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  * Copyright (C) 2002-2004 Cluster File Systems, Inc.
   5  *   Author: Peter Braam <braam@clusterfs.com>
   6  *   Author: Phil Schwan <phil@clusterfs.com>
   7  *
   8  *   This file is part of the Lustre file system, http://www.lustre.org
   9  *   Lustre is a trademark of Cluster File Systems, Inc.
  10  *
  11  *   You may have signed or agreed to another license before downloading
  12  *   this software.  If so, you are bound by the terms and conditions
  13  *   of that agreement, and the following does not apply to you.  See the
  14  *   LICENSE file included with this distribution for more information.
  15  *
  16  *   If you did not agree to a different license, then this copy of Lustre
  17  *   is open source software; you can redistribute it and/or modify it
  18  *   under the terms of version 2 of the GNU General Public License as
  19  *   published by the Free Software Foundation.
  20  *
  21  *   In either case, Lustre is distributed in the hope that it will be
  22  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  23  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  24  *   license text for more details.
  25  */
  26
  27 #ifndef EXPORT_SYMTAB
  28 # define EXPORT_SYMTAB
  29 #endif
  30 #define DEBUG_SUBSYSTEM S_LDLM
  31
  32 #ifdef __KERNEL__
  33 # include <libcfs/libcfs.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <lustre_dlm.h>
  39 #include <obd_class.h>
  40 #include <libcfs/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern cfs_mem_cache_t *ldlm_resource_slab;
  44 extern cfs_mem_cache_t *ldlm_lock_slab;
  45 extern struct list_head ldlm_namespace_list;
  46
  47 extern struct semaphore ldlm_namespace_lock;
  48 static struct semaphore ldlm_ref_sem;
  49 static int ldlm_refcount;
  50
  51 /* LDLM state */
  52
  53 static struct ldlm_state *ldlm_state;
  54
  55 inline cfs_time_t round_timeout(cfs_time_t timeout)
  56 {
  57         return cfs_time_seconds((int)cfs_duration_sec(cfs_time_sub(timeout, 0)) + 1);
  58 }
  59
  60 /* timeout for initial callback (AST) reply */
  61 static inline unsigned int ldlm_get_rq_timeout(unsigned int ldlm_timeout,
  62                                                unsigned int obd_timeout)
  63 {
  64         unsigned int timeout = min(ldlm_timeout, obd_timeout / 3);
  65
  66         return timeout < 1 ? 1 : timeout;
  67 }
  68
  69 #ifdef __KERNEL__
  70 /* w_l_spinlock protects both waiting_locks_list and expired_lock_thread */
  71 static spinlock_t waiting_locks_spinlock;   /* BH lock (timer) */
  72 static struct list_head waiting_locks_list;
  73 static cfs_timer_t waiting_locks_timer;
  74
  75 static struct expired_lock_thread {
  76         cfs_waitq_t               elt_waitq;
  77         int                       elt_state;
  78         int                       elt_dump;
  79         struct list_head          elt_expired_locks;
  80 } expired_lock_thread;
  81 #endif
  82
  83 #define ELT_STOPPED   0
  84 #define ELT_READY     1
  85 #define ELT_TERMINATE 2
  86
  87 struct ldlm_bl_pool {
  88         spinlock_t              blp_lock;
  89         struct list_head        blp_list;
  90         cfs_waitq_t             blp_waitq;
  91         atomic_t                blp_num_threads;
  92         struct completion       blp_comp;
  93 };
  94
  95 struct ldlm_bl_work_item {
  96         struct list_head        blwi_entry;
  97         struct ldlm_namespace   *blwi_ns;
  98         struct ldlm_lock_desc   blwi_ld;
  99         struct ldlm_lock        *blwi_lock;
 100         int                     blwi_flags;
 101 };
 102
 103 #ifdef __KERNEL__
 104
 105 static inline int have_expired_locks(void)
 106 {
 107         int need_to_run;
 108
 109         ENTRY;
 110         spin_lock_bh(&waiting_locks_spinlock);
 111         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 112         spin_unlock_bh(&waiting_locks_spinlock);
 113
 114         RETURN(need_to_run);
 115 }
 116
 117 static int expired_lock_main(void *arg)
 118 {
 119         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 120         struct l_wait_info lwi = { 0 };
 121         int do_dump;
 122
 123         ENTRY;
 124         cfs_daemonize("ldlm_elt");
 125
 126         expired_lock_thread.elt_state = ELT_READY;
 127         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 128
 129         while (1) {
 130                 l_wait_event(expired_lock_thread.elt_waitq,
 131                              have_expired_locks() ||
 132                              expired_lock_thread.elt_state == ELT_TERMINATE,
 133                              &lwi);
 134
 135                 spin_lock_bh(&waiting_locks_spinlock);
 136                 if (expired_lock_thread.elt_dump) {
 137                         spin_unlock_bh(&waiting_locks_spinlock);
 138
 139                         /* from waiting_locks_callback, but not in timer */
 140                         libcfs_debug_dumplog();
 141                         libcfs_run_lbug_upcall(__FILE__,
 142                                                 "waiting_locks_callback",
 143                                                 expired_lock_thread.elt_dump);
 144
 145                         spin_lock_bh(&waiting_locks_spinlock);
 146                         expired_lock_thread.elt_dump = 0;
 147                 }
 148
 149                 do_dump = 0;
 150
 151                 while (!list_empty(expired)) {
 152                         struct obd_export *export;
 153                         struct ldlm_lock *lock;
 154
 155                         lock = list_entry(expired->next, struct ldlm_lock,
 156                                           l_pending_chain);
 157                         if ((void *)lock < LP_POISON + CFS_PAGE_SIZE &&
 158                             (void *)lock >= LP_POISON) {
 159                                 spin_unlock_bh(&waiting_locks_spinlock);
 160                                 CERROR("free lock on elt list %p\n", lock);
 161                                 LBUG();
 162                         }
 163                         list_del_init(&lock->l_pending_chain);
 164                         if ((void *)lock->l_export < LP_POISON + CFS_PAGE_SIZE &&
 165                             (void *)lock->l_export >= LP_POISON) {
 166                                 CERROR("lock with free export on elt list %p\n",
 167                                        lock->l_export);
 168                                 lock->l_export = NULL;
 169                                 LDLM_ERROR(lock, "free export");
 170                                 continue;
 171                         }
 172                         export = class_export_get(lock->l_export);
 173                         spin_unlock_bh(&waiting_locks_spinlock);
 174
 175                         do_dump++;
 176                         class_fail_export(export);
 177                         class_export_put(export);
 178                         spin_lock_bh(&waiting_locks_spinlock);
 179                 }
 180                 spin_unlock_bh(&waiting_locks_spinlock);
 181
 182                 if (do_dump && obd_dump_on_eviction) {
 183                         CERROR("dump the log upon eviction\n");
 184                         libcfs_debug_dumplog();
 185                 }
 186
 187                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 188                         break;
 189         }
 190
 191         expired_lock_thread.elt_state = ELT_STOPPED;
 192         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 193         RETURN(0);
 194 }
 195
 196 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 197
 198 /* This is called from within a timer interrupt and cannot schedule */
 199 static void waiting_locks_callback(unsigned long unused)
 200 {
 201         struct ldlm_lock *lock, *last = NULL;
 202
 203 repeat:
 204         spin_lock_bh(&waiting_locks_spinlock);
 205         while (!list_empty(&waiting_locks_list)) {
 206                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 207                                   l_pending_chain);
 208
 209                 if (cfs_time_after(lock->l_callback_timeout, cfs_time_current()) ||
 210                     (lock->l_req_mode == LCK_GROUP))
 211                         break;
 212
 213                 if (ptlrpc_check_suspend()) {
 214                         /* there is a case when we talk to one mds, holding
 215                          * lock from another mds. this way we easily can get
 216                          * here, if second mds is being recovered. so, we
 217                          * suspend timeouts. bug 6019 */
 218
 219                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 220                                    lock->l_export->exp_client_uuid.uuid,
 221                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 222                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 223
 224                         list_del_init(&lock->l_pending_chain);
 225                         spin_unlock_bh(&waiting_locks_spinlock);
 226                         ldlm_add_waiting_lock(lock);
 227                         goto repeat;
 228                 }
 229
 230                 /* if timeout overlaps the activation time of suspended timeouts
 231                  * then extend it to give a chance for client to reconnect */
 232                 if (cfs_time_before(cfs_time_sub(lock->l_callback_timeout,
 233                                                  cfs_time_seconds(obd_timeout)/2),
 234                                     ptlrpc_suspend_wakeup_time())) {
 235                         LDLM_ERROR(lock, "extend timeout due to recovery: %s@%s nid %s ",
 236                                    lock->l_export->exp_client_uuid.uuid,
 237                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 238                                    libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 239
 240                         list_del_init(&lock->l_pending_chain);
 241                         spin_unlock_bh(&waiting_locks_spinlock);
 242                         ldlm_add_waiting_lock(lock);
 243                         goto repeat;
 244                 }
 245
 246                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 247                            "%s@%s nid %s\n",
 248                            lock->l_export->exp_client_uuid.uuid,
 249                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 250                            libcfs_nid2str(lock->l_export->exp_connection->c_peer.nid));
 251
 252                 last = lock;
 253
 254                 list_del(&lock->l_pending_chain);
 255                 list_add(&lock->l_pending_chain,
 256                          &expired_lock_thread.elt_expired_locks);
 257         }
 258
 259         if (!list_empty(&expired_lock_thread.elt_expired_locks)) {
 260                 if (obd_dump_on_timeout)
 261                         expired_lock_thread.elt_dump = __LINE__;
 262
 263                 cfs_waitq_signal(&expired_lock_thread.elt_waitq);
 264         }
 265
 266         /*
 267          * Make sure the timer will fire again if we have any locks
 268          * left.
