lustre/ldlm/ldlm_lockd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ldlm/ldlm_lockd.c
  40  *
  41  * Author: Peter Braam <braam@clusterfs.com>
  42  * Author: Phil Schwan <phil@clusterfs.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_LDLM
  49
  50 #ifdef __KERNEL__
  51 # include <libcfs/libcfs.h>
  52 #else
  53 # include <liblustre.h>
  54 #endif
  55
  56 #include <lustre_dlm.h>
  57 #include <obd_class.h>
  58 #include <libcfs/list.h>
  59 #include "ldlm_internal.h"
  60
  61 #ifdef __KERNEL__
  62 static int ldlm_num_threads;
  63 CFS_MODULE_PARM(ldlm_num_threads, "i", int, 0444,
  64                 "number of DLM service threads to start");
  65 #endif
  66
  67 extern cfs_mem_cache_t *ldlm_resource_slab;
  68 extern cfs_mem_cache_t *ldlm_lock_slab;
  69 static cfs_semaphore_t  ldlm_ref_sem;
  70 static int ldlm_refcount;
  71
  72 struct ldlm_cb_async_args {
  73         struct ldlm_cb_set_arg *ca_set_arg;
  74         struct ldlm_lock       *ca_lock;
  75 };
  76
  77 /* LDLM state */
  78
  79 static struct ldlm_state *ldlm_state;
  80
  81 inline cfs_time_t round_timeout(cfs_time_t timeout)
  82 {
  83         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
  84 }
  85
  86 /* timeout for initial callback (AST) reply (bz10399) */
  87 static inline unsigned int ldlm_get_rq_timeout(void)
  88 {
  89         /* Non-AT value */
  90         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
  91
  92         return timeout < 1 ? 1 : timeout;
  93 }
  94
  95 #ifdef __KERNEL__
  96 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
  97 static cfs_spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
  98 static cfs_list_t waiting_locks_list;
  99 static cfs_timer_t waiting_locks_timer;
 100
 101 static struct expired_lock_thread {
 102         cfs_waitq_t               elt_waitq;
 103         int                       elt_state;
 104         int                       elt_dump;
 105         cfs_list_t                elt_expired_locks;
 106 } expired_lock_thread;
 107 #endif
 108
 109 #define ELT_STOPPED   0
 110 #define ELT_READY     1
 111 #define ELT_TERMINATE 2
 112
 113 struct ldlm_bl_pool {
 114         cfs_spinlock_t          blp_lock;
 115
 116         /*
 117          * blp_prio_list is used for callbacks that should be handled
 118          * as a priority. It is used for LDLM_FL_DISCARD_DATA requests.
 119          * see bug 13843
 120          */
 121         cfs_list_t              blp_prio_list;
 122
 123         /*
 124          * blp_list is used for all other callbacks which are likely
 125          * to take longer to process.
 126          */
 127         cfs_list_t              blp_list;
 128
 129         cfs_waitq_t             blp_waitq;
 130         cfs_completion_t        blp_comp;
 131         cfs_atomic_t            blp_num_threads;
 132         cfs_atomic_t            blp_busy_threads;
 133         int                     blp_min_threads;
 134         int                     blp_max_threads;
 135 };
 136
 137 struct ldlm_bl_work_item {
 138         cfs_list_t              blwi_entry;
 139         struct ldlm_namespace  *blwi_ns;
 140         struct ldlm_lock_desc   blwi_ld;
 141         struct ldlm_lock       *blwi_lock;
 142         cfs_list_t              blwi_head;
 143         int                     blwi_count;
 144         cfs_completion_t        blwi_comp;
 145         int                     blwi_mode;
 146         int                     blwi_mem_pressure;
 147 };
 148
 149 #ifdef __KERNEL__
 150
 151 static inline int have_expired_locks(void)
 152 {
 153         int need_to_run;
 154
 155         ENTRY;
 156         cfs_spin_lock_bh(&waiting_locks_spinlock);
 157         need_to_run = !cfs_list_empty(&expired_lock_thread.elt_expired_locks);
 158         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 159
 160         RETURN(need_to_run);
 161 }
 162
 163 static int expired_lock_main(void *arg)
 164 {
 165         cfs_list_t *expired = &expired_lock_thread.elt_expired_locks;
 166         struct l_wait_info lwi = { 0 };
 167         int do_dump;
 168
 169         ENTRY;
 170         cfs_daemonize("ldlm_elt");
 171
 172         expired_lock_thread.elt_state = ELT_READY;
 173         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 174
 175         while (1) {
 176                 l_wait_event(expired_lock_thread.elt_waitq,
 177                              have_expired_locks() ||
 178                              expired_lock_thread.elt_state == ELT_TERMINATE,
 179                              &lwi);
 180
 181                 cfs_spin_lock_bh(&waiting_locks_spinlock);
 182                 if (expired_lock_thread.elt_dump) {
 183                         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 184
 185                         /* from waiting_locks_callback, but not in timer */
 186                         libcfs_debug_dumplog();
 187                         libcfs_run_lbug_upcall(__FILE__,
 188                                                 "waiting_locks_callback",
 189                                                 expired_lock_thread.elt_dump);
 190
 191                         cfs_spin_lock_bh(&waiting_locks_spinlock);
 192                         expired_lock_thread.elt_dump = 0;
 193                 }
 194
 195                 do_dump = 0;
 196
 197                 while (!cfs_list_empty(expired)) {
 198                         struct obd_export *export;
 199                         struct ldlm_lock *lock;
 200
 201                         lock = cfs_list_entry(expired->next, struct ldlm_lock,
 202                                           l_pending_chain);
 203                         if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
 204                             (void *)lock >= LP_POISON) {
 205                                 cfs_spin_unlock_bh(&waiting_locks_spinlock);
 206                                 CERROR("free lock on elt list %p\n", lock);
 207                                 LBUG();
 208                         }
 209                         cfs_list_del_init(&lock->l_pending_chain);
 210                         if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
 211                             (void *)lock->l_export >= LP_POISON) {
 212                                 CERROR("lock with free export on elt list %p\n",
 213                                        lock->l_export);
 214                                 lock->l_export = NULL;
 215                                 LDLM_ERROR(lock, "free export");
 216                                 /* release extra ref grabbed by
 217                                  * ldlm_add_waiting_lock() or
 218                                  * ldlm_failed_ast() */
 219                                 LDLM_LOCK_RELEASE(lock);
 220                                 continue;
 221                         }
 222                         export = class_export_lock_get(lock->l_export, lock);
 223                         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 224
 225                         do_dump++;
 226                         class_fail_export(export);
 227                         class_export_lock_put(export, lock);
 228
 229                         /* release extra ref grabbed by ldlm_add_waiting_lock()
 230                          * or ldlm_failed_ast() */
 231                         LDLM_LOCK_RELEASE(lock);
 232
 233                         cfs_spin_lock_bh(&waiting_locks_spinlock);
 234                 }
 235                 cfs_spin_unlock_bh(&waiting_locks_spinlock);
 236
 237                 if (do_dump && obd_dump_on_eviction) {
 238                         CERROR("dump the log upon eviction\n");
 239                         libcfs_debug_dumplog();
 240                 }
 241
 242                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 243                         break;
 244         }
 245
 246         expired_lock_thread.elt_state = ELT_STOPPED;
 247         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 248         RETURN(0);
 249 }
 250
 251 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 252
 253 /**
 254  * Check if there is a request in the export request list
 255  * which prevents the lock canceling.
 256  */
 257 static int ldlm_lock_busy(struct ldlm_lock *lock)
 258 {
 259         struct ptlrpc_request *req;
 260         int match = 0;
 261         ENTRY;
 262
 263         if (lock->l_export == NULL)
 264                 return 0;
 265
 266         cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
 267         cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
 268                                 rq_exp_list) {
 269                 if (req->rq_ops->hpreq_lock_match) {
 270                         match = req->rq_ops->hpreq_lock_match(req, lock);
 271                         if (match)
 272                                 break;
 273                 }
 274         }
 275         cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
 276         RETURN(match);
 277 }
 278
 279 /* This is called from within a timer interrupt and cannot schedule */
 280 static void waiting_locks_callback(unsigned long unused)
 281 {
 282         struct ldlm_lock *lock, *last = NULL;
 283
 284 repeat:
 285         cfs_spin_lock_bh(&waiting_locks_spinlock);
 286         while (!cfs_list_empty(&waiting_locks_list)) {
 287                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
 288                                       l_pending_chain);
 289                 if (cfs_time_after(lock->l_callback_timeout,
 290                                    cfs_time_current()) ||
 291                     (lock->l_req_mode == LCK_GROUP))
 292                         break;
 293
 294                 if (ptlrpc_check_suspend()) {
 295                         /* there is a case when we talk to one mds, holding
 296                          * lock from another mds. this way we easily can get
 297                          * here, if second mds is being recovered. so, we
 298                          * suspend timeouts. bug 6019 */
 299
 300                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 301                                    lock->l_export->exp_client_uuid.uuid,
 302                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 303                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 304
 305                         cfs_list_del_init(&lock->l_pending_chain);
 306                         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 307                         ldlm_add_waiting_lock(lock);
 308                         goto repeat;
 309                 }
 310
 311                 /* if timeout overlaps the activation time of suspended timeouts
 312                  * then extend it to give a chance for client to reconnect */
 313                 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
 314                                                  cfs_time_seconds(obd_timeout)/2),
 315                                     ptlrpc_suspend_wakeup_time())) {
 316                         LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
 317                                    lock->l_export->exp_client_uuid.uuid,
 318                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 319                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 320
 321                         cfs_list_del_init(&lock->l_pending_chain);
 322                         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 323                         ldlm_add_waiting_lock(lock);
 324                         goto repeat;
 325                 }
 326
 327                 /* Check if we need to prolong timeout */
 328                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT) &&
 329                     ldlm_lock_busy(lock)) {
 330                         int cont = 1;
 331
 332                         if (lock->l_pending_chain.next == &waiting_locks_list)
 333                                 cont = 0;
 334
 335                         LDLM_LOCK_GET(lock);
 336
 337                         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 338                         LDLM_DEBUG(lock, "prolong the busy lock");
 339                         ldlm_refresh_waiting_lock(lock,
 340                                                   ldlm_get_enq_timeout(lock));
 341                         cfs_spin_lock_bh(&waiting_locks_spinlock);
 342
 343                         if (!cont) {
 344                                 LDLM_LOCK_RELEASE(lock);
 345                                 break;
 346                         }
 347
 348                         LDLM_LOCK_RELEASE(lock);
 349                         continue;
 350                 }
 351                 ldlm_lock_to_ns(lock)->ns_timeouts++;
 352                 LDLM_ERROR(lock, "lock callback timer expired after %lds: "
 353                            "evicting client at %s ",
 354                            cfs_time_current_sec()- lock->l_last_activity,
 355                            libcfs_nid2str(
 356                                    lock->l_export->exp_connection->c_peer.nid));
 357
 358                 last = lock;
 359
 360                 /* no needs to take an extra ref on the lock since it was in
 361                  * the waiting_locks_list and ldlm_add_waiting_lock()
 362                  * already grabbed a ref */
 363                 cfs_list_del(&lock->l_pending_chain);
 364                 cfs_list_add(&lock->l_pending_chain,
 365                              &expired_lock_thread.elt_expired_locks);
 366         }
 367
 368         if (!cfs_list_empty(&expired_lock_thread.elt_expired_locks)) {
 369                 if (obd_dump_on_timeout)
 370                         expired_lock_thread.elt_dump = __LINE__;
 371
 372                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 373         }
 374
 375         /*
 376          * Make sure the timer will fire again if we have any locks
 377          * left.
