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                 ldlm_lock_to_ns(lock)->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         /* server namespace, doesn't need 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         }
 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(ldlm_lock_to_ns_at(lock),
 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(ldlm_lock_to_ns_at(lock)));
 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         /* server namespace, doesn't need lock */
 942         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
 943                              lock->l_resource->lr_lvb_len);
 944         res = lock->l_resource;
 945         ptlrpc_request_set_replen(req);
 946
 947
 948         req->rq_send_state = LUSTRE_IMP_FULL;
 949         /* ptlrpc_request_alloc_pack already set timeout */
 950         if (AT_OFF)
 951                 req->rq_timeout = ldlm_get_rq_timeout();
 952
 953         if (lock->l_export && lock->l_export->exp_nid_stats &&
 954             lock->l_export->exp_nid_stats->nid_ldlm_stats)
 955                 lprocfs_counter_incr(lock->l_export->exp_nid_stats->nid_ldlm_stats,
 956                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
 957
 958         rc = ptlrpc_queue_wait(req);
 959         if (rc == -ELDLM_NO_LOCK_DATA)
 960                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 961         else if (rc != 0)
 962                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 963         else
 964                 rc = ldlm_res_lvbo_update(res, req, 1);
 965
 966         ptlrpc_req_finished(req);
 967         if (rc == -ERESTART)
 968                 ldlm_reprocess_all(res);
 969
 970         RETURN(rc);
 971 }
 972
 973 #ifdef __KERNEL__
 974 extern unsigned long long lu_time_stamp_get(void);
 975 #else
 976 #define lu_time_stamp_get() time(NULL)
 977 #endif
 978
 979 static void ldlm_svc_get_eopc(const struct ldlm_request *dlm_req,
 980                        struct lprocfs_stats *srv_stats)
 981 {
 982         int lock_type = 0, op = 0;
 983
 984         lock_type = dlm_req->lock_desc.l_resource.lr_type;
 985
 986         switch (lock_type) {
 987         case LDLM_PLAIN:
 988                 op = PTLRPC_LAST_CNTR + LDLM_PLAIN_ENQUEUE;
 989                 break;
 990         case LDLM_EXTENT:
 991                 if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT)
 992                         op = PTLRPC_LAST_CNTR + LDLM_GLIMPSE_ENQUEUE;
 993                 else
 994                         op = PTLRPC_LAST_CNTR + LDLM_EXTENT_ENQUEUE;
 995                 break;
 996         case LDLM_FLOCK:
 997                 op = PTLRPC_LAST_CNTR + LDLM_FLOCK_ENQUEUE;
 998                 break;
 999         case LDLM_IBITS:
1000                 op = PTLRPC_LAST_CNTR + LDLM_IBITS_ENQUEUE;
1001                 break;
1002         default:
1003                 op = 0;
1004                 break;
1005         }
1006
1007         if (op)
1008                 lprocfs_counter_incr(srv_stats, op);
1009
1010         return;
1011 }
1012
1013 /*
1014  * Main server-side entry point into LDLM. This is called by ptlrpc service
1015  * threads to carry out client lock enqueueing requests.
1016  */
1017 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
1018                          struct ptlrpc_request *req,
1019                          const struct ldlm_request *dlm_req,
1020                          const struct ldlm_callback_suite *cbs)
1021 {
1022         struct ldlm_reply *dlm_rep;
1023         __u32 flags;
1024         ldlm_error_t err = ELDLM_OK;
1025         struct ldlm_lock *lock = NULL;
1026         void *cookie = NULL;
1027         int rc = 0;
1028         ENTRY;
1029
1030         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
1031
1032         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
1033         flags = dlm_req->lock_flags;
1034
1035         LASSERT(req->rq_export);
1036
1037         if (req->rq_rqbd->rqbd_service->srv_stats)
1038                 ldlm_svc_get_eopc(dlm_req,
1039                                   req->rq_rqbd->rqbd_service->srv_stats);
1040
1041         if (req->rq_export && req->rq_export->exp_nid_stats &&
1042             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1043                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1044                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
1045
1046         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
1047                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
1048                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
1049                           dlm_req->lock_desc.l_resource.lr_type);
1050                 GOTO(out, rc = -EFAULT);
1051         }
1052
1053         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
1054                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
1055                      dlm_req->lock_desc.l_req_mode &
1056                      (dlm_req->lock_desc.l_req_mode-1))) {
1057                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
1058                           dlm_req->lock_desc.l_req_mode);
1059                 GOTO(out, rc = -EFAULT);
1060         }
1061
1062         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
1063                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1064                              LDLM_PLAIN)) {
1065                         DEBUG_REQ(D_ERROR, req,
1066                                   "PLAIN lock request from IBITS client?");
1067                         GOTO(out, rc = -EPROTO);
1068                 }
1069         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
1070                             LDLM_IBITS)) {
1071                 DEBUG_REQ(D_ERROR, req,
1072                           "IBITS lock request from unaware client?");
1073                 GOTO(out, rc = -EPROTO);
1074         }
1075
1076 #if 0
1077         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
1078            against server's _CONNECT_SUPPORTED flags? (I don't want to use
1079            ibits for mgc/mgs) */
1080
1081         /* INODEBITS_INTEROP: Perform conversion from plain lock to
1082          * inodebits lock if client does not support them. */
1083         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
1084             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
1085                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
1086                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
1087                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1088                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
1089                         dlm_req->lock_desc.l_req_mode = LCK_CR;
1090         }
1091 #endif
1092
1093         if (unlikely(flags & LDLM_FL_REPLAY)) {
1094                 /* Find an existing lock in the per-export lock hash */
1095                 lock = cfs_hash_lookup(req->rq_export->exp_lock_hash,
1096                                        (void *)&dlm_req->lock_handle[0]);
1097                 if (lock != NULL) {
1098                         DEBUG_REQ(D_DLMTRACE, req, "found existing lock cookie "
1099                                   LPX64, lock->l_handle.h_cookie);
1100                         GOTO(existing_lock, rc = 0);
1101                 }
1102         }
1103
1104         /* The lock's callback data might be set in the policy function */
1105         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
1106                                 dlm_req->lock_desc.l_resource.lr_type,
1107                                 dlm_req->lock_desc.l_req_mode,
1108                                 cbs, NULL, 0);
1109
1110         if (!lock)
1111                 GOTO(out, rc = -ENOMEM);
1112
1113         lock->l_last_activity = cfs_time_current_sec();
1114         lock->l_remote_handle = dlm_req->lock_handle[0];
1115         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
1116
1117         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
1118         /* Don't enqueue a lock onto the export if it has already
1119          * been evicted.  Cancel it now instead. (bug 3822) */
1120         if (req->rq_export->exp_failed) {
1121                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1122                 GOTO(out, rc = -ENOTCONN);
1123         }
1124
1125         lock->l_export = class_export_lock_get(req->rq_export, lock);
1126         if (lock->l_export->exp_lock_hash)
1127                 cfs_hash_add(lock->l_export->exp_lock_hash,
1128                              &lock->l_remote_handle,
1129                              &lock->l_exp_hash);
1130
1131 existing_lock:
1132
1133         if (flags & LDLM_FL_HAS_INTENT) {
1134                 /* In this case, the reply buffer is allocated deep in
1135                  * local_lock_enqueue by the policy function. */
1136                 cookie = req;
1137         } else {
1138                 lock_res_and_lock(lock);
1139                 if (lock->l_resource->lr_lvb_len) {
1140                         req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB,
1141                                              RCL_SERVER,
1142                                              lock->l_resource->lr_lvb_len);
1143                 }
1144                 unlock_res_and_lock(lock);
1145
1146                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
1147                         GOTO(out, rc = -ENOMEM);
1148
1149                 rc = req_capsule_server_pack(&req->rq_pill);
1150                 if (rc)
1151                         GOTO(out, rc);
1152         }
1153
1154         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
1155                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1156         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
1157                 lock->l_req_extent = lock->l_policy_data.l_extent;
1158
1159         err = ldlm_lock_enqueue(ns, &lock, cookie, (int *)&flags);
1160         if (err)
1161                 GOTO(out, err);
1162
1163         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1164         dlm_rep->lock_flags = flags;
1165
1166         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
1167         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
1168
1169         /* We never send a blocking AST until the lock is granted, but
1170          * we can tell it right now */
1171         lock_res_and_lock(lock);
1172
1173         /* Now take into account flags to be inherited from original lock
1174            request both in reply to client and in our own lock flags. */
1175         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1176         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
1177
1178         /* Don't move a pending lock onto the export if it has already
1179          * been evicted.  Cancel it now instead. (bug 5683) */
1180         if (unlikely(req->rq_export->exp_failed ||
1181                      OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
1182                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
1183                 rc = -ENOTCONN;
1184         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
1185                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
1186                 if (lock->l_granted_mode == lock->l_req_mode) {
1187                         /*
1188                          * Only cancel lock if it was granted, because it would
1189                          * be destroyed immediately and would never be granted
1190                          * in the future, causing timeouts on client.  Not
1191                          * granted lock will be cancelled immediately after
1192                          * sending completion AST.
