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         LASSERTF(lock->l_export->exp_obd->obd_recovering == 0,
 403                  "BUG 6063: lock collide during recovery");
 404
 405         LASSERT(lock);
 406
 407         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 408         if (lock->l_granted_mode != lock->l_req_mode) {
 409                 /* this blocking AST will be communicated as part of the
 410                  * completion AST instead */
 411                 LDLM_DEBUG(lock, "lock not granted, not sending blocking AST");
 412                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 413                 RETURN(0);
 414         }
 415
 416         if (lock->l_destroyed) {
 417                 /* What's the point? */
 418                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 419                 RETURN(0);
 420         }
 421
 422 #if 0
 423         if (LTIME_S(CURRENT_TIME) - lock->l_export->exp_last_request_time > 30){
 424                 ldlm_failed_ast(lock, -ETIMEDOUT, "Not-attempted blocking");
 425                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 426                 RETURN(-ETIMEDOUT);
 427         }
 428 #endif
 429
 430         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 431                               LUSTRE_DLM_VERSION, LDLM_BL_CALLBACK,
 432                               1, &size, NULL);
 433         if (req == NULL) {
 434                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 435                 RETURN(-ENOMEM);
 436         }
 437
 438         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 439         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 440                sizeof(body->lock_handle1));
 441         memcpy(&body->lock_desc, desc, sizeof(*desc));
 442         body->lock_flags |= (lock->l_flags & LDLM_AST_FLAGS);
 443
 444         LDLM_DEBUG(lock, "server preparing blocking AST");
 445         req->rq_replen = lustre_msg_size(0, NULL);
 446
 447         if (lock->l_granted_mode == lock->l_req_mode)
 448                 ldlm_add_waiting_lock(lock);
 449         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 450
 451         req->rq_send_state = LUSTRE_IMP_FULL;
 452         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 453         rc = ptlrpc_queue_wait(req);
 454         if (rc != 0)
 455                 rc = ldlm_handle_ast_error(lock, req, rc, "blocking");
 456
 457         ptlrpc_req_finished(req);
 458
 459         RETURN(rc);
 460 }
 461
 462 /* XXX copied from ptlrpc/service.c */
 463 static long timeval_sub(struct timeval *large, struct timeval *small)
 464 {
 465         return (large->tv_sec - small->tv_sec) * 1000000 +
 466                 (large->tv_usec - small->tv_usec);
 467 }
 468
 469 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 470 {
 471         struct ldlm_request *body;
 472         struct ptlrpc_request *req;
 473         struct timeval granted_time;
 474         long total_enqueue_wait;
 475         int rc = 0, size[2] = {sizeof(*body)}, buffers = 1;
 476         ENTRY;
 477
 478         LASSERT(lock != NULL);
 479
 480         do_gettimeofday(&granted_time);
 481         total_enqueue_wait = timeval_sub(&granted_time, &lock->l_enqueued_time);
 482
 483         if (total_enqueue_wait / 1000000 > obd_timeout)
 484                 LDLM_ERROR(lock, "enqueue wait took %ldus", total_enqueue_wait);
 485
 486         down(&lock->l_resource->lr_lvb_sem);
 487         if (lock->l_resource->lr_lvb_len) {
 488                 buffers = 2;
 489                 size[1] = lock->l_resource->lr_lvb_len;
 490         }
 491         up(&lock->l_resource->lr_lvb_sem);
 492
 493         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 494                               LUSTRE_DLM_VERSION, LDLM_CP_CALLBACK,
 495                               buffers, size, NULL);
 496         if (req == NULL)
 497                 RETURN(-ENOMEM);
 498
 499         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*body));
 500         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 501                sizeof(body->lock_handle1));
 502         body->lock_flags = flags;
 503         ldlm_lock2desc(lock, &body->lock_desc);
 504
 505         if (buffers == 2) {
 506                 void *lvb;
 507
 508                 down(&lock->l_resource->lr_lvb_sem);
 509                 lvb = lustre_msg_buf(req->rq_reqmsg, 1,
 510                                      lock->l_resource->lr_lvb_len);
 511
 512                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 513                        lock->l_resource->lr_lvb_len);
 514                 up(&lock->l_resource->lr_lvb_sem);
 515         }
 516
 517         LDLM_DEBUG(lock, "server preparing completion AST (after %ldus wait)",
 518                    total_enqueue_wait);
 519         req->rq_replen = lustre_msg_size(0, NULL);
 520
 521         req->rq_send_state = LUSTRE_IMP_FULL;
 522         req->rq_timeout = ldlm_timeout; /* timeout for initial AST reply */
 523
 524         /* We only send real blocking ASTs after the lock is granted */
