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