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 <portals/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 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  48 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  49
  50 static DECLARE_MUTEX(ldlm_ref_sem);
  51 static int ldlm_refcount = 0;
  52
  53 /* LDLM state */
  54
  55 static struct ldlm_state *ldlm_state;
  56
  57 inline unsigned long round_timeout(unsigned long timeout)
  58 {
  59         return ((timeout / HZ) + 1) * HZ;
  60 }
  61
  62 #ifdef __KERNEL__
  63 /* XXX should this be per-ldlm? */
  64 static struct list_head waiting_locks_list;
  65 static spinlock_t waiting_locks_spinlock;
  66 static struct timer_list waiting_locks_timer;
  67
  68 static struct expired_lock_thread {
  69         wait_queue_head_t         elt_waitq;
  70         int                       elt_state;
  71         struct list_head          elt_expired_locks;
  72         spinlock_t                elt_lock;
  73 } expired_lock_thread;
  74 #endif
  75
  76 #define ELT_STOPPED   0
  77 #define ELT_READY     1
  78 #define ELT_TERMINATE 2
  79
  80 struct ldlm_bl_pool {
  81         spinlock_t              blp_lock;
  82         struct list_head        blp_list;
  83         wait_queue_head_t       blp_waitq;
  84         atomic_t                blp_num_threads;
  85         struct completion       blp_comp;
  86 };
  87
  88 struct ldlm_bl_work_item {
  89         struct list_head        blwi_entry;
  90         struct ldlm_namespace   *blwi_ns;
  91         struct ldlm_lock_desc   blwi_ld;
  92         struct ldlm_lock        *blwi_lock;
  93 };
  94
  95 #ifdef __KERNEL__
  96
  97 static inline int have_expired_locks(void)
  98 {
  99         int need_to_run;
 100
 101         spin_lock_bh(&expired_lock_thread.elt_lock);
 102         need_to_run = !list_empty(&expired_lock_thread.elt_expired_locks);
 103         spin_unlock_bh(&expired_lock_thread.elt_lock);
 104
 105         RETURN(need_to_run);
 106 }
 107
 108 static int expired_lock_main(void *arg)
 109 {
 110         struct list_head *expired = &expired_lock_thread.elt_expired_locks;
 111         struct l_wait_info lwi = { 0 };
 112         unsigned long flags;
 113
 114         ENTRY;
 115         lock_kernel();
 116         kportal_daemonize("ldlm_elt");
 117
 118         SIGNAL_MASK_LOCK(current, flags);
 119         sigfillset(&current->blocked);
 120         RECALC_SIGPENDING;
 121         SIGNAL_MASK_UNLOCK(current, flags);
 122
 123         unlock_kernel();
 124
 125         expired_lock_thread.elt_state = ELT_READY;
 126         wake_up(&expired_lock_thread.elt_waitq);
 127
 128         while (1) {
 129                 l_wait_event(expired_lock_thread.elt_waitq,
 130                              have_expired_locks() ||
 131                              expired_lock_thread.elt_state == ELT_TERMINATE,
 132                              &lwi);
 133
 134                 spin_lock_bh(&expired_lock_thread.elt_lock);
 135                 while (!list_empty(expired)) {
 136                         struct obd_export *export;
 137                         struct ldlm_lock *lock;
 138
 139                         lock = list_entry(expired->next, struct ldlm_lock,
 140                                           l_pending_chain);
 141                         if ((void *)lock < LP_POISON + PAGE_SIZE &&
 142                             (void *)lock >= LP_POISON) {
 143                                 CERROR("free lock on elt list %p\n", lock);
 144                                 LBUG();
 145                         }
 146                         list_del_init(&lock->l_pending_chain);
 147                         if ((void *)lock->l_export < LP_POISON + PAGE_SIZE &&
 148                             (void *)lock->l_export >= LP_POISON + PAGE_SIZE) {
 149                                 CERROR("lock with free export on elt list %p\n",
 150                                        export);
 151                                 lock->l_export = NULL;
 152                                 LDLM_ERROR(lock, "free export\n");
 153                                 continue;
 154                         }
 155                         export = class_export_get(lock->l_export);
 156                         spin_unlock_bh(&expired_lock_thread.elt_lock);
 157
 158                         ptlrpc_fail_export(export);
 159                         class_export_put(export);
 160                         spin_lock_bh(&expired_lock_thread.elt_lock);
 161                 }
 162                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 163
 164                 if (expired_lock_thread.elt_state == ELT_TERMINATE)
 165                         break;
 166         }
 167
 168         expired_lock_thread.elt_state = ELT_STOPPED;
 169         wake_up(&expired_lock_thread.elt_waitq);
 170         RETURN(0);
 171 }
 172
 173 static void waiting_locks_callback(unsigned long unused)
 174 {
 175         struct ldlm_lock *lock;
 176         char str[PTL_NALFMT_SIZE];
 177
 178         spin_lock_bh(&waiting_locks_spinlock);
 179         while (!list_empty(&waiting_locks_list)) {
 180                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 181                                   l_pending_chain);
 182
 183                 if ((lock->l_callback_timeout > jiffies) ||
 184                     (lock->l_req_mode == LCK_GROUP))
 185                         break;
 186
 187                 LDLM_ERROR(lock, "lock callback timer expired: evicting client "
 188                            "%s@%s nid "LPX64" (%s) ",
 189                            lock->l_export->exp_client_uuid.uuid,
 190                            lock->l_export->exp_connection->c_remote_uuid.uuid,
 191                            lock->l_export->exp_connection->c_peer.peer_nid,
 192                            portals_nid2str(lock->l_export->exp_connection->c_peer.peer_ni->pni_number,
 193                                            lock->l_export->exp_connection->c_peer.peer_nid,
 194                                            str));
 195
 196                 spin_lock_bh(&expired_lock_thread.elt_lock);
 197                 list_del(&lock->l_pending_chain);
 198                 list_add(&lock->l_pending_chain,
 199                          &expired_lock_thread.elt_expired_locks);
 200                 spin_unlock_bh(&expired_lock_thread.elt_lock);
 201                 wake_up(&expired_lock_thread.elt_waitq);
 202         }
 203
 204         /*
 205          * Make sure the timer will fire again if we have any locks
 206          * left.