 269          */
 270         if (!list_empty(&waiting_locks_list)) {
 271                 cfs_time_t timeout_rounded;
 272                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 273                                   l_pending_chain);
 274                 timeout_rounded = (cfs_time_t)round_timeout(lock->l_callback_timeout);
 275                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 276         }
 277         spin_unlock_bh(&waiting_locks_spinlock);
 278 }
 279
 280 /*
 281  * Indicate that we're waiting for a client to call us back cancelling a given
 282  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 283  * timer to fire appropriately.  (We round up to the next second, to avoid
 284  * floods of timer firings during periods of high lock contention and traffic).
 285  *
 286  * Called with the namespace lock held.
 287  */
 288 static int __ldlm_add_waiting_lock(struct ldlm_lock *lock)
 289 {
 290         cfs_time_t timeout_rounded;
 291
 292         if (!list_empty(&lock->l_pending_chain))
 293                 return 0;
 294
 295         lock->l_callback_timeout =cfs_time_add(cfs_time_current(),
 296                                                cfs_time_seconds(obd_timeout)/2);
 297
 298         timeout_rounded = round_timeout(lock->l_callback_timeout);
 299
 300         if (cfs_time_before(timeout_rounded, cfs_timer_deadline(&waiting_locks_timer)) ||
 301             !cfs_timer_is_armed(&waiting_locks_timer)) {
 302                 cfs_timer_arm(&waiting_locks_timer, timeout_rounded);
 303
 304         }
 305         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 306         return 1;
 307 }
 308
 309 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 310 {
 311         int ret;
 312
 313         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 314
 315         spin_lock_bh(&waiting_locks_spinlock);
 316         if (lock->l_destroyed) {
 317                 static cfs_time_t next;
 318                 spin_unlock_bh(&waiting_locks_spinlock);
 319                 LDLM_ERROR(lock, "not waiting on destroyed lock (bug 5653)");
 320                 if (cfs_time_after(cfs_time_current(), next)) {
 321                         next = cfs_time_shift(14400);
 322                         libcfs_debug_dumpstack(NULL);
 323                 }
 324                 return 0;
 325         }
 326
 327         ret = __ldlm_add_waiting_lock(lock);
 328         spin_unlock_bh(&waiting_locks_spinlock);
 329
 330         LDLM_DEBUG(lock, "%sadding to wait list",
 331                    ret == 0 ? "not re-" : "");
 332         return ret;
 333 }
 334
 335 /*
 336  * Remove a lock from the pending list, likely because it had its cancellation
 337  * callback arrive without incident.  This adjusts the lock-timeout timer if
 338  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 339  *
 340  * Called with namespace lock held.
 341  */
 342 int __ldlm_del_waiting_lock(struct ldlm_lock *lock)
 343 {
 344         struct list_head *list_next;
 345
 346         if (list_empty(&lock->l_pending_chain))
 347                 return 0;
 348
 349         list_next = lock->l_pending_chain.next;
 350         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 351                 /* Removing the head of the list, adjust timer. */
 352                 if (list_next == &waiting_locks_list) {
 353                         /* No more, just cancel. */
 354                         cfs_timer_disarm(&waiting_locks_timer);
 355                 } else {
 356                         struct ldlm_lock *next;
 357                         next = list_entry(list_next, struct ldlm_lock,
 358                                           l_pending_chain);
 359                         cfs_timer_arm(&waiting_locks_timer,
 360                                       round_timeout(next->l_callback_timeout));
 361                 }
 362         }
 363         list_del_init(&lock->l_pending_chain);
 364
 365         return 1;
 366 }
 367
 368 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 369 {
 370         int ret;
 371
 372         if (lock->l_export == NULL) {
 373                 /* We don't have a "waiting locks list" on clients. */
 374                 LDLM_DEBUG(lock, "client lock: no-op");
 375                 return 0;
 376         }
 377
 378         spin_lock_bh(&waiting_locks_spinlock);
 379         ret = __ldlm_del_waiting_lock(lock);
 380         spin_unlock_bh(&waiting_locks_spinlock);
 381
 382         LDLM_DEBUG(lock, "%s", ret == 0 ? "wasn't waiting" : "removed");
 383         return ret;
 384 }
 385
 386 /*
 387  * Prolong the lock
 388  *
 389  * Called with namespace lock held.
 390  */
 391 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
 392 {
 393         if (lock->l_export == NULL) {
 394                 /* We don't have a "waiting locks list" on clients. */
 395                 LDLM_DEBUG(lock, "client lock: no-op");
 396                 return 0;
 397         }
 398
 399         spin_lock_bh(&waiting_locks_spinlock);
 400
 401         if (list_empty(&lock->l_pending_chain)) {
 402                 spin_unlock_bh(&waiting_locks_spinlock);
 403                 LDLM_DEBUG(lock, "wasn't waiting");
 404                 return 0;
 405         }
 406
 407         __ldlm_del_waiting_lock(lock);
 408         __ldlm_add_waiting_lock(lock);
 409         spin_unlock_bh(&waiting_locks_spinlock);
 410
 411         LDLM_DEBUG(lock, "refreshed");
 412         return 1;
 413 }
 414
 415 #else /* !__KERNEL__ */
 416
 417 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 418 {
 419         LASSERT(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK));
 420         RETURN(1);
 421 }
 422
 423 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 424 {
 425         RETURN(0);
 426 }
 427
 428 int ldlm_refresh_waiting_lock(struct ldlm_lock *lock)
 429 {
 430         RETURN(0);
 431 }
 432 #endif /* __KERNEL__ */
 433
 434 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,
 435                             const char *ast_type)
 436 {
 437         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 438         char                     *str = libcfs_nid2str(conn->c_peer.nid);
 439
 440         LCONSOLE_ERROR_MSG(0x138, "A client on nid %s was evicted from "
 441                            "service %s.\n", str,
 442                            lock->l_export->exp_obd->obd_name);
 443
 444         LCONSOLE_ERROR_MSG(0x012, "Lock %s callback to %s timed out for "
 445                            "resource %d\n", ast_type,
 446                            obd_export_nid2str(lock->l_export), rc);
 447
 448         if (obd_dump_on_timeout)
 449                 libcfs_debug_dumplog();
 450         class_fail_export(lock->l_export);
 451 }
 452
 453 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 454                                  struct ptlrpc_request *req, int rc,
 455                                  const char *ast_type)
 456 {
 457         lnet_process_id_t peer = req->rq_import->imp_connection->c_peer;
 458
 459         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 460                 LASSERT(lock->l_export);
 461                 if (lock->l_export->exp_libclient) {
 462                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 463                                    " timeout, just cancelling lock", ast_type,
 464                                    libcfs_nid2str(peer.nid));
 465                         ldlm_lock_cancel(lock);
 466                         rc = -ERESTART;
 467                 } else if (lock->l_flags & LDLM_FL_CANCEL) {
 468                         LDLM_DEBUG(lock, "%s AST timeout from nid %s, but "
 469                                    "cancel was received (AST reply lost?)",
 470                                    ast_type, libcfs_nid2str(peer.nid));
 471                         ldlm_lock_cancel(lock);
 472                         rc = -ERESTART;
 473                 } else {
 474                         ldlm_del_waiting_lock(lock);
 475                         ldlm_failed_ast(lock, rc, ast_type);
 476                 }
 477         } else if (rc) {
 478                 if (rc == -EINVAL)
 479                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 480                                " from %s AST - normal race",
 481                                libcfs_nid2str(peer.nid),
 482                                req->rq_repmsg ?
 483                                lustre_msg_get_status(req->rq_repmsg) : -1,
 484                                ast_type);
 485                 else
 486                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 487                                    "from %s AST", libcfs_nid2str(peer.nid),
 488                                    (req->rq_repmsg != NULL) ?
 489                                    lustre_msg_get_status(req->rq_repmsg) : 0,
 490                                    ast_type);
 491                 ldlm_lock_cancel(lock);
 492                 /* Server-side AST functions are called from ldlm_reprocess_all,
 493                  * which needs to be told to please restart its reprocessing. */
 494                 rc = -ERESTART;
 495         }
 496
 497         return rc;
 498 }
 499
 500 /*
 501  * ->l_blocking_ast() method for server-side locks. This is invoked when newly
 502  * enqueued server lock conflicts with given one.
 503  *
 504  * Sends blocking ast rpc to the client owning that lock; arms timeout timer
 505  * to wait for client response.