 378          */
 379         if (!cfs_list_empty(&waiting_locks_list)) {
 380                 cfs_time_t timeout_rounded;
 381                 lock = cfs_list_entry(waiting_locks_list.next, struct ldlm_lock,
 382                                       l_pending_chain);
 383                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
 384                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 385         }
 386         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 387 }
 388
 389 /*
 390  * Indicate that we're waiting for a client to call us back cancelling a given
 391  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 392  * timer to fire appropriately.  (We round up to the next second, to avoid
 393  * floods of timer firings during periods of high lock contention and traffic).
 394  * As done by ldlm_add_waiting_lock(), the caller must grab a lock reference
 395  * if it has been added to the waiting list (1 is returned).
 396  *
 397  * Called with the namespace lock held.
 398  */
 399 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock, int seconds)
 400 {
 401         cfs_time_t timeout;
 402         cfs_time_t timeout_rounded;
 403
 404         if (!cfs_list_empty(&lock->l_pending_chain))
 405                 return 0;
 406
 407         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT) ||
 408             OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
 409                 seconds = 1;
 410
 411         timeout = cfs_time_shift(seconds);
 412         if (likely(cfs_time_after(timeout, lock->l_callback_timeout)))
 413                 lock->l_callback_timeout = timeout;
 414
 415         timeout_rounded = round_timeout(lock->l_callback_timeout);
 416
 417         if (cfs_time_before(timeout_rounded,
 418                             cfs_timer_deadline(&waiting_locks_timer)) ||
 419             !cfs_timer_is_armed(&waiting_locks_timer)) {
 420                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 421         }
 422         /* if the new lock has a shorter timeout than something earlier on
 423            the list, we'll wait the longer amount of time; no big deal. */
 424         /* FIFO */
 425         cfs_list_add_tail(&lock->l_pending_chain, &waiting_locks_list);
 426         return 1;
 427 }
 428
 429 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 430 {
 431         int ret;
 432         int timeout = ldlm_get_enq_timeout(lock);
 433
 434         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 435
 436         cfs_spin_lock_bh(&waiting_locks_spinlock);
 437         if (lock->l_destroyed) {
 438                 static cfs_time_t next;
 439                 cfs_spin_unlock_bh(&waiting_locks_spinlock);
 440                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
 441                 if (cfs_time_after(cfs_time_current(), next)) {
 442                         next = cfs_time_shift(14400);
 443                         libcfs_debug_dumpstack(NULL);
 444                 }
 445                 return 0;
 446         }
 447
 448         ret = __ldlm_add_waiting_lock(lock, timeout);
 449         if (ret)
 450                 /* grab ref on the lock if it has been added to the
 451                  * waiting list */
 452                 LDLM_LOCK_GET(lock);
 453         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 454
 455         LDLM_DEBUG(lock, "%sadding to wait list(timeout: %d, AT: %s)",
 456                    ret == 0 ? "not re-" : "", timeout,
 457                    AT_OFF ? "off" : "on");
 458         return ret;
 459 }
 460
 461 /*
 462  * Remove a lock from the pending list, likely because it had its cancellation
 463  * callback arrive without incident.  This adjusts the lock-timeout timer if
 464  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 465  * As done by ldlm_del_waiting_lock(), the caller must release the lock
 466  * reference when the lock is removed from any list (1 is returned).
 467  *
 468  * Called with namespace lock held.
 469  */
 470 static int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
 471 {
 472         cfs_list_t *list_next;
 473
 474         if (cfs_list_empty(&lock->l_pending_chain))
 475                 return 0;
 476
 477         list_next = lock->l_pending_chain.next;
 478         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 479                 /* Removing the head of the list, adjust timer. */
 480                 if (list_next == &waiting_locks_list) {
 481                         /* No more, just cancel. */
 482                         cfs_timer_disarm(&waiting_locks_timer);
 483                 } else {
 484                         struct ldlm_lock *next;
 485                         next = cfs_list_entry(list_next, struct ldlm_lock,
 486                                               l_pending_chain);
 487                         cfs_timer_arm(&waiting_locks_timer,
 488                                       round_timeout(next->l_callback_timeout));
 489                 }
 490         }
 491         cfs_list_del_init(&lock->l_pending_chain);
 492
 493         return 1;
 494 }
 495
 496 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 497 {
 498         int ret;
 499
 500         if (lock->l_export == NULL) {
 501                 /* We don't have a "waiting locks list" on clients. */
 502                 CDEBUG(D_DLMTRACE, "Client lock %p : no-op\n", lock);
 503                 return 0;
 504         }
 505
 506         cfs_spin_lock_bh(&waiting_locks_spinlock);
 507         ret = __ldlm_del_waiting_lock(lock);
 508         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 509         if (ret)
 510                 /* release lock ref if it has indeed been removed
 511                  * from a list */
 512                 LDLM_LOCK_RELEASE(lock);
 513
 514         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
 515         return ret;
 516 }
 517
 518 /*
 519  * Prolong the lock
 520  *
 521  * Called with namespace lock held.
 522  */
 523 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 524 {
 525         if (lock->l_export == NULL) {
 526                 /* We don't have a "waiting locks list" on clients. */
 527                 LDLM_DEBUG(lock, "client lock: no-op");
 528                 return 0;
 529         }
 530
 531         cfs_spin_lock_bh(&waiting_locks_spinlock);
 532
 533         if (cfs_list_empty(&lock->l_pending_chain)) {
 534                 cfs_spin_unlock_bh(&waiting_locks_spinlock);
 535                 LDLM_DEBUG(lock, "wasn't waiting");
 536                 return 0;
 537         }
 538
 539         /* we remove/add the lock to the waiting list, so no needs to
 540          * release/take a lock reference */
 541         __ldlm_del_waiting_lock(lock);
 542         __ldlm_add_waiting_lock(lock, timeout);
 543         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 544
 545         LDLM_DEBUG(lock, "refreshed");
 546         return 1;
 547 }
 548 #else /* !__KERNEL__ */
 549
 550 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 551 {
 552         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 553         RETURN(1);
 554 }
 555
 556 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 557 {
 558         RETURN(0);
 559 }
 560
 561 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock, int timeout)
 562 {
 563         RETURN(0);
 564 }
 565 #endif /* __KERNEL__ */
 566
 567 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 568                             const char *ast_type)
 569 {
 570         LCONSOLE_ERROR_MSG(0x138, "%s: A client on nid %s was evicted due "
 571                            "to a lock %s callback time out: rc %d\n",
 572                            lock->l_export->exp_obd->obd_name,
 573                            obd_export_nid2str(lock->l_export), ast_type, rc);
 574
 575         if (obd_dump_on_timeout)
 576                 libcfs_debug_dumplog();
 577 #ifdef __KERNEL__
 578         cfs_spin_lock_bh(&waiting_locks_spinlock);
 579         if (__ldlm_del_waiting_lock(lock) == 0)
 580                 /* the lock was not in any list, grab an extra ref before adding
 581                  * the lock to the expired list */
 582                 LDLM_LOCK_GET(lock);
 583         cfs_list_add(&lock->l_pending_chain,
 584                      &expired_lock_thread.elt_expired_locks);
 585         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 586         cfs_spin_unlock_bh(&waiting_locks_spinlock);
 587 #else
 588         class_fail_export(lock->l_export);
 589 #endif
 590 }
 591
 592 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 593                                  struct ptlrpc_request *req, int rc,
 594                                  const char *ast_type)
 595 {
 596         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
 597
 598         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 599                 LASSERT(lock->l_export);
 600                 if (lock->l_export->exp_libclient) {
 601                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 602                                    " timeout, just cancelling lock", ast_type,
 603                                    libcfs_nid2str(peer.nid));
 604                         ldlm_lock_cancel(lock);
 605                         rc = -ERESTART;
 606                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
 607                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
 608                                    "cancel was received (AST reply lost?)",
 609                                    ast_type, libcfs_nid2str(peer.nid));
 610                         ldlm_lock_cancel(lock);
 611                         rc = -ERESTART;
 612                 } else {
 613                         ldlm_del_waiting_lock(lock);
 614                         ldlm_failed_ast(lock, rc, ast_type);
 615                 }
 616         } else if (rc) {
 617                 if (rc == -EINVAL) {
 618                         struct ldlm_resource *res = lock->l_resource;
 619                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 620                                " from %s AST - normal race",
 621                                libcfs_nid2str(peer.nid),
 622                                req->rq_repmsg ?
 623                                lustre_msg_get_status(req->rq_repmsg) : -1,
 624                                ast_type);
 625                         if (res) {
 626                                 /* update lvbo to return proper attributes.
 627                                  * see bug 23174 */
 628                                 ldlm_resource_getref(res);
 629                                 ldlm_res_lvbo_update(res, NULL, 1);
 630                                 ldlm_resource_putref(res);
 631                         }
 632
 633                 } else {
 634                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 635                                    "from %s AST", libcfs_nid2str(peer.nid),
 636                                    (req->rq_repmsg != NULL) ?
 637                                    lustre_msg_get_status(req->rq_repmsg) : 0,
 638                                    ast_type);
 639                 }
 640                 ldlm_lock_cancel(lock);
 641                 /* Server-side AST functions are called from ldlm_reprocess_all,
 642                  * which needs to be told to please restart its reprocessing. */
 643                 rc = -ERESTART;
 644         }
 645
 646         return rc;
 647 }
 648
 649 static int ldlm_cb_interpret(const struct lu_env *env,
 650                              struct ptlrpc_request *req, void *data, int rc)
 651 {
 652         struct ldlm_cb_async_args *ca   = data;
 653         struct ldlm_lock          *lock = ca->ca_lock;
 654         struct ldlm_cb_set_arg    *arg  = ca->ca_set_arg;
 655         struct ptlrpc_request_set *set  = arg->set;
 656         ENTRY;
 657
 658         LASSERT(lock != NULL);
 659         if (rc != 0) {
 660                 rc = ldlm_handle_ast_error(lock, req, rc,
 661                                            arg->type == LDLM_BL_CALLBACK
 662                                            ? "blocking" : "completion");
 663                 if (rc == -ERESTART)
 664                         cfs_atomic_inc(&arg->restart);
 665         }
 666         LDLM_LOCK_RELEASE(lock);
 667
 668         cfs_waitq_signal(&set->set_waitq);
 669         RETURN(0);
 670 }
 671
 672 static inline int ldlm_bl_and_cp_ast_tail(struct ptlrpc_request *req,
 673                                           struct ldlm_cb_set_arg *arg,
 674                                           struct ldlm_lock *lock,
 675                                           int instant_cancel)
 676 {
 677         int rc = 0;
 678         ENTRY;
 679
 680         if (unlikely(instant_cancel)) {
 681                 rc = ptl_send_rpc(req, 1);
 682                 ptlrpc_req_finished(req);
 683                 if (rc == 0)
 684                         cfs_atomic_inc(&arg->restart);
 685         } else {
 686                 LDLM_LOCK_GET(lock);
 687                 ptlrpc_set_add_req(arg->set, req);
 688                 ++arg->rpcs;
 689         }
 690
 691         RETURN(rc);
 692 }
 693
 694 /**
 695  * Check if there are requests in the export request list which prevent
 696  * the lock canceling and make these requests high priority ones.
