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 lustre_lock ldlm_handle_lock;
  46 extern struct list_head ldlm_namespace_list;
  47
  48 static DECLARE_MUTEX(ldlm_ref_sem);
  49 static int ldlm_refcount;
  50 /* LDLM state */
  51
  52 static struct ldlm_state *ldlm_state;
  53
  54 static inline unsigned long
  55 round_timeout(unsigned long timeout)
  56 {
  57         return ((timeout / HZ) + 1) * HZ;
  58 }
  59
  60 #ifdef __KERNEL__
  61 /* XXX should this be per-ldlm? */
  62 static struct list_head waiting_locks_list;
  63 static spinlock_t waiting_locks_spinlock;
  64 static struct timer_list waiting_locks_timer;
  65
  66 static struct expired_lock_thread {
  67         wait_queue_head_t         elt_waitq;
  68         int                       elt_state;
  69         struct list_head          elt_expired_locks;
  70         spinlock_t                elt_lock;
  71 } expired_lock_thread;
  72 #endif
  73
  74 #if !defined(ENOTSUPP)
  75 #  define ENOTSUPP 524
  76 #endif
  77
  78 #define ELT_STOPPED   0
  79 #define ELT_READY     1
  80 #define ELT_TERMINATE 2
  81
  82 struct ldlm_bl_pool {
  83         spinlock_t              blp_lock;
  84         struct list_head        blp_list;
  85         wait_queue_head_t       blp_waitq;
  86         atomic_t                blp_num_threads;
  87         struct completion       blp_comp;
  88 };
  89
  90 struct ldlm_bl_work_item {
  91         struct list_head        blwi_entry;
  92         struct ldlm_namespace   *blwi_ns;
  93         struct ldlm_lock_desc   blwi_ld;
  94         struct ldlm_lock        *blwi_lock;
  95 };
  96
  97 #ifdef __KERNEL__
  98
  99 static int ldlm_add_waiting_lock(struct ldlm_lock *lock);
 100
 101 static inline int have_expired_locks(void)
 102 {
 103         int need_to_run;
 104
 105         spin_lock_bh(&expired_lock_thread.elt_lock);
 106         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 107         spin_unlock_bh(&expired_lock_thread.elt_lock);
 108
 109         RETURN(need_to_run);
 110 }
 111
 112 static int expired_lock_main(void *arg)
 113 {
 114         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 115         struct l_wait_info lwi = { 0 };
 116         unsigned long flags;
 117
 118         ENTRY;
 119         lock_kernel();
 120         kportal_daemonize("ldlm_elt");
 121
 122         SIGNAL_MASK_LOCK(current, flags);
 123         sigfillset(&current->blocked);
 124         RECALC_SIGPENDING;
 125         SIGNAL_MASK_UNLOCK(current, flags);
 126
 127         unlock_kernel();
 128
 129         expired_lock_thread.elt_state = ELT_READY;
 130         wake_up(&expired_lock_thread.elt_waitq);
 131
 132         while (1) {
 133                 l_wait_event(expired_lock_thread.elt_waitq,
 134                              have_expired_locks() ||
 135                              expired_lock_thread.elt_state == ELT_TERMINATE,
 136                              &lwi);
 137
 138                 spin_lock_bh(&expired_lock_thread.elt_lock);
 139                 while (!list_empty(expired)) {
 140                         struct obd_export *export;
 141                         struct ldlm_lock *lock;
 142
 143                         lock = list_entry(expired->next, struct ldlm_lock,
 144                                           l_pending_chain);
 145                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 146                             (void *)lock >= LP_POISON) {
 147                                 CERROR("free lock on elt list %p\n", lock);
 148                                 LBUG();
 149                         }
 150                         list_del_init(&lock->l_pending_chain);
 151                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 152                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 153                                 CERROR("lock with free export on elt list %p\n",
 154                                        export);
 155                                 lock->l_export = NULL;
 156                                 LDLM_ERROR(lock, "free export\n");
 157                                 continue;
 158                         }
 159                         export = class_export_get(lock->l_export);
 160                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 161
 162                         ptlrpc_fail_export(export);
 163                         class_export_put(export);
 164                         spin_lock_bh(&expired_lock_thread.elt_lock);
 165                 }
 166                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 167
 168                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 169                         break;
 170         }
 171
 172         expired_lock_thread.elt_state = ELT_STOPPED;
 173         wake_up(&expired_lock_thread.elt_waitq);
 174         RETURN(0);
 175 }
 176
 177 static void waiting_locks_callback(unsigned long unused)
 178 {
 179         struct ldlm_lock *lock;
 180         char str[PTL_NALFMT_SIZE];
 181
 182         if (obd_dump_on_timeout)
 183                 portals_debug_dumplog();
 184
 185 repeat:
 186         spin_lock_bh(&waiting_locks_spinlock);
 187         while (!list_empty(&waiting_locks_list)) {
 188                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 189                                   l_pending_chain);
 190
 191                 if ((lock->l_callback_timeout > jiffies) ||
 192                     (lock->l_req_mode == LCK_GROUP))
 193                         break;
 194
 195                 if (ptlrpc_check_suspend()) {
 196                         /* there is a case when we talk to one mds, holding
 197                          * lock from another mds. this way we easily can get
 198                          * here, if second mds is being recovered. so, we
 199                          * suspend timeouts. bug 6019 */
 200
 201                         LDLM_ERROR(lock, "recharge timeout: %s@%s nid %s ",
 202                                    lock->l_export->exp_client_uuid.uuid,
 203                                    lock->l_export->exp_connection->c_remote_uuid.uuid,
 204                                    ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 205
 206                         list_del_init(&lock->l_pending_chain);
 207                         spin_unlock_bh(&waiting_locks_spinlock);
 208                         ldlm_add_waiting_lock(lock);
 209
 210                         goto repeat;
 211                 }
 212
 213                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 214                            "%s@%s nid %s ",
 215                            lock->l_export->exp_client_uuid.uuid,
 216                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 217                            ptlrpc_peernid2str(&lock->l_export->exp_connection->c_peer, str));
 218
 219                 spin_lock_bh(&expired_lock_thread.elt_lock);
 220                 list_del(&lock->l_pending_chain);
 221                 list_add(&lock->l_pending_chain,
 222                          &expired_lock_thread.elt_expired_locks);
 223                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 224                 wake_up(&expired_lock_thread.elt_waitq);
 225         }
 226
 227         /*
 228          * Make sure the timer will fire again if we have any locks
 229          * left.
 230          */
 231         if (!list_empty(&waiting_locks_list)) {
 232                 unsigned long timeout_rounded;
 233                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 234                                   l_pending_chain);
 235                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 236                 mod_timer(&waiting_locks_timer, timeout_rounded);
 237         }
 238         spin_unlock_bh(&waiting_locks_spinlock);
 239 }
 240
 241 /*
 242  * Indicate that we're waiting for a client to call us back cancelling a given
 243  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 244  * timer to fire appropriately.  (We round up to the next second, to avoid
 245  * floods of timer firings during periods of high lock contention and traffic).