 506  */
 507 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 508                              struct ldlm_lock_desc *desc,
 509                              void *data, int flag)
 510 {
 511         struct ldlm_request *body;
 512         struct ptlrpc_request *req;
 513         int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 514                        [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 515         int instant_cancel = 0, rc = 0;
 516         ENTRY;
 517
 518         if (flag == LDLM_CB_CANCELING) {
 519                 /* Don't need to do anything here. */
 520                 RETURN(0);
 521         }
 522
 523         LASSERT(lock);
 524         if (lock->l_export->exp_obd->obd_recovering != 0) {
 525                 LDLM_ERROR(lock, "BUG 6063: lock collide during recovery");
 526                 ldlm_lock_dump(D_ERROR, lock, 0);
 527         }
 528
 529         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 530                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK, 2, size,
 531                               NULL);
 532         if (req == NULL)
 533                 RETURN(-ENOMEM);
 534
 535         lock_res(lock->l_resource);
 536         if (lock->l_granted_mode != lock->l_req_mode) {
 537                 /* this blocking AST will be communicated as part of the
 538                  * completion AST instead */
 539                 unlock_res(lock->l_resource);
 540                 ptlrpc_req_finished(req);
 541                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 542                 RETURN(0);
 543         }
 544
 545         if (lock->l_destroyed) {
 546                 /* What's the point? */
 547                 unlock_res(lock->l_resource);
 548                 ptlrpc_req_finished(req);
 549                 RETURN(0);
 550         }
 551
 552         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
 553                 instant_cancel = 1;
 554
 555         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 556         body->lock_handle[0] = lock->l_remote_handle;
 557         body->lock_desc = *desc;
 558         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 559
 560         LDLM_DEBUG(lock, "server preparing blocking AST");
 561
 562         ptlrpc_req_set_repsize(req, 1, NULL);
 563         if (instant_cancel) {
 564                 unlock_res(lock->l_resource);
 565                 ldlm_lock_cancel(lock);
 566         } else {
 567                 LASSERT(lock->l_granted_mode == lock->l_req_mode);
 568                 ldlm_add_waiting_lock(lock);
 569                 unlock_res(lock->l_resource);
 570         }
 571
 572         req->rq_send_state = LUSTRE_IMP_FULL;
 573         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
 574
 575         if (lock->l_export && lock->l_export->exp_ldlm_stats)
 576                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
 577                                      LDLM_BL_CALLBACK - LDLM_FIRST_OPC);
 578
 579         if (unlikely(instant_cancel)) {
 580                 rc = ptl_send_rpc(req, 1);
 581         } else {
 582                 rc = ptlrpc_queue_wait(req);
 583                 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_GLIMPSE, 2);
 584         }
 585         if (rc != 0) {
 586                 /* If client canceled the lock but the cancel has not been
 587                  * recieved yet, we need to update lvbo to have the proper
 588                  * attributes cached. */
 589                 if (rc == -EINVAL)
 590                         ldlm_res_lvbo_update(lock->l_resource, NULL, 0, 1);
 591                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 592         }
 593
 594         ptlrpc_req_finished(req);
 595
 596         /* If we cancelled the lock, we need to restart ldlm_reprocess_queue */
 597         if (!rc && instant_cancel)
 598                 rc = -ERESTART;
 599
 600         RETURN(rc);
 601 }
 602
 603 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 604 {
 605         struct ldlm_request *body;
 606         struct ptlrpc_request *req;
 607         struct timeval granted_time;
 608         long total_enqueue_wait;
 609         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 610                         [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 611         int rc = 0, buffers = 2, instant_cancel = 0;
 612         ENTRY;
 613
 614         LASSERT(lock != NULL);
 615
 616         do_gettimeofday(&granted_time);
 617         total_enqueue_wait = cfs_timeval_sub(&granted_time,
 618                                              &lock->l_enqueued_time, NULL);
 619
 620         if (total_enqueue_wait / 1000000 > obd_timeout)
 621                 LDLM_ERROR(lock, "enqueue wait took %luus from %lu",
 622                            total_enqueue_wait, lock->l_enqueued_time.tv_sec);
 623
 624         lock_res_and_lock(lock);
 625         if (lock->l_resource->lr_lvb_len) {
 626                 size[DLM_REQ_REC_OFF] = lock->l_resource->lr_lvb_len;
 627                 buffers = 3;
 628         }
 629         unlock_res_and_lock(lock);
 630
 631         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 632                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK, buffers,
 633                               size, NULL);
 634         if (req == NULL)
 635                 RETURN(-ENOMEM);
 636
 637         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 638         body->lock_handle[0] = lock->l_remote_handle;
 639         body->lock_flags = flags;
 640         ldlm_lock2desc(lock, &body->lock_desc);
 641
 642         if (buffers == 3) {
 643                 void *lvb;
 644
 645                 lvb = lustre_msg_buf(req->rq_reqmsg, DLM_REQ_REC_OFF,
 646                                      lock->l_resource->lr_lvb_len);
 647                 lock_res_and_lock(lock);
 648                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 649                        lock->l_resource->lr_lvb_len);
 650                 unlock_res_and_lock(lock);
 651         }
 652
 653         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 654                    total_enqueue_wait);
 655
 656         ptlrpc_req_set_repsize(req, 1, NULL);
 657
 658         req->rq_send_state = LUSTRE_IMP_FULL;
 659         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
 660
 661         /* We only send real blocking ASTs after the lock is granted */
 662         lock_res_and_lock(lock);
 663         if (lock->l_flags & LDLM_FL_AST_SENT) {
 664                 body->lock_flags |= LDLM_FL_AST_SENT;
 665
 666                 /* We might get here prior to ldlm_handle_enqueue setting
 667                  * LDLM_FL_CANCEL_ON_BLOCK flag. Then we will put this lock
 668                  * into waiting list, but this is safe and similar code in
 669                  * ldlm_handle_enqueue will call ldlm_lock_cancel() still,
 670                  * that would not only cancel the lock, but will also remove
 671                  * it from waiting list */
 672                 if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 673                         unlock_res_and_lock(lock);
 674                         ldlm_lock_cancel(lock);
 675                         instant_cancel = 1;
 676                         lock_res_and_lock(lock);
 677                 } else {
 678                         /* start the lock-timeout clock */
 679                         ldlm_add_waiting_lock(lock);
 680                 }
 681         }
 682         unlock_res_and_lock(lock);
 683
 684         if (lock->l_export && lock->l_export->exp_ldlm_stats)
 685                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
 686                                      LDLM_CP_CALLBACK - LDLM_FIRST_OPC);
 687
 688         rc = ptlrpc_queue_wait(req);
 689         if (rc != 0)
 690                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 691
 692         ptlrpc_req_finished(req);
 693
 694         /* If we cancelled the lock, we need to restart ldlm_reprocess_queue */
 695         if (!rc && instant_cancel)
 696                 rc = -ERESTART;
 697
 698         RETURN(rc);
 699 }
 700
 701 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 702 {
 703         struct ldlm_resource *res = lock->l_resource;
 704         struct ldlm_request *body;
 705         struct ptlrpc_request *req;
 706         int size[] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 707                        [DLM_LOCKREQ_OFF]     = sizeof(*body) };
 708         int rc = 0;
 709         ENTRY;
 710
 711         LASSERT(lock != NULL);
 712
 713         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 714                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK, 2, size,
 715                               NULL);
 716         if (req == NULL)
 717                 RETURN(-ENOMEM);
 718
 719         body = lustre_msg_buf(req->rq_reqmsg, DLM_LOCKREQ_OFF, sizeof(*body));
 720         body->lock_handle[0] = lock->l_remote_handle;
 721         ldlm_lock2desc(lock, &body->lock_desc);
 722
 723         lock_res_and_lock(lock);
 724         size[REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
 725         unlock_res_and_lock(lock);
 726         res = lock->l_resource;
 727         ptlrpc_req_set_repsize(req, 2, size);
 728
 729         req->rq_send_state = LUSTRE_IMP_FULL;
 730         req->rq_timeout = ldlm_get_rq_timeout(ldlm_timeout, obd_timeout);
 731
 732         if (lock->l_export && lock->l_export->exp_ldlm_stats)
 733                 lprocfs_counter_incr(lock->l_export->exp_ldlm_stats,
 734                                      LDLM_GL_CALLBACK - LDLM_FIRST_OPC);
 735
 736         rc = ptlrpc_queue_wait(req);
 737         if (rc == -ELDLM_NO_LOCK_DATA)
 738                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 739         else if (rc != 0)
 740                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 741         else
 742                 rc = ldlm_res_lvbo_update(res, req->rq_repmsg,
 743                                           REPLY_REC_OFF, 1);
 744         ptlrpc_req_finished(req);
 745         RETURN(rc);
 746 }
 747
 748 static struct ldlm_lock *
 749 find_existing_lock(struct obd_export *exp,
 750                    const struct lustre_handle *remote_hdl)
 751 {
 752         struct list_head *iter;
 753
 754         spin_lock(&exp->exp_ldlm_data.led_lock);
 755         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 756                 struct ldlm_lock *lock;
 757                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 758                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 759                         LDLM_LOCK_GET(lock);
 760                         spin_unlock(&exp->exp_ldlm_data.led_lock);
 761                         return lock;
 762                 }
 763         }
 764         spin_unlock(&exp->exp_ldlm_data.led_lock);
 765         return NULL;
 766 }
 767
 768 #ifdef __KERNEL__
 769 extern unsigned long long lu_time_stamp_get(void);
 770 #else
 771 #define lu_time_stamp_get() time(NULL)
 772 #endif
 773
 774 /*
 775  * Main server-side entry point into LDLM. This is called by ptlrpc service
 776  * threads to carry out client lock enqueueing requests.
