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, 2003 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 Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #ifndef EXPORT_SYMTAB
  25 # define EXPORT_SYMTAB
  26 #endif
  27 #define DEBUG_SUBSYSTEM S_LDLM
  28
  29 #ifdef __KERNEL__
  30 # include <linux/module.h>
  31 # include <linux/slab.h>
  32 # include <linux/init.h>
  33 # include <linux/wait.h>
  34 #else
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <linux/lustre_dlm.h>
  39 #include <linux/obd_class.h>
  40 #include <libcfs/list.h>
  41 #include "ldlm_internal.h"
  42
  43 extern kmem_cache_t *ldlm_resource_slab;
  44 extern kmem_cache_t *ldlm_lock_slab;
  45 extern struct list_head ldlm_namespace_list;
  46
  47 static DECLARE_MUTEX(ldlm_ref_sem);
  48 static int ldlm_refcount;
  49 /* LDLM state */
  50
  51 static struct ldlm_state *ldlm_state;
  52
  53 static inline unsigned long
  54 round_timeout(unsigned long timeout)
  55 {
  56         return ((timeout / HZ) + 1) * HZ;
  57 }
  58
  59 #ifdef __KERNEL__
  60 /* XXX should this be per-ldlm? */
  61 static struct list_head waiting_locks_list;
  62 static spinlock_t waiting_locks_spinlock;
  63 static struct timer_list waiting_locks_timer;
  64
  65 static struct expired_lock_thread {
  66         wait_queue_head_t         elt_waitq;
  67         int                       elt_state;
  68         struct list_head          elt_expired_locks;
  69         spinlock_t                elt_lock;
  70 } expired_lock_thread;
  71 #endif
  72
  73 #if !defined(ENOTSUPP)
  74 #  define ENOTSUPP 524
  75 #endif
  76
  77 #define ELT_STOPPED   0
  78 #define ELT_READY     1
  79 #define ELT_TERMINATE 2
  80
  81 struct ldlm_bl_pool {
  82         spinlock_t              blp_lock;
  83         struct list_head        blp_list;
  84         wait_queue_head_t       blp_waitq;
  85         atomic_t                blp_num_threads;
  86         struct completion       blp_comp;
  87 };
  88
  89 struct ldlm_bl_work_item {
  90         struct list_head        blwi_entry;
  91         struct ldlm_namespace   *blwi_ns;
  92         struct ldlm_lock_desc   blwi_ld;
  93         struct ldlm_lock        *blwi_lock;
  94 };
  95
  96 #ifdef __KERNEL__
  97
  98 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
  99
 100 static inline int have_expired_locks(void)
 101 {
 102         int need_to_run;
 103
 104         spin_lock_bh(&expired_lock_thread.elt_lock);
 105         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 106         spin_unlock_bh(&expired_lock_thread.elt_lock);
 107
 108         RETURN(need_to_run);
 109 }
 110
 111 static int expired_lock_main(void *arg)
 112 {
 113         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 114         struct l_wait_info lwi = { 0 };
 115         unsigned long flags;
 116
 117         ENTRY;
 118         lock_kernel();
 119         kportal_daemonize("ldlm_elt");
 120
 121         SIGNAL_MASK_LOCK(current, flags);
 122         sigfillset(&current->blocked);
 123         RECALC_SIGPENDING;
 124         SIGNAL_MASK_UNLOCK(current, flags);
 125
 126         unlock_kernel();
 127
 128         expired_lock_thread.elt_state = ELT_READY;
 129         wake_up(&expired_lock_thread.elt_waitq);
 130
 131         while (1) {
 132                 l_wait_event(expired_lock_thread.elt_waitq,
 133                              have_expired_locks() ||
 134                              expired_lock_thread.elt_state == ELT_TERMINATE,
 135                              &lwi);
 136
 137                 spin_lock_bh(&expired_lock_thread.elt_lock);
 138                 while (!list_empty(expired)) {
 139                         struct obd_export *export;
 140                         struct ldlm_lock *lock;
 141
 142                         lock = list_entry(expired->next, struct ldlm_lock,
 143                                           l_pending_chain);
 144                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 145                             (void *)lock >= LP_POISON) {
 146                                 CERROR("free lock on elt list %p\n", lock);
 147                                 LBUG();
 148                         }
 149                         list_del_init(&lock->l_pending_chain);
 150                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 151                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 152                                 CERROR("lock with free export on elt list %p\n",
 153                                        export);
 154                                 lock->l_export = NULL;
 155                                 LDLM_ERROR(lock, "free export\n");
 156                                 continue;
 157                         }
 158                         export = class_export_get(lock->l_export);
 159                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 160
 161                         ptlrpc_fail_export(export);
 162                         class_export_put(export);
 163                         spin_lock_bh(&expired_lock_thread.elt_lock);
 164                 }
 165                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 166
 167                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 168                         break;
 169         }
 170
 171         expired_lock_thread.elt_state = ELT_STOPPED;
 172         wake_up(&expired_lock_thread.elt_waitq);
 173         RETURN(0);
 174 }
 175
 176 static void waiting_locks_callback(unsigned long unused)
 177 {
 178         struct ldlm_lock *lock;
 179         char str[PTL_NALFMT_SIZE];
 180
 181         if (obd_dump_on_timeout)
 182                 portals_debug_dumplog();
 183
 184 repeat:
 185         spin_lock_bh(&waiting_locks_spinlock);
 186         while (!list_empty(&waiting_locks_list)) {
 187                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 188                                   l_pending_chain);
 189
 190                 if ((lock->l_callback_timeout > jiffies) ||
 191                     (lock->l_req_mode == LCK_GROUP))
 192                         break;
 193
 194                 if (ptlrpc_check_suspend()) {
 195                         /* there is a case when we talk to one mds, holding
 196                          * lock from another mds. this way we easily can get
 197                          * here, if second mds is being recovered. so, we
 198                          * suspend timeouts. bug 6019 */
 199
 200                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 201                                    lock->l_export->exp_client_uuid.uuid,
 202                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 203                                    ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 204
 205                         list_del_init(&lock->l_pending_chain);
 206                         spin_unlock_bh(&waiting_locks_spinlock);
 207                         ldlm_add_waiting_lock(lock);
 208
 209                         goto repeat;
 210                 }
 211
 212                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 213                            "%s@%s nid %s ",
 214                            lock->l_export->exp_client_uuid.uuid,
 215                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 216                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 217
 218                 spin_lock_bh(&expired_lock_thread.elt_lock);
 219                 list_del(&lock->l_pending_chain);
 220                 list_add(&lock->l_pending_chain,
 221                          &expired_lock_thread.elt_expired_locks);
 222                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 223                 wake_up(&expired_lock_thread.elt_waitq);
 224         }
 225
 226         /*
 227          * Make sure the timer will fire again if we have any locks
 228          * left.