 525         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 526         if (lock->l_flags & LDLM_FL_AST_SENT) {
 527                 body->lock_flags |= LDLM_FL_AST_SENT;
 528                 ldlm_add_waiting_lock(lock); /* start the lock-timeout clock */
 529         }
 530         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 531
 532         rc = ptlrpc_queue_wait(req);
 533         if (rc != 0)
 534                 rc = ldlm_handle_ast_error(lock, req, rc, "completion");
 535
 536         ptlrpc_req_finished(req);
 537
 538         RETURN(rc);
 539 }
 540
 541 int ldlm_server_glimpse_ast(struct ldlm_lock *lock, void *data)
 542 {
 543         struct ldlm_resource *res = lock->l_resource;
 544         struct ldlm_request *body;
 545         struct ptlrpc_request *req;
 546         int rc = 0, size = sizeof(*body);
 547         ENTRY;
 548
 549         LASSERT(lock != NULL);
 550
 551         req = ptlrpc_prep_req(lock->l_export->exp_imp_reverse,
 552                               LUSTRE_DLM_VERSION, LDLM_GL_CALLBACK,
 553                               1, &size, NULL);
 554         if (req == NULL)
 555                 RETURN(-ENOMEM);
 556
 557         body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
 558         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 559                sizeof(body->lock_handle1));
 560         ldlm_lock2desc(lock, &body->lock_desc);
 561
 562         down(&lock->l_resource->lr_lvb_sem);
 563         size = lock->l_resource->lr_lvb_len;
 564         up(&lock->l_resource->lr_lvb_sem);
 565         req->rq_replen = lustre_msg_size(1, &size);
 566
 567         req->rq_send_state = LUSTRE_IMP_FULL;
 568         req->rq_timeout = 2; /* 2 second timeout for initial AST reply */
 569
 570         rc = ptlrpc_queue_wait(req);
 571         if (rc == -ELDLM_NO_LOCK_DATA)
 572                 LDLM_DEBUG(lock, "lost race - client has a lock but no inode");
 573         else if (rc != 0)
 574                 rc = ldlm_handle_ast_error(lock, req, rc, "glimpse");
 575         else
 576                 rc = res->lr_namespace->ns_lvbo->lvbo_update
 577                         (res, req->rq_repmsg, 0, 1);
 578         ptlrpc_req_finished(req);
 579         RETURN(rc);
 580 }
 581
 582 static struct ldlm_lock *
 583 find_existing_lock(struct obd_export *exp, struct lustre_handle *remote_hdl)
 584 {
 585         struct obd_device *obd = exp->exp_obd;
 586         struct list_head *iter;
 587
 588         l_lock(&obd->obd_namespace->ns_lock);
 589         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
 590                 struct ldlm_lock *lock;
 591                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
 592                 if (lock->l_remote_handle.cookie == remote_hdl->cookie) {
 593                         LDLM_LOCK_GET(lock);
 594                         l_unlock(&obd->obd_namespace->ns_lock);
 595                         return lock;
 596                 }
 597         }
 598         l_unlock(&obd->obd_namespace->ns_lock);
 599         return NULL;
 600 }
 601
 602
 603 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 604                         ldlm_completion_callback completion_callback,
 605                         ldlm_blocking_callback blocking_callback,
 606                         ldlm_glimpse_callback glimpse_callback)
 607 {
 608         struct obd_device *obddev = req->rq_export->exp_obd;
 609         struct ldlm_reply *dlm_rep;
 610         struct ldlm_request *dlm_req;
 611         int rc = 0, size[2] = {sizeof(*dlm_rep)};
 612         __u32 flags;
 613         ldlm_error_t err = ELDLM_OK;
 614         struct ldlm_lock *lock = NULL;
 615         void *cookie = NULL;
 616         ENTRY;
 617
 618         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 619
 620         dlm_req = lustre_swab_reqbuf (req, MDS_REQ_INTENT_LOCKREQ_OFF,
 621                                       sizeof (*dlm_req),
 622                                       lustre_swab_ldlm_request);
 623         if (dlm_req == NULL) {
 624                 CERROR ("Can't unpack dlm_req\n");
 625                 GOTO(out, rc = -EFAULT);
 626         }
 627
 628         flags = dlm_req->lock_flags;
 629
 630         LASSERT(req->rq_export);
 631
 632         if (flags & LDLM_FL_REPLAY) {
 633                 lock = find_existing_lock(req->rq_export,
 634                                           &dlm_req->lock_handle1);
 635                 if (lock != NULL) {
 636                         DEBUG_REQ(D_HA, req, "found existing lock cookie "LPX64,
 637                                   lock->l_handle.h_cookie);
 638                         GOTO(existing_lock, rc = 0);
 639                 }
 640         }
 641
 642         /* The lock's callback data might be set in the policy function */
 643         lock = ldlm_lock_create(obddev->obd_namespace, &dlm_req->lock_handle2,
 644                                 dlm_req->lock_desc.l_resource.lr_name,
 645                                 dlm_req->lock_desc.l_resource.lr_type,