 697  */
 698 static void ldlm_lock_reorder_req(struct ldlm_lock *lock)
 699 {
 700         struct ptlrpc_request *req;
 701         ENTRY;
 702
 703         if (lock->l_export == NULL) {
 704                 LDLM_DEBUG(lock, "client lock: no-op");
 705                 RETURN_EXIT;
 706         }
 707
 708         cfs_spin_lock_bh(&lock->l_export->exp_rpc_lock);
 709         cfs_list_for_each_entry(req, &lock->l_export->exp_queued_rpc,
 710                                 rq_exp_list) {
 711                 if (!req->rq_hp && req->rq_ops->hpreq_lock_match &&
 712                     req->rq_ops->hpreq_lock_match(req, lock))
 713                         ptlrpc_hpreq_reorder(req);
 714         }
 715         cfs_spin_unlock_bh(&lock->l_export->exp_rpc_lock);
 716         EXIT;
 717 }
 718
 719 /*
 720  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
 721  * enqueued server lock conflicts with given one.
 722  *
 723  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
 724  * to wait for client response.
 725  */
 726 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 727                              struct ldlm_lock_desc *desc,
 728                              void *data, int flag)
 729 {
 730         struct ldlm_cb_async_args *ca;
 731         struct ldlm_cb_set_arg *arg = data;
 732         struct ldlm_request    *body;
 733         struct ptlrpc_request  *req;
 734         int                     instant_cancel = 0;
 735         int                     rc = 0;
 736         ENTRY;
 737
 738         if (flag == LDLM_CB_CANCELING)
 739                 /* Don't need to do anything here. */
 740                 RETURN(0);
 741
 742         LASSERT(lock);
 743         LASSERT(data != NULL);
 744         if (lock->l_export->exp_obd->obd_recovering != 0) {
 745                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
 746                 ldlm_lock_dump(D_ERROR, lock, 0);
 747         }
 748
 749         ldlm_lock_reorder_req(lock);
 750
 751         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 752                                         &RQF_LDLM_BL_CALLBACK,
 753                                         LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK);
 754         if (req == NULL)
 755                 RETURN(-ENOMEM);
 756
 757         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 758         ca = ptlrpc_req_async_args(req);
 759         ca->ca_set_arg = arg;
 760         ca->ca_lock = lock;
 761
 762         req->rq_interpret_reply = ldlm_cb_interpret;
 763         req->rq_no_resend = 1;
 764
 765         lock_res(lock->l_resource);
 766         if (lock->l_granted_mode != lock->l_req_mode) {
 767                 /* this blocking AST will be communicated as part of the
 768                  * completion AST instead */
 769                 unlock_res(lock->l_resource);
 770                 ptlrpc_req_finished(req);
 771                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 772                 RETURN(0);
 773         }
 774
 775         if (lock->l_destroyed) {
 776                 /* What's the point? */
 777                 unlock_res(lock->l_resource);
 778                 ptlrpc_req_finished(req);
 779                 RETURN(0);
 780         }
 781
 782         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
 783                 instant_cancel = 1;
 784
 785         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 786         body->lock_handle[0] = lock->l_remote_handle;
 787         body->lock_desc = *desc;
 788         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 789
 790         LDLM_DEBUG(lock, "server preparing blocking AST");
 791
 792         ptlrpc_request_set_replen(req);
 793         if (instant_cancel) {
 794                 unlock_res(lock->l_resource);
 795                 ldlm_lock_cancel(lock);
 796         } else {
 797                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
 798                 ldlm_add_waiting_lock(lock);
 799                 unlock_res(lock->l_resource);
 800         }
 801
 802         req->rq_send_state = LUSTRE_IMP_FULL;
 803         /* ptlrpc_request_alloc_pack already set timeout */
 804         if (AT_OFF)
 805                 req->rq_timeout = ldlm_get_rq_timeout();
 806
 807         if (lock->l_export && lock->l_export->exp_nid_stats &&
 808             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 809                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 810                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 811
 812         rc = ldlm_bl_and_cp_ast_tail(req, arg, lock, instant_cancel);
 813
 814         RETURN(rc);
 815 }
 816
 817 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 818 {
 819         struct ldlm_cb_set_arg *arg = data;
 820         struct ldlm_request    *body;
 821         struct ptlrpc_request  *req;
 822         struct ldlm_cb_async_args *ca;
 823         long                    total_enqueue_wait;
 824         int                     instant_cancel = 0;
 825         int                     rc = 0;
 826         ENTRY;
 827
 828         LASSERT(lock != NULL);
 829         LASSERT(data != NULL);
 830
 831         total_enqueue_wait = cfs_time_sub(cfs_time_current_sec(),
 832                                           lock->l_last_activity);
 833
 834         req = ptlrpc_request_alloc(lock->l_export->exp_imp_reverse,
 835                                     &RQF_LDLM_CP_CALLBACK);
 836         if (req == NULL)
 837                 RETURN(-ENOMEM);
 838
 839         /* server namespace, doesn't need lock */
 840         if (lock->l_resource->lr_lvb_len) {
 841                  req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_CLIENT,
 842                                       lock->l_resource->lr_lvb_len);
 843         }
 844
 845         rc = ptlrpc_request_pack(req, LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK);
 846         if (rc) {
 847                 ptlrpc_request_free(req);
 848                 RETURN(rc);
 849         }
 850
 851         CLASSERT(sizeof(*ca) <= sizeof(req->rq_async_args));
 852         ca = ptlrpc_req_async_args(req);
 853         ca->ca_set_arg = arg;
 854         ca->ca_lock = lock;
 855
 856         req->rq_interpret_reply = ldlm_cb_interpret;
 857         req->rq_no_resend = 1;
 858         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 859
 860         body->lock_handle[0] = lock->l_remote_handle;
 861         body->lock_flags = flags;
 862         ldlm_lock2desc(lock, &body->lock_desc);
 863         if (lock->l_resource->lr_lvb_len) {
 864                 void *lvb = req_capsule_client_get(&req->rq_pill, &RMF_DLM_LVB);
 865
 866                 lock_res(lock->l_resource);
 867                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 868                        lock->l_resource->lr_lvb_len);
 869                 unlock_res(lock->l_resource);
 870         }
 871
 872         LDLM_DEBUG(lock, "server preparing completion AST (after %lds wait)",
 873                    total_enqueue_wait);
 874
 875         /* Server-side enqueue wait time estimate, used in
 876             __ldlm_add_waiting_lock to set future enqueue timers */
 877         if (total_enqueue_wait < ldlm_get_enq_timeout(lock))
 878                 at_measured(ldlm_lock_to_ns_at(lock),
 879                             total_enqueue_wait);
 880         else
 881                 /* bz18618. Don't add lock enqueue time we spend waiting for a
 882                    previous callback to fail. Locks waiting legitimately will
 883                    get extended by ldlm_refresh_waiting_lock regardless of the
 884                    estimate, so it's okay to underestimate here. */
 885                 LDLM_DEBUG(lock, "lock completed after %lus; estimate was %ds. "
 886                        "It is likely that a previous callback timed out.",
 887                        total_enqueue_wait,
 888                        at_get(ldlm_lock_to_ns_at(lock)));
 889
 890         ptlrpc_request_set_replen(req);
 891
 892         req->rq_send_state = LUSTRE_IMP_FULL;
 893         /* ptlrpc_request_pack already set timeout */
 894         if (AT_OFF)
 895                 req->rq_timeout = ldlm_get_rq_timeout();
 896
 897         /* We only send real blocking ASTs after the lock is granted */
 898         lock_res_and_lock(lock);
 899         if (lock->l_flags & LDLM_FL_AST_SENT) {
 900                 body->lock_flags |= LDLM_FL_AST_SENT;
 901                 /* copy ast flags like LDLM_FL_DISCARD_DATA */
 902                 body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 903
 904                 /* We might get here prior to ldlm_handle_enqueue setting
 905                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
 906                  * into waiting list, but this is safe and similar code in
 907                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
 908                  * that would not only cancel the lock, but will also remove
 909                  * it from waiting list */
 910                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 911                         unlock_res_and_lock(lock);
 912                         ldlm_lock_cancel(lock);
 913                         instant_cancel = 1;
 914                         lock_res_and_lock(lock);
 915                 } else {
 916                         /* start the lock-timeout clock */
 917                         ldlm_add_waiting_lock(lock);
 918                 }
 919         }
 920         unlock_res_and_lock(lock);
 921
 922         if (lock->l_export && lock->l_export->exp_nid_stats &&
 923             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 924                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 925                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
 926
 927         rc = ldlm_bl_and_cp_ast_tail(req, arg, lock, instant_cancel);
 928
 929         RETURN(rc);
 930 }
 931
 932 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 933 {
 934         struct ldlm_resource  *res = lock->l_resource;
 935         struct ldlm_request   *body;
 936         struct ptlrpc_request *req;
 937         int                    rc;
 938         ENTRY;
 939
 940         LASSERT(lock != NULL);
 941
 942         req = ptlrpc_request_alloc_pack(lock->l_export->exp_imp_reverse,
 943                                         &RQF_LDLM_GL_CALLBACK,
 944                                         LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK);
 945
 946         if (req == NULL)
 947                 RETURN(-ENOMEM);
 948
 949         body = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
 950         body->lock_handle[0] = lock->l_remote_handle;
 951         ldlm_lock2desc(lock, &body->lock_desc);
 952
 953         /* server namespace, doesn't need lock */
 954         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 955                              lock->l_resource->lr_lvb_len);
 956         res = lock->l_resource;
 957         ptlrpc_request_set_replen(req);
 958
 959
 960         req->rq_send_state = LUSTRE_IMP_FULL;
 961         /* ptlrpc_request_alloc_pack already set timeout */
 962         if (AT_OFF)
 963                 req->rq_timeout = ldlm_get_rq_timeout();
 964
 965         if (lock->l_export && lock->l_export->exp_nid_stats &&
 966             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 967                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 968                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
 969
 970         rc = ptlrpc_queue_wait(req);
 971         /* Update the LVB from disk if the AST failed (this is a legal race)
 972          *
 973          * - Glimpse callback of local lock just return -ELDLM_NO_LOCK_DATA.
 974          * - Glimpse callback of remote lock might return -ELDLM_NO_LOCK_DATA
 975          *   when inode is cleared. LU-274
 976          */
 977         if (rc == -ELDLM_NO_LOCK_DATA) {
 978                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 979                 ldlm_res_lvbo_update(res, NULL, 1);
 980         } else if (rc != 0) {
 981                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 982         } else {
 983                 rc = ldlm_res_lvbo_update(res, req, 1);
 984         }
 985
 986         ptlrpc_req_finished(req);
 987         if (rc == -ERESTART)
 988                 ldlm_reprocess_all(res);
 989
 990         RETURN(rc);
 991 }
 992
 993 #ifdef __KERNEL__
 994 extern unsigned long long lu_time_stamp_get(void);
 995 #else
 996 #define lu_time_stamp_get() time(NULL)
 997 #endif
 998
 999 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
1000                        struct lprocfs_stats *srv_stats)
1001 {
1002         int lock_type = 0, op = 0;
1003
1004         lock_type = dlm_req->lock_desc.l_resource.lr_type;
1005
1006         switch (lock_type) {
1007         case LDLM_PLAIN:
1008                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
1009                 break;
1010         case LDLM_EXTENT:
1011                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
1012                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
1013                 else
1014                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
1015                 break;
1016         case LDLM_FLOCK:
1017                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
1018                 break;
1019         case LDLM_IBITS:
1020                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1021                 break;
1022         default:
1023                 op = 0;
1024                 break;
1025         }
1026
1027         if (op)
1028                 lprocfs_counter_incr(srv_stats, op);
1029
1030         return;
1031 }
1032
1033 /*
1034  * Main server-side entry point into LDLM. This is called by ptlrpc service
1035  * threads to carry out client lock enqueueing requests.
