1373                         lock_res_and_lock(lock);
1374                         ldlm_resource_unlink_lock(lock);
1375                         ldlm_lock_destroy_nolock(lock);
1376                         unlock_res_and_lock(lock);
1377                 }
1378
1379                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1380                         ldlm_reprocess_all(lock->l_resource);
1381
1382                 LDLM_LOCK_RELEASE(lock);
1383         }
1384
1385         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1386                           lock, rc);
1387
1388         return rc;
1389 }
1390
1391 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1392                         ldlm_completion_callback completion_callback,
1393                         ldlm_blocking_callback blocking_callback,
1394                         ldlm_glimpse_callback glimpse_callback)
1395 {
1396         struct ldlm_request *dlm_req;
1397         struct ldlm_callback_suite cbs = {
1398                 .lcs_completion = completion_callback,
1399                 .lcs_blocking   = blocking_callback,
1400                 .lcs_glimpse    = glimpse_callback
1401         };
1402         int rc;
1403
1404         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1405         if (dlm_req != NULL) {
1406                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1407                                           req, dlm_req, &cbs);
1408         } else {
1409                 rc = -EFAULT;
1410         }
1411         return rc;
1412 }
1413
1414 int ldlm_handle_convert0(struct ptlrpc_request *req,
1415                          const struct ldlm_request *dlm_req)
1416 {
1417         struct ldlm_reply *dlm_rep;
1418         struct ldlm_lock *lock;
1419         int rc;
1420         ENTRY;
1421
1422         if (req->rq_export && req->rq_export->exp_nid_stats &&
1423             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1424                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1425                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1426
1427         rc = req_capsule_server_pack(&req->rq_pill);
1428         if (rc)
1429                 RETURN(rc);
1430
1431         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1432         dlm_rep->lock_flags = dlm_req->lock_flags;
1433
1434         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1435         if (!lock) {
1436                 req->rq_status = EINVAL;
1437         } else {
1438                 void *res = NULL;
1439
1440                 LDLM_DEBUG(lock, "server-side convert handler START");
1441
1442                 lock->l_last_activity = cfs_time_current_sec();
1443                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1444                                         &dlm_rep->lock_flags);
1445                 if (res) {
1446                         if (ldlm_del_waiting_lock(lock))
1447                                 LDLM_DEBUG(lock, "converted waiting lock");
1448                         req->rq_status = 0;
1449                 } else {
1450                         req->rq_status = EDEADLOCK;
1451                 }
1452         }
1453
1454         if (lock) {
1455                 if (!req->rq_status)
1456                         ldlm_reprocess_all(lock->l_resource);
1457                 LDLM_DEBUG(lock, "server-side convert handler END");
1458                 LDLM_LOCK_PUT(lock);
1459         } else
1460                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1461
1462         RETURN(0);
1463 }
1464
1465 int ldlm_handle_convert(struct ptlrpc_request *req)
1466 {
1467         int rc;
1468         struct ldlm_request *dlm_req;
1469
1470         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1471         if (dlm_req != NULL) {
1472                 rc = ldlm_handle_convert0(req, dlm_req);
1473         } else {
1474                 CERROR ("Can't unpack dlm_req\n");
1475                 rc = -EFAULT;
1476         }
1477         return rc;
1478 }
1479
1480 /* Cancel all the locks whos handles are packed into ldlm_request */
1481 int ldlm_request_cancel(struct ptlrpc_request *req,
1482                         const struct ldlm_request *dlm_req, int first)
1483 {
1484         struct ldlm_resource *res, *pres = NULL;
1485         struct ldlm_lock *lock;
1486         int i, count, done = 0;
1487         ENTRY;
1488
1489         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1490         if (first >= count)
1491                 RETURN(0);
1492
1493         /* There is no lock on the server at the replay time,
1494          * skip lock cancelling to make replay tests to pass. */
1495         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1496                 RETURN(0);
1497
1498         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1499                           "starting at %d", count, first);
1500
1501         for (i = first; i < count; i++) {
1502                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1503                 if (!lock) {
1504                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1505                                           "lock (cookie "LPU64")",
1506                                           dlm_req->lock_handle[i].cookie);
1507                         continue;
1508                 }
1509
1510                 res = lock->l_resource;
1511                 done++;
1512
1513                 if (res != pres) {
1514                         if (pres != NULL) {
1515                                 ldlm_reprocess_all(pres);
1516                                 LDLM_RESOURCE_DELREF(pres);
1517                                 ldlm_resource_putref(pres);
1518                         }
1519                         if (res != NULL) {
1520                                 ldlm_resource_getref(res);
1521                                 LDLM_RESOURCE_ADDREF(res);
1522                                 ldlm_res_lvbo_update(res, NULL, 1);
1523                         }
1524                         pres = res;
1525                 }
1526                 ldlm_lock_cancel(lock);
1527                 LDLM_LOCK_PUT(lock);
1528         }
1529         if (pres != NULL) {
1530                 ldlm_reprocess_all(pres);
1531                 LDLM_RESOURCE_DELREF(pres);
1532                 ldlm_resource_putref(pres);
1533         }
1534         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1535         RETURN(done);
1536 }
1537
1538 int ldlm_handle_cancel(struct ptlrpc_request *req)
1539 {
1540         struct ldlm_request *dlm_req;
1541         int rc;
1542         ENTRY;
1543
1544         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1545         if (dlm_req == NULL) {
1546                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1547                 RETURN(-EFAULT);
1548         }
1549
1550         if (req->rq_export && req->rq_export->exp_nid_stats &&
1551             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1552                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1553                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1554
1555         rc = req_capsule_server_pack(&req->rq_pill);
1556         if (rc)
1557                 RETURN(rc);
1558
1559         if (!ldlm_request_cancel(req, dlm_req, 0))
1560                 req->rq_status = ESTALE;
1561
1562         RETURN(ptlrpc_reply(req));
1563 }
1564 #endif /* HAVE_SERVER_SUPPORT */
1565
1566 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1567                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1568 {
1569         int do_ast;
1570         ENTRY;
1571
1572         LDLM_DEBUG(lock, "client blocking AST callback handler");
1573
1574         lock_res_and_lock(lock);
1575         lock->l_flags |= LDLM_FL_CBPENDING;
1576
1577         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1578                 lock->l_flags |= LDLM_FL_CANCEL;
1579
1580         do_ast = (!lock->l_readers && !lock->l_writers);
1581         unlock_res_and_lock(lock);
1582
1583         if (do_ast) {
1584                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1585                        lock, lock->l_blocking_ast);
1586                 if (lock->l_blocking_ast != NULL)
1587                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1588                                              LDLM_CB_BLOCKING);
1589         } else {
1590                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1591                        lock);
1592         }
1593
1594         LDLM_DEBUG(lock, "client blocking callback handler END");
1595         LDLM_LOCK_RELEASE(lock);
1596         EXIT;
1597 }
1598
1599 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1600                                     struct ldlm_namespace *ns,
1601                                     struct ldlm_request *dlm_req,
1602                                     struct ldlm_lock *lock)
1603 {
1604         int lvb_len;
1605         CFS_LIST_HEAD(ast_list);
1606         ENTRY;
1607
1608         LDLM_DEBUG(lock, "client completion callback handler START");
1609
1610         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1611                 int to = cfs_time_seconds(1);
1612                 while (to > 0) {
1613                         cfs_schedule_timeout_and_set_state(
1614                                 CFS_TASK_INTERRUPTIBLE, to);
1615                         if (lock->l_granted_mode == lock->l_req_mode ||
1616                             lock->l_destroyed)
1617                                 break;
1618                 }
1619         }
1620
1621         lvb_len = req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT);
1622         if (lvb_len > 0) {
1623                 if (lock->l_lvb_len > 0) {
1624                         /* for extent lock, lvb contains ost_lvb{}. */
1625                         LASSERT(lock->l_lvb_data != NULL);
1626                         LASSERTF(lock->l_lvb_len == lvb_len,
1627                                 "preallocated %d, actual %d.\n",