 646                                 dlm_req->lock_desc.l_req_mode,
 647                                 blocking_callback, completion_callback,
 648                                 glimpse_callback, NULL, 0);
 649         if (!lock)
 650                 GOTO(out, rc = -ENOMEM);
 651
 652         do_gettimeofday(&lock->l_enqueued_time);
 653         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 654                sizeof(lock->l_remote_handle));
 655         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 656
 657         LASSERT(req->rq_export);
 658         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_BLOCKED, obd_timeout * 2);
 659         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 660         if (req->rq_export->exp_failed) {
 661                 LDLM_ERROR(lock,"lock on destroyed export %p\n",req->rq_export);
 662                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 663                 GOTO(out, err = -ENOTCONN);
 664         }
 665         lock->l_export = class_export_get(req->rq_export);
 666
 667         list_add(&lock->l_export_chain,
 668                  &lock->l_export->exp_ldlm_data.led_held_locks);
 669         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 670
 671 existing_lock:
 672
 673         if (flags & LDLM_FL_HAS_INTENT) {
 674                 /* In this case, the reply buffer is allocated deep in
 675                  * local_lock_enqueue by the policy function. */
 676                 cookie = req;
 677         } else {
 678                 int buffers = 1;
 679                 down(&lock->l_resource->lr_lvb_sem);
 680                 if (lock->l_resource->lr_lvb_len) {
 681                         size[1] = lock->l_resource->lr_lvb_len;
 682                         buffers = 2;
 683                 }
 684                 up(&lock->l_resource->lr_lvb_sem);
 685                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_LDLM_ENQUEUE_EXTENT_ERR))
 686                         GOTO(out, rc = -ENOMEM);
 687
 688                 rc = lustre_pack_reply(req, buffers, size, NULL);
 689                 if (rc)
 690                         GOTO(out, rc);
 691         }
 692
 693         if (dlm_req->lock_desc.l_resource.lr_type != LDLM_PLAIN)
 694                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 695                        sizeof(ldlm_policy_data_t));
 696         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 697                 memcpy(&lock->l_req_extent, &lock->l_policy_data.l_extent,
 698                        sizeof(lock->l_req_extent));
 699
 700         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, &flags);
 701         if (err)
 702                 GOTO(out, err);
 703
 704         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 705         dlm_rep->lock_flags = flags;
 706
 707         ldlm_lock2desc(lock, &dlm_rep->lock_desc);
 708         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 709
 710         /* We never send a blocking AST until the lock is granted, but
 711          * we can tell it right now */
 712         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 713         if (lock->l_flags & LDLM_FL_AST_SENT) {
 714                 dlm_rep->lock_flags |= LDLM_FL_AST_SENT;
 715                 if (lock->l_granted_mode == lock->l_req_mode)
 716                         ldlm_add_waiting_lock(lock);
 717         }
 718         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 719
 720         EXIT;
 721  out:
 722         req->rq_status = err;
 723         if (req->rq_reply_state == NULL) {
 724                 err = lustre_pack_reply(req, 0, NULL, NULL);
 725                 if (rc == 0)
 726                         rc = err;
 727                 req->rq_status = rc;
 728         }
 729
 730         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 731          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 732         if (lock) {
 733                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 734                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 735                            "(err=%d, rc=%d)", err, rc);
 736                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 737
 738                 if (rc == 0) {
 739                         down(&lock->l_resource->lr_lvb_sem);
 740                         size[1] = lock->l_resource->lr_lvb_len;
 741                         if (size[1] > 0) {
 742                                 void *lvb = lustre_msg_buf(req->rq_repmsg,
 743                                                            1, size[1]);
 744                                 LASSERTF(lvb != NULL, "req %p, lock %p\n",
 745                                          req, lock);
 746
 747                                 memcpy(lvb, lock->l_resource->lr_lvb_data,
 748                                        size[1]);
 749                         }
 750                         up(&lock->l_resource->lr_lvb_sem);
 751                 } else {
 752                         ldlm_lock_destroy(lock);
 753                 }