 246  */
 247 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 248 {
 249         unsigned long timeout_rounded;
 250
 251         spin_lock_bh(&waiting_locks_spinlock);
 252         if (!list_empty(&lock->l_pending_chain)) {
 253                 LDLM_DEBUG(lock, "not re-adding to wait list");
 254                 spin_unlock_bh(&waiting_locks_spinlock);
 255                 return 0;
 256         }
 257         LDLM_DEBUG(lock, "adding to wait list");
 258
 259         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 260
 261         timeout_rounded = round_timeout(lock->l_callback_timeout);
 262
 263         if (timeout_rounded < waiting_locks_timer.expires ||
 264             !timer_pending(&waiting_locks_timer)) {
 265                 mod_timer(&waiting_locks_timer, timeout_rounded);
 266         }
 267         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 268         spin_unlock_bh(&waiting_locks_spinlock);
 269         return 1;
 270 }
 271
 272 /*
 273  * Remove a lock from the pending list, likely because it had its cancellation
 274  * callback arrive without incident.  This adjusts the lock-timeout timer if
 275  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 276  */
 277 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 278 {
 279         struct list_head *list_next;
 280
 281         if (lock->l_export == NULL) {
 282                 /* We don't have a "waiting locks list" on clients. */
 283                 LDLM_DEBUG(lock, "client lock: no-op");
 284                 return 0;
 285         }
 286
 287         spin_lock_bh(&waiting_locks_spinlock);
 288
 289         if (list_empty(&lock->l_pending_chain)) {
 290                 spin_unlock_bh(&waiting_locks_spinlock);
 291                 LDLM_DEBUG(lock, "wasn't waiting");
 292                 return 0;
 293         }
 294
 295         list_next = lock->l_pending_chain.next;
 296         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 297                 /* Removing the head of the list, adjust timer. */
 298                 if (list_next == &waiting_locks_list) {
 299                         /* No more, just cancel. */
 300                         del_timer(&waiting_locks_timer);
 301                 } else {
 302                         struct ldlm_lock *next;
 303                         next = list_entry(list_next, struct ldlm_lock,
 304                                           l_pending_chain);
 305                         mod_timer(&waiting_locks_timer,
 306                                   round_timeout(next->l_callback_timeout));
 307                 }
 308         }
 309
 310         spin_lock_bh(&expired_lock_thread.elt_lock);
 311         list_del_init(&lock->l_pending_chain);
 312         spin_unlock_bh(&expired_lock_thread.elt_lock);
 313
 314         spin_unlock_bh(&waiting_locks_spinlock);
 315         LDLM_DEBUG(lock, "removed");
 316         return 1;
 317 }
 318
 319 #else /* !__KERNEL__ */
 320
 321 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 322 {
 323         RETURN(1);
 324 }
 325
 326 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 327 {
 328         RETURN(0);
 329 }
 330
 331 #endif /* __KERNEL__ */
 332
 333 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc,const char *ast_type)
 334 {
 335         struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 336         char str[PTL_NALFMT_SIZE];
 337
 338         LDLM_ERROR(lock, "%s AST failed (%d): evicting client %s@%s NID "LPX64
 339                    " (%s)", ast_type, rc, lock->l_export->exp_client_uuid.uuid,
 340                    conn->c_remote_uuid.uuid, conn->c_peer.peer_id.nid,
 341                    ptlrpc_peernid2str(&conn->c_peer, str));
 342
 343         if (obd_dump_on_timeout)
 344                 portals_debug_dumplog();
 345         ptlrpc_fail_export(lock->l_export);
 346 }
 347
 348 static int ldlm_handle_ast_error(struct ldlm_lock *lock,
 349                                  struct ptlrpc_request *req, int rc,
 350                                  const char *ast_type)
 351 {
 352         struct ptlrpc_peer *peer = &req->rq_import->imp_connection->c_peer;
 353         char str[PTL_NALFMT_SIZE];
 354
 355         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 356                 LASSERT(lock->l_export);
 357                 if (lock->l_export->exp_libclient) {
 358                         LDLM_DEBUG(lock, "%s AST to liblustre client (nid %s)"
 359                                    " timeout, just cancelling lock", ast_type,
 360                                    ptlrpc_peernid2str(peer, str));
 361                         ldlm_lock_cancel(lock);
 362                         rc = -ERESTART;
 363                 } else {
 364                         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 365                         ldlm_del_waiting_lock(lock);
 366                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 367                         ldlm_failed_ast(lock, rc, ast_type);
 368                 }
 369         } else if (rc) {
 370                 if (rc == -EINVAL)
 371                         LDLM_DEBUG(lock, "client (nid %s) returned %d"
 372                                    " from %s AST - normal race",
 373                                    ptlrpc_peernid2str(peer, str),
 374                                    req->rq_repmsg->status, ast_type);
 375                 else
 376                         LDLM_ERROR(lock, "client (nid %s) returned %d "
 377                                    "from %s AST", ptlrpc_peernid2str(peer, str),
 378                                    (req->rq_repmsg != NULL) ?