1193                          */
1194                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
1195                                 unlock_res_and_lock(lock);
1196                                 ldlm_lock_cancel(lock);
1197                                 lock_res_and_lock(lock);
1198                         } else
1199                                 ldlm_add_waiting_lock(lock);
1200                 }
1201         }
1202         /* Make sure we never ever grant usual metadata locks to liblustre
1203            clients */
1204         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
1205             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
1206              req->rq_export->exp_libclient) {
1207                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
1208                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
1209                         CERROR("Granting sync lock to libclient. "
1210                                "req fl %d, rep fl %d, lock fl "LPX64"\n",
1211                                dlm_req->lock_flags, dlm_rep->lock_flags,
1212                                lock->l_flags);
1213                         LDLM_ERROR(lock, "sync lock");
1214                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
1215                                 struct ldlm_intent *it;
1216
1217                                 it = req_capsule_client_get(&req->rq_pill,
1218                                                             &RMF_LDLM_INTENT);
1219                                 if (it != NULL) {
1220                                         CERROR("This is intent %s ("LPU64")\n",
1221                                                ldlm_it2str(it->opc), it->opc);
1222                                 }
1223                         }
1224                 }
1225         }
1226
1227         unlock_res_and_lock(lock);
1228
1229         EXIT;
1230  out:
1231         req->rq_status = rc ?: err; /* return either error - bug 11190 */
1232         if (!req->rq_packed_final) {
1233                 err = lustre_pack_reply(req, 1, NULL, NULL);
1234                 if (rc == 0)
1235                         rc = err;
1236         }
1237
1238         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
1239          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1240         if (lock) {
1241                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1242                            "(err=%d, rc=%d)", err, rc);
1243
1244                 lock_res_and_lock(lock);
1245                 if (rc == 0) {
1246                         if (lock->l_resource->lr_lvb_len > 0) {
1247                                 void *lvb;
1248
1249                                 lvb = req_capsule_server_get(&req->rq_pill,
1250                                                              &RMF_DLM_LVB);
1251                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1252                                          req, lock);
1253
1254                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
1255                                        lock->l_resource->lr_lvb_len);
1256                         }
1257                 } else {
1258                         ldlm_resource_unlink_lock(lock);
1259                         ldlm_lock_destroy_nolock(lock);
1260                 }
1261                 unlock_res_and_lock(lock);
1262
1263                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1264                         ldlm_reprocess_all(lock->l_resource);
1265
1266                 LDLM_LOCK_RELEASE(lock);
1267         }
1268
1269         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1270                           lock, rc);
1271
1272         return rc;
1273 }
1274
1275 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1276                         ldlm_completion_callback completion_callback,
1277                         ldlm_blocking_callback blocking_callback,
1278                         ldlm_glimpse_callback glimpse_callback)
1279 {
1280         struct ldlm_request *dlm_req;
1281         struct ldlm_callback_suite cbs = {
1282                 .lcs_completion = completion_callback,
1283                 .lcs_blocking   = blocking_callback,
1284                 .lcs_glimpse    = glimpse_callback
1285         };
1286         int rc;
1287
1288         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1289         if (dlm_req != NULL) {
1290                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1291                                           req, dlm_req, &cbs);
1292         } else {
1293                 rc = -EFAULT;
1294         }
1295         return rc;
1296 }
1297
1298 int ldlm_handle_convert0(struct ptlrpc_request *req,
1299                          const struct ldlm_request *dlm_req)
1300 {
1301         struct ldlm_reply *dlm_rep;
1302         struct ldlm_lock *lock;
1303         int rc;
1304         ENTRY;
1305
1306         if (req->rq_export && req->rq_export->exp_nid_stats &&
1307             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1308                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1309                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1310
1311         rc = req_capsule_server_pack(&req->rq_pill);
1312         if (rc)
1313                 RETURN(rc);
1314
1315         dlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
1316         dlm_rep->lock_flags = dlm_req->lock_flags;
1317
1318         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1319         if (!lock) {
1320                 req->rq_status = EINVAL;
1321         } else {
1322                 void *res = NULL;
1323
1324                 LDLM_DEBUG(lock, "server-side convert handler START");
1325
1326                 lock->l_last_activity = cfs_time_current_sec();
1327                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1328                                         &dlm_rep->lock_flags);
1329                 if (res) {
1330                         if (ldlm_del_waiting_lock(lock))
1331                                 LDLM_DEBUG(lock, "converted waiting lock");
1332                         req->rq_status = 0;
1333                 } else {
1334                         req->rq_status = EDEADLOCK;
1335                 }
1336         }
1337
1338         if (lock) {
1339                 if (!req->rq_status)
1340                         ldlm_reprocess_all(lock->l_resource);
1341                 LDLM_DEBUG(lock, "server-side convert handler END");
1342                 LDLM_LOCK_PUT(lock);
1343         } else
1344                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1345
1346         RETURN(0);
1347 }
1348
1349 int ldlm_handle_convert(struct ptlrpc_request *req)
1350 {
1351         int rc;
1352         struct ldlm_request *dlm_req;
1353
1354         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1355         if (dlm_req != NULL) {
1356                 rc = ldlm_handle_convert0(req, dlm_req);
1357         } else {
1358                 CERROR ("Can't unpack dlm_req\n");
1359                 rc = -EFAULT;
1360         }
1361         return rc;
1362 }
1363
1364 /* Cancel all the locks whos handles are packed into ldlm_request */
1365 int ldlm_request_cancel(struct ptlrpc_request *req,
1366                         const struct ldlm_request *dlm_req, int first)
1367 {
1368         struct ldlm_resource *res, *pres = NULL;
1369         struct ldlm_lock *lock;
1370         int i, count, done = 0;
1371         ENTRY;
1372
1373         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1374         if (first >= count)
1375                 RETURN(0);
1376
1377         /* There is no lock on the server at the replay time,
1378          * skip lock cancelling to make replay tests to pass. */
1379         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1380                 RETURN(0);
1381
1382         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1383                           "starting at %d", count, first);
1384
1385         for (i = first; i < count; i++) {
1386                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1387                 if (!lock) {
1388                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1389                                           "lock (cookie "LPU64")",
1390                                           dlm_req->lock_handle[i].cookie);
1391                         continue;
1392                 }
1393
1394                 res = lock->l_resource;
1395                 done++;
1396
1397                 if (res != pres) {
1398                         if (pres != NULL) {
1399                                 ldlm_reprocess_all(pres);
1400                                 LDLM_RESOURCE_DELREF(pres);
1401                                 ldlm_resource_putref(pres);
1402                         }
1403                         if (res != NULL) {
1404                                 ldlm_resource_getref(res);
1405                                 LDLM_RESOURCE_ADDREF(res);
1406                                 ldlm_res_lvbo_update(res, NULL, 1);
1407                         }
1408                         pres = res;
1409                 }
1410                 ldlm_lock_cancel(lock);
1411                 LDLM_LOCK_PUT(lock);
1412         }