 777  */
 778 int ldlm_handle_enqueue0(struct ldlm_namespace *ns,
 779                          struct ptlrpc_request *req,
 780                          const struct ldlm_request *dlm_req,
 781                          const struct ldlm_callback_suite *cbs)
 782 {
 783         struct ldlm_reply *dlm_rep;
 784         int size[3] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
 785                         [DLM_LOCKREPLY_OFF]   = sizeof(*dlm_rep) };
 786         int rc = 0;
 787         __u32 flags;
 788         ldlm_error_t err = ELDLM_OK;
 789         struct ldlm_lock *lock = NULL;
 790         void *cookie = NULL;
 791         ENTRY;
 792
 793         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 794
 795         ldlm_request_cancel(req, dlm_req, LDLM_ENQUEUE_CANCEL_OFF);
 796         flags = dlm_req->lock_flags;
 797
 798         LASSERT(req->rq_export);
 799
 800         if (req->rq_export->exp_ldlm_stats)
 801                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
 802                                      LDLM_ENQUEUE - LDLM_FIRST_OPC);
 803
 804         if (unlikely(dlm_req->lock_desc.l_resource.lr_type < LDLM_MIN_TYPE ||
 805                      dlm_req->lock_desc.l_resource.lr_type >= LDLM_MAX_TYPE)) {
 806                 DEBUG_REQ(D_ERROR, req, "invalid lock request type %d",
 807                           dlm_req->lock_desc.l_resource.lr_type);
 808                 GOTO(out, rc = -EFAULT);
 809         }
 810
 811         if (unlikely(dlm_req->lock_desc.l_req_mode <= LCK_MINMODE ||
 812                      dlm_req->lock_desc.l_req_mode >= LCK_MAXMODE ||
 813                      dlm_req->lock_desc.l_req_mode &
 814                      (dlm_req->lock_desc.l_req_mode-1))) {
 815                 DEBUG_REQ(D_ERROR, req, "invalid lock request mode %d",
 816                           dlm_req->lock_desc.l_req_mode);
 817                 GOTO(out, rc = -EFAULT);
 818         }
 819
 820         if (req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) {
 821                 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 822                              LDLM_PLAIN)) {
 823                         DEBUG_REQ(D_ERROR, req,
 824                                   "PLAIN lock request from IBITS client?");
 825                         GOTO(out, rc = -EPROTO);
 826                 }
 827         } else if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
 828                             LDLM_IBITS)) {
 829                 DEBUG_REQ(D_ERROR, req,
 830                           "IBITS lock request from unaware client?");
 831                 GOTO(out, rc = -EPROTO);
 832         }
 833
 834 #if 0
 835         /* FIXME this makes it impossible to use LDLM_PLAIN locks -- check
 836            against server's _CONNECT_SUPPORTED flags? (I don't want to use
 837            ibits for mgc/mgs) */
 838
 839         /* INODEBITS_INTEROP: Perform conversion from plain lock to
 840          * inodebits lock if client does not support them. */
 841         if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS) &&
 842             (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN)) {
 843                 dlm_req->lock_desc.l_resource.lr_type = LDLM_IBITS;
 844                 dlm_req->lock_desc.l_policy_data.l_inodebits.bits =
 845                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
 846                 if (dlm_req->lock_desc.l_req_mode == LCK_PR)
 847                         dlm_req->lock_desc.l_req_mode = LCK_CR;
 848         }
 849 #endif
 850
 851         if (unlikely(flags & LDLM_FL_REPLAY)) {
 852                 lock = find_existing_lock(req->rq_export,
 853                                           &dlm_req->lock_handle[0]);
 854                 if (lock != NULL) {
 855                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 856                                   lock->l_handle.h_cookie);
 857                         GOTO(existing_lock, rc = 0);
 858                 }
 859         }
 860
 861         /* The lock's callback data might be set in the policy function */
 862         lock = ldlm_lock_create(ns, &dlm_req->lock_desc.l_resource.lr_name,
 863                                 dlm_req->lock_desc.l_resource.lr_type,
 864                                 dlm_req->lock_desc.l_req_mode,
 865                                 cbs->lcs_blocking, cbs->lcs_completion,
 866                                 cbs->lcs_glimpse, NULL, 0);
 867
 868         if (!lock)
 869                 GOTO(out, rc = -ENOMEM);
 870
 871         do_gettimeofday(&lock->l_enqueued_time);
 872         lock->l_remote_handle = dlm_req->lock_handle[0];
 873         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 874
 875         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 876         /* Don't enqueue a lock onto the export if it has already
 877          * been evicted.  Cancel it now instead. (bug 3822) */
 878         if (req->rq_export->exp_failed) {
 879                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
 880                 GOTO(out, rc = -ENOTCONN);
 881         }
 882         lock->l_export = class_export_get(req->rq_export);
 883         spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 884         list_add(&lock->l_export_chain,
 885                  &lock->l_export->exp_ldlm_data.led_held_locks);
 886         spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 887
 888 existing_lock:
 889
 890         if (flags & LDLM_FL_HAS_INTENT) {
 891                 /* In this case, the reply buffer is allocated deep in
 892                  * local_lock_enqueue by the policy function. */
 893                 cookie = req;
 894         } else {
 895                 int buffers = 2;
 896
 897                 lock_res_and_lock(lock);
 898                 if (lock->l_resource->lr_lvb_len) {
 899                         size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
 900                         buffers = 3;
 901                 }
 902                 unlock_res_and_lock(lock);
 903
 904                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 905                         GOTO(out, rc = -ENOMEM);
 906
 907                 rc = lustre_pack_reply(req, buffers, size, NULL);
 908                 if (rc)
 909                         GOTO(out, rc);
 910         }
 911
 912         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 913                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
 914         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 915                 lock->l_req_extent = lock->l_policy_data.l_extent;
 916
 917         err = ldlm_lock_enqueue(ns, &lock, cookie, &flags);
 918         if (err)
 919                 GOTO(out, err);
 920
 921         dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
 922                                  sizeof(*dlm_rep));
 923         dlm_rep->lock_flags = flags;
 924
 925         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 926         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 927
 928         /* We never send a blocking AST until the lock is granted, but
 929          * we can tell it right now */
 930         lock_res_and_lock(lock);
 931
 932         /* Now take into account flags to be inherited from original lock
 933            request both in reply to client and in our own lock flags. */
 934         dlm_rep->lock_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
 935         lock->l_flags |= dlm_req->lock_flags & LDLM_INHERIT_FLAGS;
 936
 937         /* Don't move a pending lock onto the export if it has already
 938          * been evicted.  Cancel it now instead. (bug 5683) */
 939         if (unlikely(req->rq_export->exp_failed ||
 940                      OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_OLD_EXPORT))) {
 941                 LDLM_ERROR(lock, "lock on destroyed export %p", req->rq_export);
 942                 rc = -ENOTCONN;
 943         } else if (lock->l_flags & LDLM_FL_AST_SENT) {
 944                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 945                 if (lock->l_granted_mode == lock->l_req_mode) {
 946                         /*
 947                          * Only cancel lock if it was granted, because it would
 948                          * be destroyed immediatelly and would never be granted
 949                          * in the future, causing timeouts on client.  Not
 950                          * granted lock will be cancelled immediatelly after
 951                          * sending completion AST.
 952                          */
 953                         if (dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK) {
 954                                 unlock_res_and_lock(lock);
 955                                 ldlm_lock_cancel(lock);
 956                                 lock_res_and_lock(lock);
 957                         } else
 958                                 ldlm_add_waiting_lock(lock);
 959                 }
 960         }
 961         /* Make sure we never ever grant usual metadata locks to liblustre
 962            clients */
 963         if ((dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN ||
 964             dlm_req->lock_desc.l_resource.lr_type == LDLM_IBITS) &&
 965              req->rq_export->exp_libclient) {
 966                 if (unlikely(!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK) ||
 967                              !(dlm_rep->lock_flags & LDLM_FL_CANCEL_ON_BLOCK))){
 968                         CERROR("Granting sync lock to libclient. "
 969                                "req fl %d, rep fl %d, lock fl %d\n",
 970                                dlm_req->lock_flags, dlm_rep->lock_flags,
 971                                lock->l_flags);
 972                         LDLM_ERROR(lock, "sync lock");
 973                         if (dlm_req->lock_flags & LDLM_FL_HAS_INTENT) {
 974                                 struct ldlm_intent *it;
 975                                 it = lustre_msg_buf(req->rq_reqmsg,
 976                                                     DLM_INTENT_IT_OFF,
 977                                                     sizeof(*it));
 978                                 if (it != NULL) {
 979                                         CERROR("This is intent %s ("LPU64")\n",
 980                                                ldlm_it2str(it->opc), it->opc);
 981                                 }
 982                         }
 983                 }
 984         }
 985
 986         unlock_res_and_lock(lock);
 987
 988         EXIT;
 989  out:
 990         req->rq_status = err;
 991         if (req->rq_reply_state == NULL) {
 992                 err = lustre_pack_reply(req, 1, NULL, NULL);
 993                 if (rc == 0)
 994                         rc = err;
 995                 req->rq_status = rc;
 996         }
 997