 379                                    req->rq_repmsg->status : 0, ast_type);
 380                 ldlm_lock_cancel(lock);
 381                 /* Server-side AST functions are called from ldlm_reprocess_all,
 382                  * which needs to be told to please restart its reprocessing. */
 383                 rc = -ERESTART;
 384         }
 385
 386         return rc;
 387 }
 388
 389 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 390                              struct ldlm_lock_desc *desc,
 391                              void *data, int flag)
 392 {
 393         struct ldlm_request *body;
 394         struct ptlrpc_request *req;
 395         int rc = 0, size = sizeof(*body);
 396         ENTRY;
 397
 398         if (flag == LDLM_CB_CANCELING) {
 399                 /* Don't need to do anything here. */
 400                 RETURN(0);
 401         }
 402
 403         LASSERTF(lock->l_export->exp_obd->obd_recovering == 0,
 404                  "BUG 6063: lock collide during recovery");
 405
 406         LASSERT(lock);
 407
 408         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 409         if (lock->l_granted_mode != lock->l_req_mode) {
 410                 /* this blocking AST will be communicated as part of the
 411                  * completion AST instead */
 412                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 413                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 414                 RETURN(0);
 415         }
 416
 417         if (lock->l_destroyed) {
 418                 /* What's the point? */
 419                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 420                 RETURN(0);
 421         }
 422
 423 #if 0
 424         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 425                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 426                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 427                 RETURN(-ETIMEDOUT);
 428         }
 429 #endif
 430
 431         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 432                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 433                               1, &size, NULL);
 434         if (req == NULL) {
 435                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 436                 RETURN(-ENOMEM);
 437         }
 438
 439         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 440         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 441                sizeof(body->lock_handle1));
 442         memcpy(&body->lock_desc, desc, sizeof(*desc));
 443         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 444
 445         LDLM_DEBUG(lock, "server preparing blocking AST");
 446         req->rq_replen = lustre_msg_size(0, NULL);
 447
 448         if (lock->l_granted_mode == lock->l_req_mode)
 449                 ldlm_add_waiting_lock(lock);
 450         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 451
 452         req->rq_send_state = LUSTRE_IMP_FULL;
 453         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 454         rc = ptlrpc_queue_wait(req);
 455         if (rc != 0)
 456                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 457
 458         ptlrpc_req_finished(req);
 459
 460         RETURN(rc);
 461 }
 462
 463 /* XXX copied from ptlrpc/service.c */
 464 static long timeval_sub(struct timeval *large, struct timeval *small)
 465 {
 466         return (large->tv_sec - small->tv_sec) * 1000000 +
 467                 (large->tv_usec - small->tv_usec);
 468 }
 469
 470 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 471 {
 472         struct ldlm_request *body;
 473         struct ptlrpc_request *req;
 474         struct timeval granted_time;
 475         long total_enqueue_wait;
 476         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 477         ENTRY;
 478
 479         LASSERT(lock != NULL);
 480
 481         do_gettimeofday(&granted_time);
 482         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 483
 484         if (total_enqueue_wait / 1000000 > obd_timeout)
 485                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 486
 487         down(&lock->l_resource->lr_lvb_sem);
 488         if (lock->l_resource->lr_lvb_len) {
 489                 buffers = 2;
 490                 size[1] = lock->l_resource->lr_lvb_len;
 491         }
 492         up(&lock->l_resource->lr_lvb_sem);
 493
 494         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 495                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 496                               buffers, size, NULL);
 497         if (req == NULL)
 498                 RETURN(-ENOMEM);
 499
 500         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 501         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 502                sizeof(body->lock_handle1));
 503         body->lock_flags = flags;
 504         ldlm_lock2desc(lock, &body->lock_desc);
 505
 506         if (buffers == 2) {
 507                 void *lvb;
 508
 509                 down(&lock->l_resource->lr_lvb_sem);
 510                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 511                                      lock->l_resource->lr_lvb_len);
 512
 513                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 514                        lock->l_resource->lr_lvb_len);
 515                 up(&lock->l_resource->lr_lvb_sem);
 516         }
 517
 518         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 519                    total_enqueue_wait);
 520         req->rq_replen = lustre_msg_size(0, NULL);
 521
 522         req->rq_send_state = LUSTRE_IMP_FULL;
 523         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 524
 525         /* We only send real blocking ASTs after the lock is granted */
 526         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 527         if (lock->l_flags & LDLM_FL_AST_SENT) {
 528                 body->lock_flags |= LDLM_FL_AST_SENT;
 529                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 530         }
 531         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 532
 533         rc = ptlrpc_queue_wait(req);
 534         if (rc != 0)
 535                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 536
 537         ptlrpc_req_finished(req);
 538
 539         RETURN(rc);
 540 }
 541
 542 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 543 {
 544         struct ldlm_resource *res = lock->l_resource;
 545         struct ldlm_request *body;
 546         struct ptlrpc_request *req;
 547         int rc = 0, size = sizeof(*body);
 548         ENTRY;
 549
 550         LASSERT(lock != NULL);
 551
 552         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 553                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 554                               1, &size, NULL);
 555         if (req == NULL)
 556                 RETURN(-ENOMEM);
 557
 558         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 559         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 560                sizeof(body->lock_handle1));
 561         ldlm_lock2desc(lock, &body->lock_desc);
 562
 563         down(&lock->l_resource->lr_lvb_sem);
 564         size = lock->l_resource->lr_lvb_len;
 565         up(&lock->l_resource->lr_lvb_sem);
 566         req->rq_replen = lustre_msg_size(1, &size);
 567
 568         req->rq_send_state = LUSTRE_IMP_FULL;
 569         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 570
 571         rc = ptlrpc_queue_wait(req);
 572         if (rc == -ELDLM_NO_LOCK_DATA)
 573                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 574         else if (rc != 0)
 575                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 576         else
 577                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 578                         (res, req->rq_repmsg, 0, 1);
 579         ptlrpc_req_finished(req);
 580         RETURN(rc);
 581 }
 582
 583 static struct ldlm_lock *
 584 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 585 {
 586         struct obd_device *obd = exp->exp_obd;
 587         struct list_head *iter;
 588
 589         l_lock(&obd->obd_namespace->ns_lock);
 590         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 591                 struct ldlm_lock *lock;
 592                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 593                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 594                         LDLM_LOCK_GET(lock);
 595                         l_unlock(&obd->obd_namespace->ns_lock);