1413         if (pres != NULL) {
1414                 ldlm_reprocess_all(pres);
1415                 LDLM_RESOURCE_DELREF(pres);
1416                 ldlm_resource_putref(pres);
1417         }
1418         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1419         RETURN(done);
1420 }
1421
1422 int ldlm_handle_cancel(struct ptlrpc_request *req)
1423 {
1424         struct ldlm_request *dlm_req;
1425         int rc;
1426         ENTRY;
1427
1428         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1429         if (dlm_req == NULL) {
1430                 CDEBUG(D_INFO, "bad request buffer for cancel\n");
1431                 RETURN(-EFAULT);
1432         }
1433
1434         if (req->rq_export && req->rq_export->exp_nid_stats &&
1435             req->rq_export->exp_nid_stats->nid_ldlm_stats)
1436                 lprocfs_counter_incr(req->rq_export->exp_nid_stats->nid_ldlm_stats,
1437                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1438
1439         rc = req_capsule_server_pack(&req->rq_pill);
1440         if (rc)
1441                 RETURN(rc);
1442
1443         if (!ldlm_request_cancel(req, dlm_req, 0))
1444                 req->rq_status = ESTALE;
1445
1446         RETURN(ptlrpc_reply(req));
1447 }
1448
1449 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1450                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1451 {
1452         int do_ast;
1453         ENTRY;
1454
1455         LDLM_DEBUG(lock, "client blocking AST callback handler");
1456
1457         lock_res_and_lock(lock);
1458         lock->l_flags |= LDLM_FL_CBPENDING;
1459
1460         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1461                 lock->l_flags |= LDLM_FL_CANCEL;
1462
1463         do_ast = (!lock->l_readers && !lock->l_writers);
1464         unlock_res_and_lock(lock);
1465
1466         if (do_ast) {
1467                 CDEBUG(D_DLMTRACE, "Lock %p already unused, calling callback (%p)\n",
1468                        lock, lock->l_blocking_ast);
1469                 if (lock->l_blocking_ast != NULL)
1470                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1471                                              LDLM_CB_BLOCKING);
1472         } else {
1473                 CDEBUG(D_DLMTRACE, "Lock %p is referenced, will be cancelled later\n",
1474                        lock);
1475         }
1476
1477         LDLM_DEBUG(lock, "client blocking callback handler END");
1478         LDLM_LOCK_RELEASE(lock);
1479         EXIT;
1480 }
1481
1482 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1483                                     struct ldlm_namespace *ns,
1484                                     struct ldlm_request *dlm_req,
1485                                     struct ldlm_lock *lock)
1486 {
1487         CFS_LIST_HEAD(ast_list);
1488         ENTRY;
1489
1490         LDLM_DEBUG(lock, "client completion callback handler START");
1491
1492         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE)) {
1493                 int to = cfs_time_seconds(1);
1494                 while (to > 0) {
1495                         cfs_schedule_timeout_and_set_state(
1496                                 CFS_TASK_INTERRUPTIBLE, to);
1497                         if (lock->l_granted_mode == lock->l_req_mode ||
1498                             lock->l_destroyed)
1499                                 break;
1500                 }
1501         }
1502
1503         lock_res_and_lock(lock);
1504         if (lock->l_destroyed ||
1505             lock->l_granted_mode == lock->l_req_mode) {
1506                 /* bug 11300: the lock has already been granted */
1507                 unlock_res_and_lock(lock);
1508                 LDLM_DEBUG(lock, "Double grant race happened");
1509                 LDLM_LOCK_RELEASE(lock);
1510                 EXIT;
1511                 return;
1512         }
1513
1514         /* If we receive the completion AST before the actual enqueue returned,
1515          * then we might need to switch lock modes, resources, or extents. */
1516         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1517                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1518                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1519         }
1520
1521         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1522                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1523                 LDLM_DEBUG(lock, "completion AST, new policy data");
1524         }
1525
1526         ldlm_resource_unlink_lock(lock);
1527         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1528                    &lock->l_resource->lr_name,
1529                    sizeof(lock->l_resource->lr_name)) != 0) {
1530                 unlock_res_and_lock(lock);
1531                 if (ldlm_lock_change_resource(ns, lock,
1532                                 &dlm_req->lock_desc.l_resource.lr_name) != 0) {
1533                         LDLM_ERROR(lock, "Failed to allocate resource");
1534                         LDLM_LOCK_RELEASE(lock);
1535                         EXIT;
1536                         return;
1537                 }
1538                 LDLM_DEBUG(lock, "completion AST, new resource");
1539                 CERROR("change resource!\n");
1540                 lock_res_and_lock(lock);
1541         }
1542
1543         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1544                 /* BL_AST locks are not needed in lru.
1545                  * let ldlm_cancel_lru() be fast. */
1546                 ldlm_lock_remove_from_lru(lock);
1547                 lock->l_flags |= LDLM_FL_CBPENDING | LDLM_FL_BL_AST;
1548                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1549         }
1550
1551         if (lock->l_lvb_len) {
1552                 if (req_capsule_get_size(&req->rq_pill, &RMF_DLM_LVB,
1553                                          RCL_CLIENT) < lock->l_lvb_len) {
1554                         LDLM_ERROR(lock, "completion AST did not contain "
1555                                    "expected LVB!");
1556                 } else {
1557                         void *lvb = req_capsule_client_get(&req->rq_pill,
1558                                                            &RMF_DLM_LVB);
1559                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1560                 }
1561         }
1562
1563         ldlm_grant_lock(lock, &ast_list);
1564         unlock_res_and_lock(lock);
1565
1566         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1567
1568         /* Let Enqueue to call osc_lock_upcall() and initialize
1569          * l_ast_data */
1570         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 2);
1571
1572         ldlm_run_ast_work(&ast_list, LDLM_WORK_CP_AST);
1573
1574         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1575                           lock);
1576         LDLM_LOCK_RELEASE(lock);
1577         EXIT;
1578 }
1579
1580 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1581                                     struct ldlm_namespace *ns,
1582                                     struct ldlm_request *dlm_req,
1583                                     struct ldlm_lock *lock)
1584 {
1585         int rc = -ENOSYS;
1586         ENTRY;
1587
1588         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1589
1590         if (lock->l_glimpse_ast != NULL)
1591                 rc = lock->l_glimpse_ast(lock, req);
1592
1593         if (req->rq_repmsg != NULL) {
1594                 ptlrpc_reply(req);
1595         } else {
1596                 req->rq_status = rc;
1597                 ptlrpc_error(req);
1598         }
1599
1600         lock_res_and_lock(lock);
1601         if (lock->l_granted_mode == LCK_PW &&
1602             !lock->l_readers && !lock->l_writers &&
1603             cfs_time_after(cfs_time_current(),
1604                            cfs_time_add(lock->l_last_used,
1605                                         cfs_time_seconds(10)))) {
1606                 unlock_res_and_lock(lock);
1607                 if (ldlm_bl_to_thread_lock(ns, NULL, lock))
1608                         ldlm_handle_bl_callback(ns, NULL, lock);
1609
1610                 EXIT;
1611                 return;
1612         }
1613         unlock_res_and_lock(lock);
1614         LDLM_LOCK_RELEASE(lock);
1615         EXIT;
1616 }
1617
1618 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1619 {
1620         if (req->rq_no_reply)
1621                 return 0;
1622
1623         req->rq_status = rc;
1624         if (!req->rq_packed_final) {
1625                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1626                 if (rc)
1627                         return rc;
1628         }
1629         return ptlrpc_reply(req);
1630 }
1631
1632 #ifdef __KERNEL__
1633 static int __ldlm_bl_to_thread(struct ldlm_bl_work_item *blwi, int mode)
1634 {
1635         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1636         ENTRY;
1637
1638         cfs_spin_lock(&blp->blp_lock);
1639         if (blwi->blwi_lock && blwi->blwi_lock->l_flags & LDLM_FL_DISCARD_DATA) {
1640                 /* add LDLM_FL_DISCARD_DATA requests to the priority list */
1641                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_prio_list);
1642         } else {
1643                 /* other blocking callbacks are added to the regular list */