1628                                 lock->l_lvb_len, lvb_len);
1629                 } else { /* for layout lock, lvb has variable length */
1630                         void *lvb_data;
1631
1632                         OBD_ALLOC(lvb_data, lvb_len);
1633                         if (lvb_data == NULL)
1634                                 LDLM_ERROR(lock, "no memory.\n");
1635
1636                         lock_res_and_lock(lock);
1637                         if (lvb_data == NULL) {
1638                                 lock->l_flags |= LDLM_FL_FAILED;
1639                         } else {
1640                                 LASSERT(lock->l_lvb_data == NULL);
1641                                 lock->l_lvb_data = lvb_data;
1642                                 lock->l_lvb_len = lvb_len;
1643                         }
1644                         unlock_res_and_lock(lock);
1645                 }
1646         }
1647
1648         lock_res_and_lock(lock);
1649         if (lock->l_destroyed ||
1650             lock->l_granted_mode == lock->l_req_mode) {
1651                 /* bug 11300: the lock has already been granted */
1652                 unlock_res_and_lock(lock);
1653                 LDLM_DEBUG(lock, "Double grant race happened");
1654                 LDLM_LOCK_RELEASE(lock);
1655                 EXIT;
1656                 return;
1657         }
1658
1659         /* If we receive the completion AST before the actual enqueue returned,
1660          * then we might need to switch lock modes, resources, or extents. */
1661         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1662                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1663                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1664         }
1665
1666         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1667                 ldlm_convert_policy_to_local(req->rq_export,
1668                                           dlm_req->lock_desc.l_resource.lr_type,
1669                                           &dlm_req->lock_desc.l_policy_data,
1670                                           &lock->l_policy_data);
1671                 LDLM_DEBUG(lock, "completion AST, new policy data");
1672         }
1673
1674         ldlm_resource_unlink_lock(lock);
1675         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1676                    &lock->l_resource->lr_name,
1677                    sizeof(lock->l_resource->lr_name)) != 0) {
1678                 unlock_res_and_lock(lock);
1679                 if (ldlm_lock_change_resource(ns, lock,
1680                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1681                         LDLM_ERROR(lock, "Failed to allocate resource");
1682                         LDLM_LOCK_RELEASE(lock);
1683                         EXIT;
1684                         return;
1685                 }
1686                 LDLM_DEBUG(lock, "completion AST, new resource");
1687                 CERROR("change resource!\n");
1688                 lock_res_and_lock(lock);
1689         }
1690
1691         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1692                 /* BL_AST locks are not needed in lru.
1693                  * let ldlm_cancel_lru() be fast. */
1694                 ldlm_lock_remove_from_lru(lock);
1695                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1696                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1697         }
1698
1699         if (lock->l_lvb_len) {
1700                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1701                                          RCL_CLIENT) < lock->l_lvb_len) {
1702                         LDLM_ERROR(lock, "completion AST did not contain "
1703                                    "expected LVB!");
1704                 } else {
1705                         void *lvb = req_capsule_client_get(&req->rq_pill,
1706                                                            &RMF_DLM_LVB);
1707                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1708                 }
1709         }
1710
1711         ldlm_grant_lock(lock, &ast_list);
1712         unlock_res_and_lock(lock);
1713
1714         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1715
1716         /* Let Enqueue to call osc_lock_upcall() and initialize
1717          * l_ast_data */
1718         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1719
1720         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1721
1722         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1723                           lock);
1724         LDLM_LOCK_RELEASE(lock);
1725         EXIT;
1726 }
1727
1728 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1729                                     struct ldlm_namespace *ns,
1730                                     struct ldlm_request *dlm_req,
1731                                     struct ldlm_lock *lock)
1732 {
1733         int rc = -ENOSYS;
1734         ENTRY;
1735
1736         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1737
1738         if (lock->l_glimpse_ast != NULL)
1739                 rc = lock->l_glimpse_ast(lock, req);
1740
1741         if (req->rq_repmsg != NULL) {
1742                 ptlrpc_reply(req);
1743         } else {
1744                 req->rq_status = rc;
1745                 ptlrpc_error(req);
1746         }
1747
1748         lock_res_and_lock(lock);
1749         if (lock->l_granted_mode == LCK_PW &&
1750             !lock->l_readers && !lock->l_writers &&
1751             cfs_time_after(cfs_time_current(),
1752                            cfs_time_add(lock->l_last_used,
1753                                         cfs_time_seconds(10)))) {
1754                 unlock_res_and_lock(lock);
1755                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1756                         ldlm_handle_bl_callback(ns, NULL, lock);
1757
1758                 EXIT;
1759                 return;
1760         }
1761         unlock_res_and_lock(lock);
1762         LDLM_LOCK_RELEASE(lock);
1763         EXIT;
1764 }
1765
1766 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1767 {
1768         if (req->rq_no_reply)
1769                 return 0;
1770
1771         req->rq_status = rc;
1772         if (!req->rq_packed_final) {
1773                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1774                 if (rc)
1775                         return rc;
1776         }
1777         return ptlrpc_reply(req);
1778 }
1779
1780 #ifdef __KERNEL__
1781 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1782 {
1783         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1784         ENTRY;
1785
1786         cfs_spin_lock(&blp->blp_lock);
1787         if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1788                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1789                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1790         } else {
1791                 /* other blocking callbacks are added to the regular list */
1792                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1793         }
1794         cfs_spin_unlock(&blp->blp_lock);
1795
1796         cfs_waitq_signal(&blp->blp_waitq);
1797
1798         /* can not use blwi->blwi_mode as blwi could be already freed in
1799            LDLM_ASYNC mode */
1800         if (mode == LDLM_SYNC)
1801                 cfs_wait_for_completion(&blwi->blwi_comp);
1802
1803         RETURN(0);
1804 }
1805
1806 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1807                              struct ldlm_namespace *ns,
1808                              struct ldlm_lock_desc *ld,
1809                              cfs_list_t *cancels, int count,
1810                              struct ldlm_lock *lock,
1811                              int mode)
1812 {
1813         cfs_init_completion(&blwi->blwi_comp);
1814         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1815
1816         if (cfs_memory_pressure_get())
1817                 blwi->blwi_mem_pressure = 1;
1818
1819         blwi->blwi_ns = ns;
1820         blwi->blwi_mode = mode;
1821         if (ld != NULL)
1822                 blwi->blwi_ld = *ld;
1823         if (count) {
1824                 cfs_list_add(&blwi->blwi_head, cancels);
1825                 cfs_list_del_init(cancels);
1826                 blwi->blwi_count = count;
1827         } else {
1828                 blwi->blwi_lock = lock;
1829         }
1830 }
1831
1832 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1833                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1834                              cfs_list_t *cancels, int count, int mode)
1835 {
1836         ENTRY;
1837
1838         if (cancels && count == 0)
1839                 RETURN(0);
1840
1841         if (mode == LDLM_SYNC) {
1842                 /* if it is synchronous call do minimum mem alloc, as it could
1843                  * be triggered from kernel shrinker
1844                  */
1845                 struct ldlm_bl_work_item blwi;
1846                 memset(&blwi, 0, sizeof(blwi));
1847                 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1848                 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1849         } else {
1850                 struct ldlm_bl_work_item *blwi;
1851                 OBD_ALLOC(blwi, sizeof(*blwi));
1852                 if (blwi == NULL)
1853                         RETURN(-ENOMEM);
1854                 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1855
1856                 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1857         }
1858 }
1859
1860 #endif
1861
1862 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1863                            struct ldlm_lock *lock)
1864 {
1865 #ifdef __KERNEL__
1866         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1867 #else
1868         RETURN(-ENOSYS);
1869 #endif
1870 }
1871
1872 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1873                            cfs_list_t *cancels, int count, int mode)
1874 {
1875 #ifdef __KERNEL__