 596                         return lock;
 597                 }
 598         }
 599         l_unlock(&obd->obd_namespace->ns_lock);
 600         return NULL;
 601 }
 602
 603
 604 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 605                         ldlm_completion_callback completion_callback,
 606                         ldlm_blocking_callback blocking_callback,
 607                         ldlm_glimpse_callback glimpse_callback)
 608 {
 609         struct obd_device *obddev = req->rq_export->exp_obd;
 610         struct ldlm_reply *dlm_rep;
 611         struct ldlm_request *dlm_req;
 612         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 613         __u32 flags;
 614         ldlm_error_t err = ELDLM_OK;
 615         struct ldlm_lock *lock = NULL;
 616         void *cookie = NULL;
 617         ENTRY;
 618
 619         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 620
 621         dlm_req = lustre_swab_reqbuf (req, MDS_REQ_INTENT_LOCKREQ_OFF,
 622                                       sizeof (*dlm_req),
 623                                       lustre_swab_ldlm_request);
 624         if (dlm_req == NULL) {
 625                 CERROR ("Can't unpack dlm_req\n");
 626                 GOTO(out, rc = -EFAULT);
 627         }
 628
 629         flags = dlm_req->lock_flags;
 630
 631         LASSERT(req->rq_export);
 632
 633         if (flags & LDLM_FL_REPLAY) {
 634                 lock = find_existing_lock(req->rq_export,
 635                                           &dlm_req->lock_handle1);
 636                 if (lock != NULL) {
 637                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 638                                   lock->l_handle.h_cookie);
 639                         GOTO(existing_lock, rc = 0);
 640                 }
 641         }
 642
 643         /* The lock's callback data might be set in the policy function */
 644         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 645                                 dlm_req->lock_desc.l_resource.lr_name,
 646                                 dlm_req->lock_desc.l_resource.lr_type,
 647                                 dlm_req->lock_desc.l_req_mode,
 648                                 blocking_callback, completion_callback,
 649                                 glimpse_callback, NULL, 0);
 650         if (!lock)
 651                 GOTO(out, rc = -ENOMEM);
 652
 653         do_gettimeofday(&lock->l_enqueued_time);
 654         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 655                sizeof(lock->l_remote_handle));
 656         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 657
 658         LASSERT(req->rq_export);
 659         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 660         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 661         if (req->rq_export->exp_failed) {
 662                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 663                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 664                 GOTO(out, err = -ENOTCONN);
 665         }
 666         lock->l_export = class_export_get(req->rq_export);
 667
 668         list_add(&lock->l_export_chain,
 669                  &lock->l_export->exp_ldlm_data.led_held_locks);
 670         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 671
 672 existing_lock:
 673
 674         if (flags & LDLM_FL_HAS_INTENT) {
 675                 /* In this case, the reply buffer is allocated deep in
 676                  * local_lock_enqueue by the policy function. */
 677                 cookie = req;
 678         } else {
 679                 int buffers = 1;
 680                 down(&lock->l_resource->lr_lvb_sem);
 681                 if (lock->l_resource->lr_lvb_len) {
 682                         size[1] = lock->l_resource->lr_lvb_len;
 683                         buffers = 2;
 684                 }
 685                 up(&lock->l_resource->lr_lvb_sem);
 686                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 687                         GOTO(out, rc = -ENOMEM);
 688
 689                 rc = lustre_pack_reply(req, buffers, size, NULL);
 690                 if (rc)
 691                         GOTO(out, rc);
 692         }
 693
 694         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 695                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 696                        sizeof(ldlm_policy_data_t));
 697         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 698                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 699                        sizeof(lock->l_req_extent));
 700
 701         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock,
 702                                 cookie, (int *)&flags);
 703         if (err)
 704                 GOTO(out, err);
 705
 706         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*dlm_rep));
 707         dlm_rep->lock_flags = flags;
 708
 709         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 710         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 711
 712         /* We never send a blocking AST until the lock is granted, but
 713          * we can tell it right now */
 714         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 715         if (lock->l_flags & LDLM_FL_AST_SENT) {
 716                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 717                 if (lock->l_granted_mode == lock->l_req_mode)
 718                         ldlm_add_waiting_lock(lock);
 719         }
 720         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 721
 722         EXIT;
 723  out:
 724         req->rq_status = err;
 725         if (req->rq_reply_state == NULL) {
 726                 err = lustre_pack_reply(req, 0, NULL, NULL);
 727                 if (rc == 0)
 728                         rc = err;
 729                 req->rq_status = rc;
 730         }
 731
 732         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 733          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 734         if (lock) {
 735                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 736                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 737                            "(err=%d, rc=%d)", err, rc);
 738                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 739
 740                 if (rc == 0) {
 741                         down(&lock->l_resource->lr_lvb_sem);
 742                         size[1] = lock->l_resource->lr_lvb_len;
 743                         if (size[1] > 0) {
 744                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 745                                                            1, size[1]);
 746                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 747                                          req, lock);
 748
 749                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 750                                        size[1]);
 751                         }
 752                         up(&lock->l_resource->lr_lvb_sem);
 753                 } else {
 754                         ldlm_lock_destroy(lock);
 755                 }
 756
 757                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 758                         ldlm_reprocess_all(lock->l_resource);
 759                 LDLM_LOCK_PUT(lock);
 760         }
 761         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 762                           lock, rc);
 763
 764         return rc;
 765 }
 766
 767 int ldlm_handle_convert(struct ptlrpc_request *req)
 768 {
 769         struct ldlm_request *dlm_req;
 770         struct ldlm_reply *dlm_rep;
 771         struct ldlm_lock *lock;
 772         int rc, size = sizeof(*dlm_rep);
 773         ENTRY;
 774
 775         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 776                                      lustre_swab_ldlm_request);
 777         if (dlm_req == NULL) {
 778                 CERROR ("Can't unpack dlm_req\n");
 779                 RETURN (-EFAULT);
 780         }
 781
 782         rc = lustre_pack_reply(req, 1, &size, NULL);
 783         if (rc) {
 784                 CERROR("out of memory\n");
 785                 RETURN(-ENOMEM);
 786         }
 787         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 788         dlm_rep->lock_flags = dlm_req->lock_flags;
 789
 790         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 791         if (!lock) {
 792                 req->rq_status = EINVAL;
 793         } else {
 794                 void *res = NULL;
 795