1644                 cfs_list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1645         }
1646         cfs_spin_unlock(&blp->blp_lock);
1647
1648         cfs_waitq_signal(&blp->blp_waitq);
1649
1650         /* can not use blwi->blwi_mode as blwi could be already freed in
1651            LDLM_ASYNC mode */
1652         if (mode == LDLM_SYNC)
1653                 cfs_wait_for_completion(&blwi->blwi_comp);
1654
1655         RETURN(0);
1656 }
1657
1658 static inline void init_blwi(struct ldlm_bl_work_item *blwi,
1659                              struct ldlm_namespace *ns,
1660                              struct ldlm_lock_desc *ld,
1661                              cfs_list_t *cancels, int count,
1662                              struct ldlm_lock *lock,
1663                              int mode)
1664 {
1665         cfs_init_completion(&blwi->blwi_comp);
1666         CFS_INIT_LIST_HEAD(&blwi->blwi_head);
1667
1668         if (cfs_memory_pressure_get())
1669                 blwi->blwi_mem_pressure = 1;
1670
1671         blwi->blwi_ns = ns;
1672         blwi->blwi_mode = mode;
1673         if (ld != NULL)
1674                 blwi->blwi_ld = *ld;
1675         if (count) {
1676                 cfs_list_add(&blwi->blwi_head, cancels);
1677                 cfs_list_del_init(cancels);
1678                 blwi->blwi_count = count;
1679         } else {
1680                 blwi->blwi_lock = lock;
1681         }
1682 }
1683
1684 static int ldlm_bl_to_thread(struct ldlm_namespace *ns,
1685                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock,
1686                              cfs_list_t *cancels, int count, int mode)
1687 {
1688         ENTRY;
1689
1690         if (cancels && count == 0)
1691                 RETURN(0);
1692
1693         if (mode == LDLM_SYNC) {
1694                 /* if it is synchronous call do minimum mem alloc, as it could
1695                  * be triggered from kernel shrinker
1696                  */
1697                 struct ldlm_bl_work_item blwi;
1698                 memset(&blwi, 0, sizeof(blwi));
1699                 init_blwi(&blwi, ns, ld, cancels, count, lock, LDLM_SYNC);
1700                 RETURN(__ldlm_bl_to_thread(&blwi, LDLM_SYNC));
1701         } else {
1702                 struct ldlm_bl_work_item *blwi;
1703                 OBD_ALLOC(blwi, sizeof(*blwi));
1704                 if (blwi == NULL)
1705                         RETURN(-ENOMEM);
1706                 init_blwi(blwi, ns, ld, cancels, count, lock, LDLM_ASYNC);
1707
1708                 RETURN(__ldlm_bl_to_thread(blwi, LDLM_ASYNC));
1709         }
1710 }
1711
1712 #endif
1713
1714 int ldlm_bl_to_thread_lock(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1715                            struct ldlm_lock *lock)
1716 {
1717 #ifdef __KERNEL__
1718         RETURN(ldlm_bl_to_thread(ns, ld, lock, NULL, 0, LDLM_ASYNC));
1719 #else
1720         RETURN(-ENOSYS);
1721 #endif
1722 }
1723
1724 int ldlm_bl_to_thread_list(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1725                            cfs_list_t *cancels, int count, int mode)
1726 {
1727 #ifdef __KERNEL__
1728         RETURN(ldlm_bl_to_thread(ns, ld, NULL, cancels, count, mode));
1729 #else
1730         RETURN(-ENOSYS);
1731 #endif
1732 }
1733
1734 /* Setinfo coming from Server (eg MDT) to Client (eg MDC)! */
1735 static int ldlm_handle_setinfo(struct ptlrpc_request *req)
1736 {
1737         struct obd_device *obd = req->rq_export->exp_obd;
1738         char *key;
1739         void *val;
1740         int keylen, vallen;
1741         int rc = -ENOSYS;
1742         ENTRY;
1743
1744         DEBUG_REQ(D_HSM, req, "%s: handle setinfo\n", obd->obd_name);
1745
1746         req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
1747
1748         key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1749         if (key == NULL) {
1750                 DEBUG_REQ(D_IOCTL, req, "no set_info key");
1751                 RETURN(-EFAULT);
1752         }
1753         keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1754                                       RCL_CLIENT);
1755         val = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
1756         if (val == NULL) {
1757                 DEBUG_REQ(D_IOCTL, req, "no set_info val");
1758                 RETURN(-EFAULT);
1759         }
1760         vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1761                                       RCL_CLIENT);
1762
1763         /* We are responsible for swabbing contents of val */
1764
1765         if (KEY_IS(KEY_HSM_COPYTOOL_SEND))
1766                 /* Pass it on to mdc (the "export" in this case) */
1767                 rc = obd_set_info_async(req->rq_export,
1768                                         sizeof(KEY_HSM_COPYTOOL_SEND),
1769                                         KEY_HSM_COPYTOOL_SEND,
1770                                         vallen, val, NULL);
1771         else
1772                 DEBUG_REQ(D_WARNING, req, "ignoring unknown key %s", key);
1773
1774         return rc;
1775 }
1776
1777 static inline void ldlm_callback_errmsg(struct ptlrpc_request *req,
1778                                         const char *msg, int rc,
1779                                         struct lustre_handle *handle)
1780 {
1781         DEBUG_REQ((req->rq_no_reply || rc) ? D_WARNING : D_DLMTRACE, req,
1782                   "%s: [nid %s] [rc %d] [lock "LPX64"]",
1783                   msg, libcfs_id2str(req->rq_peer), rc,
1784                   handle ? handle->cookie : 0);
1785         if (req->rq_no_reply)
1786                 CWARN("No reply was sent, maybe cause bug 21636.\n");
1787         else if (rc)
1788                 CWARN("Send reply failed, maybe cause bug 21636.\n");
1789 }
1790
1791 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1792 static int ldlm_callback_handler(struct ptlrpc_request *req)
1793 {
1794         struct ldlm_namespace *ns;
1795         struct ldlm_request *dlm_req;
1796         struct ldlm_lock *lock;
1797         int rc;
1798         ENTRY;
1799
1800         /* Requests arrive in sender's byte order.  The ptlrpc service
1801          * handler has already checked and, if necessary, byte-swapped the
1802          * incoming request message body, but I am responsible for the
1803          * message buffers. */
1804
1805         /* do nothing for sec context finalize */
1806         if (lustre_msg_get_opc(req->rq_reqmsg) == SEC_CTX_FINI)
1807                 RETURN(0);
1808
1809         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1810
1811         if (req->rq_export == NULL) {
1812                 rc = ldlm_callback_reply(req, -ENOTCONN);
1813                 ldlm_callback_errmsg(req, "Operate on unconnected server",
1814                                      rc, NULL);
1815                 RETURN(0);
1816         }
1817
1818         LASSERT(req->rq_export != NULL);
1819         LASSERT(req->rq_export->exp_obd != NULL);
1820
1821         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1822         case LDLM_BL_CALLBACK:
1823                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_BL_CALLBACK))
1824                         RETURN(0);
1825                 break;
1826         case LDLM_CP_CALLBACK:
1827                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CP_CALLBACK))
1828                         RETURN(0);
1829                 break;
1830         case LDLM_GL_CALLBACK:
1831                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GL_CALLBACK))
1832                         RETURN(0);
1833                 break;
1834         case LDLM_SET_INFO:
1835                 rc = ldlm_handle_setinfo(req);
1836                 ldlm_callback_reply(req, rc);
1837                 RETURN(0);
1838         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1839                 CERROR("shouldn't be handling OBD_LOG_CANCEL on DLM thread\n");
1840                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1841                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1842                         RETURN(0);
1843                 rc = llog_origin_handle_cancel(req);
1844                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
1845                         RETURN(0);
1846                 ldlm_callback_reply(req, rc);
1847                 RETURN(0);
1848         case OBD_QC_CALLBACK:
1849                 req_capsule_set(&req->rq_pill, &RQF_QC_CALLBACK);
1850                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_QC_CALLBACK_NET))
1851                         RETURN(0);
1852                 rc = target_handle_qc_callback(req);
1853                 ldlm_callback_reply(req, rc);
1854                 RETURN(0);
1855         case QUOTA_DQACQ:
1856         case QUOTA_DQREL:
1857                 /* reply in handler */
1858                 req_capsule_set(&req->rq_pill, &RQF_MDS_QUOTA_DQACQ);
1859                 rc = target_handle_dqacq_callback(req);
1860                 RETURN(0);
1861         case LLOG_ORIGIN_HANDLE_CREATE:
1862                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_HANDLE_CREATE);
1863                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1864                         RETURN(0);
1865                 rc = llog_origin_handle_create(req);
1866                 ldlm_callback_reply(req, rc);