 998         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 999          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
1000         if (lock) {
1001                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
1002                            "(err=%d, rc=%d)", err, rc);
1003
1004                 lock_res_and_lock(lock);
1005                 if (rc == 0) {
1006                         size[DLM_REPLY_REC_OFF] = lock->l_resource->lr_lvb_len;
1007                         if (size[DLM_REPLY_REC_OFF] > 0) {
1008                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
1009                                                        DLM_REPLY_REC_OFF,
1010                                                        size[DLM_REPLY_REC_OFF]);
1011                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
1012                                          req, lock);
1013
1014                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
1015                                        size[DLM_REPLY_REC_OFF]);
1016                         }
1017                 } else {
1018                         ldlm_resource_unlink_lock(lock);
1019                         ldlm_lock_destroy_nolock(lock);
1020                 }
1021                 unlock_res_and_lock(lock);
1022
1023                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
1024                         ldlm_reprocess_all(lock->l_resource);
1025
1026                 LDLM_LOCK_PUT(lock);
1027         }
1028
1029         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
1030                           lock, rc);
1031
1032         return rc;
1033 }
1034
1035 int ldlm_handle_enqueue(struct ptlrpc_request *req,
1036                         ldlm_completion_callback completion_callback,
1037                         ldlm_blocking_callback blocking_callback,
1038                         ldlm_glimpse_callback glimpse_callback)
1039 {
1040         int rc;
1041         struct ldlm_request *dlm_req;
1042         struct ldlm_callback_suite cbs = {
1043                 .lcs_completion = completion_callback,
1044                 .lcs_blocking   = blocking_callback,
1045                 .lcs_glimpse    = glimpse_callback
1046         };
1047
1048
1049         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1050                                      sizeof *dlm_req, lustre_swab_ldlm_request);
1051         if (dlm_req != NULL) {
1052                 rc = ldlm_handle_enqueue0(req->rq_export->exp_obd->obd_namespace,
1053                                           req, dlm_req, &cbs);
1054         } else {
1055                 CERROR ("Can't unpack dlm_req\n");
1056                 rc = -EFAULT;
1057         }
1058         return rc;
1059 }
1060
1061 int ldlm_handle_convert0(struct ptlrpc_request *req,
1062                          const struct ldlm_request *dlm_req)
1063 {
1064         struct ldlm_reply *dlm_rep;
1065         struct ldlm_lock *lock;
1066         int rc;
1067         int size[2] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
1068                         [DLM_LOCKREPLY_OFF]   = sizeof(*dlm_rep) };
1069         ENTRY;
1070
1071         if (req->rq_export && req->rq_export->exp_ldlm_stats)
1072                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1073                                      LDLM_CONVERT - LDLM_FIRST_OPC);
1074
1075         rc = lustre_pack_reply(req, 2, size, NULL);
1076         if (rc) {
1077                 CERROR("out of memory\n");
1078                 RETURN(-ENOMEM);
1079         }
1080         dlm_rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
1081                                  sizeof(*dlm_rep));
1082         dlm_rep->lock_flags = dlm_req->lock_flags;
1083
1084         lock = ldlm_handle2lock(&dlm_req->lock_handle[0]);
1085         if (!lock) {
1086                 req->rq_status = EINVAL;
1087         } else {
1088                 void *res = NULL;
1089
1090                 LDLM_DEBUG(lock, "server-side convert handler START");
1091
1092                 do_gettimeofday(&lock->l_enqueued_time);
1093                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
1094                                         &dlm_rep->lock_flags);
1095                 if (res) {
1096                         if (ldlm_del_waiting_lock(lock))
1097                                 LDLM_DEBUG(lock, "converted waiting lock");
1098                         req->rq_status = 0;
1099                 } else {
1100                         req->rq_status = EDEADLOCK;
1101                 }
1102         }
1103
1104         if (lock) {
1105                 if (!req->rq_status)
1106                         ldlm_reprocess_all(lock->l_resource);
1107                 LDLM_DEBUG(lock, "server-side convert handler END");
1108                 LDLM_LOCK_PUT(lock);
1109         } else
1110                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
1111
1112         RETURN(0);
1113 }
1114
1115 int ldlm_handle_convert(struct ptlrpc_request *req)
1116 {
1117         int rc;
1118         struct ldlm_request *dlm_req;
1119
1120         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof *dlm_req,
1121                                      lustre_swab_ldlm_request);
1122         if (dlm_req != NULL) {
1123                 rc = ldlm_handle_convert0(req, dlm_req);
1124         } else {
1125                 CERROR ("Can't unpack dlm_req\n");
1126                 rc = -EFAULT;
1127         }
1128         return rc;
1129 }
1130
1131 /* Cancel all the locks, which handles are packed into ldlm_request */
1132 int ldlm_request_cancel(struct ptlrpc_request *req,
1133                         const struct ldlm_request *dlm_req, int first)
1134 {
1135         struct ldlm_resource *res, *pres = NULL;
1136         struct ldlm_lock *lock;
1137         int i, count, done = 0;
1138         ENTRY;
1139
1140         LDLM_DEBUG_NOLOCK("server-side cancel handler START: %d locks, "
1141                           "starting at %d", dlm_req->lock_count, first);
1142         count = dlm_req->lock_count ? dlm_req->lock_count : 1;
1143         if (first >= count)
1144                 RETURN(0);
1145
1146         /* There is no lock on the server at the replay time,
1147          * skip lock cancelling to make replay tests to pass. */
1148         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1149                 RETURN(0);
1150
1151         for (i = first; i < count; i++) {
1152                 lock = ldlm_handle2lock(&dlm_req->lock_handle[i]);
1153                 if (!lock) {
1154                         LDLM_DEBUG_NOLOCK("server-side cancel handler stale "
1155                                           "lock (cookie "LPU64")",
1156                                           dlm_req->lock_handle[i].cookie);
1157                         continue;
1158                 }
1159
1160                 res = lock->l_resource;
1161                 done++;
1162
1163                 if (res != pres) {
1164                         if (pres != NULL) {
1165                                 ldlm_reprocess_all(pres);
1166                                 ldlm_resource_putref(pres);
1167                         }
1168                         if (res != NULL) {
1169                                 ldlm_resource_getref(res);
1170                                 ldlm_res_lvbo_update(res, NULL, 0, 1);
1171                         }
1172                         pres = res;
1173                 }
1174                 ldlm_lock_cancel(lock);
1175                 LDLM_LOCK_PUT(lock);
1176         }
1177         if (pres != NULL) {
1178                 ldlm_reprocess_all(pres);
1179                 ldlm_resource_putref(pres);
1180         }
1181         LDLM_DEBUG_NOLOCK("server-side cancel handler END");
1182         RETURN(done);
1183 }
1184
1185 int ldlm_handle_cancel(struct ptlrpc_request *req)
1186 {
1187         struct ldlm_request *dlm_req;
1188         int rc;
1189         ENTRY;
1190
1191         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1192                                      lustre_swab_ldlm_request);
1193         if (dlm_req == NULL) {
1194                 CERROR("bad request buffer for cancel\n");
1195                 RETURN(-EFAULT);
1196         }
1197
1198         if (req->rq_export && req->rq_export->exp_ldlm_stats)
1199                 lprocfs_counter_incr(req->rq_export->exp_ldlm_stats,
1200                                      LDLM_CANCEL - LDLM_FIRST_OPC);
1201
1202         rc = lustre_pack_reply(req, 1, NULL, NULL);
1203         if (rc) {
1204                 CERROR("out of memory\n");
1205                 RETURN(-ENOMEM);
1206         }
1207
1208         if (!ldlm_request_cancel(req, dlm_req, 0))
1209                 req->rq_status = ESTALE;
1210
1211         if (ptlrpc_reply(req) != 0)
1212                 LBUG();
1213
1214         RETURN(0);
1215 }
1216
1217 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
1218                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
1219 {
1220         int do_ast;
1221         ENTRY;
1222
1223         LDLM_DEBUG(lock, "client blocking AST callback handler START");
1224
1225         lock_res_and_lock(lock);
1226         lock->l_flags |= LDLM_FL_CBPENDING;
1227
1228         if (lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK)
1229                 lock->l_flags |= LDLM_FL_CANCEL;
1230
1231         do_ast = (!lock->l_readers && !lock->l_writers);
1232         unlock_res_and_lock(lock);
1233
1234         if (do_ast) {
1235                 LDLM_DEBUG(lock, "already unused, calling "
1236                            "callback (%p)", lock->l_blocking_ast);
1237                 if (lock->l_blocking_ast != NULL)
1238                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
1239                                              LDLM_CB_BLOCKING);
1240         } else {
1241                 LDLM_DEBUG(lock, "Lock still has references, will be"
1242                            " cancelled later");
1243         }
1244
1245         LDLM_DEBUG(lock, "client blocking callback handler END");
1246         LDLM_LOCK_PUT(lock);
1247         EXIT;
1248 }
1249
1250 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
1251                                     struct ldlm_namespace *ns,
1252                                     struct ldlm_request *dlm_req,
1253                                     struct ldlm_lock *lock)
1254 {
1255         CFS_LIST_HEAD(ast_list);
1256         ENTRY;
1257
1258         LDLM_DEBUG(lock, "client completion callback handler START");
1259
1260         lock_res_and_lock(lock);
1261
1262         /* If we receive the completion AST before the actual enqueue returned,
1263          * then we might need to switch lock modes, resources, or extents. */
1264         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
1265                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
1266                 LDLM_DEBUG(lock, "completion AST, new lock mode");
1267         }
1268