1876         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1877 #else
1878         RETURN(-ENOSYS);
1879 #endif
1880 }
1881
1882 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1883 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1884 {
1885         struct obd_device *obd = req->rq_export->exp_obd;
1886         char *key;
1887         void *val;
1888         int keylen, vallen;
1889         int rc = -ENOSYS;
1890         ENTRY;
1891
1892         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1893
1894         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1895
1896         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1897         if (key == NULL) {
1898                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1899                 RETURN(-EFAULT);
1900         }
1901         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1902                                       RCL_CLIENT);
1903         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1904         if (val == NULL) {
1905                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1906                 RETURN(-EFAULT);
1907         }
1908         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1909                                       RCL_CLIENT);
1910
1911         /* We are responsible for swabbing contents of val */
1912
1913         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1914                 /* Pass it on to mdc (the "export" in this case) */
1915                 rc = obd_set_info_async(req->rq_svc_thread->t_env,
1916                                         req->rq_export,
1917                                         sizeof(KEY_HSM_COPYTOOL_SEND),
1918                                         KEY_HSM_COPYTOOL_SEND,
1919                                         vallen, val, NULL);
1920         else
1921                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1922
1923         return rc;
1924 }
1925
1926 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1927                                         const char *msg, int rc,
1928                                         struct lustre_handle *handle)
1929 {
1930         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1931                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
1932                   msg, libcfs_id2str(req->rq_peer), rc,
1933                   handle ? handle->cookie : 0);
1934         if (req->rq_no_reply)
1935                 CWARN("No reply was sent, maybe cause bug 21636.\n");
1936         else if (rc)
1937                 CWARN("Send reply failed, maybe cause bug 21636.\n");
1938 }
1939
1940 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1941 static int ldlm_callback_handler(struct ptlrpc_request *req)
1942 {
1943         struct ldlm_namespace *ns;
1944         struct ldlm_request *dlm_req;
1945         struct ldlm_lock *lock;
1946         int rc;
1947         ENTRY;
1948
1949         /* Requests arrive in sender's byte order.  The ptlrpc service
1950          * handler has already checked and, if necessary, byte-swapped the
1951          * incoming request message body, but I am responsible for the
1952          * message buffers. */
1953
1954         /* do nothing for sec context finalize */
1955         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1956                 RETURN(0);
1957
1958         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1959
1960         if (req->rq_export == NULL) {
1961                 rc = ldlm_callback_reply(req, -ENOTCONN);
1962                 ldlm_callback_errmsg(req, "Operate on unconnected server",
1963                                      rc, NULL);
1964                 RETURN(0);
1965         }
1966
1967         LASSERT(req->rq_export != NULL);
1968         LASSERT(req->rq_export->exp_obd != NULL);
1969
1970         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1971         case LDLM_BL_CALLBACK:
1972                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1973                         RETURN(0);
1974                 break;
1975         case LDLM_CP_CALLBACK:
1976                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1977                         RETURN(0);
1978                 break;
1979         case LDLM_GL_CALLBACK:
1980                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1981                         RETURN(0);
1982                 break;
1983         case LDLM_SET_INFO:
1984                 rc = ldlm_handle_setinfo(req);
1985                 ldlm_callback_reply(req, rc);
1986                 RETURN(0);
1987         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1988                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1989                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1990                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1991                         RETURN(0);
1992                 rc = llog_origin_handle_cancel(req);
1993                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1994                         RETURN(0);
1995                 ldlm_callback_reply(req, rc);
1996                 RETURN(0);
1997         case LLOG_ORIGIN_HANDLE_CREATE:
1998                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1999                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2000                         RETURN(0);
2001                 rc = llog_origin_handle_open(req);
2002                 ldlm_callback_reply(req, rc);
2003                 RETURN(0);
2004         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2005                 req_capsule_set(&req->rq_pill,
2006                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
2007                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2008                         RETURN(0);
2009                 rc = llog_origin_handle_next_block(req);
2010                 ldlm_callback_reply(req, rc);
2011                 RETURN(0);
2012         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2013                 req_capsule_set(&req->rq_pill,
2014                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
2015                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2016                         RETURN(0);
2017                 rc = llog_origin_handle_read_header(req);
2018                 ldlm_callback_reply(req, rc);
2019                 RETURN(0);
2020         case LLOG_ORIGIN_HANDLE_CLOSE:
2021                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
2022                         RETURN(0);
2023                 rc = llog_origin_handle_close(req);
2024                 ldlm_callback_reply(req, rc);
2025                 RETURN(0);
2026         default:
2027                 CERROR("unknown opcode %u\n",
2028                        lustre_msg_get_opc(req->rq_reqmsg));
2029                 ldlm_callback_reply(req, -EPROTO);
2030                 RETURN(0);
2031         }
2032
2033         ns = req->rq_export->exp_obd->obd_namespace;
2034         LASSERT(ns != NULL);
2035
2036         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2037
2038         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2039         if (dlm_req == NULL) {
2040                 rc = ldlm_callback_reply(req, -EPROTO);
2041                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
2042                                      NULL);
2043                 RETURN(0);
2044         }
2045
2046         /* Force a known safe race, send a cancel to the server for a lock
2047          * which the server has already started a blocking callback on. */
2048         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
2049             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2050                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
2051                 if (rc < 0)
2052                         CERROR("ldlm_cli_cancel: %d\n", rc);
2053         }
2054
2055         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
2056         if (!lock) {
2057                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
2058                        "disappeared\n", dlm_req->lock_handle[0].cookie);
2059                 rc = ldlm_callback_reply(req, -EINVAL);
2060                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
2061                                      &dlm_req->lock_handle[0]);
2062                 RETURN(0);
2063         }
2064
2065         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
2066             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
2067                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
2068
2069         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
2070         lock_res_and_lock(lock);
2071         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
2072         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
2073                 /* If somebody cancels lock and cache is already dropped,
2074                  * or lock is failed before cp_ast received on client,
2075                  * we can tell the server we have no lock. Otherwise, we
2076                  * should send cancel after dropping the cache. */
2077                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
2078                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
2079                     (lock->l_flags & LDLM_FL_FAILED)) {
2080                         LDLM_DEBUG(lock, "callback on lock "
2081                                    LPX64" - lock disappeared\n",
2082                                    dlm_req->lock_handle[0].cookie);
2083                         unlock_res_and_lock(lock);
2084                         LDLM_LOCK_RELEASE(lock);
2085                         rc = ldlm_callback_reply(req, -EINVAL);
2086                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
2087                                              &dlm_req->lock_handle[0]);
2088                         RETURN(0);
2089                 }
2090                 /* BL_AST locks are not needed in lru.