1867                 RETURN(0);
1868         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1869                 req_capsule_set(&req->rq_pill,
1870                                 &RQF_LLOG_ORIGIN_HANDLE_NEXT_BLOCK);
1871                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1872                         RETURN(0);
1873                 rc = llog_origin_handle_next_block(req);
1874                 ldlm_callback_reply(req, rc);
1875                 RETURN(0);
1876         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1877                 req_capsule_set(&req->rq_pill,
1878                                 &RQF_LLOG_ORIGIN_HANDLE_READ_HEADER);
1879                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1880                         RETURN(0);
1881                 rc = llog_origin_handle_read_header(req);
1882                 ldlm_callback_reply(req, rc);
1883                 RETURN(0);
1884         case LLOG_ORIGIN_HANDLE_CLOSE:
1885                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOGD_NET))
1886                         RETURN(0);
1887                 rc = llog_origin_handle_close(req);
1888                 ldlm_callback_reply(req, rc);
1889                 RETURN(0);
1890         default:
1891                 CERROR("unknown opcode %u\n",
1892                        lustre_msg_get_opc(req->rq_reqmsg));
1893                 ldlm_callback_reply(req, -EPROTO);
1894                 RETURN(0);
1895         }
1896
1897         ns = req->rq_export->exp_obd->obd_namespace;
1898         LASSERT(ns != NULL);
1899
1900         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
1901
1902         dlm_req = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
1903         if (dlm_req == NULL) {
1904                 rc = ldlm_callback_reply(req, -EPROTO);
1905                 ldlm_callback_errmsg(req, "Operate without parameter", rc,
1906                                      NULL);
1907                 RETURN(0);
1908         }
1909
1910         /* Force a known safe race, send a cancel to the server for a lock
1911          * which the server has already started a blocking callback on. */
1912         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_BL_CB_RACE) &&
1913             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1914                 rc = ldlm_cli_cancel(&dlm_req->lock_handle[0]);
1915                 if (rc < 0)
1916                         CERROR("ldlm_cli_cancel: %d\n", rc);
1917         }
1918
1919         lock = ldlm_handle2lock_long(&dlm_req->lock_handle[0], 0);
1920         if (!lock) {
1921                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1922                        "disappeared\n", dlm_req->lock_handle[0].cookie);
1923                 rc = ldlm_callback_reply(req, -EINVAL);
1924                 ldlm_callback_errmsg(req, "Operate with invalid parameter", rc,
1925                                      &dlm_req->lock_handle[0]);
1926                 RETURN(0);
1927         }
1928
1929         if ((lock->l_flags & LDLM_FL_FAIL_LOC) &&
1930             lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK)
1931                 OBD_RACE(OBD_FAIL_LDLM_CP_BL_RACE);
1932
1933         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1934         lock_res_and_lock(lock);
1935         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1936         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1937                 /* If somebody cancels lock and cache is already dropped,
1938                  * or lock is failed before cp_ast received on client,
1939                  * we can tell the server we have no lock. Otherwise, we
1940                  * should send cancel after dropping the cache. */
1941                 if (((lock->l_flags & LDLM_FL_CANCELING) &&
1942                     (lock->l_flags & LDLM_FL_BL_DONE)) ||
1943                     (lock->l_flags & LDLM_FL_FAILED)) {
1944                         LDLM_DEBUG(lock, "callback on lock "
1945                                    LPX64" - lock disappeared\n",
1946                                    dlm_req->lock_handle[0].cookie);
1947                         unlock_res_and_lock(lock);
1948                         LDLM_LOCK_RELEASE(lock);
1949                         rc = ldlm_callback_reply(req, -EINVAL);
1950                         ldlm_callback_errmsg(req, "Operate on stale lock", rc,
1951                                              &dlm_req->lock_handle[0]);
1952                         RETURN(0);
1953                 }
1954                 /* BL_AST locks are not needed in lru.
1955                  * let ldlm_cancel_lru() be fast. */
1956                 ldlm_lock_remove_from_lru(lock);
1957                 lock->l_flags |= LDLM_FL_BL_AST;
1958         }
1959         unlock_res_and_lock(lock);
1960
1961         /* We want the ost thread to get this reply so that it can respond
1962          * to ost requests (write cache writeback) that might be triggered
1963          * in the callback.
1964          *
1965          * But we'd also like to be able to indicate in the reply that we're
1966          * cancelling right now, because it's unused, or have an intent result
1967          * in the reply, so we might have to push the responsibility for sending
1968          * the reply down into the AST handlers, alas. */
1969
1970         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1971         case LDLM_BL_CALLBACK:
1972                 CDEBUG(D_INODE, "blocking ast\n");
1973                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_BL_CALLBACK);
1974                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)) {
1975                         rc = ldlm_callback_reply(req, 0);
1976                         if (req->rq_no_reply || rc)
1977                                 ldlm_callback_errmsg(req, "Normal process", rc,
1978                                                      &dlm_req->lock_handle[0]);
1979                 }
1980                 if (ldlm_bl_to_thread_lock(ns, &dlm_req->lock_desc, lock))
1981                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1982                 break;
1983         case LDLM_CP_CALLBACK:
1984                 CDEBUG(D_INODE, "completion ast\n");
1985                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_CP_CALLBACK);
1986                 ldlm_callback_reply(req, 0);
1987                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1988                 break;
1989         case LDLM_GL_CALLBACK:
1990                 CDEBUG(D_INODE, "glimpse ast\n");
1991                 req_capsule_extend(&req->rq_pill, &RQF_LDLM_GL_CALLBACK);
1992                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1993                 break;
1994         default:
1995                 LBUG();                         /* checked above */
1996         }
1997
1998         RETURN(0);
1999 }
2000
2001 static int ldlm_cancel_handler(struct ptlrpc_request *req)
2002 {
2003         int rc;
2004         ENTRY;
2005
2006         /* Requests arrive in sender's byte order.  The ptlrpc service
2007          * handler has already checked and, if necessary, byte-swapped the
2008          * incoming request message body, but I am responsible for the
2009          * message buffers. */
2010
2011         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2012
2013         if (req->rq_export == NULL) {
2014                 struct ldlm_request *dlm_req;
2015
2016                 CERROR("%s from %s arrived at %lu with bad export cookie "
2017                        LPU64"\n",
2018                        ll_opcode2str(lustre_msg_get_opc(req->rq_reqmsg)),
2019                        libcfs_nid2str(req->rq_peer.nid),
2020                        req->rq_arrival_time.tv_sec,
2021                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
2022
2023                 if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_CANCEL) {
2024                         req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2025                         dlm_req = req_capsule_client_get(&req->rq_pill,
2026                                                          &RMF_DLM_REQ);
2027                         if (dlm_req != NULL)
2028                                 ldlm_lock_dump_handle(D_ERROR,
2029                                                       &dlm_req->lock_handle[0]);
2030                 }
2031                 ldlm_callback_reply(req, -ENOTCONN);
2032                 RETURN(0);
2033         }
2034
2035         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2036
2037         /* XXX FIXME move this back to mds/handler.c, bug 249 */
2038         case LDLM_CANCEL:
2039                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2040                 CDEBUG(D_INODE, "cancel\n");
2041                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2042                         RETURN(0);
2043                 rc = ldlm_handle_cancel(req);
2044                 if (rc)
2045                         break;
2046                 RETURN(0);
2047         case OBD_LOG_CANCEL:
2048                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2049                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2050                         RETURN(0);