1036  */
1037 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1038                          struct ptlrpc_request *req,
1039                          const struct ldlm_request *dlm_req,
1040                          const struct ldlm_callback_suite *cbs)
1041 {
1042         struct ldlm_reply *dlm_rep;
1043         __u32 flags;
1044         ldlm_error_t err = ELDLM_OK;
1045         struct ldlm_lock *lock = NULL;
1046         void *cookie = NULL;
1047         int rc = 0;
1048         ENTRY;
1049
1050         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1051
1052         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1053         flags = dlm_req->lock_flags;
1054
1055         LASSERT(req->rq_export);
1056
1057         if (req->rq_rqbd->rqbd_service->srv_stats)
1058                 ldlm_svc_get_eopc(dlm_req,
1059                                   req->rq_rqbd->rqbd_service->srv_stats);
1060
1061         if (req->rq_export && req->rq_export->exp_nid_stats &&
1062             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1063                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1064                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1065
1066         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1067                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1068                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1069                           dlm_req->lock_desc.l_resource.lr_type);
1070                 GOTO(out, rc = -EFAULT);
1071         }
1072
1073         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1074                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1075                      dlm_req->lock_desc.l_req_mode &
1076                      (dlm_req->lock_desc.l_req_mode-1))) {
1077                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1078                           dlm_req->lock_desc.l_req_mode);
1079                 GOTO(out, rc = -EFAULT);
1080         }
1081
1082         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1083                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1084                              LDLM_PLAIN)) {
1085                         DEBUG_REQ(D_ERROR, req,
1086                                   "PLAIN lock request from IBITS client?");
1087                         GOTO(out, rc = -EPROTO);
1088                 }
1089         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1090                             LDLM_IBITS)) {
1091                 DEBUG_REQ(D_ERROR, req,
1092                           "IBITS lock request from unaware client?");
1093                 GOTO(out, rc = -EPROTO);
1094         }
1095
1096 #if 0
1097         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1098            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1099            ibits for mgc/mgs) */
1100
1101         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1102          * inodebits lock if client does not support them. */
1103         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1104             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1105                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1106                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1107                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1108                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1109                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1110         }
1111 #endif
1112
1113         if (unlikely(flags & LDLM_FL_REPLAY)) {
1114                 /* Find an existing lock in the per-export lock hash */
1115                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1116                                        (void *)&dlm_req->lock_handle[0]);
1117                 if (lock != NULL) {
1118                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1119                                   LPX64, lock->l_handle.h_cookie);
1120                         GOTO(existing_lock, rc = 0);
1121                 }
1122         }
1123
1124         /* The lock's callback data might be set in the policy function */
1125         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1126                                 dlm_req->lock_desc.l_resource.lr_type,
1127                                 dlm_req->lock_desc.l_req_mode,
1128                                 cbs, NULL, 0);
1129
1130         if (!lock)
1131                 GOTO(out, rc = -ENOMEM);
1132
1133         lock->l_last_activity = cfs_time_current_sec();
1134         lock->l_remote_handle = dlm_req->lock_handle[0];
1135         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1136
1137         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1138         /* Don't enqueue a lock onto the export if it is been disonnected
1139          * due to eviction (bug 3822) or server umount (bug 24324).
1140          * Cancel it now instead. */
1141         if (req->rq_export->exp_disconnected) {
1142                 LDLM_ERROR(lock, "lock on disconnected export %p",
1143                            req->rq_export);
1144                 GOTO(out, rc = -ENOTCONN);
1145         }
1146
1147         lock->l_export = class_export_lock_get(req->rq_export, lock);
1148         if (lock->l_export->exp_lock_hash)
1149                 cfs_hash_add(lock->l_export->exp_lock_hash,
1150                              &lock->l_remote_handle,
1151                              &lock->l_exp_hash);
1152
1153 existing_lock:
1154
1155         if (flags & LDLM_FL_HAS_INTENT) {
1156                 /* In this case, the reply buffer is allocated deep in
1157                  * local_lock_enqueue by the policy function. */
1158                 cookie = req;
1159         } else {
1160                 /* based on the assumption that lvb size never changes during
1161                  * resource life time otherwise it need resource->lr_lock's
1162                  * protection */
1163                 if (lock->l_resource->lr_lvb_len) {
1164                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1165                                              RCL_SERVER,
1166                                              lock->l_resource->lr_lvb_len);
1167                 }
1168
1169                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1170                         GOTO(out, rc = -ENOMEM);
1171
1172                 rc = req_capsule_server_pack(&req->rq_pill);
1173                 if (rc)
1174                         GOTO(out, rc);
1175         }
1176
1177         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1178                 ldlm_convert_policy_to_local(
1179                                           dlm_req->lock_desc.l_resource.lr_type,
1180                                           &dlm_req->lock_desc.l_policy_data,
1181                                           &lock->l_policy_data);
1182         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1183                 lock->l_req_extent = lock->l_policy_data.l_extent;
1184
1185         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1186         if (err)
1187                 GOTO(out, err);
1188
1189         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1190         dlm_rep->lock_flags = flags;
1191
1192         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1193         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1194
1195         /* We never send a blocking AST until the lock is granted, but
1196          * we can tell it right now */
1197         lock_res_and_lock(lock);
1198
1199         /* Now take into account flags to be inherited from original lock
1200            request both in reply to client and in our own lock flags. */
1201         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1202         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1203
1204         /* Don't move a pending lock onto the export if it has already been
1205          * disconnected due to eviction (bug 5683) or server umount (bug 24324).
1206          * Cancel it now instead. */
1207         if (unlikely(req->rq_export->exp_disconnected ||
1208                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1209                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1210                 rc = -ENOTCONN;
1211         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1212                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1213                 if (lock->l_granted_mode == lock->l_req_mode) {
1214                         /*
1215                          * Only cancel lock if it was granted, because it would
1216                          * be destroyed immediately and would never be granted
1217                          * in the future, causing timeouts on client.  Not
1218                          * granted lock will be cancelled immediately after
1219                          * sending completion AST.
1220                          */
1221                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1222                                 unlock_res_and_lock(lock);
1223                                 ldlm_lock_cancel(lock);
1224                                 lock_res_and_lock(lock);
1225                         } else
1226                                 ldlm_add_waiting_lock(lock);
1227                 }
1228         }
1229         /* Make sure we never ever grant usual metadata locks to liblustre
1230            clients */
1231         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1232             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1233              req->rq_export->exp_libclient) {
1234                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1235                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1236                         CERROR("Granting sync lock to libclient. "
1237                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1238                                dlm_req->lock_flags, dlm_rep->lock_flags,
1239                                lock->l_flags);
1240                         LDLM_ERROR(lock, "sync lock");
1241                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1242                                 struct ldlm_intent *it;
1243
1244                                 it = req_capsule_client_get(&req->rq_pill,
1245                                                             &RMF_LDLM_INTENT);
1246                                 if (it != NULL) {
1247                                         CERROR("This is intent %s ("LPU64")\n",
1248                                                ldlm_it2str(it->opc), it->opc);
1249                                 }
1250                         }
1251                 }
1252         }
1253
1254         unlock_res_and_lock(lock);
1255
1256         EXIT;
1257  out:
1258         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1259         if (!req->rq_packed_final) {
1260                 err = lustre_pack_reply(req, 1, NULL, NULL);
1261                 if (rc == 0)
1262                         rc = err;
1263         }
1264
1265         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1266          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1267         if (lock) {
1268                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1269                            "(err=%d, rc=%d)", err, rc);
1270
1271                 if (rc == 0) {
1272                         if (lock->l_resource->lr_lvb_len > 0) {
1273                                 /* MDT path won't handle lr_lvb_data, so
1274                                  * lock/unlock better be contained in the
1275                                  * if block */
1276                                 void *lvb;
1277
1278                                 lvb = req_capsule_server_get(&req->rq_pill,
1279                                                              &RMF_DLM_LVB);
1280                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1281                                          req, lock);
1282                                 lock_res(lock->l_resource);
1283                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
1284                                        lock->l_resource->lr_lvb_len);
1285                                 unlock_res(lock->l_resource);
1286                         }
1287                 } else {
1288                         lock_res_and_lock(lock);
1289                         ldlm_resource_unlink_lock(lock);
1290                         ldlm_lock_destroy_nolock(lock);
1291                         unlock_res_and_lock(lock);
1292                 }
1293
1294                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1295                         ldlm_reprocess_all(lock->l_resource);
1296
1297                 LDLM_LOCK_RELEASE(lock);
1298         }
1299
1300         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1301                           lock, rc);
1302
1303         return rc;
1304 }
1305
1306 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1307                         ldlm_completion_callback completion_callback,
1308                         ldlm_blocking_callback blocking_callback,
1309                         ldlm_glimpse_callback glimpse_callback)
1310 {
1311         struct ldlm_request *dlm_req;
1312         struct ldlm_callback_suite cbs = {
1313                 .lcs_completion = completion_callback,
1314                 .lcs_blocking   = blocking_callback,
1315                 .lcs_glimpse    = glimpse_callback
1316         };
1317         int rc;
1318
1319         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1320         if (dlm_req != NULL) {
1321                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1322                                           req, dlm_req, &cbs);
1323         } else {
1324                 rc = -EFAULT;
1325         }
1326         return rc;
1327 }
1328
1329 int ldlm_handle_convert0(struct ptlrpc_request *req,
1330                          const struct ldlm_request *dlm_req)
1331 {
1332         struct ldlm_reply *dlm_rep;
1333         struct ldlm_lock *lock;
1334         int rc;
1335         ENTRY;
1336
1337         if (req->rq_export && req->rq_export->exp_nid_stats &&
1338             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1339                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1340                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1341