1269         if (lock->l_resource->lr_type != LDLM_PLAIN) {
1270                 lock->l_policy_data = dlm_req->lock_desc.l_policy_data;
1271                 LDLM_DEBUG(lock, "completion AST, new policy data");
1272         }
1273
1274         ldlm_resource_unlink_lock(lock);
1275         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
1276                    &lock->l_resource->lr_name,
1277                    sizeof(lock->l_resource->lr_name)) != 0) {
1278                 unlock_res_and_lock(lock);
1279                 ldlm_lock_change_resource(ns, lock,
1280                                         &dlm_req->lock_desc.l_resource.lr_name);
1281                 LDLM_DEBUG(lock, "completion AST, new resource");
1282                 CERROR("change resource!\n");
1283                 lock_res_and_lock(lock);
1284         }
1285
1286         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
1287                 lock->l_flags |= LDLM_FL_CBPENDING;
1288                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
1289         }
1290
1291         if (lock->l_lvb_len) {
1292                 void *lvb;
1293                 lvb = lustre_swab_reqbuf(req, DLM_REQ_REC_OFF, lock->l_lvb_len,
1294                                          lock->l_lvb_swabber);
1295                 if (lvb == NULL) {
1296                         LDLM_ERROR(lock, "completion AST did not contain "
1297                                    "expected LVB!");
1298                 } else {
1299                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
1300                 }
1301         }
1302
1303         ldlm_grant_lock(lock, &ast_list);
1304         unlock_res_and_lock(lock);
1305
1306         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
1307
1308         ldlm_run_cp_ast_work(&ast_list);
1309
1310         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
1311                           lock);
1312         LDLM_LOCK_PUT(lock);
1313         EXIT;
1314 }
1315
1316 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
1317                                     struct ldlm_namespace *ns,
1318                                     struct ldlm_request *dlm_req,
1319                                     struct ldlm_lock *lock)
1320 {
1321         int rc = -ENOSYS;
1322         ENTRY;
1323
1324         LDLM_DEBUG(lock, "client glimpse AST callback handler");
1325
1326         if (lock->l_glimpse_ast != NULL)
1327                 rc = lock->l_glimpse_ast(lock, req);
1328
1329         if (req->rq_repmsg != NULL) {
1330                 ptlrpc_reply(req);
1331         } else {
1332                 req->rq_status = rc;
1333                 ptlrpc_error(req);
1334         }
1335
1336         lock_res_and_lock(lock);
1337         if (lock->l_granted_mode == LCK_PW &&
1338             !lock->l_readers && !lock->l_writers &&
1339             cfs_time_after(cfs_time_current(),
1340                            cfs_time_add(lock->l_last_used,
1341                                         cfs_time_seconds(10)))) {
1342                 unlock_res_and_lock(lock);
1343                 if (ldlm_bl_to_thread(ns, NULL, lock, 0))
1344                         ldlm_handle_bl_callback(ns, NULL, lock);
1345
1346                 EXIT;
1347                 return;
1348         }
1349         unlock_res_and_lock(lock);
1350         LDLM_LOCK_PUT(lock);
1351         EXIT;
1352 }
1353
1354 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1355 {
1356         req->rq_status = rc;
1357         if (req->rq_reply_state == NULL) {
1358                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1359                 if (rc)
1360                         return rc;
1361         }
1362         return ptlrpc_reply(req);
1363 }
1364
1365 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1366                       struct ldlm_lock *lock, int flags)
1367 {
1368 #ifdef __KERNEL__
1369         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1370         struct ldlm_bl_work_item *blwi;
1371         ENTRY;
1372
1373         OBD_ALLOC(blwi, sizeof(*blwi));
1374         if (blwi == NULL)
1375                 RETURN(-ENOMEM);
1376
1377         blwi->blwi_ns = ns;
1378         if (ld != NULL)
1379                 blwi->blwi_ld = *ld;
1380         blwi->blwi_lock = lock;
1381         blwi->blwi_flags = flags;
1382
1383         spin_lock(&blp->blp_lock);
1384         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1385         cfs_waitq_signal(&blp->blp_waitq);
1386         spin_unlock(&blp->blp_lock);
1387
1388         RETURN(0);
1389 #else
1390         RETURN(-ENOSYS);
1391 #endif
1392 }
1393
1394 static int ldlm_callback_handler(struct ptlrpc_request *req)
1395 {
1396         struct ldlm_namespace *ns;
1397         struct ldlm_request *dlm_req;
1398         struct ldlm_lock *lock;
1399         int rc;
1400         ENTRY;
1401
1402         /* Requests arrive in sender's byte order.  The ptlrpc service
1403          * handler has already checked and, if necessary, byte-swapped the
1404          * incoming request message body, but I am responsible for the
1405          * message buffers. */
1406
1407         if (req->rq_export == NULL) {
1408                 struct ldlm_request *dlm_req;
1409
1410                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1411                        "export cookie "LPX64"; this is "
1412                        "normal if this node rebooted with a lock held\n",
1413                        lustre_msg_get_opc(req->rq_reqmsg),
1414                        libcfs_id2str(req->rq_peer),
1415                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1416
1417                 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1418                                              sizeof(*dlm_req),
1419                                              lustre_swab_ldlm_request);
1420                 if (dlm_req != NULL)
1421                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1422                                dlm_req->lock_handle[0].cookie);
1423
1424                 ldlm_callback_reply(req, -ENOTCONN);
1425                 RETURN(0);
1426         }
1427
1428         LASSERT(req->rq_export != NULL);
1429         LASSERT(req->rq_export->exp_obd != NULL);
1430
1431         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1432         case LDLM_BL_CALLBACK:
1433                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1434                 break;
1435         case LDLM_CP_CALLBACK:
1436                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1437                 break;
1438         case LDLM_GL_CALLBACK:
1439                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1440                 break;
1441         case OBD_LOG_CANCEL: /* remove this eventually - for 1.4.0 compat */
1442                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1443                 rc = llog_origin_handle_cancel(req);
1444                 ldlm_callback_reply(req, rc);
1445                 RETURN(0);
1446         case OBD_QC_CALLBACK:
1447                 OBD_FAIL_RETURN(OBD_FAIL_OBD_QC_CALLBACK_NET, 0);
1448                 rc = target_handle_qc_callback(req);
1449                 ldlm_callback_reply(req, rc);
1450                 RETURN(0);
1451         case QUOTA_DQACQ:
1452         case QUOTA_DQREL:
1453                 /* reply in handler */
1454                 rc = target_handle_dqacq_callback(req);
1455                 RETURN(0);
1456         case LLOG_ORIGIN_HANDLE_CREATE:
1457                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1458                 rc = llog_origin_handle_create(req);
1459                 ldlm_callback_reply(req, rc);
1460                 RETURN(0);
1461         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1462                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1463                 rc = llog_origin_handle_next_block(req);
1464                 ldlm_callback_reply(req, rc);
1465                 RETURN(0);
1466         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1467                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1468                 rc = llog_origin_handle_read_header(req);
1469                 ldlm_callback_reply(req, rc);
1470                 RETURN(0);
1471         case LLOG_ORIGIN_HANDLE_CLOSE:
1472                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1473                 rc = llog_origin_handle_close(req);
1474                 ldlm_callback_reply(req, rc);
1475                 RETURN(0);
1476         default:
1477                 CERROR("unknown opcode %u\n",
1478                        lustre_msg_get_opc(req->rq_reqmsg));
1479                 ldlm_callback_reply(req, -EPROTO);
1480                 RETURN(0);
1481         }
1482
1483         ns = req->rq_export->exp_obd->obd_namespace;
1484         LASSERT(ns != NULL);
1485
1486         dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF, sizeof(*dlm_req),
1487                                      lustre_swab_ldlm_request);
1488         if (dlm_req == NULL) {
1489                 CERROR ("can't unpack dlm_req\n");
1490                 ldlm_callback_reply(req, -EPROTO);
1491                 RETURN (0);
1492         }
1493
1494         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle[0]);
1495         if (!lock) {
1496                 CDEBUG(D_DLMTRACE, "callback on lock "LPX64" - lock "
1497                        "disappeared\n", dlm_req->lock_handle[0].cookie);
1498                 ldlm_callback_reply(req, -EINVAL);
1499                 RETURN(0);
1500         }
1501
1502         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1503         lock_res_and_lock(lock);
1504         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1505         if (lustre_msg_get_opc(req->rq_reqmsg) == LDLM_BL_CALLBACK) {
1506                 /* If somebody cancels locks and cache is already droped,
1507                  * we can tell the server we have no lock. Otherwise, we
1508                  * should send cancel after dropping the cache. */
1509                 if ((lock->l_flags & LDLM_FL_CANCELING) &&
1510                     (lock->l_flags & LDLM_FL_BL_DONE)) {
1511                         LDLM_DEBUG(lock, "callback on lock "
1512                                    LPX64" - lock disappeared\n",
1513                                    dlm_req->lock_handle[0].cookie);
1514                         unlock_res_and_lock(lock);
1515                         LDLM_LOCK_PUT(lock);
1516                         ldlm_callback_reply(req, -EINVAL);
1517                         RETURN(0);
1518                 }
1519                 lock->l_flags |= LDLM_FL_BL_AST;
1520         }
1521         unlock_res_and_lock(lock);
1522
1523         /* We want the ost thread to get this reply so that it can respond
1524          * to ost requests (write cache writeback) that might be triggered
1525          * in the callback.