2091                  * let ldlm_cancel_lru() be fast. */
2092                 ldlm_lock_remove_from_lru(lock);
2093                 lock->l_flags |= LDLM_FL_BL_AST;
2094         }
2095         unlock_res_and_lock(lock);
2096
2097         /* We want the ost thread to get this reply so that it can respond
2098          * to ost requests (write cache writeback) that might be triggered
2099          * in the callback.
2100          *
2101          * But we'd also like to be able to indicate in the reply that we're
2102          * cancelling right now, because it's unused, or have an intent result
2103          * in the reply, so we might have to push the responsibility for sending
2104          * the reply down into the AST handlers, alas. */
2105
2106         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2107         case LDLM_BL_CALLBACK:
2108                 CDEBUG(D_INODE, "blocking ast\n");
2109                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2110                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2111                         rc = ldlm_callback_reply(req, 0);
2112                         if (req->rq_no_reply || rc)
2113                                 ldlm_callback_errmsg(req, "Normal process", rc,
2114                                                      &dlm_req->lock_handle[0]);
2115                 }
2116                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2117                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2118                 break;
2119         case LDLM_CP_CALLBACK:
2120                 CDEBUG(D_INODE, "completion ast\n");
2121                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2122                 ldlm_callback_reply(req, 0);
2123                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2124                 break;
2125         case LDLM_GL_CALLBACK:
2126                 CDEBUG(D_INODE, "glimpse ast\n");
2127                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2128                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2129                 break;
2130         default:
2131                 LBUG();                         /* checked above */
2132         }
2133
2134         RETURN(0);
2135 }
2136
2137 #ifdef HAVE_SERVER_SUPPORT
2138 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2139 {
2140         int rc;
2141         ENTRY;
2142
2143         /* Requests arrive in sender's byte order.  The ptlrpc service
2144          * handler has already checked and, if necessary, byte-swapped the
2145          * incoming request message body, but I am responsible for the
2146          * message buffers. */
2147
2148         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2149
2150         if (req->rq_export == NULL) {
2151                 struct ldlm_request *dlm_req;
2152
2153                 CERROR("%s from %s arrived at %lu with bad export cookie "
2154                        LPU64"\n",
2155                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2156                        libcfs_nid2str(req->rq_peer.nid),
2157                        req->rq_arrival_time.tv_sec,
2158                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2159
2160                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2161                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2162                         dlm_req = req_capsule_client_get(&req->rq_pill,
2163                                                          &RMF_DLM_REQ);
2164                         if (dlm_req != NULL)
2165                                 ldlm_lock_dump_handle(D_ERROR,
2166                                                       &dlm_req->lock_handle[0]);
2167                 }
2168                 ldlm_callback_reply(req, -ENOTCONN);
2169                 RETURN(0);
2170         }
2171
2172         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2173
2174         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2175         case LDLM_CANCEL:
2176                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2177                 CDEBUG(D_INODE, "cancel\n");
2178                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2179                         RETURN(0);
2180                 rc = ldlm_handle_cancel(req);
2181                 if (rc)
2182                         break;
2183                 RETURN(0);
2184         case OBD_LOG_CANCEL:
2185                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2186                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2187                         RETURN(0);
2188                 rc = llog_origin_handle_cancel(req);
2189                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2190                         RETURN(0);
2191                 ldlm_callback_reply(req, rc);
2192                 RETURN(0);
2193         default:
2194                 CERROR("invalid opcode %d\n",
2195                        lustre_msg_get_opc(req->rq_reqmsg));
2196                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2197                 ldlm_callback_reply(req, -EINVAL);
2198         }
2199
2200         RETURN(0);
2201 }
2202
2203 static int ldlm_cancel_hpreq_lock_match(struct ptlrpc_request *req,
2204                                         struct ldlm_lock *lock)
2205 {
2206         struct ldlm_request *dlm_req;
2207         struct lustre_handle lockh;
2208         int rc = 0;
2209         int i;
2210         ENTRY;
2211
2212         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2213         if (dlm_req == NULL)
2214                 RETURN(0);
2215
2216         ldlm_lock2handle(lock, &lockh);
2217         for (i = 0; i < dlm_req->lock_count; i++) {
2218                 if (lustre_handle_equal(&dlm_req->lock_handle[i],
2219                                         &lockh)) {
2220                         DEBUG_REQ(D_RPCTRACE, req,
2221                                   "Prio raised by lock "LPX64".", lockh.cookie);
2222
2223                         rc = 1;
2224                         break;
2225                 }
2226         }
2227
2228         RETURN(rc);
2229
2230 }
2231
2232 static int ldlm_cancel_hpreq_check(struct ptlrpc_request *req)
2233 {
2234         struct ldlm_request *dlm_req;
2235         int rc = 0;
2236         int i;
2237         ENTRY;
2238
2239         /* no prolong in recovery */
2240         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
2241                 RETURN(0);
2242
2243         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
2244         if (dlm_req == NULL)
2245                 RETURN(-EFAULT);
2246
2247         for (i = 0; i < dlm_req->lock_count; i++) {
2248                 struct ldlm_lock *lock;
2249
2250                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
2251                 if (lock == NULL)
2252                         continue;
2253
2254                 rc = !!(lock->l_flags & LDLM_FL_AST_SENT);
2255                 if (rc)
2256                         LDLM_DEBUG(lock, "hpreq cancel lock");
2257                 LDLM_LOCK_PUT(lock);
2258
2259                 if (rc)
2260                         break;
2261         }
2262
2263         RETURN(rc);
2264 }
2265
2266 static struct ptlrpc_hpreq_ops ldlm_cancel_hpreq_ops = {
2267         .hpreq_lock_match = ldlm_cancel_hpreq_lock_match,