1342         rc = req_capsule_server_pack(&req->rq_pill);
1343         if (rc)
1344                 RETURN(rc);
1345
1346         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1347         dlm_rep->lock_flags = dlm_req->lock_flags;
1348
1349         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1350         if (!lock) {
1351                 req->rq_status = EINVAL;
1352         } else {
1353                 void *res = NULL;
1354
1355                 LDLM_DEBUG(lock, "server-side convert handler START");
1356
1357                 lock->l_last_activity = cfs_time_current_sec();
1358                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1359                                         &dlm_rep->lock_flags);
1360                 if (res) {
1361                         if (ldlm_del_waiting_lock(lock))
1362                                 LDLM_DEBUG(lock, "converted waiting lock");
1363                         req->rq_status = 0;
1364                 } else {
1365                         req->rq_status = EDEADLOCK;
1366                 }
1367         }
1368
1369         if (lock) {
1370                 if (!req->rq_status)
1371                         ldlm_reprocess_all(lock->l_resource);
1372                 LDLM_DEBUG(lock, "server-side convert handler END");
1373                 LDLM_LOCK_PUT(lock);
1374         } else
1375                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1376
1377         RETURN(0);
1378 }
1379
1380 int ldlm_handle_convert(struct ptlrpc_request *req)
1381 {
1382         int rc;
1383         struct ldlm_request *dlm_req;
1384
1385         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1386         if (dlm_req != NULL) {
1387                 rc = ldlm_handle_convert0(req, dlm_req);
1388         } else {
1389                 CERROR ("Can't unpack dlm_req\n");
1390                 rc = -EFAULT;
1391         }
1392         return rc;
1393 }
1394
1395 /* Cancel all the locks whos handles are packed into ldlm_request */
1396 int ldlm_request_cancel(struct ptlrpc_request *req,
1397                         const struct ldlm_request *dlm_req, int first)
1398 {
1399         struct ldlm_resource *res, *pres = NULL;
1400         struct ldlm_lock *lock;
1401         int i, count, done = 0;
1402         ENTRY;
1403
1404         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1405         if (first >= count)
1406                 RETURN(0);
1407
1408         /* There is no lock on the server at the replay time,
1409          * skip lock cancelling to make replay tests to pass. */
1410         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1411                 RETURN(0);
1412
1413         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1414                           "starting at %d", count, first);
1415
1416         for (i = first; i < count; i++) {
1417                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1418                 if (!lock) {
1419                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1420                                           "lock (cookie "LPU64")",
1421                                           dlm_req->lock_handle[i].cookie);
1422                         continue;
1423                 }
1424
1425                 res = lock->l_resource;
1426                 done++;
1427
1428                 if (res != pres) {
1429                         if (pres != NULL) {
1430                                 ldlm_reprocess_all(pres);
1431                                 LDLM_RESOURCE_DELREF(pres);
1432                                 ldlm_resource_putref(pres);
1433                         }
1434                         if (res != NULL) {
1435                                 ldlm_resource_getref(res);
1436                                 LDLM_RESOURCE_ADDREF(res);
1437                                 ldlm_res_lvbo_update(res, NULL, 1);
1438                         }
1439                         pres = res;
1440                 }
1441                 ldlm_lock_cancel(lock);
1442                 LDLM_LOCK_PUT(lock);
1443         }
1444         if (pres != NULL) {
1445                 ldlm_reprocess_all(pres);
1446                 LDLM_RESOURCE_DELREF(pres);
1447                 ldlm_resource_putref(pres);
1448         }
1449         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1450         RETURN(done);
1451 }
1452
1453 int ldlm_handle_cancel(struct ptlrpc_request *req)
1454 {
1455         struct ldlm_request *dlm_req;
1456         int rc;
1457         ENTRY;
1458
1459         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1460         if (dlm_req == NULL) {
1461                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1462                 RETURN(-EFAULT);
1463         }
1464
1465         if (req->rq_export && req->rq_export->exp_nid_stats &&
1466             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1467                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1468                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1469
1470         rc = req_capsule_server_pack(&req->rq_pill);
1471         if (rc)
1472                 RETURN(rc);
1473
1474         if (!ldlm_request_cancel(req, dlm_req, 0))
1475                 req->rq_status = ESTALE;
1476
1477         RETURN(ptlrpc_reply(req));
1478 }
1479
1480 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1481                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1482 {
1483         int do_ast;
1484         ENTRY;
1485
1486         LDLM_DEBUG(lock, "client blocking AST callback handler");
1487
1488         lock_res_and_lock(lock);
1489         lock->l_flags |= LDLM_FL_CBPENDING;
1490
1491         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1492                 lock->l_flags |= LDLM_FL_CANCEL;
1493
1494         do_ast = (!lock->l_readers && !lock->l_writers);
1495         unlock_res_and_lock(lock);
1496
1497         if (do_ast) {
1498                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1499                        lock, lock->l_blocking_ast);
1500                 if (lock->l_blocking_ast != NULL)
1501                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1502                                              LDLM_CB_BLOCKING);
1503         } else {
1504                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1505                        lock);
1506         }
1507
1508         LDLM_DEBUG(lock, "client blocking callback handler END");
1509         LDLM_LOCK_RELEASE(lock);
1510         EXIT;
1511 }
1512
1513 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1514                                     struct ldlm_namespace *ns,
1515                                     struct ldlm_request *dlm_req,
1516                                     struct ldlm_lock *lock)
1517 {
1518         CFS_LIST_HEAD(ast_list);
1519         ENTRY;
1520
1521         LDLM_DEBUG(lock, "client completion callback handler START");
1522
1523         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1524                 int to = cfs_time_seconds(1);
1525                 while (to > 0) {
1526                         cfs_schedule_timeout_and_set_state(
1527                                 CFS_TASK_INTERRUPTIBLE, to);
1528                         if (lock->l_granted_mode == lock->l_req_mode ||
1529                             lock->l_destroyed)
1530                                 break;
1531                 }
1532         }
1533
1534         lock_res_and_lock(lock);
1535         if (lock->l_destroyed ||
1536             lock->l_granted_mode == lock->l_req_mode) {
1537                 /* bug 11300: the lock has already been granted */
1538                 unlock_res_and_lock(lock);
1539                 LDLM_DEBUG(lock, "Double grant race happened");
1540                 LDLM_LOCK_RELEASE(lock);
1541                 EXIT;
1542                 return;
1543         }
1544
1545         /* If we receive the completion AST before the actual enqueue returned,
1546          * then we might need to switch lock modes, resources, or extents. */
1547         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1548                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1549                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1550         }
1551
1552         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1553                 ldlm_convert_policy_to_local(
1554                                           dlm_req->lock_desc.l_resource.lr_type,
1555                                           &dlm_req->lock_desc.l_policy_data,
1556                                           &lock->l_policy_data);
1557                 LDLM_DEBUG(lock, "completion AST, new policy data");
1558         }
1559
1560         ldlm_resource_unlink_lock(lock);
1561         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1562                    &lock->l_resource->lr_name,
1563                    sizeof(lock->l_resource->lr_name)) != 0) {
1564                 unlock_res_and_lock(lock);
1565                 if (ldlm_lock_change_resource(ns, lock,
1566                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1567                         LDLM_ERROR(lock, "Failed to allocate resource");
1568                         LDLM_LOCK_RELEASE(lock);
1569                         EXIT;
1570                         return;
1571                 }
1572                 LDLM_DEBUG(lock, "completion AST, new resource");
1573                 CERROR("change resource!\n");
1574                 lock_res_and_lock(lock);
1575         }
1576
1577         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1578                 /* BL_AST locks are not needed in lru.
1579                  * let ldlm_cancel_lru() be fast. */
1580                 ldlm_lock_remove_from_lru(lock);
1581                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1582                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1583         }
1584
1585         if (lock->l_lvb_len) {
1586                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1587                                          RCL_CLIENT) < lock->l_lvb_len) {
1588                         LDLM_ERROR(lock, "completion AST did not contain "
1589                                    "expected LVB!");
1590                 } else {
1591                         void *lvb = req_capsule_client_get(&req->rq_pill,
1592                                                            &RMF_DLM_LVB);
1593                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1594                 }
1595         }
1596
1597         ldlm_grant_lock(lock, &ast_list);
1598         unlock_res_and_lock(lock);
1599
1600         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1601
1602         /* Let Enqueue to call osc_lock_upcall() and initialize
1603          * l_ast_data */
1604         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1605
1606         ldlm_run_ast_work(ns, &ast_list, LDLM_WORK_CP_AST);
1607
1608         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1609                           lock);
1610         LDLM_LOCK_RELEASE(lock);
1611         EXIT;
1612 }
1613
1614 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1615                                     struct ldlm_namespace *ns,
1616                                     struct ldlm_request *dlm_req,
1617                                     struct ldlm_lock *lock)
1618 {
1619         int rc = -ENOSYS;
1620         ENTRY;
1621
1622         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1623
1624         if (lock->l_glimpse_ast != NULL)
1625                 rc = lock->l_glimpse_ast(lock, req);
1626
1627         if (req->rq_repmsg != NULL) {
1628                 ptlrpc_reply(req);
1629         } else {
1630                 req->rq_status = rc;
1631                 ptlrpc_error(req);
1632         }
1633
1634         lock_res_and_lock(lock);
1635         if (lock->l_granted_mode == LCK_PW &&
1636             !lock->l_readers && !lock->l_writers &&
1637             cfs_time_after(cfs_time_current(),
1638                            cfs_time_add(lock->l_last_used,
1639                                         cfs_time_seconds(10)))) {
1640                 unlock_res_and_lock(lock);
1641                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1642                         ldlm_handle_bl_callback(ns, NULL, lock);
1643
1644                 EXIT;
1645                 return;
1646         }
1647         unlock_res_and_lock(lock);
1648         LDLM_LOCK_RELEASE(lock);
1649         EXIT;
1650 }
1651
1652 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1653 {
1654         if (req->rq_no_reply)
1655                 return 0;
1656
1657         req->rq_status = rc;
1658         if (!req->rq_packed_final) {
1659                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1660                 if (rc)
1661                         return rc;
1662         }
1663         return ptlrpc_reply(req);
1664 }
1665
1666 #ifdef __KERNEL__
1667 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1668 {
1669         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1670         ENTRY;
1671
1672         cfs_spin_lock(&blp->blp_lock);
1673         if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1674                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1675                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1676         } else {
1677                 /* other blocking callbacks are added to the regular list */
1678                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1679         }
1680         cfs_spin_unlock(&blp->blp_lock);
1681
1682         cfs_waitq_signal(&blp->blp_waitq);
1683
1684         /* can not use blwi->blwi_mode as blwi could be already freed in
1685            LDLM_ASYNC mode */
1686         if (mode == LDLM_SYNC)
1687                 cfs_wait_for_completion(&blwi->blwi_comp);
1688
1689         RETURN(0);
1690 }
1691
1692 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1693                              struct ldlm_namespace *ns,
1694                              struct ldlm_lock_desc *ld,
1695                              cfs_list_t *cancels, int count,
1696                              struct ldlm_lock *lock,
1697                              int mode)
1698 {
1699         cfs_init_completion(&blwi->blwi_comp);
1700         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1701
1702         if (cfs_memory_pressure_get())
1703                 blwi->blwi_mem_pressure = 1;