 796                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 797                 LDLM_DEBUG(lock, "server-side convert handler START");
 798                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 799
 800                 do_gettimeofday(&lock->l_enqueued_time);
 801                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 802                                         (int *)&dlm_rep->lock_flags);
 803                 if (res) {
 804                         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 805                         if (ldlm_del_waiting_lock(lock))
 806                                 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 807                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 808                         req->rq_status = 0;
 809                 } else {
 810                         req->rq_status = EDEADLOCK;
 811                 }
 812         }
 813
 814         if (lock) {
 815                 if (!req->rq_status)
 816                         ldlm_reprocess_all(lock->l_resource);
 817                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 818                 LDLM_DEBUG(lock, "server-side convert handler END");
 819                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 820                 LDLM_LOCK_PUT(lock);
 821         } else
 822                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 823
 824         RETURN(0);
 825 }
 826
 827 int ldlm_handle_cancel(struct ptlrpc_request *req)
 828 {
 829         struct ldlm_request *dlm_req;
 830         struct ldlm_lock *lock;
 831         struct ldlm_resource *res;
 832         int rc;
 833         ENTRY;
 834
 835         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 836                                       lustre_swab_ldlm_request);
 837         if (dlm_req == NULL) {
 838                 CERROR("bad request buffer for cancel\n");
 839                 RETURN(-EFAULT);
 840         }
 841
 842         rc = lustre_pack_reply(req, 0, NULL, NULL);
 843         if (rc) {
 844                 CERROR("out of memory\n");
 845                 RETURN(-ENOMEM);
 846         }
 847
 848         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 849         if (!lock) {
 850                 CERROR("received cancel for unknown lock cookie "LPX64
 851                        " from client %s id %s\n",
 852                        dlm_req->lock_handle1.cookie,
 853                        req->rq_export->exp_client_uuid.uuid,
 854                        req->rq_peerstr);
 855                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 856                                   "(cookie "LPU64")",
 857                                   dlm_req->lock_handle1.cookie);
 858                 req->rq_status = ESTALE;
 859         } else {
 860                 LDLM_DEBUG(lock, "server-side cancel handler START");
 861                 res = lock->l_resource;
 862                 if (res && res->lr_namespace->ns_lvbo &&
 863                     res->lr_namespace->ns_lvbo->lvbo_update) {
 864                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 865                                 (res, NULL, 0, 0);
 866                                 //(res, req->rq_reqmsg, 1);
 867                 }
 868
 869                 l_lock(&res->lr_namespace->ns_lock);
 870                 ldlm_lock_cancel(lock);
 871                 if (ldlm_del_waiting_lock(lock))
 872                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 873                 l_unlock(&res->lr_namespace->ns_lock);
 874                 req->rq_status = rc;
 875         }
 876
 877         if (ptlrpc_reply(req) != 0)
 878                 LBUG();
 879
 880         if (lock) {
 881                 ldlm_reprocess_all(lock->l_resource);
 882                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 883                 LDLM_DEBUG(lock, "server-side cancel handler END");
 884                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 885                 LDLM_LOCK_PUT(lock);
 886         }
 887
 888         RETURN(0);
 889 }
 890
 891 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 892                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 893 {
 894         int do_ast;
 895         ENTRY;
 896
 897         l_lock(&ns->ns_lock);
 898         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 899
 900         lock->l_flags |= LDLM_FL_CBPENDING;
 901         do_ast = (!lock->l_readers && !lock->l_writers);
 902
 903         if (do_ast) {
 904                 LDLM_DEBUG(lock, "already unused, calling "
 905                            "callback (%p)", lock->l_blocking_ast);
 906                 if (lock->l_blocking_ast != NULL) {
 907                         l_unlock(&ns->ns_lock);
 908                         l_check_no_ns_lock(ns);
 909                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 910                                              LDLM_CB_BLOCKING);
 911                         l_lock(&ns->ns_lock);
 912                 }
 913         } else {
 914                 LDLM_DEBUG(lock, "Lock still has references, will be"
 915                            " cancelled later");
 916         }
 917
 918         LDLM_DEBUG(lock, "client blocking callback handler END");
 919         l_unlock(&ns->ns_lock);
 920         LDLM_LOCK_PUT(lock);
 921         EXIT;
 922 }
 923
 924 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 925                                     struct ldlm_namespace *ns,
 926                                     struct ldlm_request *dlm_req,
 927                                     struct ldlm_lock *lock)
 928 {
 929         LIST_HEAD(ast_list);
 930         ENTRY;
 931
 932         l_lock(&ns->ns_lock);
 933         LDLM_DEBUG(lock, "client completion callback handler START");
 934
 935         /* If we receive the completion AST before the actual enqueue returned,
 936          * then we might need to switch lock modes, resources, or extents. */
 937         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 938                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 939                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 940         }
 941
 942         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 943                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 944                        sizeof(lock->l_policy_data));
 945                 LDLM_DEBUG(lock, "completion AST, new policy data");
 946         }
 947
 948         ldlm_resource_unlink_lock(lock);
 949         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 950                    &lock->l_resource->lr_name,
 951                    sizeof(lock->l_resource->lr_name)) != 0) {
 952                 ldlm_lock_change_resource(ns, lock,
 953                                          dlm_req->lock_desc.l_resource.lr_name);
 954                 LDLM_DEBUG(lock, "completion AST, new resource");
 955         }
 956
 957         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 958                 lock->l_flags |= LDLM_FL_CBPENDING;
 959                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 960         }
 961
 962         if (lock->l_lvb_len) {
 963                 void *lvb;
 964                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 965                                          lock->l_lvb_swabber);
 966                 if (lvb == NULL) {
 967                         LDLM_ERROR(lock, "completion AST did not contain "
 968                                    "expected LVB!");
 969                 } else {
 970                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 971                 }
 972         }
 973
 974         lock->l_resource->lr_tmp = &ast_list;
 975         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 976         lock->l_resource->lr_tmp = NULL;
 977         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 978         l_unlock(&ns->ns_lock);
 979         LDLM_LOCK_PUT(lock);
 980
 981         ldlm_run_ast_work(ns, &ast_list);
 982
 983         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 984                           lock);
 985         EXIT;
 986 }
 987
 988 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 989                                     struct ldlm_namespace *ns,
 990                                     struct ldlm_request *dlm_req,
 991                                     struct ldlm_lock *lock)
 992 {
 993         int rc = -ENOSYS;
 994         ENTRY;
 995
 996         l_lock(&ns->ns_lock);