 754
 755                 if (!err && dlm_req->lock_desc.l_resource.lr_type != LDLM_FLOCK)
 756                         ldlm_reprocess_all(lock->l_resource);
 757                 LDLM_LOCK_PUT(lock);
 758         }
 759         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p, rc %d)",
 760                           lock, rc);
 761
 762         return rc;
 763 }
 764
 765 int ldlm_handle_convert(struct ptlrpc_request *req)
 766 {
 767         struct ldlm_request *dlm_req;
 768         struct ldlm_reply *dlm_rep;
 769         struct ldlm_lock *lock;
 770         int rc, size = sizeof(*dlm_rep);
 771         ENTRY;
 772
 773         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 774                                      lustre_swab_ldlm_request);
 775         if (dlm_req == NULL) {
 776                 CERROR ("Can't unpack dlm_req\n");
 777                 RETURN (-EFAULT);
 778         }
 779
 780         rc = lustre_pack_reply(req, 1, &size, NULL);
 781         if (rc) {
 782                 CERROR("out of memory\n");
 783                 RETURN(-ENOMEM);
 784         }
 785         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
 786         dlm_rep->lock_flags = dlm_req->lock_flags;
 787
 788         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 789         if (!lock) {
 790                 req->rq_status = EINVAL;
 791         } else {
 792                 void *res = NULL;
 793
 794                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 795                 LDLM_DEBUG(lock, "server-side convert handler START");
 796                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 797
 798                 res = ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 799                                         &dlm_rep->lock_flags);
 800                 if (res) {
 801                         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 802                         if (ldlm_del_waiting_lock(lock))
 803                                 CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 804                         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 805                         req->rq_status = 0;
 806                 } else {
 807                         req->rq_status = EDEADLOCK;
 808                 }
 809         }
 810
 811         if (lock) {
 812                 if (!req->rq_status)
 813                         ldlm_reprocess_all(lock->l_resource);
 814                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 815                 LDLM_DEBUG(lock, "server-side convert handler END");
 816                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 817                 LDLM_LOCK_PUT(lock);
 818         } else
 819                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 820
 821         RETURN(0);
 822 }
 823
 824 int ldlm_handle_cancel(struct ptlrpc_request *req)
 825 {
 826         struct ldlm_request *dlm_req;
 827         struct ldlm_lock *lock;
 828         struct ldlm_resource *res;
 829         int rc;
 830         ENTRY;
 831
 832         dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
 833                                       lustre_swab_ldlm_request);
 834         if (dlm_req == NULL) {
 835                 CERROR("bad request buffer for cancel\n");
 836                 RETURN(-EFAULT);
 837         }
 838
 839         rc = lustre_pack_reply(req, 0, NULL, NULL);
 840         if (rc) {
 841                 CERROR("out of memory\n");
 842                 RETURN(-ENOMEM);
 843         }
 844
 845         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 846         if (!lock) {
 847                 CERROR("received cancel for unknown lock cookie "LPX64
 848                        " from client %s id %s\n",
 849                        dlm_req->lock_handle1.cookie,
 850                        req->rq_export->exp_client_uuid.uuid,
 851                        req->rq_peerstr);
 852                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 853                                   "(cookie "LPU64")",
 854                                   dlm_req->lock_handle1.cookie);
 855                 req->rq_status = ESTALE;
 856         } else {
 857                 LDLM_DEBUG(lock, "server-side cancel handler START");
 858                 res = lock->l_resource;
 859                 if (res && res->lr_namespace->ns_lvbo &&
 860                     res->lr_namespace->ns_lvbo->lvbo_update) {
 861                         (void)res->lr_namespace->ns_lvbo->lvbo_update
 862                                 (res, NULL, 0, 0);
 863                                 //(res, req->rq_reqmsg, 1);
 864                 }
 865
 866                 l_lock(&res->lr_namespace->ns_lock);
 867                 ldlm_lock_cancel(lock);
 868                 if (ldlm_del_waiting_lock(lock))
 869                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 870                 l_unlock(&res->lr_namespace->ns_lock);
 871                 req->rq_status = rc;
 872         }
 873
 874         if (ptlrpc_reply(req) != 0)