1704
1705         blwi->blwi_ns = ns;
1706         blwi->blwi_mode = mode;
1707         if (ld != NULL)
1708                 blwi->blwi_ld = *ld;
1709         if (count) {
1710                 cfs_list_add(&blwi->blwi_head, cancels);
1711                 cfs_list_del_init(cancels);
1712                 blwi->blwi_count = count;
1713         } else {
1714                 blwi->blwi_lock = lock;
1715         }
1716 }
1717
1718 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1719                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1720                              cfs_list_t *cancels, int count, int mode)
1721 {
1722         ENTRY;
1723
1724         if (cancels && count == 0)
1725                 RETURN(0);
1726
1727         if (mode == LDLM_SYNC) {
1728                 /* if it is synchronous call do minimum mem alloc, as it could
1729                  * be triggered from kernel shrinker
1730                  */
1731                 struct ldlm_bl_work_item blwi;
1732                 memset(&blwi, 0, sizeof(blwi));
1733                 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1734                 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1735         } else {
1736                 struct ldlm_bl_work_item *blwi;
1737                 OBD_ALLOC(blwi, sizeof(*blwi));
1738                 if (blwi == NULL)
1739                         RETURN(-ENOMEM);
1740                 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1741
1742                 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1743         }
1744 }
1745
1746 #endif
1747
1748 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1749                            struct ldlm_lock *lock)
1750 {
1751 #ifdef __KERNEL__
1752         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1753 #else
1754         RETURN(-ENOSYS);
1755 #endif
1756 }
1757
1758 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1759                            cfs_list_t *cancels, int count, int mode)
1760 {
1761 #ifdef __KERNEL__
1762         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1763 #else
1764         RETURN(-ENOSYS);
1765 #endif
1766 }
1767
1768 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1769 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1770 {
1771         struct obd_device *obd = req->rq_export->exp_obd;
1772         char *key;
1773         void *val;
1774         int keylen, vallen;
1775         int rc = -ENOSYS;
1776         ENTRY;
1777
1778         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1779
1780         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1781
1782         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1783         if (key == NULL) {
1784                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1785                 RETURN(-EFAULT);
1786         }
1787         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1788                                       RCL_CLIENT);
1789         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1790         if (val == NULL) {
1791                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1792                 RETURN(-EFAULT);
1793         }
1794         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1795                                       RCL_CLIENT);
1796
1797         /* We are responsible for swabbing contents of val */
1798
1799         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1800                 /* Pass it on to mdc (the "export" in this case) */
1801                 rc = obd_set_info_async(req->rq_export,
1802                                         sizeof(KEY_HSM_COPYTOOL_SEND),
1803                                         KEY_HSM_COPYTOOL_SEND,
1804                                         vallen, val, NULL);
1805         else
1806                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1807
1808         return rc;
1809 }
1810
1811 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1812                                         const char *msg, int rc,
1813                                         struct lustre_handle *handle)
1814 {
1815         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1816                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
1817                   msg, libcfs_id2str(req->rq_peer), rc,
1818                   handle ? handle->cookie : 0);
1819         if (req->rq_no_reply)
1820                 CWARN("No reply was sent, maybe cause bug 21636.\n");
1821         else if (rc)
1822                 CWARN("Send reply failed, maybe cause bug 21636.\n");
1823 }
1824
1825 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1826 static int ldlm_callback_handler(struct ptlrpc_request *req)
1827 {
1828         struct ldlm_namespace *ns;
1829         struct ldlm_request *dlm_req;
1830         struct ldlm_lock *lock;
1831         int rc;
1832         ENTRY;
1833
1834         /* Requests arrive in sender's byte order.  The ptlrpc service
1835          * handler has already checked and, if necessary, byte-swapped the
1836          * incoming request message body, but I am responsible for the
1837          * message buffers. */
1838
1839         /* do nothing for sec context finalize */
1840         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1841                 RETURN(0);
1842
1843         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1844
1845         if (req->rq_export == NULL) {
1846                 rc = ldlm_callback_reply(req, -ENOTCONN);
1847                 ldlm_callback_errmsg(req, "Operate on unconnected server",
1848                                      rc, NULL);
1849                 RETURN(0);
1850         }
1851
1852         LASSERT(req->rq_export != NULL);
1853         LASSERT(req->rq_export->exp_obd != NULL);
1854
1855         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1856         case LDLM_BL_CALLBACK:
1857                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1858                         RETURN(0);
1859                 break;
1860         case LDLM_CP_CALLBACK:
1861                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1862                         RETURN(0);
1863                 break;
1864         case LDLM_GL_CALLBACK:
1865                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1866                         RETURN(0);
1867                 break;
1868         case LDLM_SET_INFO:
1869                 rc = ldlm_handle_setinfo(req);
1870                 ldlm_callback_reply(req, rc);
1871                 RETURN(0);
1872         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1873                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1874                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1875                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1876                         RETURN(0);
1877                 rc = llog_origin_handle_cancel(req);
1878                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1879                         RETURN(0);
1880                 ldlm_callback_reply(req, rc);
1881                 RETURN(0);
1882         case OBD_QC_CALLBACK:
1883                 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1884                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1885                         RETURN(0);
1886                 rc = target_handle_qc_callback(req);
1887                 ldlm_callback_reply(req, rc);
1888                 RETURN(0);
1889         case QUOTA_DQACQ:
1890         case QUOTA_DQREL:
1891                 /* reply in handler */
1892                 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1893                 rc = target_handle_dqacq_callback(req);
1894                 RETURN(0);
1895         case LLOG_ORIGIN_HANDLE_CREATE:
1896                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1897                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1898                         RETURN(0);
1899                 rc = llog_origin_handle_create(req);
1900                 ldlm_callback_reply(req, rc);
1901                 RETURN(0);
1902         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1903                 req_capsule_set(&req->rq_pill,
1904                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1905                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1906                         RETURN(0);
1907                 rc = llog_origin_handle_next_block(req);
1908                 ldlm_callback_reply(req, rc);
1909                 RETURN(0);
1910         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1911                 req_capsule_set(&req->rq_pill,
1912                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1913                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1914                         RETURN(0);
1915                 rc = llog_origin_handle_read_header(req);
1916                 ldlm_callback_reply(req, rc);
1917                 RETURN(0);
1918         case LLOG_ORIGIN_HANDLE_CLOSE:
1919                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1920                         RETURN(0);
1921                 rc = llog_origin_handle_close(req);
1922                 ldlm_callback_reply(req, rc);
1923                 RETURN(0);
1924         default:
1925                 CERROR("unknown opcode %u\n",
1926                        lustre_msg_get_opc(req->rq_reqmsg));
1927                 ldlm_callback_reply(req, -EPROTO);
1928                 RETURN(0);
1929         }
1930
1931         ns = req->rq_export->exp_obd->obd_namespace;
1932         LASSERT(ns != NULL);
1933
1934         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1935
1936         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1937         if (dlm_req == NULL) {
1938                 rc = ldlm_callback_reply(req, -EPROTO);
1939                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
1940                                      NULL);
1941                 RETURN(0);
1942         }
1943
1944         /* Force a known safe race, send a cancel to the server for a lock
1945          * which the server has already started a blocking callback on. */
1946         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1947             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1948                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1949                 if (rc < 0)
1950                         CERROR("ldlm_cli_cancel: %d\n", rc);
1951         }
1952
1953         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
1954         if (!lock) {
1955                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1956                        "disappeared\n", dlm_req->lock_handle[0].cookie);
1957                 rc = ldlm_callback_reply(req, -EINVAL);
1958                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
1959                                      &dlm_req->lock_handle[0]);
1960                 RETURN(0);
1961         }
1962
1963         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
1964             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
1965                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
1966
1967         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1968         lock_res_and_lock(lock);
1969         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1970         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1971                 /* If somebody cancels lock and cache is already dropped,
1972                  * or lock is failed before cp_ast received on client,
1973                  * we can tell the server we have no lock. Otherwise, we
1974                  * should send cancel after dropping the cache. */
1975                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
1976                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
1977                     (lock->l_flags & LDLM_FL_FAILED)) {
1978                         LDLM_DEBUG(lock, "callback on lock "
1979                                    LPX64" - lock disappeared\n",
1980                                    dlm_req->lock_handle[0].cookie);
1981                         unlock_res_and_lock(lock);
1982                         LDLM_LOCK_RELEASE(lock);
1983                         rc = ldlm_callback_reply(req, -EINVAL);
1984                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
1985                                              &dlm_req->lock_handle[0]);
1986                         RETURN(0);
1987                 }
1988                 /* BL_AST locks are not needed in lru.
1989                  * let ldlm_cancel_lru() be fast. */
1990                 ldlm_lock_remove_from_lru(lock);
1991                 lock->l_flags |= LDLM_FL_BL_AST;
1992         }
1993         unlock_res_and_lock(lock);
1994
1995         /* We want the ost thread to get this reply so that it can respond
1996          * to ost requests (write cache writeback) that might be triggered
1997          * in the callback.
1998          *
1999          * But we'd also like to be able to indicate in the reply that we're
2000          * cancelling right now, because it's unused, or have an intent result
2001          * in the reply, so we might have to push the responsibility for sending
2002          * the reply down into the AST handlers, alas. */
2003
2004         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2005         case LDLM_BL_CALLBACK:
2006                 CDEBUG(D_INODE, "blocking ast\n");
2007                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
2008                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
2009                         rc = ldlm_callback_reply(req, 0);
2010                         if (req->rq_no_reply || rc)
2011                                 ldlm_callback_errmsg(req, "Normal process", rc,
2012                                                      &dlm_req->lock_handle[0]);
2013                 }
2014                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
2015                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
2016                 break;
2017         case LDLM_CP_CALLBACK:
2018                 CDEBUG(D_INODE, "completion ast\n");
2019                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
2020                 ldlm_callback_reply(req, 0);
2021                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
2022                 break;