 997         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 998
 999         if (lock->l_glimpse_ast != NULL) {
1000                 l_unlock(&ns->ns_lock);
1001                 l_check_no_ns_lock(ns);
1002                 rc = lock->l_glimpse_ast(lock, req);
1003                 l_lock(&ns->ns_lock);
1004         }
1005
1006         if (req->rq_repmsg != NULL) {
1007                 ptlrpc_reply(req);
1008         } else {
1009                 req->rq_status = rc;
1010                 ptlrpc_error(req);
1011         }
1012
1013         l_unlock(&ns->ns_lock);
1014         if (lock->l_granted_mode == LCK_PW &&
1015             !lock->l_readers && !lock->l_writers &&
1016             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
1017                 if (ldlm_bl_to_thread(ns, NULL, lock))
1018                         ldlm_handle_bl_callback(ns, NULL, lock);
1019
1020                 EXIT;
1021                 return;
1022         }
1023         LDLM_LOCK_PUT(lock);
1024         EXIT;
1025 }
1026
1027 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1028 {
1029         req->rq_status = rc;
1030         if (req->rq_reply_state == NULL) {
1031                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1032                 if (rc)
1033                         return rc;
1034         }
1035         return ptlrpc_reply(req);
1036 }
1037
1038 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1039                       struct ldlm_lock *lock)
1040 {
1041 #ifdef __KERNEL__
1042         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1043         struct ldlm_bl_work_item *blwi;
1044         ENTRY;
1045
1046         OBD_ALLOC(blwi, sizeof(*blwi));
1047         if (blwi == NULL)
1048                 RETURN(-ENOMEM);
1049
1050         blwi->blwi_ns = ns;
1051         if (ld != NULL)
1052                 blwi->blwi_ld = *ld;
1053         blwi->blwi_lock = lock;
1054
1055         spin_lock(&blp->blp_lock);
1056         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1057         wake_up(&blp->blp_waitq);
1058         spin_unlock(&blp->blp_lock);
1059
1060         RETURN(0);
1061 #else
1062         RETURN(-ENOSYS);
1063 #endif
1064
1065 }
1066
1067 static int ldlm_msg_check_version(struct lustre_msg *msg)
1068 {
1069         int rc;
1070
1071         switch (msg->opc) {
1072         case LDLM_ENQUEUE:
1073         case LDLM_CONVERT:
1074         case LDLM_CANCEL:
1075         case LDLM_BL_CALLBACK:
1076         case LDLM_CP_CALLBACK:
1077         case LDLM_GL_CALLBACK:
1078                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1079                 if (rc)
1080                         CERROR("bad opc %u version %08x, expecting %08x\n",
1081                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1082                 break;
1083         case OBD_LOG_CANCEL:
1084         case LLOG_ORIGIN_HANDLE_OPEN:
1085         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1086         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1087         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1088         case LLOG_ORIGIN_HANDLE_CLOSE:
1089         case LLOG_CATINFO:
1090                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1091                 if (rc)
1092                         CERROR("bad opc %u version %08x, expecting %08x\n",
1093                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1094                 break;
1095         default:
1096                 CERROR("LDLM unknown opcode %d\n", msg->opc);
1097                 rc = -ENOTSUPP;
1098                 break;
1099         }
1100
1101         return rc;
1102 }
1103
1104 static int ldlm_callback_handler(struct ptlrpc_request *req)
1105 {
1106         struct ldlm_namespace *ns;
1107         struct ldlm_request *dlm_req;
1108         struct ldlm_lock *lock;
1109         int rc;
1110         ENTRY;
1111
1112         rc = ldlm_msg_check_version(req->rq_reqmsg);
1113         if (rc) {
1114                 CERROR("LDLM_CB drop mal-formed request\n");
1115                 RETURN(rc);
1116         }
1117
1118         /* Requests arrive in sender's byte order.  The ptlrpc service
1119          * handler has already checked and, if necessary, byte-swapped the
1120          * incoming request message body, but I am responsible for the
1121          * message buffers. */
1122
1123         if (req->rq_export == NULL) {
1124                 struct ldlm_request *dlm_req;
1125
1126                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1127                        "export cookie "LPX64"; this is "
1128                        "normal if this node rebooted with a lock held\n",
1129                        req->rq_reqmsg->opc,
1130                        req->rq_peerstr,
1131                        req->rq_reqmsg->handle.cookie);
1132                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1133                                              lustre_swab_ldlm_request);
1134                 if (dlm_req != NULL)
1135                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1136                                dlm_req->lock_handle1.cookie);
1137
1138                 ldlm_callback_reply(req, -ENOTCONN);
1139                 RETURN(0);
1140         }
1141
1142         LASSERT(req->rq_export != NULL);
1143         LASSERT(req->rq_export->exp_obd != NULL);
1144
1145         switch(req->rq_reqmsg->opc) {
1146         case LDLM_BL_CALLBACK:
1147                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1148                 break;
1149         case LDLM_CP_CALLBACK:
1150                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1151                 break;
1152         case LDLM_GL_CALLBACK:
1153                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1154                 break;
1155         case OBD_LOG_CANCEL:
1156                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1157                 rc = llog_origin_handle_cancel(req);
1158                 ldlm_callback_reply(req, rc);
1159                 RETURN(0);
1160         case LLOG_ORIGIN_HANDLE_OPEN:
1161                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1162                 rc = llog_origin_handle_open(req);
1163                 ldlm_callback_reply(req, rc);
1164                 RETURN(0);
1165         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1166                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1167                 rc = llog_origin_handle_next_block(req);
1168                 ldlm_callback_reply(req, rc);
1169                 RETURN(0);
1170         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1171                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1172                 rc = llog_origin_handle_prev_block(req);
1173                 ldlm_callback_reply(req, rc);
1174                 RETURN(0);
1175         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1176                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1177                 rc = llog_origin_handle_read_header(req);
1178                 ldlm_callback_reply(req, rc);
1179                 RETURN(0);
1180         case LLOG_ORIGIN_HANDLE_CLOSE:
1181                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1182                 rc = llog_origin_handle_close(req);
1183                 ldlm_callback_reply(req, rc);
1184                 RETURN(0);
1185         default:
1186                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1187                 ldlm_callback_reply(req, -EPROTO);
1188                 RETURN(0);
1189         }
1190
1191         ns = req->rq_export->exp_obd->obd_namespace;
1192         LASSERT(ns != NULL);
1193
1194         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1195                                       lustre_swab_ldlm_request);
1196         if (dlm_req == NULL) {
1197                 CERROR ("can't unpack dlm_req\n");
1198                 ldlm_callback_reply (req, -EPROTO);
1199                 RETURN (0);
1200         }
1201
1202         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1203         if (!lock) {
1204                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1205                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1206                 ldlm_callback_reply(req, -EINVAL);
1207                 RETURN(0);
1208         }
1209
1210         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1211         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1212
1213         /* We want the ost thread to get this reply so that it can respond