 207          */
 208         if (!list_empty(&waiting_locks_list)) {
 209                 unsigned long timeout_rounded;
 210                 lock = list_entry(waiting_locks_list.next, struct ldlm_lock,
 211                                   l_pending_chain);
 212                 timeout_rounded = round_timeout(lock->l_callback_timeout);
 213                 mod_timer(&waiting_locks_timer, timeout_rounded);
 214         }
 215         spin_unlock_bh(&waiting_locks_spinlock);
 216 }
 217
 218 /*
 219  * Indicate that we're waiting for a client to call us back cancelling a given
 220  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
 221  * timer to fire appropriately.  (We round up to the next second, to avoid
 222  * floods of timer firings during periods of high lock contention and traffic).
 223  */
 224 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 225 {
 226         unsigned long timeout_rounded;
 227
 228         spin_lock_bh(&waiting_locks_spinlock);
 229         if (!list_empty(&lock->l_pending_chain)) {
 230                 LDLM_DEBUG(lock, "not re-adding to wait list");
 231                 spin_unlock_bh(&waiting_locks_spinlock);
 232                 return 0;
 233         }
 234         LDLM_DEBUG(lock, "adding to wait list");
 235
 236         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
 237
 238         timeout_rounded = round_timeout(lock->l_callback_timeout);
 239
 240         if (timeout_rounded < waiting_locks_timer.expires ||
 241             !timer_pending(&waiting_locks_timer)) {
 242                 mod_timer(&waiting_locks_timer, timeout_rounded);
 243         }
 244         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
 245         spin_unlock_bh(&waiting_locks_spinlock);
 246         return 1;
 247 }
 248
 249 /*
 250  * Remove a lock from the pending list, likely because it had its cancellation
 251  * callback arrive without incident.  This adjusts the lock-timeout timer if
 252  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 253  */
 254 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 255 {
 256         struct list_head *list_next;
 257
 258         if (lock->l_export == NULL) {
 259                 /* We don't have a "waiting locks list" on clients. */
 260                 LDLM_DEBUG(lock, "client lock: no-op");
 261                 return 0;
 262         }
 263
 264         spin_lock_bh(&waiting_locks_spinlock);
 265
 266         if (list_empty(&lock->l_pending_chain)) {
 267                 spin_unlock_bh(&waiting_locks_spinlock);
 268                 LDLM_DEBUG(lock, "wasn't waiting");
 269                 return 0;
 270         }
 271
 272         list_next = lock->l_pending_chain.next;
 273         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 274                 /* Removing the head of the list, adjust timer. */
 275                 if (list_next == &waiting_locks_list) {
 276                         /* No more, just cancel. */
 277                         del_timer(&waiting_locks_timer);
 278                 } else {
 279                         struct ldlm_lock *next;
 280                         next = list_entry(list_next, struct ldlm_lock,
 281                                           l_pending_chain);
 282                         mod_timer(&waiting_locks_timer,
 283                                   round_timeout(next->l_callback_timeout));
 284                 }
 285         }
 286
 287         spin_lock_bh(&expired_lock_thread.elt_lock);
 288         list_del_init(&lock->l_pending_chain);
 289         spin_unlock_bh(&expired_lock_thread.elt_lock);
 290
 291         spin_unlock_bh(&waiting_locks_spinlock);
 292         LDLM_DEBUG(lock, "removed");
 293         return 1;
 294 }
 295
 296 #else /* !__KERNEL__ */
 297
 298 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
 299 {
 300         RETURN(1);
 301 }
 302
 303 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 304 {
 305         RETURN(0);
 306 }
 307
 308 #endif /* __KERNEL__ */
 309
 310 static void ldlm_failed_ast(struct ldlm_lock *lock, int rc, char *ast_type)
 311 {
 312         const struct ptlrpc_connection *conn = lock->l_export->exp_connection;
 313         char str[PTL_NALFMT_SIZE];
 314
 315         CERROR("%s AST failed (%d) for res "LPU64"/"LPU64
 316                ", mode %s: evicting client %s@%s NID "LPX64" (%s)\n",
 317                ast_type, rc,
 318                lock->l_resource->lr_name.name[0],
 319                lock->l_resource->lr_name.name[1],
 320                ldlm_lockname[lock->l_granted_mode],
 321                lock->l_export->exp_client_uuid.uuid,
 322                conn->c_remote_uuid.uuid, conn->c_peer.peer_nid,
 323                portals_nid2str(conn->c_peer.peer_ni->pni_number,
 324                                conn->c_peer.peer_nid, str));
 325         ptlrpc_fail_export(lock->l_export);
 326 }
 327
 328 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 329                              struct ldlm_lock_desc *desc,
 330                              void *data, int flag)
 331 {
 332         struct ldlm_request *body;
 333         struct ptlrpc_request *req;
 334         int rc = 0, size = sizeof(*body);
 335         ENTRY;
 336
 337         if (flag == LDLM_CB_CANCELING) {
 338                 /* Don't need to do anything here. */
 339                 RETURN(0);
 340         }
 341
 342         LASSERT(lock);
 343
 344         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 345         if (lock->l_granted_mode != lock->l_req_mode) {
 346                 /* this blocking AST will be communicated as part of the
 347                  * completion AST instead */
 348                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 349                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");                 RETURN(0);
 350         }
 351
 352         if (lock->l_destroyed) {
 353                 /* What's the point? */
 354                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 355                 RETURN(0);
 356         }
 357
 358 #if 0
 359         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 360                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 361                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 362                 RETURN(-ETIMEDOUT);
 363         }
 364 #endif
 365
 366         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 367                               LDLM_BL_CALLBACK, 1, &size, NULL);
 368         if (req == NULL) {
 369                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 370                 RETURN(-ENOMEM);
 371         }
 372
 373         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 374         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 375                sizeof(body->lock_handle1));
 376         memcpy(&body->lock_desc, desc, sizeof(*desc));
 377         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 378
 379         LDLM_DEBUG(lock, "server preparing blocking AST");
 380         req->rq_replen = lustre_msg_size(0, NULL);
 381
 382         if (lock->l_granted_mode == lock->l_req_mode)
 383                 ldlm_add_waiting_lock(lock);
 384         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 385
 386         req->rq_send_state = LUSTRE_IMP_FULL;
 387         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 388         rc = ptlrpc_queue_wait(req);
 389         if (rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) {
 390                 LASSERT(lock->l_export);
 391                 if (lock->l_export->exp_libclient) {
 392                         CDEBUG(D_HA, "BLOCKING AST to liblustre client (nid "
 393                                LPU64") timeout, simply cancel lock 0x%p\n",
 394                                req->rq_peer.peer_nid, lock);
 395                         ldlm_lock_cancel(lock);
 396                         rc = -ERESTART;
 397                 } else {
 398                         ldlm_del_waiting_lock(lock);
 399                         ldlm_failed_ast(lock, rc, "blocking");
 400                 }
 401         } else if (rc) {
 402                 if (rc == -EINVAL)
 403                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 404                                "from blocking AST for lock %p--normal race\n",
 405                                req->rq_peer.peer_nid,
 406                                req->rq_repmsg->status, lock);
 407                 else if (rc == -ENOTCONN)
 408                         CDEBUG(D_DLMTRACE, "client (nid "LPU64") returned %d "
 409                                "from blocking AST for lock %p--this client was "
 410                                "probably rebooted while it held a lock, nothing"
 411                                " serious\n",req->rq_peer.peer_nid,
 412                                req->rq_repmsg->status, lock);
 413                 else
 414                         CDEBUG(D_ERROR, "client (nid "LPU64") returned %d "
 415                                "from blocking AST for lock %p\n",
 416                                req->rq_peer.peer_nid,
 417                                (req->rq_repmsg != NULL)?