1526          *
1527          * But we'd also like to be able to indicate in the reply that we're
1528          * cancelling right now, because it's unused, or have an intent result
1529          * in the reply, so we might have to push the responsibility for sending
1530          * the reply down into the AST handlers, alas. */
1531
1532         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1533         case LDLM_BL_CALLBACK:
1534                 CDEBUG(D_INODE, "blocking ast\n");
1535                 if (!(lock->l_flags & LDLM_FL_CANCEL_ON_BLOCK))
1536                         ldlm_callback_reply(req, 0);
1537                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock, 0))
1538                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1539                 break;
1540         case LDLM_CP_CALLBACK:
1541                 CDEBUG(D_INODE, "completion ast\n");
1542                 ldlm_callback_reply(req, 0);
1543                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1544                 break;
1545         case LDLM_GL_CALLBACK:
1546                 CDEBUG(D_INODE, "glimpse ast\n");
1547                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1548                 break;
1549         default:
1550                 LBUG();                         /* checked above */
1551         }
1552
1553         RETURN(0);
1554 }
1555
1556 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1557 {
1558         int rc;
1559         ENTRY;
1560
1561         /* Requests arrive in sender's byte order.  The ptlrpc service
1562          * handler has already checked and, if necessary, byte-swapped the
1563          * incoming request message body, but I am responsible for the
1564          * message buffers. */
1565
1566         if (req->rq_export == NULL) {
1567                 struct ldlm_request *dlm_req;
1568
1569                 CERROR("operation %d from %s with bad export cookie "LPU64"\n",
1570                        lustre_msg_get_opc(req->rq_reqmsg),
1571                        libcfs_id2str(req->rq_peer),
1572                        lustre_msg_get_handle(req->rq_reqmsg)->cookie);
1573
1574                 dlm_req = lustre_swab_reqbuf(req, DLM_LOCKREQ_OFF,
1575                                              sizeof(*dlm_req),
1576                                              lustre_swab_ldlm_request);
1577                 if (dlm_req != NULL)
1578                         ldlm_lock_dump_handle(D_ERROR,
1579                                               &dlm_req->lock_handle[0]);
1580                 ldlm_callback_reply(req, -ENOTCONN);
1581                 RETURN(0);
1582         }
1583
1584         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1585
1586         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1587         case LDLM_CANCEL:
1588                 CDEBUG(D_INODE, "cancel\n");
1589                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1590                 rc = ldlm_handle_cancel(req);
1591                 if (rc)
1592                         break;
1593                 RETURN(0);
1594         case OBD_LOG_CANCEL:
1595                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1596                 rc = llog_origin_handle_cancel(req);
1597                 ldlm_callback_reply(req, rc);
1598                 RETURN(0);
1599         default:
1600                 CERROR("invalid opcode %d\n",
1601                        lustre_msg_get_opc(req->rq_reqmsg));
1602                 ldlm_callback_reply(req, -EINVAL);
1603         }
1604
1605         RETURN(0);
1606 }
1607
1608 void ldlm_revoke_export_locks(struct obd_export *exp)
1609 {
1610         struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
1611         struct list_head  rpc_list;
1612         struct ldlm_lock *lock, *next;
1613         struct ldlm_lock_desc desc;
1614
1615         ENTRY;
1616         INIT_LIST_HEAD(&rpc_list);
1617
1618         spin_lock(&exp->exp_ldlm_data.led_lock);
1619         list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
1620                 lock_res_and_lock(lock);
1621
1622                 if (lock->l_req_mode != lock->l_granted_mode) {
1623                         unlock_res_and_lock(lock);
1624                         continue;
1625                 }
1626
1627                 LASSERT(lock->l_resource);
1628                 if (lock->l_resource->lr_type != LDLM_IBITS &&
1629                     lock->l_resource->lr_type != LDLM_PLAIN) {
1630                         unlock_res_and_lock(lock);
1631                         continue;
1632                 }
1633
1634                 if (lock->l_flags & LDLM_FL_AST_SENT) {
1635                         unlock_res_and_lock(lock);
1636                         continue;
1637                 }
1638
1639                 LASSERT(lock->l_blocking_ast);
1640                 LASSERT(!lock->l_blocking_lock);
1641
1642                 lock->l_flags |= LDLM_FL_AST_SENT;
1643                 list_move(&lock->l_export_chain, &rpc_list);
1644
1645                 unlock_res_and_lock(lock);
1646         }
1647         spin_unlock(&exp->exp_ldlm_data.led_lock);
1648
1649         while (!list_empty(&rpc_list)) {
1650                 lock = list_entry(rpc_list.next, struct ldlm_lock,
1651                                   l_export_chain);
1652                 list_del_init(&lock->l_export_chain);
1653
1654                 /* the desc just pretend to exclusive */
1655                 ldlm_lock2desc(lock, &desc);
1656                 desc.l_req_mode = LCK_EX;
1657                 desc.l_granted_mode = 0;
1658
1659                 LDLM_LOCK_GET(lock);
1660                 lock->l_blocking_ast(lock, &desc, lock->l_ast_data,
1661                                      LDLM_CB_BLOCKING);
1662                 LDLM_LOCK_PUT(lock);
1663         }
1664         EXIT;
1665 }
1666
1667 #ifdef __KERNEL__
1668 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1669 {
1670         struct ldlm_bl_work_item *blwi = NULL;
1671
1672         spin_lock(&blp->blp_lock);
1673         if (!list_empty(&blp->blp_list)) {
1674                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1675                                   blwi_entry);
1676                 list_del(&blwi->blwi_entry);
1677         }
1678         spin_unlock(&blp->blp_lock);
1679
1680         return blwi;
1681 }
1682
1683 struct ldlm_bl_thread_data {
1684         int                     bltd_num;
1685         struct ldlm_bl_pool     *bltd_blp;
1686 };
1687
1688 static int ldlm_bl_thread_main(void *arg)
1689 {
1690         struct ldlm_bl_thread_data *bltd = arg;
1691         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1692         ENTRY;
1693
1694         {
1695                 char name[CFS_CURPROC_COMM_MAX];
1696                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1697                          bltd->bltd_num);
1698                 cfs_daemonize(name);
1699         }
1700
1701         atomic_inc(&blp->blp_num_threads);
1702         complete(&blp->blp_comp);
1703
1704         while(1) {
1705                 struct l_wait_info lwi = { 0 };
1706                 struct ldlm_bl_work_item *blwi = NULL;
1707
1708                 l_wait_event_exclusive(blp->blp_waitq,
1709                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1710                                        &lwi);
1711
1712                 if (blwi->blwi_ns == NULL)
1713                         break;
1714
1715                 if (blwi->blwi_flags == LDLM_FL_CANCELING) {
1716                         /* The special case when we cancel locks in lru
1717                          * asynchronously, then we first remove the lock from
1718                          * l_bl_ast explicitely in ldlm_cancel_lru before
1719                          * sending it to this thread. Thus lock is marked
1720                          * LDLM_FL_CANCELING, and already cancelled locally. */
1721                         CFS_LIST_HEAD(head);
1722                         LASSERT(list_empty(&blwi->blwi_lock->l_bl_ast));
1723                         list_add(&blwi->blwi_lock->l_bl_ast, &head);
1724                         ldlm_cli_cancel_req(blwi->blwi_lock->l_conn_export,
1725                                             &head, 1, 0);
1726                         LDLM_LOCK_PUT(blwi->blwi_lock);
1727                 } else {
1728                         ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1729                                                 blwi->blwi_lock);
1730                 }
1731                 OBD_FREE(blwi, sizeof(*blwi));
1732         }
1733
1734         atomic_dec(&blp->blp_num_threads);
1735         complete(&blp->blp_comp);
1736         RETURN(0);
1737 }
1738
1739 #endif
1740
1741 static int ldlm_setup(void);
1742 static int ldlm_cleanup(int force);
1743
1744 int ldlm_get_ref(void)
1745 {
1746         int rc = 0;
1747         ENTRY;
1748         mutex_down(&ldlm_ref_sem);
1749         if (++ldlm_refcount == 1) {
1750                 rc = ldlm_setup();
1751                 if (rc)
1752                         ldlm_refcount--;
1753         }
1754         mutex_up(&ldlm_ref_sem);
1755
1756         RETURN(rc);
1757 }
1758
1759 void ldlm_put_ref(int force)
1760 {
1761         ENTRY;
1762         mutex_down(&ldlm_ref_sem);
1763         if (ldlm_refcount == 1) {
1764                 int rc = ldlm_cleanup(force);
1765                 if (rc)
1766                         CERROR("ldlm_cleanup failed: %d\n", rc);
1767                 else
1768                         ldlm_refcount--;
1769         } else {
1770                 ldlm_refcount--;
1771         }
1772         mutex_up(&ldlm_ref_sem);
1773
1774         EXIT;
1775 }
1776
1777 static int ldlm_setup(void)
1778 {
1779         struct ldlm_bl_pool *blp;
1780         int rc = 0;
1781 #ifdef __KERNEL__
1782         int i;
1783 #endif
1784         ENTRY;
1785
1786         if (ldlm_state != NULL)
1787                 RETURN(-EALREADY);
1788
1789         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1790         if (ldlm_state == NULL)
1791                 RETURN(-ENOMEM);
1792
1793 #ifdef LPROCFS
1794         rc = ldlm_proc_setup();
1795         if (rc != 0)
1796                 GOTO(out_free, rc);
1797 #endif
1798
1799         ldlm_state->ldlm_cb_service =
1800                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1801                                 LDLM_MAXREPSIZE, LDLM_CB_REQUEST_PORTAL,
1802                                 LDLM_CB_REPLY_PORTAL, ldlm_timeout * 900,
1803                                 ldlm_callback_handler, "ldlm_cbd",