2051                 rc = llog_origin_handle_cancel(req);
2052                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2053                         RETURN(0);
2054                 ldlm_callback_reply(req, rc);
2055                 RETURN(0);
2056         default:
2057                 CERROR("invalid opcode %d\n",
2058                        lustre_msg_get_opc(req->rq_reqmsg));
2059                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CALLBACK);
2060                 ldlm_callback_reply(req, -EINVAL);
2061         }
2062
2063         RETURN(0);
2064 }
2065
2066 int ldlm_revoke_lock_cb(cfs_hash_t *hs, cfs_hash_bd_t *bd,
2067                         cfs_hlist_node_t *hnode, void *data)
2068
2069 {
2070         cfs_list_t         *rpc_list = data;
2071         struct ldlm_lock   *lock = cfs_hash_object(hs, hnode);
2072
2073         lock_res_and_lock(lock);
2074
2075         if (lock->l_req_mode != lock->l_granted_mode) {
2076                 unlock_res_and_lock(lock);
2077                 return 0;
2078         }
2079
2080         LASSERT(lock->l_resource);
2081         if (lock->l_resource->lr_type != LDLM_IBITS &&
2082             lock->l_resource->lr_type != LDLM_PLAIN) {
2083                 unlock_res_and_lock(lock);
2084                 return 0;
2085         }
2086
2087         if (lock->l_flags & LDLM_FL_AST_SENT) {
2088                 unlock_res_and_lock(lock);
2089                 return 0;
2090         }
2091
2092         LASSERT(lock->l_blocking_ast);
2093         LASSERT(!lock->l_blocking_lock);
2094
2095         lock->l_flags |= LDLM_FL_AST_SENT;
2096         if (lock->l_export && lock->l_export->exp_lock_hash &&
2097             !cfs_hlist_unhashed(&lock->l_exp_hash))
2098                 cfs_hash_del(lock->l_export->exp_lock_hash,
2099                              &lock->l_remote_handle, &lock->l_exp_hash);
2100         cfs_list_add_tail(&lock->l_rk_ast, rpc_list);
2101         LDLM_LOCK_GET(lock);
2102
2103         unlock_res_and_lock(lock);
2104         return 0;
2105 }
2106
2107 void ldlm_revoke_export_locks(struct obd_export *exp)
2108 {
2109         cfs_list_t  rpc_list;
2110         ENTRY;
2111
2112         CFS_INIT_LIST_HEAD(&rpc_list);
2113         cfs_hash_for_each_empty(exp->exp_lock_hash,
2114                                 ldlm_revoke_lock_cb, &rpc_list);
2115         ldlm_run_ast_work(&rpc_list, LDLM_WORK_REVOKE_AST);
2116
2117         EXIT;
2118 }
2119
2120 #ifdef __KERNEL__
2121 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
2122 {
2123         struct ldlm_bl_work_item *blwi = NULL;
2124         static unsigned int num_bl = 0;
2125
2126         cfs_spin_lock(&blp->blp_lock);
2127         /* process a request from the blp_list at least every blp_num_threads */
2128         if (!cfs_list_empty(&blp->blp_list) &&
2129             (cfs_list_empty(&blp->blp_prio_list) || num_bl == 0))
2130                 blwi = cfs_list_entry(blp->blp_list.next,
2131                                       struct ldlm_bl_work_item, blwi_entry);
2132         else
2133                 if (!cfs_list_empty(&blp->blp_prio_list))
2134                         blwi = cfs_list_entry(blp->blp_prio_list.next,
2135                                               struct ldlm_bl_work_item,
2136                                               blwi_entry);
2137
2138         if (blwi) {
2139                 if (++num_bl >= cfs_atomic_read(&blp->blp_num_threads))
2140                         num_bl = 0;
2141                 cfs_list_del(&blwi->blwi_entry);
2142         }
2143         cfs_spin_unlock(&blp->blp_lock);
2144
2145         return blwi;
2146 }
2147
2148 /* This only contains temporary data until the thread starts */
2149 struct ldlm_bl_thread_data {
2150         char                    bltd_name[CFS_CURPROC_COMM_MAX];
2151         struct ldlm_bl_pool     *bltd_blp;
2152         cfs_completion_t        bltd_comp;
2153         int                     bltd_num;
2154 };
2155
2156 static int ldlm_bl_thread_main(void *arg);
2157
2158 static int ldlm_bl_thread_start(struct ldlm_bl_pool *blp)
2159 {
2160         struct ldlm_bl_thread_data bltd = { .bltd_blp = blp };
2161         int rc;
2162
2163         cfs_init_completion(&bltd.bltd_comp);
2164         rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
2165         if (rc < 0) {
2166                 CERROR("cannot start LDLM thread ldlm_bl_%02d: rc %d\n",
2167                        cfs_atomic_read(&blp->blp_num_threads), rc);
2168                 return rc;
2169         }
2170         cfs_wait_for_completion(&bltd.bltd_comp);
2171
2172         return 0;
2173 }
2174
2175 static int ldlm_bl_thread_main(void *arg)
2176 {
2177         struct ldlm_bl_pool *blp;
2178         ENTRY;
2179
2180         {
2181                 struct ldlm_bl_thread_data *bltd = arg;
2182
2183                 blp = bltd->bltd_blp;
2184
2185                 bltd->bltd_num =
2186                         cfs_atomic_inc_return(&blp->blp_num_threads) - 1;
2187                 cfs_atomic_inc(&blp->blp_busy_threads);
2188
2189                 snprintf(bltd->bltd_name, sizeof(bltd->bltd_name) - 1,
2190                         "ldlm_bl_%02d", bltd->bltd_num);
2191                 cfs_daemonize(bltd->bltd_name);
2192
2193                 cfs_complete(&bltd->bltd_comp);
2194                 /* cannot use bltd after this, it is only on caller's stack */
2195         }
2196
2197         while (1) {
2198                 struct l_wait_info lwi = { 0 };
2199                 struct ldlm_bl_work_item *blwi = NULL;
2200
2201                 blwi = ldlm_bl_get_work(blp);
2202
2203                 if (blwi == NULL) {
2204                         int busy;
2205
2206                         cfs_atomic_dec(&blp->blp_busy_threads);
2207                         l_wait_event_exclusive(blp->blp_waitq,
2208                                          (blwi = ldlm_bl_get_work(blp)) != NULL,
2209                                          &lwi);
2210                         busy = cfs_atomic_inc_return(&blp->blp_busy_threads);
2211
2212                         if (blwi->blwi_ns == NULL)
2213                                 /* added by ldlm_cleanup() */
2214                                 break;
2215
2216                         /* Not fatal if racy and have a few too many threads */
2217                         if (unlikely(busy < blp->blp_max_threads &&
2218                             busy >= cfs_atomic_read(&blp->blp_num_threads)))
2219                                 /* discard the return value, we tried */
2220                                 ldlm_bl_thread_start(blp);
2221                 } else {
2222                         if (blwi->blwi_ns == NULL)
2223                                 /* added by ldlm_cleanup() */
2224                                 break;
2225                 }
2226                 if (blwi->blwi_mem_pressure)
2227                         cfs_memory_pressure_set();
2228
2229                 if (blwi->blwi_count) {
2230                         int count;
2231                         /* The special case when we cancel locks in lru
2232                          * asynchronously, we pass the list of locks here.
2233                          * Thus locks are marked LDLM_FL_CANCELING, but NOT
2234                          * canceled locally yet. */
2235                         count = ldlm_cli_cancel_list_local(&blwi->blwi_head,
2236                                                            blwi->blwi_count,
2237                                                            LCF_BL_AST);
2238                         ldlm_cli_cancel_list(&blwi->blwi_head, count, NULL, 0);
2239                 } else {
2240                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
2241                                                 blwi->blwi_lock);
2242                 }
2243                 if (blwi->blwi_mem_pressure)
2244                         cfs_memory_pressure_clr();
2245
2246                 if (blwi->blwi_mode == LDLM_ASYNC)
2247                         OBD_FREE(blwi, sizeof(*blwi));
2248                 else
2249                         cfs_complete(&blwi->blwi_comp);
2250         }
2251
2252         cfs_atomic_dec(&blp->blp_busy_threads);
2253         cfs_atomic_dec(&blp->blp_num_threads);
2254         cfs_complete(&blp->blp_comp);
2255         RETURN(0);
2256 }
2257
2258 #endif
2259
2260 static int ldlm_setup(void);
2261 static int ldlm_cleanup(void);
2262
2263 int ldlm_get_ref(void)
2264 {
2265         int rc = 0;
2266         ENTRY;
2267         cfs_mutex_down(&ldlm_ref_sem);
2268         if (++ldlm_refcount == 1) {
2269                 rc = ldlm_setup();
2270                 if (rc)
2271                         ldlm_refcount--;
2272         }
2273         cfs_mutex_up(&ldlm_ref_sem);
2274
2275         RETURN(rc);
2276 }
2277
2278 void ldlm_put_ref(void)
2279 {
2280         ENTRY;
2281         cfs_mutex_down(&ldlm_ref_sem);
2282         if (ldlm_refcount == 1) {
2283                 int rc = ldlm_cleanup();
2284                 if (rc)
2285                         CERROR("ldlm_cleanup failed: %d\n", rc);
2286                 else
2287                         ldlm_refcount--;
2288         } else {
2289                 ldlm_refcount--;
2290         }
2291         cfs_mutex_up(&ldlm_ref_sem);
2292
2293         EXIT;
2294 }
2295
2296 /*
2297  * Export handle<->lock hash operations.