 418                                req->rq_repmsg->status : 0,
 419                                lock);
 420                 LDLM_DEBUG(lock, "client sent rc %d rq_status %d from blocking "
 421                            "AST", rc, req->rq_status);
 422                 ldlm_lock_cancel(lock);
 423                 /* Server-side AST functions are called from ldlm_reprocess_all,
 424                  * which needs to be told to please restart its reprocessing. */
 425                 rc = -ERESTART;
 426         }
 427
 428         ptlrpc_req_finished(req);
 429
 430         RETURN(rc);
 431 }
 432
 433 /* XXX copied from ptlrpc/service.c */
 434 static long timeval_sub(struct timeval *large, struct timeval *small)
 435 {
 436         return (large->tv_sec - small->tv_sec) * 1000000 +
 437                 (large->tv_usec - small->tv_usec);
 438 }
 439
 440 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 441 {
 442         struct ldlm_request *body;
 443         struct ptlrpc_request *req;
 444         struct timeval granted_time;
 445         long total_enqueue_wait;
 446         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 447         ENTRY;
 448
 449         LASSERT(lock != NULL);
 450
 451         do_gettimeofday(&granted_time);
 452         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 453
 454         if (total_enqueue_wait / 1000000 > obd_timeout)
 455                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 456
 457         if (lock->l_resource->lr_lvb_len) {
 458                 buffers = 2;
 459                 size[1] = lock->l_resource->lr_lvb_len;
 460         }
 461
 462         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 463                               LDLM_CP_CALLBACK, buffers, size, NULL);
 464         if (req == NULL)
 465                 RETURN(-ENOMEM);
 466
 467         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 468         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 469                sizeof(body->lock_handle1));
 470         body->lock_flags = flags;
 471         ldlm_lock2desc(lock, &body->lock_desc);
 472
 473         if (buffers == 2) {
 474                 void *lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 475                                            lock->l_resource->lr_lvb_len);
 476                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 477                        lock->l_resource->lr_lvb_len);
 478         }
 479
 480         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 481                    total_enqueue_wait);
 482         req->rq_replen = lustre_msg_size(0, NULL);
 483
 484         req->rq_send_state = LUSTRE_IMP_FULL;
 485         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 486
 487         /* We only send real blocking ASTs after the lock is granted */
 488         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 489         if (lock->l_flags & LDLM_FL_AST_SENT) {
 490                 body->lock_flags |= LDLM_FL_AST_SENT;
 491                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 492         }
 493         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 494
 495         rc = ptlrpc_queue_wait(req);
 496         if ((rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) &&
 497              !lock->l_export->exp_libclient) {
 498                 ldlm_del_waiting_lock(lock);
 499                 ldlm_failed_ast(lock, rc, "completion");
 500         } else if (rc == -EINVAL) {
 501                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 502                            "lock");
 503         } else if (rc) {
 504                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 505                            "completion AST", rc, req->rq_status);
 506                 ldlm_lock_cancel(lock);
 507                 /* Server-side AST functions are called from ldlm_reprocess_all,
 508                  * which needs to be told to please restart its reprocessing. */
 509                 rc = -ERESTART;
 510         }
 511         ptlrpc_req_finished(req);
 512
 513         RETURN(rc);
 514 }
 515
 516 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 517 {
 518         struct ldlm_resource *res = lock->l_resource;
 519         struct ldlm_request *body;
 520         struct ptlrpc_request *req;
 521         int rc = 0, size = sizeof(*body);
 522         ENTRY;
 523
 524         LASSERT(lock != NULL);
 525
 526         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 527                               LDLM_GL_CALLBACK, 1, &size, NULL);
 528         if (req == NULL)
 529                 RETURN(-ENOMEM);
 530
 531         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 532         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 533                sizeof(body->lock_handle1));
 534         ldlm_lock2desc(lock, &body->lock_desc);
 535
 536         size = lock->l_resource->lr_lvb_len;
 537         req->rq_replen = lustre_msg_size(1, &size);
 538
 539         req->rq_send_state = LUSTRE_IMP_FULL;
 540         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 541
 542         rc = ptlrpc_queue_wait(req);
 543         if ((rc == -ETIMEDOUT || rc == -EINTR || rc == -ENOTCONN) &&
 544             !lock->l_export->exp_libclient) {
 545                 ldlm_del_waiting_lock(lock);
 546                 ldlm_failed_ast(lock, rc, "glimpse");
 547         } else if (rc == -EINVAL) {
 548                 LDLM_DEBUG(lock, "lost the race -- client no longer has this "
 549                            "lock");
 550         } else if (rc == -ELDLM_NO_LOCK_DATA) {
 551                 LDLM_DEBUG(lock, "lost a race -- client has a lock, but no "
 552                            "inode");
 553         } else if (rc) {
 554                 LDLM_ERROR(lock, "client sent rc %d rq_status %d from "
 555                            "glimpse AST", rc, req->rq_status);
 556         } else {
 557                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 558                         (res, req->rq_repmsg, 0, 1);
 559         }
 560         ptlrpc_req_finished(req);
 561         RETURN(rc);
 562 }
 563
 564 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 565                         ldlm_completion_callback completion_callback,
 566                         ldlm_blocking_callback blocking_callback,
 567                         ldlm_glimpse_callback glimpse_callback)
 568 {
 569         struct obd_device *obddev = req->rq_export->exp_obd;
 570         struct ldlm_reply *dlm_rep;
 571         struct ldlm_request *dlm_req;
 572         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 573         __u32 flags;
 574         ldlm_error_t err = ELDLM_OK;
 575         struct ldlm_lock *lock = NULL;
 576         void *cookie = NULL;
 577         ENTRY;
 578
 579         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 580
 581         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 582                                       lustre_swab_ldlm_request);
 583         if (dlm_req == NULL) {
 584                 CERROR ("Can't unpack dlm_req\n");
 585                 GOTO(out, rc = -EFAULT);
 586         }
 587
 588         flags = dlm_req->lock_flags;
 589
 590         /* The lock's callback data might be set in the policy function */
 591         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 592                                 dlm_req->lock_desc.l_resource.lr_name,