1804                                 ldlm_svc_proc_dir, NULL,
1805                                 LDLM_THREADS_AUTO_MIN, LDLM_THREADS_AUTO_MAX,
1806                                 "ldlm_cb",
1807                                 LCT_MD_THREAD|LCT_DT_THREAD);
1808
1809         if (!ldlm_state->ldlm_cb_service) {
1810                 CERROR("failed to start service\n");
1811                 GOTO(out_proc, rc = -ENOMEM);
1812         }
1813
1814         ldlm_state->ldlm_cancel_service =
1815                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1816                                 LDLM_MAXREPSIZE, LDLM_CANCEL_REQUEST_PORTAL,
1817                                 LDLM_CANCEL_REPLY_PORTAL, ldlm_timeout * 6000,
1818                                 ldlm_cancel_handler, "ldlm_canceld",
1819                                 ldlm_svc_proc_dir, NULL,
1820                                 LDLM_THREADS_AUTO_MIN, LDLM_THREADS_AUTO_MAX,
1821                                 "ldlm_cn",
1822                                 LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD);
1823
1824         if (!ldlm_state->ldlm_cancel_service) {
1825                 CERROR("failed to start service\n");
1826                 GOTO(out_proc, rc = -ENOMEM);
1827         }
1828
1829         OBD_ALLOC(blp, sizeof(*blp));
1830         if (blp == NULL)
1831                 GOTO(out_proc, rc = -ENOMEM);
1832         ldlm_state->ldlm_bl_pool = blp;
1833
1834         atomic_set(&blp->blp_num_threads, 0);
1835         cfs_waitq_init(&blp->blp_waitq);
1836         spin_lock_init(&blp->blp_lock);
1837
1838         CFS_INIT_LIST_HEAD(&blp->blp_list);
1839
1840 #ifdef __KERNEL__
1841         for (i = 0; i < LDLM_BL_THREADS; i++) {
1842                 struct ldlm_bl_thread_data bltd = {
1843                         .bltd_num = i,
1844                         .bltd_blp = blp,
1845                 };
1846                 init_completion(&blp->blp_comp);
1847                 rc = cfs_kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1848                 if (rc < 0) {
1849                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1850                         GOTO(out_thread, rc);
1851                 }
1852                 wait_for_completion(&blp->blp_comp);
1853         }
1854
1855         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cancel_service);
1856         if (rc)
1857                 GOTO(out_thread, rc);
1858
1859         rc = ptlrpc_start_threads(NULL, ldlm_state->ldlm_cb_service);
1860         if (rc)
1861                 GOTO(out_thread, rc);
1862
1863         CFS_INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1864         expired_lock_thread.elt_state = ELT_STOPPED;
1865         cfs_waitq_init(&expired_lock_thread.elt_waitq);
1866
1867         CFS_INIT_LIST_HEAD(&waiting_locks_list);
1868         spin_lock_init(&waiting_locks_spinlock);
1869         cfs_timer_init(&waiting_locks_timer, waiting_locks_callback, 0);
1870
1871         rc = cfs_kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FILES);
1872         if (rc < 0) {
1873                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1874                 GOTO(out_thread, rc);
1875         }
1876
1877         wait_event(expired_lock_thread.elt_waitq,
1878                    expired_lock_thread.elt_state == ELT_READY);
1879 #endif
1880
1881         RETURN(0);
1882
1883 #ifdef __KERNEL__
1884  out_thread:
1885         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1886         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1887 #endif
1888
1889  out_proc:
1890 #ifdef LPROCFS
1891         ldlm_proc_cleanup();
1892  out_free:
1893 #endif
1894         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1895         ldlm_state = NULL;
1896         return rc;
1897 }
1898
1899 static int ldlm_cleanup(int force)
1900 {
1901 #ifdef __KERNEL__
1902         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1903 #endif
1904         ENTRY;
1905
1906         if (!list_empty(&ldlm_namespace_list)) {
1907                 CERROR("ldlm still has namespaces; clean these up first.\n");
1908                 ldlm_dump_all_namespaces(D_DLMTRACE);
1909                 RETURN(-EBUSY);
1910         }
1911
1912 #ifdef __KERNEL__
1913         while (atomic_read(&blp->blp_num_threads) > 0) {
1914                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1915
1916                 init_completion(&blp->blp_comp);
1917
1918                 spin_lock(&blp->blp_lock);
1919                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1920                 cfs_waitq_signal(&blp->blp_waitq);
1921                 spin_unlock(&blp->blp_lock);
1922
1923                 wait_for_completion(&blp->blp_comp);
1924         }
1925         OBD_FREE(blp, sizeof(*blp));
1926
1927         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1928         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1929         ldlm_proc_cleanup();
1930
1931         expired_lock_thread.elt_state = ELT_TERMINATE;
1932         cfs_waitq_signal(&expired_lock_thread.elt_waitq);
1933         wait_event(expired_lock_thread.elt_waitq,
1934                    expired_lock_thread.elt_state == ELT_STOPPED);
1935 #else
1936         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1937         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1938 #endif
1939
1940         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1941         ldlm_state = NULL;
1942
1943         RETURN(0);
1944 }
1945
1946 int __init ldlm_init(void)
1947 {
1948         init_mutex(&ldlm_ref_sem);
1949         init_mutex(&ldlm_namespace_lock);
1950         ldlm_resource_slab = cfs_mem_cache_create("ldlm_resources",
1951                                                sizeof(struct ldlm_resource), 0,
1952                                                SLAB_HWCACHE_ALIGN);
1953         if (ldlm_resource_slab == NULL)
1954                 return -ENOMEM;
1955
1956         ldlm_lock_slab = cfs_mem_cache_create("ldlm_locks",
1957                                            sizeof(struct ldlm_lock), 0,
1958                                            SLAB_HWCACHE_ALIGN);
1959         if (ldlm_lock_slab == NULL) {
1960                 cfs_mem_cache_destroy(ldlm_resource_slab);
1961                 return -ENOMEM;
1962         }
1963
1964         return 0;
1965 }
1966
1967 void __exit ldlm_exit(void)
1968 {
1969         int rc;
1970         if (ldlm_refcount)
1971                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1972         rc = cfs_mem_cache_destroy(ldlm_resource_slab);
1973         LASSERTF(rc == 0, "couldn't free ldlm resource slab\n");
1974         rc = cfs_mem_cache_destroy(ldlm_lock_slab);
1975         LASSERTF(rc == 0, "couldn't free ldlm lock slab\n");
1976 }
1977
1978 /* ldlm_extent.c */
1979 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1980
1981 /* ldlm_lock.c */
1982 EXPORT_SYMBOL(ldlm_get_processing_policy);
1983 EXPORT_SYMBOL(ldlm_lock2desc);
1984 EXPORT_SYMBOL(ldlm_register_intent);
1985 EXPORT_SYMBOL(ldlm_lockname);
1986 EXPORT_SYMBOL(ldlm_typename);
1987 EXPORT_SYMBOL(ldlm_lock2handle);
1988 EXPORT_SYMBOL(__ldlm_handle2lock);
1989 EXPORT_SYMBOL(ldlm_lock_get);
1990 EXPORT_SYMBOL(ldlm_lock_put);
1991 EXPORT_SYMBOL(ldlm_lock_match);
1992 EXPORT_SYMBOL(ldlm_lock_cancel);
1993 EXPORT_SYMBOL(ldlm_lock_addref);
1994 EXPORT_SYMBOL(ldlm_lock_decref);
1995 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1996 EXPORT_SYMBOL(ldlm_lock_change_resource);
1997 EXPORT_SYMBOL(ldlm_lock_set_data);
1998 EXPORT_SYMBOL(ldlm_it2str);
1999 EXPORT_SYMBOL(ldlm_lock_dump);
2000 EXPORT_SYMBOL(ldlm_lock_dump_handle);
2001 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
2002 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
2003 EXPORT_SYMBOL(ldlm_lock_allow_match);
2004
2005 /* ldlm_request.c */
2006 EXPORT_SYMBOL(ldlm_completion_ast);
2007 EXPORT_SYMBOL(ldlm_blocking_ast);
2008 EXPORT_SYMBOL(ldlm_glimpse_ast);
2009 EXPORT_SYMBOL(ldlm_expired_completion_wait);
2010 EXPORT_SYMBOL(ldlm_prep_enqueue_req);
2011 EXPORT_SYMBOL(ldlm_cli_convert);
2012 EXPORT_SYMBOL(ldlm_cli_enqueue);
2013 EXPORT_SYMBOL(ldlm_cli_enqueue_fini);
2014 EXPORT_SYMBOL(ldlm_cli_enqueue_local);
2015 EXPORT_SYMBOL(ldlm_cli_cancel);
2016 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
2017 EXPORT_SYMBOL(ldlm_cli_cancel_unused_resource);
2018 EXPORT_SYMBOL(ldlm_cli_cancel_req);
2019 EXPORT_SYMBOL(ldlm_cli_join_lru);
2020 EXPORT_SYMBOL(ldlm_replay_locks);
2021 EXPORT_SYMBOL(ldlm_resource_foreach);
2022 EXPORT_SYMBOL(ldlm_namespace_foreach);
2023 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
2024 EXPORT_SYMBOL(ldlm_resource_iterate);
2025 EXPORT_SYMBOL(ldlm_cancel_resource_local);
2026 EXPORT_SYMBOL(ldlm_cli_cancel_list);
2027
2028 /* ldlm_lockd.c */
2029 EXPORT_SYMBOL(ldlm_server_blocking_ast);
2030 EXPORT_SYMBOL(ldlm_server_completion_ast);
2031 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
2032 EXPORT_SYMBOL(ldlm_handle_enqueue);
2033 EXPORT_SYMBOL(ldlm_handle_enqueue0);
2034 EXPORT_SYMBOL(ldlm_handle_cancel);
2035 EXPORT_SYMBOL(ldlm_request_cancel);
2036 EXPORT_SYMBOL(ldlm_handle_convert);
2037 EXPORT_SYMBOL(ldlm_handle_convert0);
2038 EXPORT_SYMBOL(ldlm_del_waiting_lock);
2039 EXPORT_SYMBOL(ldlm_get_ref);
2040 EXPORT_SYMBOL(ldlm_put_ref);
2041 EXPORT_SYMBOL(ldlm_refresh_waiting_lock);
2042 EXPORT_SYMBOL(ldlm_revoke_export_locks);
2043
2044 /* ldlm_resource.c */
2045 EXPORT_SYMBOL(ldlm_namespace_new);
2046 EXPORT_SYMBOL(ldlm_namespace_cleanup);
2047 EXPORT_SYMBOL(ldlm_namespace_free);
2048 EXPORT_SYMBOL(ldlm_namespace_dump);
2049 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
2050 EXPORT_SYMBOL(ldlm_resource_get);
2051 EXPORT_SYMBOL(ldlm_resource_putref);
2052 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
2053
2054 /* ldlm_lib.c */
2055 EXPORT_SYMBOL(client_import_add_conn);
2056 EXPORT_SYMBOL(client_import_del_conn);
2057 EXPORT_SYMBOL(client_obd_setup);
2058 EXPORT_SYMBOL(client_obd_cleanup);
2059 EXPORT_SYMBOL(client_connect_import);
2060 EXPORT_SYMBOL(client_disconnect_export);
2061 EXPORT_SYMBOL(target_start_recovery_thread);
2062 EXPORT_SYMBOL(target_stop_recovery_thread);
2063 EXPORT_SYMBOL(target_handle_connect);
2064 EXPORT_SYMBOL(target_cleanup_recovery);
2065 EXPORT_SYMBOL(target_destroy_export);
2066 EXPORT_SYMBOL(target_cancel_recovery_timer);
2067 EXPORT_SYMBOL(target_send_reply);
2068 EXPORT_SYMBOL(target_queue_recovery_request);
2069 EXPORT_SYMBOL(target_handle_ping);
2070 EXPORT_SYMBOL(target_handle_disconnect);
2071
2072 /* l_lock.c */
2073 EXPORT_SYMBOL(lock_res_and_lock);
2074 EXPORT_SYMBOL(unlock_res_and_lock);
2075