2023         case LDLM_GL_CALLBACK:
2024                 CDEBUG(D_INODE, "glimpse ast\n");
2025                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
2026                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
2027                 break;
2028         default:
2029                 LBUG();                         /* checked above */
2030         }
2031
2032         RETURN(0);
2033 }
2034
2035 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2036 {
2037         int rc;
2038         ENTRY;
2039
2040         /* Requests arrive in sender's byte order.  The ptlrpc service
2041          * handler has already checked and, if necessary, byte-swapped the
2042          * incoming request message body, but I am responsible for the
2043          * message buffers. */
2044
2045         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2046
2047         if (req->rq_export == NULL) {
2048                 struct ldlm_request *dlm_req;
2049
2050                 CERROR("%s from %s arrived at %lu with bad export cookie "
2051                        LPU64"\n",
2052                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2053                        libcfs_nid2str(req->rq_peer.nid),
2054                        req->rq_arrival_time.tv_sec,
2055                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2056
2057                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2058                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2059                         dlm_req = req_capsule_client_get(&req->rq_pill,
2060                                                          &RMF_DLM_REQ);
2061                         if (dlm_req != NULL)
2062                                 ldlm_lock_dump_handle(D_ERROR,
2063                                                       &dlm_req->lock_handle[0]);
2064                 }
2065                 ldlm_callback_reply(req, -ENOTCONN);
2066                 RETURN(0);
2067         }
2068
2069         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2070
2071         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2072         case LDLM_CANCEL:
2073                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2074                 CDEBUG(D_INODE, "cancel\n");
2075                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2076                         RETURN(0);
2077                 rc = ldlm_handle_cancel(req);
2078                 if (rc)
2079                         break;
2080                 RETURN(0);
2081         case OBD_LOG_CANCEL:
2082                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2083                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2084                         RETURN(0);
2085                 rc = llog_origin_handle_cancel(req);
2086                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2087                         RETURN(0);
2088                 ldlm_callback_reply(req, rc);
2089                 RETURN(0);
2090         default:
2091                 CERROR("invalid opcode %d\n",
2092                        lustre_msg_get_opc(req->rq_reqmsg));
2093                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2094                 ldlm_callback_reply(req, -EINVAL);
2095         }
2096
2097         RETURN(0);
2098 }
2099
2100 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2101                         cfs_hlist_node_t *hnode, void *data)
2102
2103 {
2104         cfs_list_t         *rpc_list = data;
2105         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2106
2107         lock_res_and_lock(lock);
2108
2109         if (lock->l_req_mode != lock->l_granted_mode) {
2110                 unlock_res_and_lock(lock);
2111                 return 0;
2112         }
2113
2114         LASSERT(lock->l_resource);
2115         if (lock->l_resource->lr_type != LDLM_IBITS &&
2116             lock->l_resource->lr_type != LDLM_PLAIN) {
2117                 unlock_res_and_lock(lock);
2118                 return 0;
2119         }
2120
2121         if (lock->l_flags & LDLM_FL_AST_SENT) {
2122                 unlock_res_and_lock(lock);
2123                 return 0;
2124         }
2125
2126         LASSERT(lock->l_blocking_ast);
2127         LASSERT(!lock->l_blocking_lock);
2128
2129         lock->l_flags |= LDLM_FL_AST_SENT;
2130         if (lock->l_export && lock->l_export->exp_lock_hash &&
2131             !cfs_hlist_unhashed(&lock->l_exp_hash))
2132                 cfs_hash_del(lock->l_export->exp_lock_hash,
2133                              &lock->l_remote_handle, &lock->l_exp_hash);
2134         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2135         LDLM_LOCK_GET(lock);
2136
2137         unlock_res_and_lock(lock);
2138         return 0;
2139 }
2140
2141 void ldlm_revoke_export_locks(struct obd_export *exp)
2142 {
2143         cfs_list_t  rpc_list;
2144         ENTRY;
2145
2146         CFS_INIT_LIST_HEAD(&rpc_list);
2147         cfs_hash_for_each_empty(exp->exp_lock_hash,
2148                                 ldlm_revoke_lock_cb, &rpc_list);
2149         ldlm_run_ast_work(exp->exp_obd->obd_namespace, &rpc_list,
2150                           LDLM_WORK_REVOKE_AST);
2151
2152         EXIT;
2153 }
2154
2155 #ifdef __KERNEL__
2156 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2157 {
2158         struct ldlm_bl_work_item *blwi = NULL;
2159         static unsigned int num_bl = 0;
2160
2161         cfs_spin_lock(&blp->blp_lock);
2162         /* process a request from the blp_list at least every blp_num_threads */
2163         if (!cfs_list_empty(&blp->blp_list) &&
2164             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2165                 blwi = cfs_list_entry(blp->blp_list.next,
2166                                       struct ldlm_bl_work_item, blwi_entry);
2167         else
2168                 if (!cfs_list_empty(&blp->blp_prio_list))
2169                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2170                                               struct ldlm_bl_work_item,
2171                                               blwi_entry);
2172
2173         if (blwi) {
2174                 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2175                         num_bl = 0;
2176                 cfs_list_del(&blwi->blwi_entry);
2177         }
2178         cfs_spin_unlock(&blp->blp_lock);
2179
2180         return blwi;
2181 }
2182
2183 /* This only contains temporary data until the thread starts */
2184 struct ldlm_bl_thread_data {
2185         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2186         struct ldlm_bl_pool     *bltd_blp;
2187         cfs_completion_t        bltd_comp;
2188         int                     bltd_num;
2189 };
2190
2191 static int ldlm_bl_thread_main(void *arg);
2192
2193 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2194 {
2195         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2196         int rc;
2197
2198         cfs_init_completion(&bltd.bltd_comp);
2199         rc = cfs_create_thread(ldlm_bl_thread_main, &bltd, 0);
2200         if (rc < 0) {
2201                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2202                        cfs_atomic_read(&blp->blp_num_threads), rc);
2203                 return rc;
2204         }
2205         cfs_wait_for_completion(&bltd.bltd_comp);
2206
2207         return 0;
2208 }
2209
2210 static int ldlm_bl_thread_main(void *arg)
2211 {
2212         struct ldlm_bl_pool *blp;
2213         ENTRY;
2214
2215         {
2216                 struct ldlm_bl_thread_data *bltd = arg;
2217
2218                 blp = bltd->bltd_blp;
2219
2220                 bltd->bltd_num =
2221                         cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2222                 cfs_atomic_inc(&blp->blp_busy_threads);
2223
2224                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2225                         "ldlm_bl_%02d", bltd->bltd_num);
2226                 cfs_daemonize(bltd->bltd_name);
2227
2228                 cfs_complete(&bltd->bltd_comp);
2229                 /* cannot use bltd after this, it is only on caller's stack */
2230         }
2231
2232         while (1) {
2233                 struct l_wait_info lwi = { 0 };
2234                 struct ldlm_bl_work_item *blwi = NULL;
2235
2236                 blwi = ldlm_bl_get_work(blp);
2237
2238                 if (blwi == NULL) {
2239                         int busy;
2240
2241                         cfs_atomic_dec(&blp->blp_busy_threads);
2242                         l_wait_event_exclusive(blp->blp_waitq,
2243                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2244                                          &lwi);
2245                         busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2246
2247                         if (blwi->blwi_ns == NULL)
2248                                 /* added by ldlm_cleanup() */
2249                                 break;
2250
2251                         /* Not fatal if racy and have a few too many threads */
2252                         if (unlikely(busy < blp->blp_max_threads &&
2253                             busy >= cfs_atomic_read(&blp->blp_num_threads) &&
2254                             !blwi->blwi_mem_pressure))
2255                                 /* discard the return value, we tried */
2256                                 ldlm_bl_thread_start(blp);
2257                 } else {
2258                         if (blwi->blwi_ns == NULL)
2259                                 /* added by ldlm_cleanup() */
2260                                 break;
2261                 }
2262                 if (blwi->blwi_mem_pressure)
2263                         cfs_memory_pressure_set();
2264
2265                 if (blwi->blwi_count) {
2266                         int count;
2267                         /* The special case when we cancel locks in lru
2268                          * asynchronously, we pass the list of locks here.
2269                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2270                          * canceled locally yet. */
2271                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2272                                                            blwi->blwi_count,
2273                                                            LCF_BL_AST);
2274                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2275                 } else {
2276                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2277                                                 blwi->blwi_lock);
2278                 }
2279                 if (blwi->blwi_mem_pressure)
2280                         cfs_memory_pressure_clr();
2281
2282                 if (blwi->blwi_mode == LDLM_ASYNC)
2283                         OBD_FREE(blwi, sizeof(*blwi));
2284                 else
2285                         cfs_complete(&blwi->blwi_comp);
2286         }
2287
2288         cfs_atomic_dec(&blp->blp_busy_threads);
2289         cfs_atomic_dec(&blp->blp_num_threads);
2290         cfs_complete(&blp->blp_comp);
2291         RETURN(0);
2292 }
2293
2294 #endif
2295
2296 static int ldlm_setup(void);
2297 static int ldlm_cleanup(void);
2298
2299 int ldlm_get_ref(void)
2300 {
2301         int rc = 0;
2302         ENTRY;
2303         cfs_mutex_down(&ldlm_ref_sem);
2304         if (++ldlm_refcount == 1) {
2305                 rc = ldlm_setup();
2306                 if (rc)
2307                         ldlm_refcount--;
2308         }
2309         cfs_mutex_up(&ldlm_ref_sem);
2310
2311         RETURN(rc);
2312 }
2313
2314 void ldlm_put_ref(void)
2315 {
2316         ENTRY;
2317         cfs_mutex_down(&ldlm_ref_sem);
2318         if (ldlm_refcount == 1) {
2319                 int rc = ldlm_cleanup();
2320                 if (rc)
2321                         CERROR("ldlm_cleanup failed: %d\n", rc);
2322                 else
2323                         ldlm_refcount--;
2324         } else {
2325                 ldlm_refcount--;
2326         }
2327         cfs_mutex_up(&ldlm_ref_sem);
2328
2329         EXIT;
2330 }
2331
2332 /*
2333  * Export handle<->lock hash operations.
2334  */
2335 static unsigned
2336 ldlm_export_lock_hash(cfs_hash_t *hs, const void *key, unsigned mask)
2337 {
2338         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2339 }
2340
2341 static void *
2342 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2343 {
2344         struct ldlm_lock *lock;
2345
2346         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2347         return &lock->l_remote_handle;
2348 }
2349
2350 static void
2351 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2352 {
2353         struct ldlm_lock     *lock;
2354
2355         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2356         lock->l_remote_handle = *(struct lustre_handle *)key;
2357 }
2358
2359 static int
2360 ldlm_export_lock_keycmp(const void *key, cfs_hlist_node_t *hnode)
2361 {
2362         return lustre_handle_equal(ldlm_export_lock_key(hnode), key);
2363 }
2364
2365 static void *
2366 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2367 {
2368         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2369 }
2370
2371 static void
2372 ldlm_export_lock_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2373 {
2374         struct ldlm_lock *lock;
2375
2376         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2377         LDLM_LOCK_GET(lock);
2378 }
2379
2380 static void
2381 ldlm_export_lock_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
2382 {
2383         struct ldlm_lock *lock;
2384
2385         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2386         LDLM_LOCK_RELEASE(lock);
2387 }
2388
2389 static cfs_hash_ops_t ldlm_export_lock_ops = {
2390         .hs_hash        = ldlm_export_lock_hash,
2391         .hs_key         = ldlm_export_lock_key,
2392         .hs_keycmp      = ldlm_export_lock_keycmp,
2393         .hs_keycpy      = ldlm_export_lock_keycpy,
2394         .hs_object      = ldlm_export_lock_object,
2395         .hs_get         = ldlm_export_lock_get,
2396         .hs_put         = ldlm_export_lock_put,
2397         .hs_put_locked  = ldlm_export_lock_put,
2398 };
2399
2400 int ldlm_init_export(struct obd_export *exp)
2401 {
2402         ENTRY;
2403
2404         exp->exp_lock_hash =
2405                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2406                                 HASH_EXP_LOCK_CUR_BITS,
2407                                 HASH_EXP_LOCK_MAX_BITS,
2408                                 HASH_EXP_LOCK_BKT_BITS, 0,