1214          * to ost requests (write cache writeback) that might be triggered
1215          * in the callback.
1216          *
1217          * But we'd also like to be able to indicate in the reply that we're
1218          * cancelling right now, because it's unused, or have an intent result
1219          * in the reply, so we might have to push the responsibility for sending
1220          * the reply down into the AST handlers, alas. */
1221
1222         switch (req->rq_reqmsg->opc) {
1223         case LDLM_BL_CALLBACK:
1224                 CDEBUG(D_INODE, "blocking ast\n");
1225                 ldlm_callback_reply(req, 0);
1226                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1227                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1228
1229                 break;
1230         case LDLM_CP_CALLBACK:
1231                 CDEBUG(D_INODE, "completion ast\n");
1232                 ldlm_callback_reply(req, 0);
1233                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1234                 break;
1235         case LDLM_GL_CALLBACK:
1236                 CDEBUG(D_INODE, "glimpse ast\n");
1237                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1238                 break;
1239         default:
1240                 LBUG();                         /* checked above */
1241         }
1242
1243         RETURN(0);
1244 }
1245
1246 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1247 {
1248         int rc = 0;
1249         ENTRY;
1250
1251         rc = ldlm_msg_check_version(req->rq_reqmsg);
1252         if (rc) {
1253                 CERROR("LDLM_CL drop mal-formed request\n");
1254                 RETURN(rc);
1255         }
1256
1257         /* Requests arrive in sender's byte order.  The ptlrpc service
1258          * handler has already checked and, if necessary, byte-swapped the
1259          * incoming request message body, but I am responsible for the
1260          * message buffers. */
1261
1262         if (req->rq_export == NULL) {
1263                 struct ldlm_request *dlm_req;
1264                 CERROR("operation %d with bad export from %s\n",
1265                        req->rq_reqmsg->opc,
1266                        req->rq_peerstr);
1267                 CERROR("--> export cookie: "LPX64"\n",
1268                        req->rq_reqmsg->handle.cookie);
1269                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1270                                              lustre_swab_ldlm_request);
1271                 if (dlm_req != NULL)
1272                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1273                 ldlm_callback_reply(req, -ENOTCONN);
1274                 RETURN(0);
1275         }
1276
1277         switch (req->rq_reqmsg->opc) {
1278
1279         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1280         case LDLM_CANCEL:
1281                 CDEBUG(D_INODE, "cancel\n");
1282                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1283                 rc = ldlm_handle_cancel(req);
1284                 break;
1285         default:
1286                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1287                 ldlm_callback_reply(req, -EINVAL);
1288         }
1289
1290         RETURN(rc);
1291 }
1292
1293 #ifdef __KERNEL__
1294 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1295 {
1296         struct ldlm_bl_work_item *blwi = NULL;
1297
1298         spin_lock(&blp->blp_lock);
1299         if (!list_empty(&blp->blp_list)) {
1300                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1301                                   blwi_entry);
1302                 list_del(&blwi->blwi_entry);
1303         }
1304         spin_unlock(&blp->blp_lock);
1305
1306         return blwi;
1307 }
1308
1309 struct ldlm_bl_thread_data {
1310         int                     bltd_num;
1311         struct ldlm_bl_pool     *bltd_blp;
1312 };
1313
1314 static int ldlm_bl_thread_main(void *arg)
1315 {
1316         struct ldlm_bl_thread_data *bltd = arg;
1317         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1318         unsigned long flags;
1319         ENTRY;
1320
1321         /* XXX boiler-plate */
1322         {
1323                 char name[sizeof(current->comm)];
1324                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1325                          bltd->bltd_num);
1326                 kportal_daemonize(name);
1327         }
1328         SIGNAL_MASK_LOCK(current, flags);
1329         sigfillset(&current->blocked);
1330         RECALC_SIGPENDING;
1331         SIGNAL_MASK_UNLOCK(current, flags);
1332
1333         atomic_inc(&blp->blp_num_threads);
1334         complete(&blp->blp_comp);
1335
1336         while(1) {
1337                 struct l_wait_info lwi = { 0 };
1338                 struct ldlm_bl_work_item *blwi = NULL;
1339
1340                 l_wait_event_exclusive(blp->blp_waitq,
1341                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1342                                        &lwi);
1343
1344                 if (blwi->blwi_ns == NULL)
1345                         break;
1346
1347                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1348                                         blwi->blwi_lock);
1349                 OBD_FREE(blwi, sizeof(*blwi));
1350         }
1351
1352         atomic_dec(&blp->blp_num_threads);
1353         complete(&blp->blp_comp);
1354         RETURN(0);
1355 }
1356
1357 #endif
1358
1359 static int ldlm_setup(void);
1360 static int ldlm_cleanup(int force);
1361
1362 int ldlm_get_ref(void)
1363 {
1364         int rc = 0;
1365         down(&ldlm_ref_sem);
1366         if (++ldlm_refcount == 1) {
1367                 rc = ldlm_setup();
1368                 if (rc)
1369                         ldlm_refcount--;
1370         }
1371         up(&ldlm_ref_sem);
1372
1373         RETURN(rc);
1374 }
1375
1376 void ldlm_put_ref(int force)
1377 {
1378         down(&ldlm_ref_sem);
1379         if (ldlm_refcount == 1) {
1380                 int rc = ldlm_cleanup(force);
1381                 if (rc)
1382                         CERROR("ldlm_cleanup failed: %d\n", rc);
1383                 else
1384                         ldlm_refcount--;
1385         } else {
1386                 ldlm_refcount--;
1387         }
1388         up(&ldlm_ref_sem);
1389
1390         EXIT;
1391 }
1392
1393 static int ldlm_setup(void)
1394 {
1395         struct ldlm_bl_pool *blp;
1396         int rc = 0;
1397 #ifdef __KERNEL__
1398         int i;
1399 #endif
1400         ENTRY;
1401
1402         if (ldlm_state != NULL)
1403                 RETURN(-EALREADY);
1404
1405         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1406         if (ldlm_state == NULL)
1407                 RETURN(-ENOMEM);
1408
1409 #ifdef __KERNEL__
1410         rc = ldlm_proc_setup();
1411         if (rc != 0)
1412                 GOTO(out_free, rc);
1413 #endif
1414
1415         ldlm_state->ldlm_cb_service =
1416                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1417                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1418                                 1500, ldlm_callback_handler, "ldlm_cbd",
1419                                 ldlm_svc_proc_dir);
1420
1421         if (!ldlm_state->ldlm_cb_service) {
1422                 CERROR("failed to start service\n");
1423                 GOTO(out_proc, rc = -ENOMEM);
1424         }
1425
1426         ldlm_state->ldlm_cancel_service =
1427                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1428                                 LDLM_CANCEL_REQUEST_PORTAL,
1429                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1430                                 ldlm_cancel_handler, "ldlm_canceld",
1431                                 ldlm_svc_proc_dir);
1432
1433         if (!ldlm_state->ldlm_cancel_service) {
1434                 CERROR("failed to start service\n");
1435                 GOTO(out_proc, rc = -ENOMEM);
1436         }
1437
1438         OBD_ALLOC(blp, sizeof(*blp));
1439         if (blp == NULL)
1440                 GOTO(out_proc, rc = -ENOMEM);
1441         ldlm_state->ldlm_bl_pool = blp;
1442
1443         atomic_set(&blp->blp_num_threads, 0);
1444         init_waitqueue_head(&blp->blp_waitq);
1445         spin_lock_init(&blp->blp_lock);
1446
1447         INIT_LIST_HEAD(&blp->blp_list);
1448
1449 #ifdef __KERNEL__
1450         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1451                 struct ldlm_bl_thread_data bltd = {
1452                         .bltd_num = i,
1453                         .bltd_blp = blp,