 593                                 dlm_req->lock_desc.l_resource.lr_type,
 594                                 dlm_req->lock_desc.l_req_mode,
 595                                 blocking_callback, completion_callback,
 596                                 glimpse_callback, NULL, 0);
 597         if (!lock)
 598                 GOTO(out, rc = -ENOMEM);
 599
 600         do_gettimeofday(&lock->l_enqueued_time);
 601         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 602                sizeof(lock->l_remote_handle));
 603         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 604
 605         LASSERT(req->rq_export);
 606         lock->l_export = class_export_get(req->rq_export);
 607         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 608         list_add(&lock->l_export_chain,
 609                  &lock->l_export->exp_ldlm_data.led_held_locks);
 610         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 611
 612         if (flags & LDLM_FL_HAS_INTENT) {
 613                 /* In this case, the reply buffer is allocated deep in
 614                  * local_lock_enqueue by the policy function. */
 615                 cookie = req;
 616         } else {
 617                 int buffers = 1;
 618                 if (lock->l_resource->lr_lvb_len) {
 619                         size[1] = lock->l_resource->lr_lvb_len;
 620                         buffers = 2;
 621                 }
 622
 623                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 624                         GOTO(out, rc = -ENOMEM);
 625
 626                 rc = lustre_pack_reply(req, buffers, size, NULL);
 627                 if (rc)
 628                         GOTO(out, rc);
 629         }
 630
 631         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 632                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 633                        sizeof(ldlm_policy_data_t));
 634         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 635                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 636                        sizeof(lock->l_req_extent));
 637
 638         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 639         if (err)
 640                 GOTO(out, err);
 641
 642         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 643         dlm_rep->lock_flags = flags;
 644
 645         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 646         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 647
 648         /* We never send a blocking AST until the lock is granted, but
 649          * we can tell it right now */
 650         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 651         if (lock->l_flags & LDLM_FL_AST_SENT) {
 652                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 653                 if (lock->l_granted_mode == lock->l_req_mode)
 654                         ldlm_add_waiting_lock(lock);
 655         }
 656         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 657
 658         EXIT;
 659  out:
 660         req->rq_status = err;
 661         if (req->rq_reply_state == NULL) {
 662                 err = lustre_pack_reply(req, 0, NULL, NULL);
 663                 if (rc == 0)
 664                         rc = err;
 665                 req->rq_status = rc;
 666         }
 667
 668         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 669          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 670         if (lock) {
 671                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 672                            "(err=%d, rc=%d)", err, rc);
 673
 674                 if (lock->l_resource->lr_lvb_len > 0 && rc == 0) {
 675                         void *lvb = lustre_msg_buf(req->rq_repmsg, 1,
 676                                                   lock->l_resource->lr_lvb_len);
 677                         LASSERT(lvb != NULL);
 678                         memcpy(lvb, lock->l_resource->lr_lvb_data,
 679                                lock->l_resource->lr_lvb_len);
 680                 }
 681
 682                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 683                         ldlm_reprocess_all(lock->l_resource);
 684                 LDLM_LOCK_PUT(lock);
 685         }
 686         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 687                           lock, rc);
 688
 689         return rc;
 690 }
 691
 692 int ldlm_handle_convert(struct ptlrpc_request *req)
 693 {
 694         struct ldlm_request *dlm_req;
 695         struct ldlm_reply *dlm_rep;
 696         struct ldlm_lock *lock;
 697         int rc, size = sizeof(*dlm_rep);
 698         ENTRY;
 699
 700         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
 701                                       lustre_swab_ldlm_request);
 702         if (dlm_req == NULL) {
 703                 CERROR ("Can't unpack dlm_req\n");
 704                 RETURN (-EFAULT);
 705         }
 706
 707         rc = lustre_pack_reply(req, 1, &size, NULL);
 708         if (rc) {
 709                 CERROR("out of memory\n");
 710                 RETURN(-ENOMEM);
 711         }
 712         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 713         dlm_rep->lock_flags = dlm_req->lock_flags;
 714
 715         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 716         if (!lock) {
 717                 req->rq_status = EINVAL;
 718         } else {
 719                 LDLM_DEBUG(lock, "server-side convert handler START");
 720                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 721                                   &dlm_rep->lock_flags);
 722                 if (ldlm_del_waiting_lock(lock))
 723                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 724                 req->rq_status = 0;
 725         }
 726
 727         if (lock) {
 728                 ldlm_reprocess_all(lock->l_resource);
 729                 LDLM_DEBUG(lock, "server-side convert handler END");
 730                 LDLM_LOCK_PUT(lock);
 731         } else
 732                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 733
 734         RETURN(0);
 735 }
 736
 737 int ldlm_handle_cancel(struct ptlrpc_request *req)
 738 {
 739         struct ldlm_request *dlm_req;
 740         struct ldlm_lock *lock;
 741         struct ldlm_resource *res;
 742         char str[PTL_NALFMT_SIZE];
 743         int rc;
 744         ENTRY;
 745
 746         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 747                                       lustre_swab_ldlm_request);
 748         if (dlm_req == NULL) {
 749                 CERROR("bad request buffer for cancel\n");
 750                 RETURN(-EFAULT);
 751         }
 752
 753         rc = lustre_pack_reply(req, 0, NULL, NULL);
 754         if (rc) {
 755                 CERROR("out of memory\n");
 756                 RETURN(-ENOMEM);
 757         }
 758
 759         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 760         if (!lock) {
 761                 CERROR("received cancel for unknown lock cookie "LPX64
 762                        " from client %s nid "LPX64" (%s)\n",
 763                        dlm_req->lock_handle1.cookie,
 764                        req->rq_export->exp_client_uuid.uuid,
 765                        req->rq_peer.peer_nid,
 766                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 767                                        req->rq_peer.peer_nid, str));