2298  */
2299 static unsigned
2300 ldlm_export_lock_hash(cfs_hash_t *hs, void *key, unsigned mask)
2301 {
2302         return cfs_hash_u64_hash(((struct lustre_handle *)key)->cookie, mask);
2303 }
2304
2305 static void *
2306 ldlm_export_lock_key(cfs_hlist_node_t *hnode)
2307 {
2308         struct ldlm_lock *lock;
2309         ENTRY;
2310
2311         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2312         RETURN(&lock->l_remote_handle);
2313 }
2314
2315 static void
2316 ldlm_export_lock_keycpy(cfs_hlist_node_t *hnode, void *key)
2317 {
2318         struct ldlm_lock     *lock;
2319
2320         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2321         lock->l_remote_handle = *(struct lustre_handle *)key;
2322 }
2323
2324 static int
2325 ldlm_export_lock_keycmp(void *key, cfs_hlist_node_t *hnode)
2326 {
2327         ENTRY;
2328         RETURN(lustre_handle_equal(ldlm_export_lock_key(hnode), key));
2329 }
2330
2331 static void *
2332 ldlm_export_lock_object(cfs_hlist_node_t *hnode)
2333 {
2334         return cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2335 }
2336
2337 static void *
2338 ldlm_export_lock_get(cfs_hlist_node_t *hnode)
2339 {
2340         struct ldlm_lock *lock;
2341         ENTRY;
2342
2343         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2344         LDLM_LOCK_GET(lock);
2345
2346         RETURN(lock);
2347 }
2348
2349 static void *
2350 ldlm_export_lock_put(cfs_hlist_node_t *hnode)
2351 {
2352         struct ldlm_lock *lock;
2353         ENTRY;
2354
2355         lock = cfs_hlist_entry(hnode, struct ldlm_lock, l_exp_hash);
2356         LDLM_LOCK_RELEASE(lock);
2357
2358         RETURN(lock);
2359 }
2360
2361 static cfs_hash_ops_t ldlm_export_lock_ops = {
2362         .hs_hash        = ldlm_export_lock_hash,
2363         .hs_key         = ldlm_export_lock_key,
2364         .hs_keycmp      = ldlm_export_lock_keycmp,
2365         .hs_keycpy      = ldlm_export_lock_keycpy,
2366         .hs_object      = ldlm_export_lock_object,
2367         .hs_get         = ldlm_export_lock_get,
2368         .hs_put         = ldlm_export_lock_put,
2369         .hs_put_locked  = ldlm_export_lock_put,
2370 };
2371
2372 int ldlm_init_export(struct obd_export *exp)
2373 {
2374         ENTRY;
2375
2376         exp->exp_lock_hash =
2377                 cfs_hash_create(obd_uuid2str(&exp->exp_client_uuid),
2378                                 HASH_EXP_LOCK_CUR_BITS,
2379                                 HASH_EXP_LOCK_MAX_BITS,
2380                                 HASH_EXP_LOCK_BKT_BITS, 0,
2381                                 CFS_HASH_MIN_THETA, CFS_HASH_MAX_THETA,
2382                                 &ldlm_export_lock_ops,
2383                                 CFS_HASH_DEFAULT | CFS_HASH_REHASH_KEY |
2384                                 CFS_HASH_NBLK_CHANGE);
2385
2386         if (!exp->exp_lock_hash)
2387                 RETURN(-ENOMEM);
2388
2389         RETURN(0);
2390 }
2391 EXPORT_SYMBOL(ldlm_init_export);
2392
2393 void ldlm_destroy_export(struct obd_export *exp)
2394 {
2395         ENTRY;
2396         cfs_hash_putref(exp->exp_lock_hash);
2397         exp->exp_lock_hash = NULL;
2398         EXIT;
2399 }
2400 EXPORT_SYMBOL(ldlm_destroy_export);
2401
2402 static int ldlm_setup(void)
2403 {
2404         struct ldlm_bl_pool *blp;
2405         int rc = 0;
2406         int ldlm_min_threads = LDLM_THREADS_AUTO_MIN;
2407         int ldlm_max_threads = LDLM_THREADS_AUTO_MAX;
2408 #ifdef __KERNEL__
2409         int i;
2410 #endif
2411         ENTRY;
2412
2413         if (ldlm_state != NULL)
2414                 RETURN(-EALREADY);
2415
2416         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
2417         if (ldlm_state == NULL)
2418                 RETURN(-ENOMEM);
2419
2420 #ifdef LPROCFS
2421         rc = ldlm_proc_setup();
2422         if (rc != 0)
2423                 GOTO(out_free, rc);
2424 #endif
2425
2426 #ifdef __KERNEL__
2427         if (ldlm_num_threads) {
2428                 /* If ldlm_num_threads is set, it is the min and the max. */
2429                 if (ldlm_num_threads > LDLM_THREADS_AUTO_MAX)
2430                         ldlm_num_threads = LDLM_THREADS_AUTO_MAX;
2431                 if (ldlm_num_threads < LDLM_THREADS_AUTO_MIN)
2432                         ldlm_num_threads = LDLM_THREADS_AUTO_MIN;
2433                 ldlm_min_threads = ldlm_max_threads = ldlm_num_threads;
2434         }
2435 #endif
2436
2437         ldlm_state->ldlm_cb_service =
2438                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2439                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
2440                                 LDLM_CB_REPLY_PORTAL, 2,
2441                                 ldlm_callback_handler, "ldlm_cbd",
2442                                 ldlm_svc_proc_dir, NULL,
2443                                 ldlm_min_threads, ldlm_max_threads,
2444                                 "ldlm_cb",
2445                                 LCT_MD_THREAD|LCT_DT_THREAD, NULL);
2446
2447         if (!ldlm_state->ldlm_cb_service) {
2448                 CERROR("failed to start service\n");
2449                 GOTO(out_proc, rc = -ENOMEM);
2450         }
2451
2452         ldlm_state->ldlm_cancel_service =
2453                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
2454                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
2455                                 LDLM_CANCEL_REPLY_PORTAL, 6,
2456                                 ldlm_cancel_handler, "ldlm_canceld",
2457                                 ldlm_svc_proc_dir, NULL,
2458                                 ldlm_min_threads, ldlm_max_threads,
2459                                 "ldlm_cn",
2460                                 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
2461                                 NULL);
2462
2463         if (!ldlm_state->ldlm_cancel_service) {
2464                 CERROR("failed to start service\n");
2465                 GOTO(out_proc, rc = -ENOMEM);
2466         }
2467
2468         OBD_ALLOC(blp, sizeof(*blp));
2469         if (blp == NULL)
2470                 GOTO(out_proc, rc = -ENOMEM);
2471         ldlm_state->ldlm_bl_pool = blp;
2472
2473         cfs_spin_lock_init(&blp->blp_lock);
2474         CFS_INIT_LIST_HEAD(&blp->blp_list);
2475         CFS_INIT_LIST_HEAD(&blp->blp_prio_list);
2476         cfs_waitq_init(&blp->blp_waitq);
2477         cfs_atomic_set(&blp->blp_num_threads, 0);
2478         cfs_atomic_set(&blp->blp_busy_threads, 0);
2479         blp->blp_min_threads = ldlm_min_threads;
2480         blp->blp_max_threads = ldlm_max_threads;
2481
2482 #ifdef __KERNEL__
2483         for (i = 0; i < blp->blp_min_threads; i++) {
2484                 rc = ldlm_bl_thread_start(blp);
2485                 if (rc < 0)
2486                         GOTO(out_thread, rc);
2487         }
2488
2489         rc = ptlrpc_start_threads(ldlm_state->ldlm_cancel_service);
2490         if (rc)
2491                 GOTO(out_thread, rc);
2492
2493         rc = ptlrpc_start_threads(ldlm_state->ldlm_cb_service);
2494         if (rc)
2495                 GOTO(out_thread, rc);
2496
2497         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
2498         expired_lock_thread.elt_state = ELT_STOPPED;
2499         cfs_waitq_init(&expired_lock_thread.elt_waitq);
2500
2501         CFS_INIT_LIST_HEAD(&waiting_locks_list);
2502         cfs_spin_lock_init(&waiting_locks_spinlock);
2503         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
2504
2505         rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
2506         if (rc < 0) {
2507                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
2508                 GOTO(out_thread, rc);
2509         }
2510
2511         cfs_wait_event(expired_lock_thread.elt_waitq,
2512                        expired_lock_thread.elt_state == ELT_READY);
2513 #endif
2514
2515 #ifdef __KERNEL__
2516         rc = ldlm_pools_init();
2517         if (rc)
2518                 GOTO(out_thread, rc);
2519 #endif
2520         RETURN(0);
2521
2522 #ifdef __KERNEL__
2523  out_thread:
2524         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2525         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2526 #endif
2527
2528  out_proc:
2529 #ifdef LPROCFS
2530         ldlm_proc_cleanup();
2531  out_free:
2532 #endif
2533         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2534         ldlm_state = NULL;
2535         return rc;
2536 }
2537
2538 static int ldlm_cleanup(void)
2539 {
2540 #ifdef __KERNEL__
2541         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