 229          */
 230         if (!list_empty(&waiting_locks_list)) {
 231                 unsigned long timeout_rounded;
 232                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 233                                   l_pending_chain);
 234                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 235                 mod_timer(&waiting_locks_timer, timeout_rounded);
 236         }
 237         spin_unlock_bh(&waiting_locks_spinlock);
 238 }
 239
 240 /*
 241  * Indicate that we're waiting for a client to call us back cancelling a given
 242  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 243  * timer to fire appropriately.  (We round up to the next second, to avoid
 244  * floods of timer firings during periods of high lock contention and traffic).
 245  */
 246 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 247 {
 248         unsigned long timeout_rounded;
 249
 250         spin_lock_bh(&waiting_locks_spinlock);
 251         if (!list_empty(&lock->l_pending_chain)) {
 252                 LDLM_DEBUG(lock, "not re-adding to wait list");
 253                 spin_unlock_bh(&waiting_locks_spinlock);
 254                 return 0;
 255         }
 256         LDLM_DEBUG(lock, "adding to wait list");
 257
 258         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 259
 260         timeout_rounded = round_timeout(lock->l_callback_timeout);
 261
 262         if (timeout_rounded < waiting_locks_timer.expires ||
 263             !timer_pending(&waiting_locks_timer)) {
 264                 mod_timer(&waiting_locks_timer, timeout_rounded);
 265         }
 266         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 267         spin_unlock_bh(&waiting_locks_spinlock);
 268         return 1;
 269 }
 270
 271 /*
 272  * Remove a lock from the pending list, likely because it had its cancellation
 273  * callback arrive without incident.  This adjusts the lock-timeout timer if
 274  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 275  */
 276 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 277 {
 278         struct list_head *list_next;
 279
 280         if (lock->l_export == NULL) {
 281                 /* We don't have a "waiting locks list" on clients. */
 282                 LDLM_DEBUG(lock, "client lock: no-op");
 283                 return 0;
 284         }
 285
 286         spin_lock_bh(&waiting_locks_spinlock);
 287
 288         if (list_empty(&lock->l_pending_chain)) {
 289                 spin_unlock_bh(&waiting_locks_spinlock);
 290                 LDLM_DEBUG(lock, "wasn't waiting");
 291                 return 0;
 292         }
 293
 294         list_next = lock->l_pending_chain.next;
 295         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 296                 /* Removing the head of the list, adjust timer. */
 297                 if (list_next == &waiting_locks_list) {
 298                         /* No more, just cancel. */
 299                         del_timer(&waiting_locks_timer);
 300                 } else {
 301                         struct ldlm_lock *next;
 302                         next = list_entry(list_next, struct ldlm_lock,
 303                                           l_pending_chain);
 304                         mod_timer(&waiting_locks_timer,
 305                                   round_timeout(next->l_callback_timeout));
 306                 }
 307         }
 308
 309         spin_lock_bh(&expired_lock_thread.elt_lock);
 310         list_del_init(&lock->l_pending_chain);
 311         spin_unlock_bh(&expired_lock_thread.elt_lock);
 312
 313         spin_unlock_bh(&waiting_locks_spinlock);
 314         LDLM_DEBUG(lock, "removed");
 315         return 1;
 316 }
 317
 318 #else /* !__KERNEL__ */
 319
 320 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 321 {
 322         RETURN(1);
 323 }
 324
 325 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 326 {
 327         RETURN(0);
 328 }
 329
 330 #endif /* __KERNEL__ */
 331
 332 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 333 {
 334         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 335         char str[PTL_NALFMT_SIZE];
 336
 337         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 338                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 339                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 340                    ptlrpc_peernid2str(&conn->c_peer, str));
 341
 342         if (obd_dump_on_timeout)
 343                 portals_debug_dumplog();
 344         ptlrpc_fail_export(lock->l_export);
 345 }
 346
 347 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 348                                  struct ptlrpc_request *req, int rc,
 349                                  const char *ast_type)
 350 {
 351         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 352         char str[PTL_NALFMT_SIZE];
 353
 354         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 355                 LASSERT(lock->l_export);
 356                 if (lock->l_export->exp_libclient) {
 357                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 358                                    " timeout, just cancelling lock", ast_type,
 359                                    ptlrpc_peernid2str(peer, str));
 360                         ldlm_lock_cancel(lock);
 361                         rc = -ERESTART;
 362                 } else {
 363                         ldlm_del_waiting_lock(lock);
 364                         ldlm_failed_ast(lock, rc, ast_type);
 365                 }
 366         } else if (rc) {
 367                 if (rc == -EINVAL)
 368                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 369                                    " from %s AST - normal race",
 370                                    ptlrpc_peernid2str(peer, str),
 371                                    req->rq_repmsg->status, ast_type);
 372                 else
 373                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 374                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 375                                    (req->rq_repmsg != NULL) ?
