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