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