 376                                    req->rq_repmsg->status : 0, ast_type);
 377                 ldlm_lock_cancel(lock);
 378                 /* Server-side AST functions are called from ldlm_reprocess_all,
 379                  * which needs to be told to please restart its reprocessing. */
 380                 rc = -ERESTART;
 381         }
 382
 383         return rc;
 384 }
 385
 386 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 387                              struct ldlm_lock_desc *desc,
 388                              void *data, int flag)
 389 {
 390         struct ldlm_request *body;
 391         struct ptlrpc_request *req;
 392         int rc = 0, size = sizeof(*body);
 393         ENTRY;
 394
 395         if (flag == LDLM_CB_CANCELING) {
 396                 /* Don't need to do anything here. */
 397                 RETURN(0);
 398         }
 399
 400         LASSERTF(lock->l_export->exp_obd->obd_recovering == 0,
 401                  "BUG 6063: lock collide during recovery");
 402
 403         LASSERT(lock);
 404
 405 #if 0
 406         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 407                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 408                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 409                 RETURN(-ETIMEDOUT);
 410         }
 411 #endif
 412
 413         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 414                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 415                               1, &size, NULL);
 416         if (req == NULL)
 417                 RETURN(-ENOMEM);
 418
 419         lock_res_and_lock(lock);
 420         if (lock->l_granted_mode != lock->l_req_mode) {
 421                 /* this blocking AST will be communicated as part of the
 422                  * completion AST instead */
 423                 unlock_res_and_lock(lock);
 424                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 425                 ptlrpc_req_finished(req);
 426                 RETURN(0);
 427         }
 428
 429         if (lock->l_destroyed) {
 430                 /* What's the point? */
 431                 unlock_res_and_lock(lock);
 432                 ptlrpc_req_finished(req);
 433                 RETURN(0);
 434         }
 435
 436         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 437         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 438                sizeof(body->lock_handle1));
 439         memcpy(&body->lock_desc, desc, sizeof(*desc));
 440         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 441
 442         LDLM_DEBUG(lock, "server preparing blocking AST");
 443         req->rq_replen = lustre_msg_size(0, NULL);
 444
 445         if (lock->l_granted_mode == lock->l_req_mode)
 446                 ldlm_add_waiting_lock(lock);
 447         unlock_res_and_lock(lock);
 448
 449         req->rq_send_state = LUSTRE_IMP_FULL;
 450         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 451         rc = ptlrpc_queue_wait(req);
 452         if (rc != 0)
 453                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 454
 455         ptlrpc_req_finished(req);
 456
 457         RETURN(rc);
 458 }
 459
 460 /* XXX copied from ptlrpc/service.c */
 461 static long timeval_sub(struct timeval *large, struct timeval *small)
 462 {
 463         return (large->tv_sec - small->tv_sec) * 1000000 +
 464                 (large->tv_usec - small->tv_usec);
 465 }
 466
 467 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 468 {
 469         struct ldlm_request *body;
 470         struct ptlrpc_request *req;
 471         struct timeval granted_time;
 472         long total_enqueue_wait;
 473         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 474         ENTRY;
 475
 476         LASSERT(lock != NULL);
 477
 478         do_gettimeofday(&granted_time);
 479         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 480
 481         if (((flags & LDLM_FL_NO_TIMEOUT) == 0) &&
 482             ((total_enqueue_wait / 1000000) > obd_timeout))
 483                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 484
 485         lock_res_and_lock(lock);
 486         if (lock->l_resource->lr_lvb_len) {
 487                 buffers = 2;
 488                 size[1] = lock->l_resource->lr_lvb_len;
 489         }
 490         unlock_res_and_lock(lock);
 491
 492         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 493                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 494                               buffers, size, NULL);
 495         if (req == NULL)
 496                 RETURN(-ENOMEM);
 497
 498         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 499         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 500                sizeof(body->lock_handle1));
 501         body->lock_flags = flags;
 502         ldlm_lock2desc(lock, &body->lock_desc);
 503
 504         if (buffers == 2) {
 505                 void *lvb;
 506
 507                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 508                                      lock->l_resource->lr_lvb_len);
 509                 lock_res_and_lock(lock);
 510                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 511                        lock->l_resource->lr_lvb_len);
 512                 unlock_res_and_lock(lock);
 513         }
 514
 515         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 516                    total_enqueue_wait);
 517         req->rq_replen = lustre_msg_size(0, NULL);
 518
 519         req->rq_send_state = LUSTRE_IMP_FULL;
 520         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 521
 522         /* We only send real blocking ASTs after the lock is granted */
 523         lock_res_and_lock(lock);
 524         if (lock->l_flags & LDLM_FL_AST_SENT) {
 525                 body->lock_flags |= LDLM_FL_AST_SENT;
 526                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 527         }
 528         unlock_res_and_lock(lock);
 529
 530         rc = ptlrpc_queue_wait(req);
 531         if (rc != 0)
 532                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 533
 534         ptlrpc_req_finished(req);
 535
 536         RETURN(rc);
 537 }
 538
 539 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 540 {
 541         struct ldlm_resource *res;
 542         struct ldlm_request *body;
 543         struct ptlrpc_request *req;
 544         int rc = 0, size = sizeof(*body);
 545         ENTRY;
 546
 547         LASSERT(lock != NULL);
 548
 549         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 550                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 551                               1, &size, NULL);
 552         if (req == NULL)
 553                 RETURN(-ENOMEM);
 554
 555         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 556         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 557                sizeof(body->lock_handle1));
 558         ldlm_lock2desc(lock, &body->lock_desc);
 559
 560         lock_res_and_lock(lock);
 561         size = lock->l_resource->lr_lvb_len;
 562         unlock_res_and_lock(lock);
 563         res = lock->l_resource;
 564         req->rq_replen = lustre_msg_size(1, &size);
 565
 566         req->rq_send_state = LUSTRE_IMP_FULL;
 567         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 568
 569         rc = ptlrpc_queue_wait(req);
 570         if (rc == -ELDLM_NO_LOCK_DATA)
 571                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 572         else if (rc != 0)
 573                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 574         else
 575                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 576                         (res, req->rq_repmsg, 0, 1);
 577         ptlrpc_req_finished(req);
 578         RETURN(rc);
 579 }
 580
 581 static struct ldlm_lock *
 582 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 583 {
 584         struct list_head *iter;
 585
 586         spin_lock(&exp->exp_ldlm_data.led_lock);
 587         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 588                 struct ldlm_lock *lock;
 589                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 590                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 591                         LDLM_LOCK_GET(lock);