2268         .hpreq_check      = ldlm_cancel_hpreq_check,
2269         .hpreq_fini       = NULL,
2270 };
2271
2272 static int ldlm_hpreq_handler(struct ptlrpc_request *req)
2273 {
2274         ENTRY;
2275
2276         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2277
2278         if (req->rq_export == NULL)
2279                 RETURN(0);
2280
2281         if (LDLM_CANCEL == lustre_msg_get_opc(req->rq_reqmsg)) {
2282                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2283                 req->rq_ops = &ldlm_cancel_hpreq_ops;
2284         }
2285         RETURN(0);
2286 }
2287
2288 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2289                         cfs_hlist_node_t *hnode, void *data)
2290
2291 {
2292         cfs_list_t         *rpc_list = data;
2293         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2294
2295         lock_res_and_lock(lock);
2296
2297         if (lock->l_req_mode != lock->l_granted_mode) {
2298                 unlock_res_and_lock(lock);
2299                 return 0;
2300         }
2301
2302         LASSERT(lock->l_resource);
2303         if (lock->l_resource->lr_type != LDLM_IBITS &&
2304             lock->l_resource->lr_type != LDLM_PLAIN) {
2305                 unlock_res_and_lock(lock);
2306                 return 0;
2307         }
2308
2309         if (lock->l_flags & LDLM_FL_AST_SENT) {
2310                 unlock_res_and_lock(lock);
2311                 return 0;
2312         }
2313
2314         LASSERT(lock->l_blocking_ast);
2315         LASSERT(!lock->l_blocking_lock);
2316
2317         lock->l_flags |= LDLM_FL_AST_SENT;
2318         if (lock->l_export && lock->l_export->exp_lock_hash) {
2319                 /* NB: it's safe to call cfs_hash_del() even lock isn't
2320                  * in exp_lock_hash. */
2321                 /* In the function below, .hs_keycmp resolves to
2322                  * ldlm_export_lock_keycmp() */
2323                 /* coverity[overrun-buffer-val] */
2324                 cfs_hash_del(lock->l_export->exp_lock_hash,
2325                              &lock->l_remote_handle, &lock->l_exp_hash);
2326         }
2327
2328         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2329         LDLM_LOCK_GET(lock);
2330
2331         unlock_res_and_lock(lock);
2332         return 0;
2333 }
2334
2335 void ldlm_revoke_export_locks(struct obd_export *exp)
2336 {
2337         cfs_list_t  rpc_list;
2338         ENTRY;
2339
2340         CFS_INIT_LIST_HEAD(&rpc_list);
2341         cfs_hash_for_each_empty(exp->exp_lock_hash,
2342                                 ldlm_revoke_lock_cb, &rpc_list);
2343         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2344                           LDLM_WORK_REVOKE_AST);
2345
2346         EXIT;
2347 }
2348 #endif /* HAVE_SERVER_SUPPORT */
2349
2350 #ifdef __KERNEL__
2351 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2352 {
2353         struct ldlm_bl_work_item *blwi = NULL;
2354         static unsigned int num_bl = 0;
2355
2356         cfs_spin_lock(&blp->blp_lock);
2357         /* process a request from the blp_list at least every blp_num_threads */
2358         if (!cfs_list_empty(&blp->blp_list) &&
2359             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2360                 blwi = cfs_list_entry(blp->blp_list.next,
2361                                       struct ldlm_bl_work_item, blwi_entry);
2362         else
2363                 if (!cfs_list_empty(&blp->blp_prio_list))
2364                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2365                                               struct ldlm_bl_work_item,
2366                                               blwi_entry);
2367
2368         if (blwi) {
2369                 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2370                         num_bl = 0;
2371                 cfs_list_del(&blwi->blwi_entry);
2372         }
2373         cfs_spin_unlock(&blp->blp_lock);
2374
2375         return blwi;
2376 }
2377
2378 /* This only contains temporary data until the thread starts */
2379 struct ldlm_bl_thread_data {
2380         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2381         struct ldlm_bl_pool     *bltd_blp;
2382         cfs_completion_t        bltd_comp;
2383         int                     bltd_num;
2384 };
2385
2386 static int ldlm_bl_thread_main(void *arg);
2387
2388 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2389 {
2390         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2391         int rc;
2392
2393         cfs_init_completion(&bltd.bltd_comp);
2394         rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2395         if (rc < 0) {
2396                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2397                        cfs_atomic_read(&blp->blp_num_threads), rc);
2398                 return rc;
2399         }
2400         cfs_wait_for_completion(&bltd.bltd_comp);
2401
2402         return 0;
2403 }
2404
2405 static int ldlm_bl_thread_main(void *arg)
2406 {
2407         struct ldlm_bl_pool *blp;
2408         ENTRY;
2409
2410         {
2411                 struct ldlm_bl_thread_data *bltd = arg;
2412
2413                 blp = bltd->bltd_blp;
2414
2415                 bltd->bltd_num =
2416                         cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2417                 cfs_atomic_inc(&blp->blp_busy_threads);
2418
2419                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2420                         "ldlm_bl_%02d", bltd->bltd_num);
2421                 cfs_daemonize(bltd->bltd_name);
2422
2423                 cfs_complete(&bltd->bltd_comp);
2424                 /* cannot use bltd after this, it is only on caller's stack */
2425         }
2426
2427         while (1) {
2428                 struct l_wait_info lwi = { 0 };
2429                 struct ldlm_bl_work_item *blwi = NULL;
2430                 int busy;
2431
2432                 blwi = ldlm_bl_get_work(blp);
2433
2434                 if (blwi == NULL) {
2435                         cfs_atomic_dec(&blp->blp_busy_threads);
2436                         l_wait_event_exclusive(blp->blp_waitq,
2437                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2438                                          &lwi);
2439                         busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2440                 } else {
2441                         busy = cfs_atomic_read(&blp->blp_busy_threads);
2442                 }
2443
2444                 if (blwi->blwi_ns == NULL)
2445                         /* added by ldlm_cleanup() */
2446                         break;
2447
2448                 /* Not fatal if racy and have a few too many threads */
2449                 if (unlikely(busy < blp->blp_max_threads &&
2450                              busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2451                              !blwi->blwi_mem_pressure))
2452                         /* discard the return value, we tried */
2453                         ldlm_bl_thread_start(blp);
2454
2455                 if (blwi->blwi_mem_pressure)
2456                         cfs_memory_pressure_set();
2457
2458                 if (blwi->blwi_count) {
2459                         int count;
2460                         /* The special case when we cancel locks in lru
2461                          * asynchronously, we pass the list of locks here.