1454                 };
1455                 init_completion(&blp->blp_comp);
1456                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1457                 if (rc < 0) {
1458                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1459                         GOTO(out_thread, rc);
1460                 }
1461                 wait_for_completion(&blp->blp_comp);
1462         }
1463
1464         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1465                                     LDLM_NUM_THREADS, "ldlm_cn");
1466         if (rc)
1467                 GOTO(out_thread, rc);
1468
1469         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1470                                     LDLM_NUM_THREADS, "ldlm_cb");
1471         if (rc)
1472                 GOTO(out_thread, rc);
1473
1474         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1475         spin_lock_init(&expired_lock_thread.elt_lock);
1476         expired_lock_thread.elt_state = ELT_STOPPED;
1477         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1478
1479         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1480         if (rc < 0) {
1481                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1482                 GOTO(out_thread, rc);
1483         }
1484
1485         wait_event(expired_lock_thread.elt_waitq,
1486                    expired_lock_thread.elt_state == ELT_READY);
1487
1488         INIT_LIST_HEAD(&waiting_locks_list);
1489         spin_lock_init(&waiting_locks_spinlock);
1490         waiting_locks_timer.function = waiting_locks_callback;
1491         waiting_locks_timer.data = 0;
1492         init_timer(&waiting_locks_timer);
1493 #endif
1494
1495         RETURN(0);
1496
1497 #ifdef __KERNEL__
1498  out_thread:
1499         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1500         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1501 #endif
1502
1503  out_proc:
1504 #ifdef __KERNEL__
1505         ldlm_proc_cleanup();
1506  out_free:
1507 #endif
1508         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1509         ldlm_state = NULL;
1510         return rc;
1511 }
1512
1513 static int ldlm_cleanup(int force)
1514 {
1515 #ifdef __KERNEL__
1516         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1517 #endif
1518         ENTRY;
1519
1520         if (!list_empty(&ldlm_namespace_list)) {
1521                 CERROR("ldlm still has namespaces; clean these up first.\n");
1522                 ldlm_dump_all_namespaces(D_DLMTRACE);
1523                 RETURN(-EBUSY);
1524         }
1525
1526 #ifdef __KERNEL__
1527         while (atomic_read(&blp->blp_num_threads) > 0) {
1528                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1529
1530                 init_completion(&blp->blp_comp);
1531
1532                 spin_lock(&blp->blp_lock);
1533                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1534                 wake_up(&blp->blp_waitq);
1535                 spin_unlock(&blp->blp_lock);
1536
1537                 wait_for_completion(&blp->blp_comp);
1538         }
1539         OBD_FREE(blp, sizeof(*blp));
1540
1541         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1542         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1543         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1544         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1545         ldlm_proc_cleanup();
1546
1547         expired_lock_thread.elt_state = ELT_TERMINATE;
1548         wake_up(&expired_lock_thread.elt_waitq);
1549         wait_event(expired_lock_thread.elt_waitq,
1550                    expired_lock_thread.elt_state == ELT_STOPPED);
1551 #else
1552         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1553         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1554 #endif
1555
1556         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1557         ldlm_state = NULL;
1558
1559         RETURN(0);
1560 }
1561
1562 int __init ldlm_init(void)
1563 {
1564         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1565                                                sizeof(struct ldlm_resource), 0,
1566                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1567         if (ldlm_resource_slab == NULL)
1568                 return -ENOMEM;
1569
1570         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1571                                            sizeof(struct ldlm_lock), 0,
1572                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1573         if (ldlm_lock_slab == NULL) {
1574                 kmem_cache_destroy(ldlm_resource_slab);
1575                 return -ENOMEM;
1576         }
1577
1578         l_lock_init(&ldlm_handle_lock);
1579
1580         return 0;
1581 }
1582
1583 void __exit ldlm_exit(void)
1584 {
1585         if ( ldlm_refcount )
1586                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1587         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1588                  "couldn't free ldlm resource slab\n");
1589         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1590                  "couldn't free ldlm lock slab\n");
1591
1592 }
1593
1594 /* ldlm_flock.c */
1595 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1596
1597 /* ldlm_extent.c */
1598 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1599
1600 /* ldlm_lock.c */
1601 EXPORT_SYMBOL(ldlm_get_processing_policy);
1602 EXPORT_SYMBOL(ldlm_lock2desc);
1603 EXPORT_SYMBOL(ldlm_register_intent);
1604 EXPORT_SYMBOL(ldlm_lockname);
1605 EXPORT_SYMBOL(ldlm_typename);
1606 EXPORT_SYMBOL(ldlm_lock2handle);
1607 EXPORT_SYMBOL(__ldlm_handle2lock);
1608 EXPORT_SYMBOL(ldlm_lock_get);
1609 EXPORT_SYMBOL(ldlm_lock_put);
1610 EXPORT_SYMBOL(ldlm_lock_match);
1611 EXPORT_SYMBOL(ldlm_lock_cancel);
1612 EXPORT_SYMBOL(ldlm_lock_addref);
1613 EXPORT_SYMBOL(ldlm_lock_decref);
1614 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1615 EXPORT_SYMBOL(ldlm_lock_change_resource);
1616 EXPORT_SYMBOL(ldlm_lock_set_data);
1617 EXPORT_SYMBOL(ldlm_it2str);
1618 EXPORT_SYMBOL(ldlm_lock_dump);
1619 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1620 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1621 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1622 EXPORT_SYMBOL(ldlm_lock_allow_match);
1623
1624 /* ldlm_request.c */
1625 EXPORT_SYMBOL(ldlm_completion_ast);
1626 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1627 EXPORT_SYMBOL(ldlm_cli_convert);
1628 EXPORT_SYMBOL(ldlm_cli_enqueue);
1629 EXPORT_SYMBOL(ldlm_cli_cancel);
1630 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1631 EXPORT_SYMBOL(ldlm_replay_locks);
1632 EXPORT_SYMBOL(ldlm_resource_foreach);
1633 EXPORT_SYMBOL(ldlm_namespace_foreach);
1634 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1635 EXPORT_SYMBOL(ldlm_change_cbdata);
1636
1637 /* ldlm_lockd.c */
1638 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1639 EXPORT_SYMBOL(ldlm_server_completion_ast);
1640 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1641 EXPORT_SYMBOL(ldlm_handle_enqueue);
1642 EXPORT_SYMBOL(ldlm_handle_cancel);
1643 EXPORT_SYMBOL(ldlm_handle_convert);
1644 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1645 EXPORT_SYMBOL(ldlm_get_ref);
1646 EXPORT_SYMBOL(ldlm_put_ref);
1647
1648 #if 0
1649 /* ldlm_test.c */
1650 EXPORT_SYMBOL(ldlm_test);
1651 EXPORT_SYMBOL(ldlm_regression_start);
1652 EXPORT_SYMBOL(ldlm_regression_stop);
1653 #endif
1654
1655 /* ldlm_resource.c */
1656 EXPORT_SYMBOL(ldlm_namespace_new);
1657 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1658 EXPORT_SYMBOL(ldlm_namespace_free);
1659 EXPORT_SYMBOL(ldlm_namespace_dump);
1660 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1661 EXPORT_SYMBOL(ldlm_resource_get);
1662 EXPORT_SYMBOL(ldlm_resource_putref);
1663
1664 /* l_lock.c */
1665 EXPORT_SYMBOL(l_lock);
1666 EXPORT_SYMBOL(l_unlock);
1667
1668 /* ldlm_lib.c */
1669 EXPORT_SYMBOL(client_import_add_conn);
1670 EXPORT_SYMBOL(client_import_del_conn);
1671 EXPORT_SYMBOL(client_obd_setup);
1672 EXPORT_SYMBOL(client_obd_cleanup);
1673 EXPORT_SYMBOL(client_connect_import);
1674 EXPORT_SYMBOL(client_disconnect_export);
1675 EXPORT_SYMBOL(target_start_recovery_thread);
1676 EXPORT_SYMBOL(target_stop_recovery_thread);
1677 EXPORT_SYMBOL(target_handle_connect);
1678 EXPORT_SYMBOL(target_cleanup_recovery);
1679 EXPORT_SYMBOL(target_destroy_export);
1680 EXPORT_SYMBOL(target_cancel_recovery_timer);
1681 EXPORT_SYMBOL(target_send_reply);
1682 EXPORT_SYMBOL(target_queue_recovery_request);
1683 EXPORT_SYMBOL(target_handle_ping);
1684 EXPORT_SYMBOL(target_handle_disconnect);