 592                         spin_unlock(&exp->exp_ldlm_data.led_lock);
 593                         return lock;
 594                 }
 595         }
 596         spin_unlock(&exp->exp_ldlm_data.led_lock);
 597         return NULL;
 598 }
 599
 600 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 601                         ldlm_completion_callback completion_callback,
 602                         ldlm_blocking_callback blocking_callback,
 603                         ldlm_glimpse_callback glimpse_callback)
 604 {
 605         struct obd_device *obddev = req->rq_export->exp_obd;
 606         struct ldlm_reply *dlm_rep;
 607         struct ldlm_request *dlm_req;
 608         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 609         __u32 flags;
 610         ldlm_error_t err = ELDLM_OK;
 611         struct ldlm_lock *lock = NULL;
 612         void *cookie = NULL;
 613         ENTRY;
 614
 615         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 616
 617         dlm_req = lustre_swab_reqbuf (req, MDS_REQ_INTENT_LOCKREQ_OFF,
 618                                       sizeof (*dlm_req),
 619                                       lustre_swab_ldlm_request);
 620         if (dlm_req == NULL) {
 621                 CERROR ("Can't unpack dlm_req\n");
 622                 GOTO(out, rc = -EFAULT);
 623         }
 624
 625         flags = dlm_req->lock_flags;
 626
 627         LASSERT(req->rq_export);
 628
 629         if (flags & LDLM_FL_REPLAY) {
 630                 lock = find_existing_lock(req->rq_export,
 631                                           &dlm_req->lock_handle1);
 632                 if (lock != NULL) {
 633                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 634                                   lock->l_handle.h_cookie);
 635                         GOTO(existing_lock, rc = 0);
 636                 }
 637         }
 638
 639         /* The lock's callback data might be set in the policy function */
 640         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 641                                 dlm_req->lock_desc.l_resource.lr_name,
 642                                 dlm_req->lock_desc.l_resource.lr_type,
 643                                 dlm_req->lock_desc.l_req_mode,
 644                                 blocking_callback, completion_callback,
 645                                 glimpse_callback, NULL, 0);
 646         if (!lock)
 647                 GOTO(out, rc = -ENOMEM);
 648
 649         do_gettimeofday(&lock->l_enqueued_time);
 650         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 651                sizeof(lock->l_remote_handle));
 652         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 653
 654         LASSERT(req->rq_export);
 655         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 656         if (req->rq_export->exp_failed) {
 657                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 658                 GOTO(out, err = -ENOTCONN);
 659         }
 660
 661         lock->l_export = class_export_get(req->rq_export);
 662         spin_lock(&lock->l_export->exp_ldlm_data.led_lock);
 663         list_add(&lock->l_export_chain,
 664                  &lock->l_export->exp_ldlm_data.led_held_locks);
 665         spin_unlock(&lock->l_export->exp_ldlm_data.led_lock);
 666
 667 existing_lock:
 668
 669         if (flags & LDLM_FL_HAS_INTENT) {
 670                 /* In this case, the reply buffer is allocated deep in
 671                  * local_lock_enqueue by the policy function. */
 672                 cookie = req;
 673         } else {
 674                 int buffers = 1;
 675                 lock_res_and_lock(lock);
 676                 if (lock->l_resource->lr_lvb_len) {
 677                         size[1] = lock->l_resource->lr_lvb_len;
 678                         buffers = 2;
 679                 }
 680                 unlock_res_and_lock(lock);
 681                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 682                         GOTO(out, rc = -ENOMEM);
 683
 684                 rc = lustre_pack_reply(req, buffers, size, NULL);
 685                 if (rc)
 686                         GOTO(out, rc);
 687         }
 688
 689         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 690                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 691                        sizeof(ldlm_policy_data_t));
 692         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 693                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 694                        sizeof(lock->l_req_extent));
 695
 696         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock,
 697                                 cookie, (int *)&flags);
 698         if (err)
 699                 GOTO(out, err);
 700
 701         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*dlm_rep));
 702         dlm_rep->lock_flags = flags;
 703
 704         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 705         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 706
 707         /* We never send a blocking AST until the lock is granted, but
 708          * we can tell it right now */
 709         lock_res_and_lock(lock);
 710         if (lock->l_flags & LDLM_FL_AST_SENT) {
 711                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 712                 if (lock->l_granted_mode == lock->l_req_mode)
 713                         ldlm_add_waiting_lock(lock);
 714         }
 715         unlock_res_and_lock(lock);
 716
 717         EXIT;
 718  out:
 719         req->rq_status = err;
 720         if (req->rq_reply_state == NULL) {
 721                 err = lustre_pack_reply(req, 0, NULL, NULL);
 722                 if (rc == 0)
 723                         rc = err;
 724                 req->rq_status = rc;
 725         }
 726
 727         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 728          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 729         if (lock) {
 730                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 731                            "(err=%d, rc=%d)", err, rc);
 732
 733                 if (rc == 0) {
 734                         lock_res_and_lock(lock);
 735                         size[1] = lock->l_resource->lr_lvb_len;
 736                         if (size[1] > 0) {
 737                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 738                                                            1, size[1]);
 739                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 740                                          req, lock);
 741
 742                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 743                                        size[1]);
 744                         }
 745                         unlock_res_and_lock(lock);
 746                 } else {
 747                         ldlm_lock_destroy(lock);
 748                 }
 749
 750                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 751                         ldlm_reprocess_all(lock->l_resource);
 752                 LDLM_LOCK_PUT(lock);
 753         }
 754         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 755                           lock, rc);
 756
 757         return rc;
 758 }
 759
 760 int ldlm_handle_convert(struct ptlrpc_request *req)
 761 {
 762         struct ldlm_request *dlm_req;
 763         struct ldlm_reply *dlm_rep;
 764         struct ldlm_lock *lock;
 765         int rc, size = sizeof(*dlm_rep);
 766         ENTRY;
 767
 768         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 769                                      lustre_swab_ldlm_request);
 770         if (dlm_req == NULL) {
 771                 CERROR ("Can't unpack dlm_req\n");
 772                 RETURN (-EFAULT);
 773         }
 774
 775         rc = lustre_pack_reply(req, 1, &size, NULL);
 776         if (rc) {
 777                 CERROR("out of memory\n");
 778                 RETURN(-ENOMEM);
 779         }
 780         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 781         dlm_rep->lock_flags = dlm_req->lock_flags;
 782
 783         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 784         if (!lock) {
 785                 req->rq_status = EINVAL;
 786         } else {
 787                 void *res = NULL;
 788
 789                 LDLM_DEBUG(lock, "server-side convert handler START");
