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