2409                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2410                                 &ldlm_export_lock_ops,
2411                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2412                                 CFS_HASH_NBLK_CHANGE);
2413
2414         if (!exp->exp_lock_hash)
2415                 RETURN(-ENOMEM);
2416
2417         RETURN(0);
2418 }
2419 EXPORT_SYMBOL(ldlm_init_export);
2420
2421 void ldlm_destroy_export(struct obd_export *exp)
2422 {
2423         ENTRY;
2424         cfs_hash_putref(exp->exp_lock_hash);
2425         exp->exp_lock_hash = NULL;
2426         EXIT;
2427 }
2428 EXPORT_SYMBOL(ldlm_destroy_export);
2429
2430 static int ldlm_setup(void)
2431 {
2432         struct ldlm_bl_pool *blp;
2433         int rc = 0;
2434         int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2435         int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2436 #ifdef __KERNEL__
2437         int i;
2438 #endif
2439         ENTRY;
2440
2441         if (ldlm_state != NULL)
2442                 RETURN(-EALREADY);
2443
2444         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2445         if (ldlm_state == NULL)
2446                 RETURN(-ENOMEM);
2447
2448 #ifdef LPROCFS
2449         rc = ldlm_proc_setup();
2450         if (rc != 0)
2451                 GOTO(out_free, rc);
2452 #endif
2453
2454 #ifdef __KERNEL__
2455         if (ldlm_num_threads) {
2456                 /* If ldlm_num_threads is set, it is the min and the max. */
2457                 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2458                         ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2459                 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2460                         ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2461                 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2462         }
2463 #endif
2464
2465         ldlm_state->ldlm_cb_service =
2466                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2467                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2468                                 LDLM_CB_REPLY_PORTAL, 2,
2469                                 ldlm_callback_handler, "ldlm_cbd",
2470                                 ldlm_svc_proc_dir, NULL,
2471                                 ldlm_min_threads, ldlm_max_threads,
2472                                 "ldlm_cb",
2473                                 LCT_MD_THREAD|LCT_DT_THREAD, NULL);
2474
2475         if (!ldlm_state->ldlm_cb_service) {
2476                 CERROR("failed to start service\n");
2477                 GOTO(out_proc, rc = -ENOMEM);
2478         }
2479
2480         ldlm_state->ldlm_cancel_service =
2481                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2482                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2483                                 LDLM_CANCEL_REPLY_PORTAL, 6,
2484                                 ldlm_cancel_handler, "ldlm_canceld",
2485                                 ldlm_svc_proc_dir, NULL,
2486                                 ldlm_min_threads, ldlm_max_threads,
2487                                 "ldlm_cn",
2488                                 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
2489                                 NULL);
2490
2491         if (!ldlm_state->ldlm_cancel_service) {
2492                 CERROR("failed to start service\n");
2493                 GOTO(out_proc, rc = -ENOMEM);
2494         }
2495
2496         OBD_ALLOC(blp, sizeof(*blp));
2497         if (blp == NULL)
2498                 GOTO(out_proc, rc = -ENOMEM);
2499         ldlm_state->ldlm_bl_pool = blp;
2500
2501         cfs_spin_lock_init(&blp->blp_lock);
2502         CFS_INIT_LIST_HEAD(&blp->blp_list);
2503         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2504         cfs_waitq_init(&blp->blp_waitq);
2505         cfs_atomic_set(&blp->blp_num_threads, 0);
2506         cfs_atomic_set(&blp->blp_busy_threads, 0);
2507         blp->blp_min_threads = ldlm_min_threads;
2508         blp->blp_max_threads = ldlm_max_threads;
2509
2510 #ifdef __KERNEL__
2511         for (i = 0; i < blp->blp_min_threads; i++) {
2512                 rc = ldlm_bl_thread_start(blp);
2513                 if (rc < 0)
2514                         GOTO(out_thread, rc);
2515         }
2516
2517         rc = ptlrpc_start_threads(ldlm_state->ldlm_cancel_service);
2518         if (rc)
2519                 GOTO(out_thread, rc);
2520
2521         rc = ptlrpc_start_threads(ldlm_state->ldlm_cb_service);
2522         if (rc)
2523                 GOTO(out_thread, rc);
2524
2525         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2526         expired_lock_thread.elt_state = ELT_STOPPED;
2527         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2528
2529         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2530         cfs_spin_lock_init(&waiting_locks_spinlock);
2531         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2532
2533         rc = cfs_create_thread(expired_lock_main, NULL, CFS_DAEMON_FLAGS);
2534         if (rc < 0) {
2535                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2536                 GOTO(out_thread, rc);
2537         }
2538
2539         cfs_wait_event(expired_lock_thread.elt_waitq,
2540                        expired_lock_thread.elt_state == ELT_READY);
2541 #endif
2542
2543 #ifdef __KERNEL__
2544         rc = ldlm_pools_init();
2545         if (rc)
2546                 GOTO(out_thread, rc);
2547 #endif
2548         RETURN(0);
2549
2550 #ifdef __KERNEL__
2551  out_thread:
2552         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2553         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2554 #endif
2555
2556  out_proc:
2557 #ifdef LPROCFS
2558         ldlm_proc_cleanup();
2559  out_free:
2560 #endif
2561         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2562         ldlm_state = NULL;
2563         return rc;
2564 }
2565
2566 static int ldlm_cleanup(void)
2567 {
2568 #ifdef __KERNEL__
2569         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2570 #endif
2571         ENTRY;
2572
2573         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2574             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2575                 CERROR("ldlm still has namespaces; clean these up first.\n");
2576                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2577                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2578                 RETURN(-EBUSY);
2579         }
2580
2581 #ifdef __KERNEL__
2582         ldlm_pools_fini();
2583 #endif
2584
2585 #ifdef __KERNEL__
2586         while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2587                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2588
2589                 cfs_init_completion(&blp->blp_comp);
2590
2591                 cfs_spin_lock(&blp->blp_lock);
2592                 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2593                 cfs_waitq_signal(&blp->blp_waitq);
2594                 cfs_spin_unlock(&blp->blp_lock);
2595
2596                 cfs_wait_for_completion(&blp->blp_comp);
2597         }
2598         OBD_FREE(blp, sizeof(*blp));
2599
2600         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2601         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2602         ldlm_proc_cleanup();
2603
2604         expired_lock_thread.elt_state = ELT_TERMINATE;
2605         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2606         cfs_wait_event(expired_lock_thread.elt_waitq,
2607                        expired_lock_thread.elt_state == ELT_STOPPED);
2608 #else
2609         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2610         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2611 #endif
2612
2613         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2614         ldlm_state = NULL;
2615
2616         RETURN(0);
2617 }
2618
2619 int __init ldlm_init(void)
2620 {
2621         cfs_init_mutex(&ldlm_ref_sem);
2622         cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2623         cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2624         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2625                                                sizeof(struct ldlm_resource), 0,
2626                                                CFS_SLAB_HWCACHE_ALIGN);
2627         if (ldlm_resource_slab == NULL)
2628                 return -ENOMEM;
2629
2630         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2631                               sizeof(struct ldlm_lock), 0,
2632                               CFS_SLAB_HWCACHE_ALIGN | CFS_SLAB_DESTROY_BY_RCU);
2633         if (ldlm_lock_slab == NULL) {
2634                 cfs_mem_cache_destroy(ldlm_resource_slab);
2635                 return -ENOMEM;
2636         }
2637
2638         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2639                                         sizeof(struct ldlm_interval),
2640                                         0, CFS_SLAB_HWCACHE_ALIGN);
2641         if (ldlm_interval_slab == NULL) {
2642                 cfs_mem_cache_destroy(ldlm_resource_slab);
2643                 cfs_mem_cache_destroy(ldlm_lock_slab);
2644                 return -ENOMEM;
2645         }
2646 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2647         class_export_dump_hook = ldlm_dump_export_locks;
2648 #endif
2649         return 0;
2650 }
2651
2652 void __exit ldlm_exit(void)
2653 {
2654         int rc;
2655         if (ldlm_refcount)
2656                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2657         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2658         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2659 #ifdef __KERNEL__
2660         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2661          * synchronize_rcu() to wait a grace period elapsed, so that
2662          * ldlm_lock_free() get a chance to be called. */
2663         synchronize_rcu();
2664 #endif
2665         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2666         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2667         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2668         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2669 }
2670
2671 /* ldlm_extent.c */
2672 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2673
2674 /* ldlm_lock.c */
2675 EXPORT_SYMBOL(ldlm_get_processing_policy);
2676 EXPORT_SYMBOL(ldlm_lock2desc);
2677 EXPORT_SYMBOL(ldlm_register_intent);
2678 EXPORT_SYMBOL(ldlm_lockname);
2679 EXPORT_SYMBOL(ldlm_typename);
2680 EXPORT_SYMBOL(ldlm_lock2handle);
2681 EXPORT_SYMBOL(__ldlm_handle2lock);
2682 EXPORT_SYMBOL(ldlm_lock_get);
2683 EXPORT_SYMBOL(ldlm_lock_put);
2684 EXPORT_SYMBOL(ldlm_lock_match);
2685 EXPORT_SYMBOL(ldlm_lock_cancel);
2686 EXPORT_SYMBOL(ldlm_lock_addref);
2687 EXPORT_SYMBOL(ldlm_lock_addref_try);
2688 EXPORT_SYMBOL(ldlm_lock_decref);
2689 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2690 EXPORT_SYMBOL(ldlm_lock_change_resource);
2691 EXPORT_SYMBOL(ldlm_it2str);
2692 EXPORT_SYMBOL(ldlm_lock_dump);
2693 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2694 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2695 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2696 EXPORT_SYMBOL(ldlm_lock_allow_match);
2697 EXPORT_SYMBOL(ldlm_lock_downgrade);
2698 EXPORT_SYMBOL(ldlm_lock_convert);
2699
2700 /* ldlm_request.c */
2701 EXPORT_SYMBOL(ldlm_completion_ast_async);
2702 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2703 EXPORT_SYMBOL(ldlm_completion_ast);
2704 EXPORT_SYMBOL(ldlm_blocking_ast);
2705 EXPORT_SYMBOL(ldlm_glimpse_ast);
2706 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2707 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2708 EXPORT_SYMBOL(ldlm_prep_elc_req);
2709 EXPORT_SYMBOL(ldlm_cli_convert);
2710 EXPORT_SYMBOL(ldlm_cli_enqueue);
2711 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2712 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2713 EXPORT_SYMBOL(ldlm_cli_cancel);
2714 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2715 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2716 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2717 EXPORT_SYMBOL(ldlm_replay_locks);
2718 EXPORT_SYMBOL(ldlm_resource_foreach);
2719 EXPORT_SYMBOL(ldlm_namespace_foreach);
2720 EXPORT_SYMBOL(ldlm_resource_iterate);
2721 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2722 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2723 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2724
2725 /* ldlm_lockd.c */
2726 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2727 EXPORT_SYMBOL(ldlm_server_completion_ast);
2728 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2729 EXPORT_SYMBOL(ldlm_handle_enqueue);
2730 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2731 EXPORT_SYMBOL(ldlm_handle_cancel);
2732 EXPORT_SYMBOL(ldlm_request_cancel);
2733 EXPORT_SYMBOL(ldlm_handle_convert);
2734 EXPORT_SYMBOL(ldlm_handle_convert0);
2735 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2736 EXPORT_SYMBOL(ldlm_get_ref);
2737 EXPORT_SYMBOL(ldlm_put_ref);
2738 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2739 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2740
2741 /* ldlm_resource.c */
2742 EXPORT_SYMBOL(ldlm_namespace_new);
2743 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2744 EXPORT_SYMBOL(ldlm_namespace_free);
2745 EXPORT_SYMBOL(ldlm_namespace_dump);
2746 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2747 EXPORT_SYMBOL(ldlm_resource_get);
2748 EXPORT_SYMBOL(ldlm_resource_putref);
2749 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2750
2751 /* ldlm_lib.c */
2752 EXPORT_SYMBOL(client_import_add_conn);
2753 EXPORT_SYMBOL(client_import_del_conn);
2754 EXPORT_SYMBOL(client_obd_setup);
2755 EXPORT_SYMBOL(client_obd_cleanup);
2756 EXPORT_SYMBOL(client_connect_import);
2757 EXPORT_SYMBOL(client_disconnect_export);
2758 EXPORT_SYMBOL(server_disconnect_export);
2759 EXPORT_SYMBOL(target_stop_recovery_thread);
2760 EXPORT_SYMBOL(target_handle_connect);
2761 EXPORT_SYMBOL(target_cleanup_recovery);
2762 EXPORT_SYMBOL(target_destroy_export);
2763 EXPORT_SYMBOL(target_cancel_recovery_timer);
2764 EXPORT_SYMBOL(target_send_reply);
2765 EXPORT_SYMBOL(target_queue_recovery_request);
2766 EXPORT_SYMBOL(target_handle_ping);
2767 EXPORT_SYMBOL(target_pack_pool_reply);
2768 EXPORT_SYMBOL(target_handle_disconnect);
2769
2770 /* l_lock.c */
2771 EXPORT_SYMBOL(lock_res_and_lock);
2772 EXPORT_SYMBOL(unlock_res_and_lock);