 875                 LBUG();
 876
 877         if (lock) {
 878                 ldlm_reprocess_all(lock->l_resource);
 879                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 880                 LDLM_DEBUG(lock, "server-side cancel handler END");
 881                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 882                 LDLM_LOCK_PUT(lock);
 883         }
 884
 885         RETURN(0);
 886 }
 887
 888 void ldlm_handle_bl_callback(struct ldlm_namespace *ns,
 889                              struct ldlm_lock_desc *ld, struct ldlm_lock *lock)
 890 {
 891         int do_ast;
 892         ENTRY;
 893
 894         l_lock(&ns->ns_lock);
 895         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 896
 897         lock->l_flags |= LDLM_FL_CBPENDING;
 898         do_ast = (!lock->l_readers && !lock->l_writers);
 899
 900         if (do_ast) {
 901                 LDLM_DEBUG(lock, "already unused, calling "
 902                            "callback (%p)", lock->l_blocking_ast);
 903                 if (lock->l_blocking_ast != NULL) {
 904                         l_unlock(&ns->ns_lock);
 905                         l_check_no_ns_lock(ns);
 906                         lock->l_blocking_ast(lock, ld, lock->l_ast_data,
 907                                              LDLM_CB_BLOCKING);
 908                         l_lock(&ns->ns_lock);
 909                 }
 910         } else {
 911                 LDLM_DEBUG(lock, "Lock still has references, will be"
 912                            " cancelled later");
 913         }
 914
 915         LDLM_DEBUG(lock, "client blocking callback handler END");
 916         l_unlock(&ns->ns_lock);
 917         LDLM_LOCK_PUT(lock);
 918         EXIT;
 919 }
 920
 921 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 922                                     struct ldlm_namespace *ns,
 923                                     struct ldlm_request *dlm_req,
 924                                     struct ldlm_lock *lock)
 925 {
 926         LIST_HEAD(ast_list);
 927         ENTRY;
 928
 929         l_lock(&ns->ns_lock);
 930         LDLM_DEBUG(lock, "client completion callback handler START");
 931
 932         /* If we receive the completion AST before the actual enqueue returned,
 933          * then we might need to switch lock modes, resources, or extents. */
 934         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 935                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 936                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 937         }
 938
 939         if (lock->l_resource->lr_type != LDLM_PLAIN) {
 940                 memcpy(&lock->l_policy_data, &dlm_req->lock_desc.l_policy_data,
 941                        sizeof(lock->l_policy_data));
 942                 LDLM_DEBUG(lock, "completion AST, new policy data");
 943         }
 944
 945         ldlm_resource_unlink_lock(lock);
 946         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 947                    &lock->l_resource->lr_name,
 948                    sizeof(lock->l_resource->lr_name)) != 0) {
 949                 ldlm_lock_change_resource(ns, lock,
 950                                          dlm_req->lock_desc.l_resource.lr_name);
 951                 LDLM_DEBUG(lock, "completion AST, new resource");
 952         }
 953
 954         if (dlm_req->lock_flags & LDLM_FL_AST_SENT) {
 955                 lock->l_flags |= LDLM_FL_CBPENDING;
 956                 LDLM_DEBUG(lock, "completion AST includes blocking AST");
 957         }
 958
 959         if (lock->l_lvb_len) {
 960                 void *lvb;
 961                 lvb = lustre_swab_reqbuf(req, 1, lock->l_lvb_len,
 962                                          lock->l_lvb_swabber);
 963                 if (lvb == NULL) {
 964                         LDLM_ERROR(lock, "completion AST did not contain "
 965                                    "expected LVB!");
 966                 } else {
 967                         memcpy(lock->l_lvb_data, lvb, lock->l_lvb_len);
 968                 }
 969         }
 970
 971         lock->l_resource->lr_tmp = &ast_list;
 972         ldlm_grant_lock(lock, req, sizeof(*req), 1);
 973         lock->l_resource->lr_tmp = NULL;
 974         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 975         l_unlock(&ns->ns_lock);
 976         LDLM_LOCK_PUT(lock);
 977
 978         ldlm_run_ast_work(ns, &ast_list);
 979
 980         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 981                           lock);
 982         EXIT;
 983 }
 984
 985 static void ldlm_handle_gl_callback(struct ptlrpc_request *req,
 986                                     struct ldlm_namespace *ns,
 987                                     struct ldlm_request *dlm_req,
 988                                     struct ldlm_lock *lock)
 989 {
 990         int rc = -ENOSYS;
 991         ENTRY;
 992
 993         l_lock(&ns->ns_lock);
 994         LDLM_DEBUG(lock, "client glimpse AST callback handler");
 995
 996         if (lock->l_glimpse_ast != NULL) {