 790
 791                 do_gettimeofday(&lock->l_enqueued_time);
 792                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 793                                         (int *)&dlm_rep->lock_flags);
 794                 if (res) {
 795                         if (ldlm_del_waiting_lock(lock))
 796                                 LDLM_DEBUG(lock, "converted waiting lock");
 797                         req->rq_status = 0;
 798                 } else {
 799                         req->rq_status = EDEADLOCK;
 800                 }
 801         }
 802
 803         if (lock) {
 804                 if (!req->rq_status)
 805                         ldlm_reprocess_all(lock->l_resource);
 806                 LDLM_DEBUG(lock, "server-side convert handler END");
 807                 LDLM_LOCK_PUT(lock);
 808         } else
 809                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 810
 811         RETURN(0);
 812 }
 813
 814 int ldlm_handle_cancel(struct ptlrpc_request *req)
 815 {
 816         struct ldlm_request *dlm_req;
 817         struct ldlm_lock *lock;
 818         struct ldlm_resource *res;
 819         int rc;
 820         ENTRY;
 821
 822         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 823                                       lustre_swab_ldlm_request);
 824         if (dlm_req == NULL) {
 825                 CERROR("bad request buffer for cancel\n");
 826                 RETURN(-EFAULT);
 827         }
 828
 829         rc = lustre_pack_reply(req, 0, NULL, NULL);
 830         if (rc) {
 831                 CERROR("out of memory\n");
 832                 RETURN(-ENOMEM);
 833         }
 834
 835         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 836         if (!lock) {
 837                 CERROR("received cancel for unknown lock cookie "LPX64
 838                        " from client %s id %s\n",
 839                        dlm_req->lock_handle1.cookie,
 840                        req->rq_export->exp_client_uuid.uuid,
 841                        req->rq_peerstr);
 842                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 843                                   "(cookie "LPU64")",
 844                                   dlm_req->lock_handle1.cookie);
 845                 req->rq_status = ESTALE;
 846         } else {
 847                 LDLM_DEBUG(lock, "server-side cancel handler START");
 848                 res = lock->l_resource;
 849                 if (res && res->lr_namespace->ns_lvbo &&
 850                     res->lr_namespace->ns_lvbo->lvbo_update) {
 851                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 852                                 (res, NULL, 0, 0);
 853                                 //(res, req->rq_reqmsg, 1);
 854                 }
 855
 856                 ldlm_lock_cancel(lock);
 857                 if (ldlm_del_waiting_lock(lock))
 858                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 859                 req->rq_status = rc;
 860         }
 861
 862         if (ptlrpc_reply(req) != 0)
 863                 LBUG();
 864
 865         if (lock) {
 866                 ldlm_reprocess_all(lock->l_resource);
 867                 LDLM_DEBUG(lock, "server-side cancel handler END");
 868                 LDLM_LOCK_PUT(lock);
 869         }
 870
 871         RETURN(0);
 872 }
 873
 874 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 875                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 876 {
 877         int do_ast;
 878         ENTRY;
 879
 880         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 881
 882         lock_res_and_lock(lock);
 883         lock->l_flags |= LDLM_FL_CBPENDING;
 884         do_ast = (!lock->l_readers && !lock->l_writers);
 885         unlock_res_and_lock(lock);
 886
 887         if (do_ast) {
 888                 LDLM_DEBUG(lock, "already unused, calling "
 889                            "callback (%p)", lock->l_blocking_ast);
 890                 if (lock->l_blocking_ast != NULL)
 891                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 892                                              LDLM_CB_BLOCKING);
 893         } else {
 894                 LDLM_DEBUG(lock, "Lock still has references, will be"
 895                            " cancelled later");
 896         }
 897
 898         LDLM_DEBUG(lock, "client blocking callback handler END");
 899         LDLM_LOCK_PUT(lock);
 900         EXIT;
 901 }
 902
 903 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 904                                     struct ldlm_namespace *ns,
 905                                     struct ldlm_request *dlm_req,
 906                                     struct ldlm_lock *lock)
 907 {
 908         LIST_HEAD(ast_list);
 909         ENTRY;
 910
 911         LDLM_DEBUG(lock, "client completion callback handler START");
 912
 913         lock_res_and_lock(lock);
 914
 915         /* If we receive the completion AST before the actual enqueue returned,
 916          * then we might need to switch lock modes, resources, or extents. */
 917         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 918                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 919                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 920         }
 921
 922         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 923                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 924                        sizeof(lock->l_policy_data));
 925                 LDLM_DEBUG(lock, "completion AST, new policy data");
 926         }
 927
 928         ldlm_resource_unlink_lock(lock);
 929         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 930                    &lock->l_resource->lr_name,
 931                    sizeof(lock->l_resource->lr_name)) != 0) {
 932                 unlock_res_and_lock(lock);
 933                 ldlm_lock_change_resource(ns, lock,
 934                                          dlm_req->lock_desc.l_resource.lr_name);
 935                 LDLM_DEBUG(lock, "completion AST, new resource");
 936                 CERROR("change resource!\n");
 937                 lock_res_and_lock(lock);
 938         }
 939
 940         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 941                 lock->l_flags |= LDLM_FL_CBPENDING;
 942                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 943         }
 944
 945         if (lock->l_lvb_len) {
 946                 void *lvb;
 947                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 948                                          lock->l_lvb_swabber);
 949                 if (lvb == NULL) {
 950                         LDLM_ERROR(lock, "completion AST did not contain "
 951                                    "expected LVB!");
 952                 } else {
 953                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 954                 }
 955         }
 956
 957         ldlm_grant_lock(lock, &ast_list);
 958         unlock_res_and_lock(lock);
 959
 960         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 961         LDLM_LOCK_PUT(lock);
 962
 963         ldlm_run_cp_ast_work(&ast_list);
 964
 965         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 966                           lock);
 967         EXIT;
 968 }
 969
 970 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 971                                     struct ldlm_namespace *ns,
 972                                     struct ldlm_request *dlm_req,
 973                                     struct ldlm_lock *lock)
 974 {
 975         int rc = -ENOSYS;
 976         ENTRY;
 977
 978         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 979
 980         if (lock->l_glimpse_ast != NULL)
 981                 rc = lock->l_glimpse_ast(lock, req);
 982
 983         if (req->rq_repmsg != NULL) {
 984                 ptlrpc_reply(req);
 985         } else {
 986                 req->rq_status = rc;
 987                 ptlrpc_error(req);
 988         }
 989
 990         lock_res_and_lock(lock);
 991         if (lock->l_granted_mode == LCK_PW &&
 992             !lock->l_readers && !lock->l_writers &&
 993             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 994                 unlock_res_and_lock(lock);
 995                 if (ldlm_bl_to_thread(ns, NULL, lock))
 996                         ldlm_handle_bl_callback(ns, NULL, lock);