2462                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2463                          * canceled locally yet. */
2464                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2465                                                            blwi->blwi_count,
2466                                                            LCF_BL_AST);
2467                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2468                 } else {
2469                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2470                                                 blwi->blwi_lock);
2471                 }
2472                 if (blwi->blwi_mem_pressure)
2473                         cfs_memory_pressure_clr();
2474
2475                 if (blwi->blwi_mode == LDLM_ASYNC)
2476                         OBD_FREE(blwi, sizeof(*blwi));
2477                 else
2478                         cfs_complete(&blwi->blwi_comp);
2479         }
2480
2481         cfs_atomic_dec(&blp->blp_busy_threads);
2482         cfs_atomic_dec(&blp->blp_num_threads);
2483         cfs_complete(&blp->blp_comp);
2484         RETURN(0);
2485 }
2486
2487 #endif
2488
2489 static int ldlm_setup(void);
2490 static int ldlm_cleanup(void);
2491
2492 int ldlm_get_ref(void)
2493 {
2494         int rc = 0;
2495         ENTRY;
2496         cfs_mutex_lock(&ldlm_ref_mutex);
2497         if (++ldlm_refcount == 1) {
2498                 rc = ldlm_setup();
2499                 if (rc)
2500                         ldlm_refcount--;
2501         }
2502         cfs_mutex_unlock(&ldlm_ref_mutex);
2503
2504         RETURN(rc);
2505 }
2506
2507 void ldlm_put_ref(void)
2508 {
2509         ENTRY;
2510         cfs_mutex_lock(&ldlm_ref_mutex);
2511         if (ldlm_refcount == 1) {
2512                 int rc = ldlm_cleanup();
2513                 if (rc)
2514                         CERROR("ldlm_cleanup failed: %d\n", rc);
2515                 else
2516                         ldlm_refcount--;
2517         } else {
2518                 ldlm_refcount--;
2519         }
2520         cfs_mutex_unlock(&ldlm_ref_mutex);
2521
2522         EXIT;
2523 }
2524
2525 /*
2526  * Export handle<->lock hash operations.
2527  */
2528 static unsigned
2529 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2530 {
2531         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2532 }
2533
2534 static void *
2535 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2536 {
2537         struct ldlm_lock *lock;
2538
2539         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2540         return &lock->l_remote_handle;
2541 }
2542
2543 static void
2544 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2545 {
2546         struct ldlm_lock     *lock;
2547
2548         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2549         lock->l_remote_handle = *(struct lustre_handle *)key;
2550 }
2551
2552 static int
2553 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2554 {
2555         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2556 }
2557
2558 static void *
2559 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2560 {
2561         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2562 }
2563
2564 static void
2565 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2566 {
2567         struct ldlm_lock *lock;
2568
2569         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2570         LDLM_LOCK_GET(lock);
2571 }
2572
2573 static void
2574 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2575 {
2576         struct ldlm_lock *lock;
2577
2578         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2579         LDLM_LOCK_RELEASE(lock);
2580 }
2581
2582 static cfs_hash_ops_t ldlm_export_lock_ops = {
2583         .hs_hash        = ldlm_export_lock_hash,
2584         .hs_key         = ldlm_export_lock_key,
2585         .hs_keycmp      = ldlm_export_lock_keycmp,
2586         .hs_keycpy      = ldlm_export_lock_keycpy,
2587         .hs_object      = ldlm_export_lock_object,
2588         .hs_get         = ldlm_export_lock_get,
2589         .hs_put         = ldlm_export_lock_put,
2590         .hs_put_locked  = ldlm_export_lock_put,
2591 };
2592
2593 int ldlm_init_export(struct obd_export *exp)
2594 {
2595         ENTRY;
2596
2597         exp->exp_lock_hash =
2598                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2599                                 HASH_EXP_LOCK_CUR_BITS,
2600                                 HASH_EXP_LOCK_MAX_BITS,
2601                                 HASH_EXP_LOCK_BKT_BITS, 0,
2602                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2603                                 &ldlm_export_lock_ops,
2604                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2605                                 CFS_HASH_NBLK_CHANGE);
2606
2607         if (!exp->exp_lock_hash)
2608                 RETURN(-ENOMEM);
2609
2610         RETURN(0);
2611 }
2612 EXPORT_SYMBOL(ldlm_init_export);
2613
2614 void ldlm_destroy_export(struct obd_export *exp)
2615 {
2616         ENTRY;
2617         cfs_hash_putref(exp->exp_lock_hash);
2618         exp->exp_lock_hash = NULL;
2619
2620         ldlm_destroy_flock_export(exp);
2621         EXIT;
2622 }
2623 EXPORT_SYMBOL(ldlm_destroy_export);
2624
2625 static int ldlm_setup(void)
2626 {
2627         static struct ptlrpc_service_conf       conf;
2628         struct ldlm_bl_pool                     *blp = NULL;
2629         int rc = 0;
2630 #ifdef __KERNEL__
2631         int i;
2632 #endif
2633         ENTRY;
2634
2635         if (ldlm_state != NULL)
2636                 RETURN(-EALREADY);
2637
2638         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2639         if (ldlm_state == NULL)
2640                 RETURN(-ENOMEM);
2641
2642 #ifdef LPROCFS
2643         rc = ldlm_proc_setup();
2644         if (rc != 0)
2645                 GOTO(out, rc);
2646 #endif
2647
2648         memset(&conf, 0, sizeof(conf));
2649         conf = (typeof(conf)) {
2650                 .psc_name               = "ldlm_cbd",
2651                 .psc_watchdog_factor    = 2,
2652                 .psc_buf                = {
2653                         .bc_nbufs               = LDLM_NBUFS,
2654                         .bc_buf_size            = LDLM_BUFSIZE,
2655                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2656                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2657                         .bc_req_portal          = LDLM_CB_REQUEST_PORTAL,
2658                         .bc_rep_portal          = LDLM_CB_REPLY_PORTAL,
2659                 },
2660                 .psc_thr                = {
2661                         .tc_thr_name            = "ldlm_cb",
2662                         .tc_thr_factor          = LDLM_THR_FACTOR,
2663                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2664                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2665                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2666                         .tc_nthrs_user          = ldlm_num_threads,
2667                         .tc_cpu_affinity        = 1,
2668                         .tc_ctx_tags            = LCT_MD_THREAD | LCT_DT_THREAD,