 997                 l_unlock(&ns->ns_lock);
 998                 l_check_no_ns_lock(ns);
 999                 rc = lock->l_glimpse_ast(lock, req);
1000                 l_lock(&ns->ns_lock);
1001         }
1002
1003         if (req->rq_repmsg != NULL) {
1004                 ptlrpc_reply(req);
1005         } else {
1006                 req->rq_status = rc;
1007                 ptlrpc_error(req);
1008         }
1009
1010         l_unlock(&ns->ns_lock);
1011         if (lock->l_granted_mode == LCK_PW &&
1012             !lock->l_readers && !lock->l_writers &&
1013             time_after(jiffies, lock->l_last_used + 10 * HZ)) {
1014                 if (ldlm_bl_to_thread(ns, NULL, lock))
1015                         ldlm_handle_bl_callback(ns, NULL, lock);
1016
1017                 EXIT;
1018                 return;
1019         }
1020         LDLM_LOCK_PUT(lock);
1021         EXIT;
1022 }
1023
1024 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
1025 {
1026         req->rq_status = rc;
1027         if (req->rq_reply_state == NULL) {
1028                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1029                 if (rc)
1030                         return rc;
1031         }
1032         return ptlrpc_reply(req);
1033 }
1034
1035 int ldlm_bl_to_thread(struct ldlm_namespace *ns, struct ldlm_lock_desc *ld,
1036                       struct ldlm_lock *lock)
1037 {
1038 #ifdef __KERNEL__
1039         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1040         struct ldlm_bl_work_item *blwi;
1041         ENTRY;
1042
1043         OBD_ALLOC(blwi, sizeof(*blwi));
1044         if (blwi == NULL)
1045                 RETURN(-ENOMEM);
1046
1047         blwi->blwi_ns = ns;
1048         if (ld != NULL)
1049                 blwi->blwi_ld = *ld;
1050         blwi->blwi_lock = lock;
1051
1052         spin_lock(&blp->blp_lock);
1053         list_add_tail(&blwi->blwi_entry, &blp->blp_list);
1054         wake_up(&blp->blp_waitq);
1055         spin_unlock(&blp->blp_lock);
1056
1057         RETURN(0);
1058 #else
1059         RETURN(-ENOSYS);
1060 #endif
1061
1062 }
1063
1064 static int ldlm_msg_check_version(struct lustre_msg *msg)
1065 {
1066         int rc;
1067
1068         switch (msg->opc) {
1069         case LDLM_ENQUEUE:
1070         case LDLM_CONVERT:
1071         case LDLM_CANCEL:
1072         case LDLM_BL_CALLBACK:
1073         case LDLM_CP_CALLBACK:
1074         case LDLM_GL_CALLBACK:
1075                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1076                 if (rc)
1077                         CERROR("bad opc %u version %08x, expecting %08x\n",
1078                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
1079                 break;
1080         case OBD_LOG_CANCEL:
1081         case LLOG_ORIGIN_HANDLE_OPEN:
1082         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1083         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1084         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1085         case LLOG_ORIGIN_HANDLE_CLOSE:
1086         case LLOG_CATINFO:
1087                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1088                 if (rc)
1089                         CERROR("bad opc %u version %08x, expecting %08x\n",
1090                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
1091                 break;
1092         default:
1093                 CERROR("LDLM unknown opcode %d\n", msg->opc);
1094                 rc = -ENOTSUPP;
1095                 break;
1096         }
1097
1098         return rc;
1099 }
1100
1101 static int ldlm_callback_handler(struct ptlrpc_request *req)
1102 {
1103         struct ldlm_namespace *ns;
1104         struct ldlm_request *dlm_req;
1105         struct ldlm_lock *lock;
1106         int rc;
1107         ENTRY;
1108
1109         rc = ldlm_msg_check_version(req->rq_reqmsg);
1110         if (rc) {
1111                 CERROR("LDLM_CB drop mal-formed request\n");
1112                 RETURN(rc);
1113         }
1114
1115         /* Requests arrive in sender's byte order.  The ptlrpc service
1116          * handler has already checked and, if necessary, byte-swapped the
1117          * incoming request message body, but I am responsible for the
1118          * message buffers. */
1119
1120         if (req->rq_export == NULL) {
1121                 struct ldlm_request *dlm_req;
1122
1123                 CDEBUG(D_RPCTRACE, "operation %d from %s with bad "
1124                        "export cookie "LPX64"; this is "
1125                        "normal if this node rebooted with a lock held\n",
1126                        req->rq_reqmsg->opc,
1127                        req->rq_peerstr,
1128                        req->rq_reqmsg->handle.cookie);
1129                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1130                                              lustre_swab_ldlm_request);
1131                 if (dlm_req != NULL)
1132                         CDEBUG(D_RPCTRACE, "--> lock cookie: "LPX64"\n",
1133                                dlm_req->lock_handle1.cookie);
1134