 997
 998                 EXIT;
 999                 return;
1000         }
1001         unlock_res_and_lock(lock);
1002         LDLM_LOCK_PUT(lock);
1003         EXIT;
1004 }
1005
1006 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1007 {
1008         req->rq_status = rc;
1009         if (req->rq_reply_state == NULL) {
1010                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1011                 if (rc)
1012                         return rc;
1013         }
1014         return ptlrpc_reply(req);
1015 }
1016
1017 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1018                       struct ldlm_lock *lock)
1019 {
1020 #ifdef __KERNEL__
1021         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1022         struct ldlm_bl_work_item *blwi;
1023         ENTRY;
1024
1025         OBD_ALLOC(blwi, sizeof(*blwi));
1026         if (blwi == NULL)
1027                 RETURN(-ENOMEM);
1028
1029         blwi->blwi_ns = ns;
1030         if (ld != NULL)
1031                 blwi->blwi_ld = *ld;
1032         blwi->blwi_lock = lock;
1033
1034         spin_lock(&blp->blp_lock);
1035         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1036         wake_up(&blp->blp_waitq);
1037         spin_unlock(&blp->blp_lock);
1038
1039         RETURN(0);
1040 #else
1041         RETURN(-ENOSYS);
1042 #endif
1043
1044 }
1045
1046 static int ldlm_msg_check_version(struct lustre_msg *msg)
1047 {
1048         int rc;
1049
1050         switch (msg->opc) {
1051         case LDLM_ENQUEUE:
1052         case LDLM_CONVERT:
1053         case LDLM_CANCEL:
1054         case LDLM_BL_CALLBACK:
1055         case LDLM_CP_CALLBACK:
1056         case LDLM_GL_CALLBACK:
1057         case LDLM_FLK_DEADLOCK_CHK:
1058                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1059                 if (rc)
1060                         CERROR("bad opc %u version %08x, expecting %08x\n",
1061                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1062                 break;
1063         case OBD_LOG_CANCEL:
1064         case LLOG_ORIGIN_HANDLE_OPEN:
1065         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1066         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1067         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1068         case LLOG_ORIGIN_HANDLE_CLOSE:
1069         case LLOG_CATINFO:
1070                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1071                 if (rc)
1072                         CERROR("bad opc %u version %08x, expecting %08x\n",
1073                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1074                 break;
1075         default:
1076                 CERROR("LDLM unknown opcode %d\n", msg->opc);
1077                 rc = -ENOTSUPP;
1078                 break;
1079         }
1080
1081         return rc;
1082 }
1083
1084 static int ldlm_callback_handler(struct ptlrpc_request *req)
1085 {
1086         struct ldlm_namespace *ns;
1087         struct ldlm_request *dlm_req;
1088         struct ldlm_lock *lock;
1089         int rc;
1090         ENTRY;
1091
1092         rc = ldlm_msg_check_version(req->rq_reqmsg);
1093         if (rc) {
1094                 CERROR("LDLM_CB drop mal-formed request\n");
1095                 RETURN(rc);
1096         }
1097
1098         /* Requests arrive in sender's byte order.  The ptlrpc service
1099          * handler has already checked and, if necessary, byte-swapped the
1100          * incoming request message body, but I am responsible for the
1101          * message buffers. */
1102
1103         if (req->rq_export == NULL) {
1104                 struct ldlm_request *dlm_req;
1105
1106                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1107                        "export cookie "LPX64"; this is "
1108                        "normal if this node rebooted with a lock held\n",
1109                        req->rq_reqmsg->opc,
1110                        req->rq_peerstr,
1111                        req->rq_reqmsg->handle.cookie);
1112                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1113                                              lustre_swab_ldlm_request);
1114                 if (dlm_req != NULL)
1115                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1116                                dlm_req->lock_handle1.cookie);
1117
1118                 ldlm_callback_reply(req, -ENOTCONN);
1119                 RETURN(0);
1120         }
1121
1122         LASSERT(req->rq_export != NULL);
1123         LASSERT(req->rq_export->exp_obd != NULL);
1124
1125         switch(req->rq_reqmsg->opc) {
1126         case LDLM_BL_CALLBACK:
1127                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1128                 break;
1129         case LDLM_CP_CALLBACK:
1130                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1131                 break;
1132         case LDLM_GL_CALLBACK:
1133                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1134                 break;
1135         case OBD_LOG_CANCEL:
1136                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1137                 rc = llog_origin_handle_cancel(req);
1138                 ldlm_callback_reply(req, rc);
1139                 RETURN(0);
1140         case LLOG_ORIGIN_HANDLE_OPEN:
1141                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1142                 rc = llog_origin_handle_open(req);
1143                 ldlm_callback_reply(req, rc);
1144                 RETURN(0);
1145         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1146                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1147                 rc = llog_origin_handle_next_block(req);
1148                 ldlm_callback_reply(req, rc);
1149                 RETURN(0);
1150         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1151                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1152                 rc = llog_origin_handle_prev_block(req);
1153                 ldlm_callback_reply(req, rc);
1154                 RETURN(0);
1155         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1156                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1157                 rc = llog_origin_handle_read_header(req);
1158                 ldlm_callback_reply(req, rc);
1159                 RETURN(0);
1160         case LLOG_ORIGIN_HANDLE_CLOSE:
1161                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1162                 rc = llog_origin_handle_close(req);
1163                 ldlm_callback_reply(req, rc);
1164                 RETURN(0);
1165         default:
1166                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1167                 ldlm_callback_reply(req, -EPROTO);
1168                 RETURN(0);
1169         }
1170
1171         ns = req->rq_export->exp_obd->obd_namespace;
1172         LASSERT(ns != NULL);
1173
1174         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1175                                       lustre_swab_ldlm_request);
1176         if (dlm_req == NULL) {
1177                 CERROR ("can't unpack dlm_req\n");
1178                 ldlm_callback_reply (req, -EPROTO);
1179                 RETURN (0);
1180         }
1181
1182         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1183         if (!lock) {
1184                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1185                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1186                 ldlm_callback_reply(req, -EINVAL);
1187                 RETURN(0);
1188         }
1189
1190         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1191         lock_res_and_lock(lock);
1192         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1193         unlock_res_and_lock(lock);
1194
1195         /* We want the ost thread to get this reply so that it can respond
1196          * to ost requests (write cache writeback) that might be triggered
1197          * in the callback.
1198          *
1199          * But we'd also like to be able to indicate in the reply that we're
1200          * cancelling right now, because it's unused, or have an intent result
1201          * in the reply, so we might have to push the responsibility for sending
1202          * the reply down into the AST handlers, alas. */
1203