2669                 },
2670                 .psc_cpt                = {
2671                         .cc_pattern             = ldlm_cpts,
2672                 },
2673                 .psc_ops                = {
2674                         .so_req_handler         = ldlm_callback_handler,
2675                 },
2676         };
2677         ldlm_state->ldlm_cb_service = \
2678                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2679         if (IS_ERR(ldlm_state->ldlm_cb_service)) {
2680                 CERROR("failed to start service\n");
2681                 rc = PTR_ERR(ldlm_state->ldlm_cb_service);
2682                 ldlm_state->ldlm_cb_service = NULL;
2683                 GOTO(out, rc);
2684         }
2685
2686 #ifdef HAVE_SERVER_SUPPORT
2687         memset(&conf, 0, sizeof(conf));
2688         conf = (typeof(conf)) {
2689                 .psc_name               = "ldlm_canceld",
2690                 .psc_watchdog_factor    = 6,
2691                 .psc_buf                = {
2692                         .bc_nbufs               = LDLM_NBUFS,
2693                         .bc_buf_size            = LDLM_BUFSIZE,
2694                         .bc_req_max_size        = LDLM_MAXREQSIZE,
2695                         .bc_rep_max_size        = LDLM_MAXREPSIZE,
2696                         .bc_req_portal          = LDLM_CANCEL_REQUEST_PORTAL,
2697                         .bc_rep_portal          = LDLM_CANCEL_REPLY_PORTAL,
2698
2699                 },
2700                 .psc_thr                = {
2701                         .tc_thr_name            = "ldlm_cn",
2702                         .tc_thr_factor          = LDLM_THR_FACTOR,
2703                         .tc_nthrs_init          = LDLM_NTHRS_INIT,
2704                         .tc_nthrs_base          = LDLM_NTHRS_BASE,
2705                         .tc_nthrs_max           = LDLM_NTHRS_MAX,
2706                         .tc_nthrs_user          = ldlm_num_threads,
2707                         .tc_cpu_affinity        = 1,
2708                         .tc_ctx_tags            = LCT_MD_THREAD | \
2709                                                   LCT_DT_THREAD | \
2710                                                   LCT_CL_THREAD,
2711                 },
2712                 .psc_cpt                = {
2713                         .cc_pattern             = ldlm_cpts,
2714                 },
2715                 .psc_ops                = {
2716                         .so_req_handler         = ldlm_cancel_handler,
2717                         .so_hpreq_handler       = ldlm_hpreq_handler,
2718                 },
2719         };
2720         ldlm_state->ldlm_cancel_service = \
2721                         ptlrpc_register_service(&conf, ldlm_svc_proc_dir);
2722         if (IS_ERR(ldlm_state->ldlm_cancel_service)) {
2723                 CERROR("failed to start service\n");
2724                 rc = PTR_ERR(ldlm_state->ldlm_cancel_service);
2725                 ldlm_state->ldlm_cancel_service = NULL;
2726                 GOTO(out, rc);
2727         }
2728 #endif
2729
2730         OBD_ALLOC(blp, sizeof(*blp));
2731         if (blp == NULL)
2732                 GOTO(out, rc = -ENOMEM);
2733         ldlm_state->ldlm_bl_pool = blp;
2734
2735         cfs_spin_lock_init(&blp->blp_lock);
2736         CFS_INIT_LIST_HEAD(&blp->blp_list);
2737         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2738         cfs_waitq_init(&blp->blp_waitq);
2739         cfs_atomic_set(&blp->blp_num_threads, 0);
2740         cfs_atomic_set(&blp->blp_busy_threads, 0);
2741
2742 #ifdef __KERNEL__
2743         if (ldlm_num_threads == 0) {
2744                 blp->blp_min_threads = LDLM_NTHRS_INIT;
2745                 blp->blp_max_threads = LDLM_NTHRS_MAX;
2746         } else {
2747                 blp->blp_min_threads = blp->blp_max_threads = \
2748                         min_t(int, LDLM_NTHRS_MAX, max_t(int, LDLM_NTHRS_INIT,
2749                                                          ldlm_num_threads));
2750         }
2751
2752         for (i = 0; i < blp->blp_min_threads; i++) {
2753                 rc = ldlm_bl_thread_start(blp);
2754                 if (rc < 0)
2755                         GOTO(out, rc);
2756         }
2757
2758 # ifdef HAVE_SERVER_SUPPORT
2759         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2760         expired_lock_thread.elt_state = ELT_STOPPED;
2761         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2762
2763         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2764         cfs_spin_lock_init(&waiting_locks_spinlock);
2765         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2766
2767         rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2768         if (rc < 0) {
2769                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2770                 GOTO(out, rc);
2771         }
2772
2773         cfs_wait_event(expired_lock_thread.elt_waitq,
2774                        expired_lock_thread.elt_state == ELT_READY);
2775 # endif /* HAVE_SERVER_SUPPORT */
2776
2777         rc = ldlm_pools_init();
2778         if (rc) {
2779                 CERROR("Failed to initialize LDLM pools: %d\n", rc);
2780                 GOTO(out, rc);
2781         }
2782 #endif
2783         RETURN(0);
2784
2785  out:
2786         ldlm_cleanup();
2787         RETURN(rc);
2788 }
2789
2790 static int ldlm_cleanup(void)
2791 {
2792         ENTRY;
2793
2794         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2795             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2796                 CERROR("ldlm still has namespaces; clean these up first.\n");
2797                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2798                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2799                 RETURN(-EBUSY);
2800         }
2801
2802 #ifdef __KERNEL__
2803         ldlm_pools_fini();
2804
2805         if (ldlm_state->ldlm_bl_pool != NULL) {
2806                 struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2807
2808                 while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2809                         struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2810
2811                         cfs_init_completion(&blp->blp_comp);
2812
2813                         cfs_spin_lock(&blp->blp_lock);
2814                         cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2815                         cfs_waitq_signal(&blp->blp_waitq);
2816                         cfs_spin_unlock(&blp->blp_lock);
2817
2818                         cfs_wait_for_completion(&blp->blp_comp);
2819                 }
2820
2821                 OBD_FREE(blp, sizeof(*blp));
2822         }
2823 #endif /* __KERNEL__ */
2824
2825         if (ldlm_state->ldlm_cb_service != NULL)
2826                 ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2827 # ifdef HAVE_SERVER_SUPPORT
2828         if (ldlm_state->ldlm_cancel_service != NULL)
2829                 ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2830 # endif
2831
2832 #ifdef __KERNEL__
2833         ldlm_proc_cleanup();
2834
2835 # ifdef HAVE_SERVER_SUPPORT
2836         if (expired_lock_thread.elt_state != ELT_STOPPED) {