1135                 ldlm_callback_reply(req, -ENOTCONN);
1136                 RETURN(0);
1137         }
1138
1139         LASSERT(req->rq_export != NULL);
1140         LASSERT(req->rq_export->exp_obd != NULL);
1141
1142         switch(req->rq_reqmsg->opc) {
1143         case LDLM_BL_CALLBACK:
1144                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1145                 break;
1146         case LDLM_CP_CALLBACK:
1147                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
1148                 break;
1149         case LDLM_GL_CALLBACK:
1150                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_GL_CALLBACK, 0);
1151                 break;
1152         case OBD_LOG_CANCEL:
1153                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1154                 rc = llog_origin_handle_cancel(req);
1155                 ldlm_callback_reply(req, rc);
1156                 RETURN(0);
1157         case LLOG_ORIGIN_HANDLE_OPEN:
1158                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1159                 rc = llog_origin_handle_open(req);
1160                 ldlm_callback_reply(req, rc);
1161                 RETURN(0);
1162         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1163                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1164                 rc = llog_origin_handle_next_block(req);
1165                 ldlm_callback_reply(req, rc);
1166                 RETURN(0);
1167         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1168                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1169                 rc = llog_origin_handle_prev_block(req);
1170                 ldlm_callback_reply(req, rc);
1171                 RETURN(0);
1172         case LLOG_ORIGIN_HANDLE_READ_HEADER:
1173                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1174                 rc = llog_origin_handle_read_header(req);
1175                 ldlm_callback_reply(req, rc);
1176                 RETURN(0);
1177         case LLOG_ORIGIN_HANDLE_CLOSE:
1178                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1179                 rc = llog_origin_handle_close(req);
1180                 ldlm_callback_reply(req, rc);
1181                 RETURN(0);
1182         default:
1183                 CERROR("unknown opcode %u\n", req->rq_reqmsg->opc);
1184                 ldlm_callback_reply(req, -EPROTO);
1185                 RETURN(0);
1186         }
1187
1188         ns = req->rq_export->exp_obd->obd_namespace;
1189         LASSERT(ns != NULL);
1190
1191         dlm_req = lustre_swab_reqbuf (req, 0, sizeof (*dlm_req),
1192                                       lustre_swab_ldlm_request);
1193         if (dlm_req == NULL) {
1194                 CERROR ("can't unpack dlm_req\n");
1195                 ldlm_callback_reply (req, -EPROTO);
1196                 RETURN (0);
1197         }
1198
1199         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
1200         if (!lock) {
1201                 LDLM_DEBUG_NOLOCK("callback on lock "LPX64" - lock "
1202                                   "disappeared\n",dlm_req->lock_handle1.cookie);
1203                 ldlm_callback_reply(req, -EINVAL);
1204                 RETURN(0);
1205         }
1206
1207         /* Copy hints/flags (e.g. LDLM_FL_DISCARD_DATA) from AST. */
1208         lock->l_flags |= (dlm_req->lock_flags & LDLM_AST_FLAGS);
1209
1210         /* We want the ost thread to get this reply so that it can respond
1211          * to ost requests (write cache writeback) that might be triggered
1212          * in the callback.
1213          *
1214          * But we'd also like to be able to indicate in the reply that we're
1215          * cancelling right now, because it's unused, or have an intent result
1216          * in the reply, so we might have to push the responsibility for sending
1217          * the reply down into the AST handlers, alas. */
1218
1219         switch (req->rq_reqmsg->opc) {
1220         case LDLM_BL_CALLBACK:
1221                 CDEBUG(D_INODE, "blocking ast\n");
1222                 ldlm_callback_reply(req, 0);
1223                 if (ldlm_bl_to_thread(ns, &dlm_req->lock_desc, lock))
1224                         ldlm_handle_bl_callback(ns, &dlm_req->lock_desc, lock);
1225
1226                 break;
1227         case LDLM_CP_CALLBACK:
1228                 CDEBUG(D_INODE, "completion ast\n");
1229                 ldlm_callback_reply(req, 0);
1230                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
1231                 break;
1232         case LDLM_GL_CALLBACK:
1233                 CDEBUG(D_INODE, "glimpse ast\n");
1234                 ldlm_handle_gl_callback(req, ns, dlm_req, lock);
1235                 break;
1236         default:
1237                 LBUG();                         /* checked above */
1238         }
1239
1240         RETURN(0);
1241 }
1242
1243 static int ldlm_cancel_handler(struct ptlrpc_request *req)
1244 {
1245         int rc = 0;
1246         ENTRY;
1247
1248         rc = ldlm_msg_check_version(req->rq_reqmsg);