 768                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 769                                   "(cookie "LPU64")",
 770                                   dlm_req->lock_handle1.cookie);
 771                 req->rq_status = ESTALE;
 772         } else {
 773                 LDLM_DEBUG(lock, "server-side cancel handler START");
 774                 res = lock->l_resource;
 775                 if (res && res->lr_namespace->ns_lvbo &&
 776                     res->lr_namespace->ns_lvbo->lvbo_update) {
 777                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 778                                 (res, NULL, 0, 0);
 779                                 //(res, req->rq_reqmsg, 1);
 780                 }
 781
 782                 ldlm_lock_cancel(lock);
 783                 if (ldlm_del_waiting_lock(lock))
 784                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 785                 req->rq_status = rc;
 786         }
 787
 788         if (ptlrpc_reply(req) != 0)
 789                 LBUG();
 790
 791         if (lock) {
 792                 ldlm_reprocess_all(lock->l_resource);
 793                 LDLM_DEBUG(lock, "server-side cancel handler END");
 794                 LDLM_LOCK_PUT(lock);
 795         }
 796
 797         RETURN(0);
 798 }
 799
 800 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 801                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 802 {
 803         int do_ast;
 804         ENTRY;
 805
 806         l_lock(&ns->ns_lock);
 807         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 808
 809         lock->l_flags |= LDLM_FL_CBPENDING;
 810         do_ast = (!lock->l_readers && !lock->l_writers);
 811
 812         if (do_ast) {
 813                 LDLM_DEBUG(lock, "already unused, calling "
 814                            "callback (%p)", lock->l_blocking_ast);
 815                 if (lock->l_blocking_ast != NULL) {
 816                         l_unlock(&ns->ns_lock);
 817                         l_check_no_ns_lock(ns);
 818                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 819                                              LDLM_CB_BLOCKING);
 820                         l_lock(&ns->ns_lock);
 821                 }
 822         } else {
 823                 LDLM_DEBUG(lock, "Lock still has references, will be"
 824                            " cancelled later");
 825         }
 826
 827         LDLM_DEBUG(lock, "client blocking callback handler END");
 828         l_unlock(&ns->ns_lock);
 829         LDLM_LOCK_PUT(lock);
 830         EXIT;
 831 }
 832
 833 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 834                                     struct ldlm_namespace *ns,
 835                                     struct ldlm_request *dlm_req,
 836                                     struct ldlm_lock *lock)
 837 {
 838         LIST_HEAD(ast_list);
 839         ENTRY;
 840
 841         l_lock(&ns->ns_lock);
 842         LDLM_DEBUG(lock, "client completion callback handler START");
 843
 844         /* If we receive the completion AST before the actual enqueue returned,
 845          * then we might need to switch lock modes, resources, or extents. */
 846         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 847                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 848                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 849         }
 850
 851         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 852                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 853                        sizeof(lock->l_policy_data));
 854                 LDLM_DEBUG(lock, "completion AST, new policy data");
 855         }
 856
 857         ldlm_resource_unlink_lock(lock);
 858         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 859                    &lock->l_resource->lr_name,
 860                    sizeof(lock->l_resource->lr_name)) != 0) {
 861                 ldlm_lock_change_resource(ns, lock,
 862                                          dlm_req->lock_desc.l_resource.lr_name);
 863                 LDLM_DEBUG(lock, "completion AST, new resource");
 864         }
 865
 866         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 867                 lock->l_flags |= LDLM_FL_CBPENDING;
 868                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 869         }
 870
 871         if (lock->l_lvb_len) {
 872                 void *lvb;
 873                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 874                                          lock->l_lvb_swabber);
 875                 if (lvb == NULL) {
 876                         LDLM_ERROR(lock, "completion AST did not contain "
 877                                    "expected LVB!");
 878                 } else {
 879                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 880                 }
 881         }
 882
 883         lock->l_resource->lr_tmp = &ast_list;
 884         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 885         lock->l_resource->lr_tmp = NULL;
 886         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 887         l_unlock(&ns->ns_lock);
 888         LDLM_LOCK_PUT(lock);
 889
 890         ldlm_run_ast_work(ns, &ast_list);
 891
 892         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 893                           lock);
 894         EXIT;
 895 }
 896
 897 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 898                                     struct ldlm_namespace *ns,
 899                                     struct ldlm_request *dlm_req,
 900                                     struct ldlm_lock *lock)
 901 {
 902         int rc = -ENOSYS;
 903         ENTRY;
 904
 905         l_lock(&ns->ns_lock);
 906         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 907
 908         if (lock->l_glimpse_ast != NULL) {
 909                 l_unlock(&ns->ns_lock);
 910                 l_check_no_ns_lock(ns);
 911                 rc = lock->l_glimpse_ast(lock, req);
 912                 l_lock(&ns->ns_lock);
 913         }
 914
 915         if (req->rq_repmsg != NULL) {
 916                 ptlrpc_reply(req);
 917         } else {
 918                 req->rq_status = rc;
 919                 ptlrpc_error(req);
 920         }
 921
 922         if (lock->l_granted_mode == LCK_PW &&
 923             !lock->l_readers && !lock->l_writers &&
 924             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
 925                 l_unlock(&ns->ns_lock);
 926                 ldlm_handle_bl_callback(ns, NULL, lock);
 927                 EXIT;
 928                 return;
 929         }
 930
 931         l_unlock(&ns->ns_lock);
 932         LDLM_LOCK_PUT(lock);
 933         EXIT;
 934 }
 935
 936 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 937 {
 938         req->rq_status = rc;
 939         if (req->rq_reply_state == NULL) {
 940                 rc = lustre_pack_reply(req, 0, NULL, NULL);
 941                 if (rc)
 942                         return rc;
 943         }
 944         return ptlrpc_reply(req);
 945 }
 946
 947 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
 948                       struct ldlm_lock *lock)
 949 {
 950 #ifdef __KERNEL__
 951         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
 952         struct ldlm_bl_work_item *blwi;
 953         ENTRY;
 954