2542 #endif
2543         ENTRY;
2544
2545         if (!cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_SERVER)) ||
2546             !cfs_list_empty(ldlm_namespace_list(LDLM_NAMESPACE_CLIENT))) {
2547                 CERROR("ldlm still has namespaces; clean these up first.\n");
2548                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
2549                 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
2550                 RETURN(-EBUSY);
2551         }
2552
2553 #ifdef __KERNEL__
2554         ldlm_pools_fini();
2555 #endif
2556
2557 #ifdef __KERNEL__
2558         while (cfs_atomic_read(&blp->blp_num_threads) > 0) {
2559                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
2560
2561                 cfs_init_completion(&blp->blp_comp);
2562
2563                 cfs_spin_lock(&blp->blp_lock);
2564                 cfs_list_add_tail(&blwi.blwi_entry, &blp->blp_list);
2565                 cfs_waitq_signal(&blp->blp_waitq);
2566                 cfs_spin_unlock(&blp->blp_lock);
2567
2568                 cfs_wait_for_completion(&blp->blp_comp);
2569         }
2570         OBD_FREE(blp, sizeof(*blp));
2571
2572         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2573         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2574         ldlm_proc_cleanup();
2575
2576         expired_lock_thread.elt_state = ELT_TERMINATE;
2577         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
2578         cfs_wait_event(expired_lock_thread.elt_waitq,
2579                        expired_lock_thread.elt_state == ELT_STOPPED);
2580 #else
2581         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
2582         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
2583 #endif
2584
2585         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
2586         ldlm_state = NULL;
2587
2588         RETURN(0);
2589 }
2590
2591 int __init ldlm_init(void)
2592 {
2593         cfs_init_mutex(&ldlm_ref_sem);
2594         cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_SERVER));
2595         cfs_init_mutex(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
2596         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
2597                                                sizeof(struct ldlm_resource), 0,
2598                                                CFS_SLAB_HWCACHE_ALIGN);
2599         if (ldlm_resource_slab == NULL)
2600                 return -ENOMEM;
2601
2602         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
2603                               sizeof(struct ldlm_lock), 0,
2604                               CFS_SLAB_HWCACHE_ALIGN | CFS_SLAB_DESTROY_BY_RCU);
2605         if (ldlm_lock_slab == NULL) {
2606                 cfs_mem_cache_destroy(ldlm_resource_slab);
2607                 return -ENOMEM;
2608         }
2609
2610         ldlm_interval_slab = cfs_mem_cache_create("interval_node",
2611                                         sizeof(struct ldlm_interval),
2612                                         0, CFS_SLAB_HWCACHE_ALIGN);
2613         if (ldlm_interval_slab == NULL) {
2614                 cfs_mem_cache_destroy(ldlm_resource_slab);
2615                 cfs_mem_cache_destroy(ldlm_lock_slab);
2616                 return -ENOMEM;
2617         }
2618 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2619         class_export_dump_hook = ldlm_dump_export_locks;
2620 #endif
2621         return 0;
2622 }
2623
2624 void __exit ldlm_exit(void)
2625 {
2626         int rc;
2627         if (ldlm_refcount)
2628                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
2629         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
2630         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
2631 #ifdef __KERNEL__
2632         /* ldlm_lock_put() use RCU to call ldlm_lock_free, so need call
2633          * synchronize_rcu() to wait a grace period elapsed, so that
2634          * ldlm_lock_free() get a chance to be called. */
2635         synchronize_rcu();
2636 #endif
2637         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
2638         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
2639         rc = cfs_mem_cache_destroy(ldlm_interval_slab);
2640         LASSERTF(rc == 0, "couldn't free interval node slab\n");
2641 }
2642
2643 /* ldlm_extent.c */
2644 EXPORT_SYMBOL(ldlm_extent_shift_kms);
2645
2646 /* ldlm_lock.c */
2647 EXPORT_SYMBOL(ldlm_get_processing_policy);
2648 EXPORT_SYMBOL(ldlm_lock2desc);
2649 EXPORT_SYMBOL(ldlm_register_intent);
2650 EXPORT_SYMBOL(ldlm_lockname);
2651 EXPORT_SYMBOL(ldlm_typename);
2652 EXPORT_SYMBOL(ldlm_lock2handle);
2653 EXPORT_SYMBOL(__ldlm_handle2lock);
2654 EXPORT_SYMBOL(ldlm_lock_get);
2655 EXPORT_SYMBOL(ldlm_lock_put);
2656 EXPORT_SYMBOL(ldlm_lock_match);
2657 EXPORT_SYMBOL(ldlm_lock_cancel);
2658 EXPORT_SYMBOL(ldlm_lock_addref);
2659 EXPORT_SYMBOL(ldlm_lock_addref_try);
2660 EXPORT_SYMBOL(ldlm_lock_decref);
2661 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
2662 EXPORT_SYMBOL(ldlm_lock_change_resource);
2663 EXPORT_SYMBOL(ldlm_it2str);
2664 EXPORT_SYMBOL(ldlm_lock_dump);
2665 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2666 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2667 EXPORT_SYMBOL(ldlm_lock_allow_match_locked);
2668 EXPORT_SYMBOL(ldlm_lock_allow_match);
2669 EXPORT_SYMBOL(ldlm_lock_downgrade);
2670 EXPORT_SYMBOL(ldlm_lock_convert);
2671
2672 /* ldlm_request.c */
2673 EXPORT_SYMBOL(ldlm_completion_ast_async);
2674 EXPORT_SYMBOL(ldlm_blocking_ast_nocheck);
2675 EXPORT_SYMBOL(ldlm_completion_ast);
2676 EXPORT_SYMBOL(ldlm_blocking_ast);
2677 EXPORT_SYMBOL(ldlm_glimpse_ast);
2678 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2679 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2680 EXPORT_SYMBOL(ldlm_prep_elc_req);
2681 EXPORT_SYMBOL(ldlm_cli_convert);
2682 EXPORT_SYMBOL(ldlm_cli_enqueue);
2683 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2684 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2685 EXPORT_SYMBOL(ldlm_cli_cancel);
2686 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2687 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2688 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2689 EXPORT_SYMBOL(ldlm_replay_locks);
2690 EXPORT_SYMBOL(ldlm_resource_foreach);
2691 EXPORT_SYMBOL(ldlm_namespace_foreach);
2692 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2693 EXPORT_SYMBOL(ldlm_resource_iterate);
2694 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2695 EXPORT_SYMBOL(ldlm_cli_cancel_list_local);
2696 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2697
2698 /* ldlm_lockd.c */
2699 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2700 EXPORT_SYMBOL(ldlm_server_completion_ast);
2701 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2702 EXPORT_SYMBOL(ldlm_handle_enqueue);
2703 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2704 EXPORT_SYMBOL(ldlm_handle_cancel);
2705 EXPORT_SYMBOL(ldlm_request_cancel);
2706 EXPORT_SYMBOL(ldlm_handle_convert);
2707 EXPORT_SYMBOL(ldlm_handle_convert0);
2708 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2709 EXPORT_SYMBOL(ldlm_get_ref);
2710 EXPORT_SYMBOL(ldlm_put_ref);
2711 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2712 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2713
2714 /* ldlm_resource.c */
2715 EXPORT_SYMBOL(ldlm_namespace_new);
2716 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2717 EXPORT_SYMBOL(ldlm_namespace_free);
2718 EXPORT_SYMBOL(ldlm_namespace_dump);
2719 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2720 EXPORT_SYMBOL(ldlm_resource_get);
2721 EXPORT_SYMBOL(ldlm_resource_putref);
2722 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2723
2724 /* ldlm_lib.c */
2725 EXPORT_SYMBOL(client_import_add_conn);
2726 EXPORT_SYMBOL(client_import_del_conn);
2727 EXPORT_SYMBOL(client_obd_setup);
2728 EXPORT_SYMBOL(client_obd_cleanup);
2729 EXPORT_SYMBOL(client_connect_import);
2730 EXPORT_SYMBOL(client_disconnect_export);
2731 EXPORT_SYMBOL(server_disconnect_export);
2732 EXPORT_SYMBOL(target_stop_recovery_thread);
2733 EXPORT_SYMBOL(target_handle_connect);
2734 EXPORT_SYMBOL(target_cleanup_recovery);
2735 EXPORT_SYMBOL(target_destroy_export);
2736 EXPORT_SYMBOL(target_cancel_recovery_timer);
2737 EXPORT_SYMBOL(target_send_reply);
2738 EXPORT_SYMBOL(target_queue_recovery_request);
2739 EXPORT_SYMBOL(target_handle_ping);
2740 EXPORT_SYMBOL(target_pack_pool_reply);
2741 EXPORT_SYMBOL(target_handle_disconnect);
2742
2743 /* l_lock.c */
2744 EXPORT_SYMBOL(lock_res_and_lock);
2745 EXPORT_SYMBOL(unlock_res_and_lock);