2837                 expired_lock_thread.elt_state = ELT_TERMINATE;
2838                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2839                 cfs_wait_event(expired_lock_thread.elt_waitq,
2840                                expired_lock_thread.elt_state == ELT_STOPPED);
2841         }
2842 # endif
2843 #endif /* __KERNEL__ */
2844
2845         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2846         ldlm_state = NULL;
2847
2848         RETURN(0);
2849 }
2850
2851 int ldlm_init(void)
2852 {
2853         cfs_mutex_init(&ldlm_ref_mutex);
2854         cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2855         cfs_mutex_init(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2856         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2857                                                sizeof(struct ldlm_resource), 0,
2858                                                CFS_SLAB_HWCACHE_ALIGN);
2859         if (ldlm_resource_slab == NULL)
2860                 return -ENOMEM;
2861
2862         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2863                               sizeof(struct ldlm_lock), 0,
2864                               CFS_SLAB_HWCACHE_ALIGN | SLAB_DESTROY_BY_RCU);
2865         if (ldlm_lock_slab == NULL) {
2866                 cfs_mem_cache_destroy(ldlm_resource_slab);
2867                 return -ENOMEM;
2868         }
2869
2870         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2871                                         sizeof(struct ldlm_interval),
2872                                         0, CFS_SLAB_HWCACHE_ALIGN);
2873         if (ldlm_interval_slab == NULL) {
2874                 cfs_mem_cache_destroy(ldlm_resource_slab);
2875                 cfs_mem_cache_destroy(ldlm_lock_slab);
2876                 return -ENOMEM;
2877         }
2878 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2879         class_export_dump_hook = ldlm_dump_export_locks;
2880 #endif
2881         return 0;
2882 }
2883
2884 void ldlm_exit(void)
2885 {
2886         int rc;
2887         if (ldlm_refcount)
2888                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2889         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2890         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2891 #ifdef __KERNEL__
2892         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2893          * synchronize_rcu() to wait a grace period elapsed, so that
2894          * ldlm_lock_free() get a chance to be called. */
2895         synchronize_rcu();
2896 #endif
2897         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2898         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2899         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2900         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2901 }
2902
2903 /* ldlm_extent.c */
2904 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2905
2906 /* ldlm_lock.c */
2907 #ifdef HAVE_SERVER_SUPPORT
2908 EXPORT_SYMBOL(ldlm_get_processing_policy);
2909 #endif
2910 EXPORT_SYMBOL(ldlm_lock2desc);
2911 EXPORT_SYMBOL(ldlm_register_intent);
2912 EXPORT_SYMBOL(ldlm_lockname);
2913 EXPORT_SYMBOL(ldlm_typename);
2914 EXPORT_SYMBOL(ldlm_lock2handle);
2915 EXPORT_SYMBOL(__ldlm_handle2lock);
2916 EXPORT_SYMBOL(ldlm_lock_get);
2917 EXPORT_SYMBOL(ldlm_lock_put);
2918 EXPORT_SYMBOL(ldlm_lock_match);
2919 EXPORT_SYMBOL(ldlm_lock_cancel);
2920 EXPORT_SYMBOL(ldlm_lock_addref);
2921 EXPORT_SYMBOL(ldlm_lock_addref_try);
2922 EXPORT_SYMBOL(ldlm_lock_decref);
2923 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2924 EXPORT_SYMBOL(ldlm_lock_change_resource);
2925 EXPORT_SYMBOL(ldlm_it2str);
2926 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2927 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2928 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2929 EXPORT_SYMBOL(ldlm_lock_allow_match);
2930 EXPORT_SYMBOL(ldlm_lock_downgrade);
2931 EXPORT_SYMBOL(ldlm_lock_convert);
2932
2933 /* ldlm_request.c */
2934 EXPORT_SYMBOL(ldlm_completion_ast_async);
2935 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2936 EXPORT_SYMBOL(ldlm_completion_ast);
2937 EXPORT_SYMBOL(ldlm_blocking_ast);
2938 EXPORT_SYMBOL(ldlm_glimpse_ast);
2939 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2940 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2941 EXPORT_SYMBOL(ldlm_prep_elc_req);
2942 EXPORT_SYMBOL(ldlm_cli_convert);
2943 EXPORT_SYMBOL(ldlm_cli_enqueue);
2944 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2945 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2946 EXPORT_SYMBOL(ldlm_cli_cancel);
2947 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2948 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2949 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2950 EXPORT_SYMBOL(ldlm_replay_locks);
2951 EXPORT_SYMBOL(ldlm_resource_foreach);
2952 EXPORT_SYMBOL(ldlm_namespace_foreach);
2953 EXPORT_SYMBOL(ldlm_resource_iterate);
2954 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2955 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2956 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2957
2958 /* ldlm_lockd.c */
2959 #ifdef HAVE_SERVER_SUPPORT
2960 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2961 EXPORT_SYMBOL(ldlm_server_completion_ast);
2962 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2963 EXPORT_SYMBOL(ldlm_glimpse_locks);
2964 EXPORT_SYMBOL(ldlm_handle_enqueue);
2965 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2966 EXPORT_SYMBOL(ldlm_handle_cancel);
2967 EXPORT_SYMBOL(ldlm_request_cancel);
2968 EXPORT_SYMBOL(ldlm_handle_convert);
2969 EXPORT_SYMBOL(ldlm_handle_convert0);
2970 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2971 #endif
2972 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2973 EXPORT_SYMBOL(ldlm_get_ref);
2974 EXPORT_SYMBOL(ldlm_put_ref);
2975 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2976
2977 /* ldlm_resource.c */
2978 EXPORT_SYMBOL(ldlm_namespace_new);
2979 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2980 EXPORT_SYMBOL(ldlm_namespace_free);
2981 EXPORT_SYMBOL(ldlm_namespace_dump);
2982 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2983 EXPORT_SYMBOL(ldlm_resource_get);
2984 EXPORT_SYMBOL(ldlm_resource_putref);
2985 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2986
2987 /* ldlm_lib.c */
2988 EXPORT_SYMBOL(client_import_add_conn);
2989 EXPORT_SYMBOL(client_import_del_conn);
2990 EXPORT_SYMBOL(client_obd_setup);
2991 EXPORT_SYMBOL(client_obd_cleanup);
2992 EXPORT_SYMBOL(client_connect_import);
2993 EXPORT_SYMBOL(client_disconnect_export);
2994 EXPORT_SYMBOL(target_send_reply);
2995 EXPORT_SYMBOL(target_pack_pool_reply);
2996
2997 #ifdef HAVE_SERVER_SUPPORT
2998 EXPORT_SYMBOL(server_disconnect_export);
2999 EXPORT_SYMBOL(target_stop_recovery_thread);
3000 EXPORT_SYMBOL(target_handle_connect);
3001 EXPORT_SYMBOL(target_cleanup_recovery);
3002 EXPORT_SYMBOL(target_destroy_export);
3003 EXPORT_SYMBOL(target_cancel_recovery_timer);
3004 EXPORT_SYMBOL(target_queue_recovery_request);
3005 EXPORT_SYMBOL(target_handle_ping);
3006 EXPORT_SYMBOL(target_handle_disconnect);
3007 #endif
3008
3009 /* l_lock.c */
3010 EXPORT_SYMBOL(lock_res_and_lock);
3011 EXPORT_SYMBOL(unlock_res_and_lock);