 955         OBD_ALLOC(blwi, sizeof(*blwi));
 956         if (blwi == NULL)
 957                 RETURN(-ENOMEM);
 958
 959         blwi->blwi_ns = ns;
 960         if (ld != NULL)
 961                 blwi->blwi_ld = *ld;
 962         blwi->blwi_lock = lock;
 963
 964         spin_lock(&blp->blp_lock);
 965         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
 966         wake_up(&blp->blp_waitq);
 967         spin_unlock(&blp->blp_lock);
 968 #else
 969         LBUG();
 970 #endif
 971
 972         RETURN(0);
 973 }
 974
 975 static int ldlm_callback_handler(struct ptlrpc_request *req)
 976 {
 977         struct ldlm_namespace *ns;
 978         struct ldlm_request *dlm_req;
 979         struct ldlm_lock *lock;
 980         char str[PTL_NALFMT_SIZE];
 981         int rc;
 982         ENTRY;
 983
 984         /* Requests arrive in sender's byte order.  The ptlrpc service
 985          * handler has already checked and, if necessary, byte-swapped the
 986          * incoming request message body, but I am responsible for the
 987          * message buffers. */
 988
 989         if (req->rq_export == NULL) {
 990                 struct ldlm_request *dlm_req;
 991
 992                 CDEBUG(D_RPCTRACE, "operation %d from nid "LPX64" (%s) with bad "
 993                        "export cookie "LPX64" (ptl req %d/rep %d); this is "
 994                        "normal if this node rebooted with a lock held\n",
 995                        req->rq_reqmsg->opc, req->rq_peer.peer_nid,
 996                        portals_nid2str(req->rq_peer.peer_ni->pni_number,
 997                                        req->rq_peer.peer_nid, str),
 998                        req->rq_reqmsg->handle.cookie,
 999                        req->rq_request_portal, req->rq_reply_portal);
1000
1001                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1002                                              lustre_swab_ldlm_request);
1003                 if (dlm_req != NULL)
1004                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1005                                dlm_req->lock_handle1.cookie);
1006
1007                 ldlm_callback_reply(req, -ENOTCONN);
1008                 RETURN(0);
1009         }
1010
1011         LASSERT(req->rq_export != NULL);
1012         LASSERT(req->rq_export->exp_obd != NULL);
1013
1014         switch(req->rq_reqmsg->opc) {
1015         case LDLM_BL_CALLBACK:
1016                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1017                 break;
1018         case LDLM_CP_CALLBACK:
1019                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1020                 break;
1021         case LDLM_GL_CALLBACK:
1022                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1023                 break;
1024         case OBD_LOG_CANCEL:
1025                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1026                 rc = llog_origin_handle_cancel(req);
1027                 ldlm_callback_reply(req, rc);
1028                 RETURN(0);
1029         case LLOG_ORIGIN_HANDLE_CREATE:
1030                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1031                 rc = llog_origin_handle_create(req);
1032                 ldlm_callback_reply(req, rc);
1033                 RETURN(0);
1034         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1035                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1036                 rc = llog_origin_handle_next_block(req);
1037                 ldlm_callback_reply(req, rc);
1038                 RETURN(0);
1039         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1040                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1041                 rc = llog_origin_handle_read_header(req);
1042                 ldlm_callback_reply(req, rc);
1043                 RETURN(0);
1044         case LLOG_ORIGIN_HANDLE_CLOSE:
1045                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1046                 rc = llog_origin_handle_close(req);
1047                 ldlm_callback_reply(req, rc);
1048                 RETURN(0);
1049         default:
1050                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1051                 ldlm_callback_reply(req, -EPROTO);
1052                 RETURN(0);
1053         }
1054
1055         ns = req->rq_export->exp_obd->obd_namespace;
1056         LASSERT(ns != NULL);
1057
1058         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1059                                       lustre_swab_ldlm_request);
1060         if (dlm_req == NULL) {
1061                 CERROR ("can't unpack dlm_req\n");
1062                 ldlm_callback_reply (req, -EPROTO);
1063                 RETURN (0);
1064         }
1065
1066         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1067         if (!lock) {
1068                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
1069                        dlm_req->lock_handle1.cookie);
1070                 ldlm_callback_reply(req, -EINVAL);
1071                 RETURN(0);
1072         }
1073
1074         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1075         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1076
1077         /* We want the ost thread to get this reply so that it can respond
1078          * to ost requests (write cache writeback) that might be triggered
1079          * in the callback.
1080          *
1081          * But we'd also like to be able to indicate in the reply that we're
1082          * cancelling right now, because it's unused, or have an intent result
1083          * in the reply, so we might have to push the responsibility for sending
1084          * the reply down into the AST handlers, alas. */
1085
1086         switch (req->rq_reqmsg->opc) {
1087         case LDLM_BL_CALLBACK:
1088                 CDEBUG(D_INODE, "blocking ast\n");
1089 #ifdef __KERNEL__
1090                 rc = ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock);
1091                 ldlm_callback_reply(req, rc);
1092 #else
1093                 rc = 0;
1094                 ldlm_callback_reply(req, rc);
1095                 ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1096 #endif
1097                 break;
1098         case LDLM_CP_CALLBACK:
1099                 CDEBUG(D_INODE, "completion ast\n");
1100                 ldlm_callback_reply(req, 0);
1101                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1102                 break;
1103         case LDLM_GL_CALLBACK:
1104                 CDEBUG(D_INODE, "glimpse ast\n");
1105                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1106                 break;
1107         default:
1108                 LBUG();                         /* checked above */
1109         }
1110
1111         RETURN(0);
1112 }
1113
1114 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1115 {
1116         int rc;
1117         ENTRY;
1118
1119         /* Requests arrive in sender's byte order.  The ptlrpc service
1120          * handler has already checked and, if necessary, byte-swapped the
1121          * incoming request message body, but I am responsible for the
1122          * message buffers. */
1123
1124         if (req->rq_export == NULL) {
1125                 struct ldlm_request *dlm_req;
1126                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
1127                        req->rq_reqmsg->opc, req->rq_request_portal,