1204         switch (req->rq_reqmsg->opc) {
1205         case LDLM_BL_CALLBACK:
1206                 CDEBUG(D_INODE, "blocking ast\n");
1207                 ldlm_callback_reply(req, 0);
1208                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1209                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1210
1211                 break;
1212         case LDLM_CP_CALLBACK:
1213                 CDEBUG(D_INODE, "completion ast\n");
1214                 ldlm_callback_reply(req, 0);
1215                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1216                 break;
1217         case LDLM_GL_CALLBACK:
1218                 CDEBUG(D_INODE, "glimpse ast\n");
1219                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1220                 break;
1221         default:
1222                 LBUG();                         /* checked above */
1223         }
1224
1225         RETURN(0);
1226 }
1227
1228 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1229 {
1230         int rc = 0;
1231         ENTRY;
1232
1233         rc = ldlm_msg_check_version(req->rq_reqmsg);
1234         if (rc) {
1235                 CERROR("LDLM_CL drop mal-formed request\n");
1236                 RETURN(rc);
1237         }
1238
1239         /* Requests arrive in sender's byte order.  The ptlrpc service
1240          * handler has already checked and, if necessary, byte-swapped the
1241          * incoming request message body, but I am responsible for the
1242          * message buffers. */
1243
1244         if (req->rq_export == NULL) {
1245                 struct ldlm_request *dlm_req;
1246                 CERROR("operation %d with bad export from %s\n",
1247                        req->rq_reqmsg->opc,
1248                        req->rq_peerstr);
1249                 CERROR("--> export cookie: "LPX64"\n",
1250                        req->rq_reqmsg->handle.cookie);
1251                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1252                                              lustre_swab_ldlm_request);
1253                 if (dlm_req != NULL)
1254                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1255                 ldlm_callback_reply(req, -ENOTCONN);
1256                 RETURN(0);
1257         }
1258
1259         switch (req->rq_reqmsg->opc) {
1260
1261         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1262         case LDLM_CANCEL:
1263                 CDEBUG(D_INODE, "cancel\n");
1264                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1265                 rc = ldlm_handle_cancel(req);
1266                 break;
1267         default:
1268                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1269                 ldlm_callback_reply(req, -EINVAL);
1270         }
1271
1272         RETURN(rc);
1273 }
1274
1275 #ifdef __KERNEL__
1276 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1277 {
1278         struct ldlm_bl_work_item *blwi = NULL;
1279
1280         spin_lock(&blp->blp_lock);
1281         if (!list_empty(&blp->blp_list)) {
1282                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1283                                   blwi_entry);
1284                 list_del(&blwi->blwi_entry);
1285         }
1286         spin_unlock(&blp->blp_lock);
1287
1288         return blwi;
1289 }
1290
1291 struct ldlm_bl_thread_data {
1292         int                     bltd_num;
1293         struct ldlm_bl_pool     *bltd_blp;
1294 };
1295
1296 static int ldlm_bl_thread_main(void *arg)
1297 {
1298         struct ldlm_bl_thread_data *bltd = arg;
1299         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1300         unsigned long flags;
1301         ENTRY;
1302
1303         /* XXX boiler-plate */
1304         {
1305                 char name[sizeof(current->comm)];
1306                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1307                          bltd->bltd_num);
1308                 kportal_daemonize(name);
1309         }
1310         SIGNAL_MASK_LOCK(current, flags);
1311         sigfillset(&current->blocked);
1312         RECALC_SIGPENDING;
1313         SIGNAL_MASK_UNLOCK(current, flags);
1314
1315         atomic_inc(&blp->blp_num_threads);
1316         complete(&blp->blp_comp);
1317
1318         while(1) {
1319                 struct l_wait_info lwi = { 0 };
1320                 struct ldlm_bl_work_item *blwi = NULL;
1321
1322                 l_wait_event_exclusive(blp->blp_waitq,
1323                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1324                                        &lwi);
1325
1326                 if (blwi->blwi_ns == NULL)
1327                         break;
1328
1329                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1330                                         blwi->blwi_lock);
1331                 OBD_FREE(blwi, sizeof(*blwi));
1332         }
1333
1334         atomic_dec(&blp->blp_num_threads);
1335         complete(&blp->blp_comp);
1336         RETURN(0);
1337 }
1338
1339 #endif
1340
1341 static int ldlm_setup(void);
1342 static int ldlm_cleanup(int force);
1343
1344 int ldlm_get_ref(void)
1345 {
1346         int rc = 0;
1347         down(&ldlm_ref_sem);
1348         if (++ldlm_refcount == 1) {
1349                 rc = ldlm_setup();
1350                 if (rc)
1351                         ldlm_refcount--;
1352         }
1353         up(&ldlm_ref_sem);
1354
1355         RETURN(rc);
1356 }
1357
1358 void ldlm_put_ref(int force)
1359 {
1360         down(&ldlm_ref_sem);
1361         if (ldlm_refcount == 1) {
1362                 int rc = ldlm_cleanup(force);
1363                 if (rc)
1364                         CERROR("ldlm_cleanup failed: %d\n", rc);
1365                 else
1366                         ldlm_refcount--;
1367         } else {
1368                 ldlm_refcount--;
1369         }
1370         up(&ldlm_ref_sem);
1371
1372         EXIT;
1373 }
1374
1375 static int ldlm_setup(void)
1376 {
1377         struct ldlm_bl_pool *blp;
1378         int rc = 0;
1379 #ifdef __KERNEL__
1380         int i;
1381 #endif
1382         ENTRY;
1383
1384         if (ldlm_state != NULL)
1385                 RETURN(-EALREADY);
1386
1387         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1388         if (ldlm_state == NULL)
1389                 RETURN(-ENOMEM);
1390
1391 #ifdef __KERNEL__
1392         rc = ldlm_proc_setup();
1393         if (rc != 0)
1394                 GOTO(out_free, rc);
1395 #endif
1396
1397         ldlm_state->ldlm_cb_service =
1398                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1399                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1400                                 1500, ldlm_callback_handler, "ldlm_cbd",
1401                                 ldlm_svc_proc_dir);
1402
1403         if (!ldlm_state->ldlm_cb_service) {
1404                 CERROR("failed to start service\n");
1405                 GOTO(out_proc, rc = -ENOMEM);
1406         }
1407
1408         ldlm_state->ldlm_cancel_service =
1409                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1410                                 LDLM_CANCEL_REQUEST_PORTAL,
1411                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1412                                 ldlm_cancel_handler, "ldlm_canceld",
1413                                 ldlm_svc_proc_dir);
1414
1415         if (!ldlm_state->ldlm_cancel_service) {
1416                 CERROR("failed to start service\n");
1417                 GOTO(out_proc, rc = -ENOMEM);
1418         }
1419
1420         OBD_ALLOC(blp, sizeof(*blp));
1421         if (blp == NULL)
1422                 GOTO(out_proc, rc = -ENOMEM);
1423         ldlm_state->ldlm_bl_pool = blp;
1424
1425         atomic_set(&blp->blp_num_threads, 0);
1426         init_waitqueue_head(&blp->blp_waitq);
1427         spin_lock_init(&blp->blp_lock);
1428
1429         INIT_LIST_HEAD(&blp->blp_list);
1430
1431 #ifdef __KERNEL__
1432         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1433                 struct ldlm_bl_thread_data bltd = {
1434                         .bltd_num = i,
1435                         .bltd_blp = blp,
1436                 };
1437                 init_completion(&blp->blp_comp);
1438                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1439                 if (rc < 0) {
1440                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1441                         GOTO(out_thread, rc);