1128                        req->rq_reply_portal);
1129                 CERROR("--> export cookie: "LPX64"\n",
1130                        req->rq_reqmsg->handle.cookie);
1131                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1132                                              lustre_swab_ldlm_request);
1133                 if (dlm_req != NULL)
1134                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1135                 RETURN(-ENOTCONN);
1136         }
1137
1138         switch (req->rq_reqmsg->opc) {
1139
1140         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1141         case LDLM_CANCEL:
1142                 CDEBUG(D_INODE, "cancel\n");
1143                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1144                 rc = ldlm_handle_cancel(req);
1145                 if (rc)
1146                         break;
1147                 RETURN(0);
1148
1149         default:
1150                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1151                 RETURN(-EINVAL);
1152         }
1153
1154         RETURN(0);
1155 }
1156
1157 #ifdef __KERNEL__
1158 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1159 {
1160         struct ldlm_bl_work_item *blwi = NULL;
1161
1162         spin_lock(&blp->blp_lock);
1163         if (!list_empty(&blp->blp_list)) {
1164                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1165                                   blwi_entry);
1166                 list_del(&blwi->blwi_entry);
1167         }
1168         spin_unlock(&blp->blp_lock);
1169
1170         return blwi;
1171 }
1172
1173 struct ldlm_bl_thread_data {
1174         int                     bltd_num;
1175         struct ldlm_bl_pool     *bltd_blp;
1176 };
1177
1178 static int ldlm_bl_thread_main(void *arg)
1179 {
1180         struct ldlm_bl_thread_data *bltd = arg;
1181         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1182         unsigned long flags;
1183         ENTRY;
1184
1185         /* XXX boiler-plate */
1186         {
1187                 char name[sizeof(current->comm)];
1188                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1189                          bltd->bltd_num);
1190                 kportal_daemonize(name);
1191         }
1192         SIGNAL_MASK_LOCK(current, flags);
1193         sigfillset(&current->blocked);
1194         RECALC_SIGPENDING;
1195         SIGNAL_MASK_UNLOCK(current, flags);
1196
1197         atomic_inc(&blp->blp_num_threads);
1198         complete(&blp->blp_comp);
1199
1200         while(1) {
1201                 struct l_wait_info lwi = { 0 };
1202                 struct ldlm_bl_work_item *blwi = NULL;
1203
1204                 l_wait_event_exclusive(blp->blp_waitq,
1205                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1206                                        &lwi);
1207
1208                 if (blwi->blwi_ns == NULL)
1209                         break;
1210
1211                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1212                                         blwi->blwi_lock);
1213                 OBD_FREE(blwi, sizeof(*blwi));
1214         }
1215
1216         atomic_dec(&blp->blp_num_threads);
1217         complete(&blp->blp_comp);
1218         RETURN(0);
1219 }
1220
1221 #endif
1222
1223 static int ldlm_setup(void);
1224 static int ldlm_cleanup(int force);
1225
1226 int ldlm_get_ref(void)
1227 {
1228         int rc = 0;
1229         down(&ldlm_ref_sem);
1230         if (++ldlm_refcount == 1) {
1231                 rc = ldlm_setup();
1232                 if (rc)
1233                         ldlm_refcount--;
1234         }
1235         up(&ldlm_ref_sem);
1236
1237         RETURN(rc);
1238 }
1239
1240 void ldlm_put_ref(int force)
1241 {
1242         down(&ldlm_ref_sem);
1243         if (ldlm_refcount == 1) {
1244                 int rc = ldlm_cleanup(force);
1245                 if (rc)
1246                         CERROR("ldlm_cleanup failed: %d\n", rc);
1247                 else
1248                         ldlm_refcount--;
1249         } else {
1250                 ldlm_refcount--;
1251         }
1252         up(&ldlm_ref_sem);
1253
1254         EXIT;
1255 }
1256
1257 static int ldlm_setup(void)
1258 {
1259         struct ldlm_bl_pool *blp;
1260         int rc = 0;
1261 #ifdef __KERNEL__
1262         int i;
1263 #endif
1264         ENTRY;
1265
1266         if (ldlm_state != NULL)
1267                 RETURN(-EALREADY);
1268
1269         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1270         if (ldlm_state == NULL)
1271                 RETURN(-ENOMEM);
1272
1273 #ifdef __KERNEL__
1274         rc = ldlm_proc_setup();
1275         if (rc != 0)
1276                 GOTO(out_free, rc);
1277 #endif
1278
1279         ldlm_state->ldlm_cb_service =
1280                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1281                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1282                                 ldlm_callback_handler, "ldlm_cbd",
1283                                 ldlm_svc_proc_dir);
1284
1285         if (!ldlm_state->ldlm_cb_service) {
1286                 CERROR("failed to start service\n");
1287                 GOTO(out_proc, rc = -ENOMEM);
1288         }
1289
1290         ldlm_state->ldlm_cancel_service =
1291                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1292                                 LDLM_CANCEL_REQUEST_PORTAL,
1293                                 LDLM_CANCEL_REPLY_PORTAL,
1294                                 ldlm_cancel_handler, "ldlm_canceld",
1295                                 ldlm_svc_proc_dir);
1296
1297         if (!ldlm_state->ldlm_cancel_service) {
1298                 CERROR("failed to start service\n");
1299                 GOTO(out_proc, rc = -ENOMEM);
1300         }
1301
1302         OBD_ALLOC(blp, sizeof(*blp));
1303         if (blp == NULL)
1304                 GOTO(out_proc, rc = -ENOMEM);
1305         ldlm_state->ldlm_bl_pool = blp;
1306
1307         atomic_set(&blp->blp_num_threads, 0);
1308         init_waitqueue_head(&blp->blp_waitq);
1309         spin_lock_init(&blp->blp_lock);
1310
1311         INIT_LIST_HEAD(&blp->blp_list);
1312
1313 #ifdef __KERNEL__
1314         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1315                 struct ldlm_bl_thread_data bltd = {
1316                         .bltd_num = i,
1317                         .bltd_blp = blp,
1318                 };
1319                 init_completion(&blp->blp_comp);
1320                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1321                 if (rc < 0) {
1322                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1323                         LBUG();
1324                         GOTO(out_thread, rc);
1325                 }
1326                 wait_for_completion(&blp->blp_comp);
1327         }
1328
1329         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1330                                     LDLM_NUM_THREADS, "ldlm_cn");
1331         if (rc) {
1332                 LBUG();
1333                 GOTO(out_thread, rc);
1334         }
1335
1336         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1337                                     LDLM_NUM_THREADS, "ldlm_cb");
1338         if (rc) {
1339                 LBUG();
1340                 GOTO(out_thread, rc);
1341         }
1342
1343         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1344         spin_lock_init(&expired_lock_thread.elt_lock);