1442                 }
1443                 wait_for_completion(&blp->blp_comp);
1444         }
1445
1446         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1447                                     LDLM_NUM_THREADS, "ldlm_cn");
1448         if (rc)
1449                 GOTO(out_thread, rc);
1450
1451         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1452                                     LDLM_NUM_THREADS, "ldlm_cb");
1453         if (rc)
1454                 GOTO(out_thread, rc);
1455
1456         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1457         spin_lock_init(&expired_lock_thread.elt_lock);
1458         expired_lock_thread.elt_state = ELT_STOPPED;
1459         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1460
1461         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1462         if (rc < 0) {
1463                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1464                 GOTO(out_thread, rc);
1465         }
1466
1467         wait_event(expired_lock_thread.elt_waitq,
1468                    expired_lock_thread.elt_state == ELT_READY);
1469
1470         INIT_LIST_HEAD(&waiting_locks_list);
1471         spin_lock_init(&waiting_locks_spinlock);
1472         waiting_locks_timer.function = waiting_locks_callback;
1473         waiting_locks_timer.data = 0;
1474         init_timer(&waiting_locks_timer);
1475 #endif
1476
1477         RETURN(0);
1478
1479 #ifdef __KERNEL__
1480  out_thread:
1481         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1482         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1483 #endif
1484
1485  out_proc:
1486 #ifdef __KERNEL__
1487         ldlm_proc_cleanup();
1488  out_free:
1489 #endif
1490         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1491         ldlm_state = NULL;
1492         return rc;
1493 }
1494
1495 static int ldlm_cleanup(int force)
1496 {
1497 #ifdef __KERNEL__
1498         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1499 #endif
1500         ENTRY;
1501
1502         if (!list_empty(&ldlm_namespace_list)) {
1503                 CERROR("ldlm still has namespaces; clean these up first.\n");
1504                 ldlm_dump_all_namespaces(D_DLMTRACE);
1505                 RETURN(-EBUSY);
1506         }
1507
1508 #ifdef __KERNEL__
1509         while (atomic_read(&blp->blp_num_threads) > 0) {
1510                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1511
1512                 init_completion(&blp->blp_comp);
1513
1514                 spin_lock(&blp->blp_lock);
1515                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1516                 wake_up(&blp->blp_waitq);
1517                 spin_unlock(&blp->blp_lock);
1518
1519                 wait_for_completion(&blp->blp_comp);
1520         }
1521         OBD_FREE(blp, sizeof(*blp));
1522
1523         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1524         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1525         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1526         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1527         ldlm_proc_cleanup();
1528
1529         expired_lock_thread.elt_state = ELT_TERMINATE;
1530         wake_up(&expired_lock_thread.elt_waitq);
1531         wait_event(expired_lock_thread.elt_waitq,
1532                    expired_lock_thread.elt_state == ELT_STOPPED);
1533 #else
1534         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1535         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1536 #endif
1537
1538         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1539         ldlm_state = NULL;
1540
1541         RETURN(0);
1542 }
1543
1544 int __init ldlm_init(void)
1545 {
1546         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1547                                                sizeof(struct ldlm_resource), 0,
1548                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1549         if (ldlm_resource_slab == NULL)
1550                 return -ENOMEM;
1551
1552         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1553                                            sizeof(struct ldlm_lock), 0,
1554                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1555         if (ldlm_lock_slab == NULL) {
1556                 kmem_cache_destroy(ldlm_resource_slab);
1557                 return -ENOMEM;
1558         }
1559
1560         return 0;
1561 }
1562
1563 void __exit ldlm_exit(void)
1564 {
1565         if ( ldlm_refcount )
1566                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1567         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1568                  "couldn't free ldlm resource slab\n");
1569         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1570                  "couldn't free ldlm lock slab\n");
1571
1572 }
1573
1574 /* ldlm_flock.c */
1575 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1576 EXPORT_SYMBOL(ldlm_handle_flock_deadlock_check);
1577
1578 /* ldlm_extent.c */
1579 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1580
1581 /* ldlm_lock.c */
1582 EXPORT_SYMBOL(ldlm_get_processing_policy);
1583 EXPORT_SYMBOL(ldlm_lock2desc);
1584 EXPORT_SYMBOL(ldlm_register_intent);
1585 EXPORT_SYMBOL(ldlm_lockname);
1586 EXPORT_SYMBOL(ldlm_typename);
1587 EXPORT_SYMBOL(ldlm_lock2handle);
1588 EXPORT_SYMBOL(__ldlm_handle2lock);
1589 EXPORT_SYMBOL(ldlm_lock_get);
1590 EXPORT_SYMBOL(ldlm_lock_put);
1591 EXPORT_SYMBOL(ldlm_lock_match);
1592 EXPORT_SYMBOL(ldlm_lock_cancel);
1593 EXPORT_SYMBOL(ldlm_lock_addref);
1594 EXPORT_SYMBOL(ldlm_lock_decref);
1595 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1596 EXPORT_SYMBOL(ldlm_lock_change_resource);
1597 EXPORT_SYMBOL(ldlm_lock_set_data);
1598 EXPORT_SYMBOL(ldlm_it2str);
1599 EXPORT_SYMBOL(ldlm_lock_dump);
1600 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1601 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1602 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1603 EXPORT_SYMBOL(ldlm_lock_allow_match);
1604
1605 /* ldlm_request.c */
1606 EXPORT_SYMBOL(ldlm_completion_ast);
1607 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1608 EXPORT_SYMBOL(ldlm_cli_convert);
1609 EXPORT_SYMBOL(ldlm_cli_enqueue);
1610 EXPORT_SYMBOL(ldlm_cli_cancel);
1611 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1612 EXPORT_SYMBOL(ldlm_replay_locks);
1613 EXPORT_SYMBOL(ldlm_resource_foreach);
1614 EXPORT_SYMBOL(ldlm_namespace_foreach);
1615 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1616 EXPORT_SYMBOL(ldlm_change_cbdata);
1617
1618 /* ldlm_lockd.c */
1619 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1620 EXPORT_SYMBOL(ldlm_server_completion_ast);
1621 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1622 EXPORT_SYMBOL(ldlm_handle_enqueue);
1623 EXPORT_SYMBOL(ldlm_handle_cancel);
1624 EXPORT_SYMBOL(ldlm_handle_convert);
1625 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1626 EXPORT_SYMBOL(ldlm_get_ref);
1627 EXPORT_SYMBOL(ldlm_put_ref);
1628
1629 #if 0
1630 /* ldlm_test.c */
1631 EXPORT_SYMBOL(ldlm_test);
1632 EXPORT_SYMBOL(ldlm_regression_start);
1633 EXPORT_SYMBOL(ldlm_regression_stop);
1634 #endif
1635
1636 /* ldlm_resource.c */
1637 EXPORT_SYMBOL(ldlm_namespace_new);
1638 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1639 EXPORT_SYMBOL(ldlm_namespace_free);
1640 EXPORT_SYMBOL(ldlm_namespace_dump);
1641 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1642 EXPORT_SYMBOL(ldlm_resource_get);
1643 EXPORT_SYMBOL(ldlm_resource_putref);
1644
1645 /* ldlm_lib.c */
1646 EXPORT_SYMBOL(client_import_add_conn);
1647 EXPORT_SYMBOL(client_import_del_conn);
1648 EXPORT_SYMBOL(client_obd_setup);
1649 EXPORT_SYMBOL(client_obd_cleanup);
1650 EXPORT_SYMBOL(client_connect_import);
1651 EXPORT_SYMBOL(client_disconnect_export);
1652 EXPORT_SYMBOL(target_start_recovery_thread);
1653 EXPORT_SYMBOL(target_stop_recovery_thread);
1654 EXPORT_SYMBOL(target_handle_connect);
1655 EXPORT_SYMBOL(target_cleanup_recovery);
1656 EXPORT_SYMBOL(target_destroy_export);
1657 EXPORT_SYMBOL(target_cancel_recovery_timer);
1658 EXPORT_SYMBOL(target_send_reply);
1659 EXPORT_SYMBOL(target_queue_recovery_request);
1660 EXPORT_SYMBOL(target_handle_ping);
1661 EXPORT_SYMBOL(target_handle_disconnect);
1662
1663 /* l_lock.c */
1664 EXPORT_SYMBOL(lock_res_and_lock);
1665 EXPORT_SYMBOL(unlock_res_and_lock);
1666
1667