1345         expired_lock_thread.elt_state = ELT_STOPPED;
1346         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1347
1348         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1349         if (rc < 0) {
1350                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1351                 GOTO(out_thread, rc);
1352         }
1353
1354         wait_event(expired_lock_thread.elt_waitq,
1355                    expired_lock_thread.elt_state == ELT_READY);
1356
1357         INIT_LIST_HEAD(&waiting_locks_list);
1358         spin_lock_init(&waiting_locks_spinlock);
1359         waiting_locks_timer.function = waiting_locks_callback;
1360         waiting_locks_timer.data = 0;
1361         init_timer(&waiting_locks_timer);
1362 #endif
1363
1364         RETURN(0);
1365
1366 #ifdef __KERNEL__
1367  out_thread:
1368         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1369         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1370 #endif
1371
1372  out_proc:
1373 #ifdef __KERNEL__
1374         ldlm_proc_cleanup();
1375  out_free:
1376 #endif
1377         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1378         ldlm_state = NULL;
1379         return rc;
1380 }
1381
1382 static int ldlm_cleanup(int force)
1383 {
1384 #ifdef __KERNEL__
1385         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1386 #endif
1387         ENTRY;
1388
1389         if (!list_empty(&ldlm_namespace_list)) {
1390                 CERROR("ldlm still has namespaces; clean these up first.\n");
1391                 ldlm_dump_all_namespaces();
1392                 RETURN(-EBUSY);
1393         }
1394
1395 #ifdef __KERNEL__
1396         while (atomic_read(&blp->blp_num_threads) > 0) {
1397                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1398
1399                 init_completion(&blp->blp_comp);
1400
1401                 spin_lock(&blp->blp_lock);
1402                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1403                 wake_up(&blp->blp_waitq);
1404                 spin_unlock(&blp->blp_lock);
1405
1406                 wait_for_completion(&blp->blp_comp);
1407         }
1408         OBD_FREE(blp, sizeof(*blp));
1409
1410         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1411         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1412         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1413         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1414         ldlm_proc_cleanup();
1415
1416         expired_lock_thread.elt_state = ELT_TERMINATE;
1417         wake_up(&expired_lock_thread.elt_waitq);
1418         wait_event(expired_lock_thread.elt_waitq,
1419                    expired_lock_thread.elt_state == ELT_STOPPED);
1420
1421 #endif
1422
1423         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1424         ldlm_state = NULL;
1425
1426         RETURN(0);
1427 }
1428
1429 int __init ldlm_init(void)
1430 {
1431         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1432                                                sizeof(struct ldlm_resource), 0,
1433                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1434         if (ldlm_resource_slab == NULL)
1435                 return -ENOMEM;
1436
1437         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1438                                            sizeof(struct ldlm_lock), 0,
1439                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1440         if (ldlm_lock_slab == NULL) {
1441                 kmem_cache_destroy(ldlm_resource_slab);
1442                 return -ENOMEM;
1443         }
1444
1445         l_lock_init(&ldlm_handle_lock);
1446
1447         return 0;
1448 }
1449
1450 void __exit ldlm_exit(void)
1451 {
1452         if ( ldlm_refcount )
1453                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1454         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
1455                 CERROR("couldn't free ldlm resource slab\n");
1456         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
1457                 CERROR("couldn't free ldlm lock slab\n");
1458 }
1459
1460 /* ldlm_flock.c */
1461 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1462
1463 /* ldlm_extent.c */
1464 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1465
1466 /* ldlm_lock.c */
1467 EXPORT_SYMBOL(ldlm_get_processing_policy);
1468 EXPORT_SYMBOL(ldlm_lock2desc);
1469 EXPORT_SYMBOL(ldlm_register_intent);
1470 EXPORT_SYMBOL(ldlm_lockname);
1471 EXPORT_SYMBOL(ldlm_typename);
1472 EXPORT_SYMBOL(ldlm_lock2handle);
1473 EXPORT_SYMBOL(__ldlm_handle2lock);
1474 EXPORT_SYMBOL(ldlm_lock_get);
1475 EXPORT_SYMBOL(ldlm_lock_put);
1476 EXPORT_SYMBOL(ldlm_lock_match);
1477 EXPORT_SYMBOL(ldlm_lock_cancel);
1478 EXPORT_SYMBOL(ldlm_lock_addref);
1479 EXPORT_SYMBOL(ldlm_lock_decref);
1480 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1481 EXPORT_SYMBOL(ldlm_lock_change_resource);
1482 EXPORT_SYMBOL(ldlm_lock_set_data);
1483 EXPORT_SYMBOL(ldlm_it2str);
1484 EXPORT_SYMBOL(ldlm_lock_dump);
1485 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1486 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1487 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1488 EXPORT_SYMBOL(ldlm_lock_allow_match);
1489
1490 /* ldlm_request.c */
1491 EXPORT_SYMBOL(ldlm_completion_ast);
1492 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1493 EXPORT_SYMBOL(ldlm_cli_convert);
1494 EXPORT_SYMBOL(ldlm_cli_enqueue);
1495 EXPORT_SYMBOL(ldlm_cli_cancel);
1496 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1497 EXPORT_SYMBOL(ldlm_replay_locks);
1498 EXPORT_SYMBOL(ldlm_resource_foreach);
1499 EXPORT_SYMBOL(ldlm_namespace_foreach);
1500 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1501 EXPORT_SYMBOL(ldlm_change_cbdata);
1502
1503 /* ldlm_lockd.c */
1504 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1505 EXPORT_SYMBOL(ldlm_server_completion_ast);
1506 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1507 EXPORT_SYMBOL(ldlm_handle_enqueue);
1508 EXPORT_SYMBOL(ldlm_handle_cancel);
1509 EXPORT_SYMBOL(ldlm_handle_convert);
1510 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1511 EXPORT_SYMBOL(ldlm_get_ref);
1512 EXPORT_SYMBOL(ldlm_put_ref);
1513
1514 #if 0
1515 /* ldlm_test.c */
1516 EXPORT_SYMBOL(ldlm_test);
1517 EXPORT_SYMBOL(ldlm_regression_start);
1518 EXPORT_SYMBOL(ldlm_regression_stop);
1519 #endif
1520
1521 /* ldlm_resource.c */
1522 EXPORT_SYMBOL(ldlm_namespace_new);
1523 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1524 EXPORT_SYMBOL(ldlm_namespace_free);
1525 EXPORT_SYMBOL(ldlm_namespace_dump);
1526 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1527 EXPORT_SYMBOL(ldlm_resource_get);
1528 EXPORT_SYMBOL(ldlm_resource_putref);
1529
1530 /* l_lock.c */
1531 EXPORT_SYMBOL(l_lock);
1532 EXPORT_SYMBOL(l_unlock);
1533
1534 /* ldlm_lib.c */
1535 EXPORT_SYMBOL(client_obd_setup);
1536 EXPORT_SYMBOL(client_obd_cleanup);
1537 EXPORT_SYMBOL(client_connect_import);
1538 EXPORT_SYMBOL(client_disconnect_export);
1539 EXPORT_SYMBOL(target_abort_recovery);
1540 EXPORT_SYMBOL(target_handle_connect);
1541 EXPORT_SYMBOL(target_destroy_export);
1542 EXPORT_SYMBOL(target_cancel_recovery_timer);
1543 EXPORT_SYMBOL(target_send_reply);
1544 EXPORT_SYMBOL(target_queue_recovery_request);
1545 EXPORT_SYMBOL(target_handle_ping);
1546 EXPORT_SYMBOL(target_handle_disconnect);
1547 EXPORT_SYMBOL(target_queue_final_reply);