1249         if (rc) {
1250                 CERROR("LDLM_CL drop mal-formed request\n");
1251                 RETURN(rc);
1252         }
1253
1254         /* Requests arrive in sender's byte order.  The ptlrpc service
1255          * handler has already checked and, if necessary, byte-swapped the
1256          * incoming request message body, but I am responsible for the
1257          * message buffers. */
1258
1259         if (req->rq_export == NULL) {
1260                 struct ldlm_request *dlm_req;
1261                 CERROR("operation %d with bad export from %s\n",
1262                        req->rq_reqmsg->opc,
1263                        req->rq_peerstr);
1264                 CERROR("--> export cookie: "LPX64"\n",
1265                        req->rq_reqmsg->handle.cookie);
1266                 dlm_req = lustre_swab_reqbuf(req, 0, sizeof (*dlm_req),
1267                                              lustre_swab_ldlm_request);
1268                 if (dlm_req != NULL)
1269                         ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
1270                 ldlm_callback_reply(req, -ENOTCONN);
1271                 RETURN(0);
1272         }
1273
1274         switch (req->rq_reqmsg->opc) {
1275
1276         /* XXX FIXME move this back to mds/handler.c, bug 249 */
1277         case LDLM_CANCEL:
1278                 CDEBUG(D_INODE, "cancel\n");
1279                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
1280                 rc = ldlm_handle_cancel(req);
1281                 break;
1282         default:
1283                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
1284                 ldlm_callback_reply(req, -EINVAL);
1285         }
1286
1287         RETURN(rc);
1288 }
1289
1290 #ifdef __KERNEL__
1291 static struct ldlm_bl_work_item *ldlm_bl_get_work(struct ldlm_bl_pool *blp)
1292 {
1293         struct ldlm_bl_work_item *blwi = NULL;
1294
1295         spin_lock(&blp->blp_lock);
1296         if (!list_empty(&blp->blp_list)) {
1297                 blwi = list_entry(blp->blp_list.next, struct ldlm_bl_work_item,
1298                                   blwi_entry);
1299                 list_del(&blwi->blwi_entry);
1300         }
1301         spin_unlock(&blp->blp_lock);
1302
1303         return blwi;
1304 }
1305
1306 struct ldlm_bl_thread_data {
1307         int                     bltd_num;
1308         struct ldlm_bl_pool     *bltd_blp;
1309 };
1310
1311 static int ldlm_bl_thread_main(void *arg)
1312 {
1313         struct ldlm_bl_thread_data *bltd = arg;
1314         struct ldlm_bl_pool *blp = bltd->bltd_blp;
1315         unsigned long flags;
1316         ENTRY;
1317
1318         /* XXX boiler-plate */
1319         {
1320                 char name[sizeof(current->comm)];
1321                 snprintf(name, sizeof(name) - 1, "ldlm_bl_%02d",
1322                          bltd->bltd_num);
1323                 kportal_daemonize(name);
1324         }
1325         SIGNAL_MASK_LOCK(current, flags);
1326         sigfillset(&current->blocked);
1327         RECALC_SIGPENDING;
1328         SIGNAL_MASK_UNLOCK(current, flags);
1329
1330         atomic_inc(&blp->blp_num_threads);
1331         complete(&blp->blp_comp);
1332
1333         while(1) {
1334                 struct l_wait_info lwi = { 0 };
1335                 struct ldlm_bl_work_item *blwi = NULL;
1336
1337                 l_wait_event_exclusive(blp->blp_waitq,
1338                                        (blwi = ldlm_bl_get_work(blp)) != NULL,
1339                                        &lwi);
1340
1341                 if (blwi->blwi_ns == NULL)
1342                         break;
1343
1344                 ldlm_handle_bl_callback(blwi->blwi_ns, &blwi->blwi_ld,
1345                                         blwi->blwi_lock);
1346                 OBD_FREE(blwi, sizeof(*blwi));
1347         }
1348
1349         atomic_dec(&blp->blp_num_threads);
1350         complete(&blp->blp_comp);
1351         RETURN(0);
1352 }
1353
1354 #endif
1355
1356 static int ldlm_setup(void);
1357 static int ldlm_cleanup(int force);
1358
1359 int ldlm_get_ref(void)
1360 {
1361         int rc = 0;
1362         down(&ldlm_ref_sem);
1363         if (++ldlm_refcount == 1) {
1364                 rc = ldlm_setup();
1365                 if (rc)
1366                         ldlm_refcount--;
1367         }
1368         up(&ldlm_ref_sem);
1369
1370         RETURN(rc);
1371 }
1372
1373 void ldlm_put_ref(int force)
1374 {
1375         down(&ldlm_ref_sem);
1376         if (ldlm_refcount == 1) {
1377                 int rc = ldlm_cleanup(force);
1378                 if (rc)
1379                         CERROR("ldlm_cleanup failed: %d\n", rc);
1380                 else
1381                         ldlm_refcount--;
1382         } else {
1383                 ldlm_refcount--;
1384         }
1385         up(&ldlm_ref_sem);
1386
1387         EXIT;
1388 }
1389
1390 static int ldlm_setup(void)
1391 {
1392         struct ldlm_bl_pool *blp;
1393         int rc = 0;
1394 #ifdef __KERNEL__
1395         int i;
1396 #endif
1397         ENTRY;
1398
1399         if (ldlm_state != NULL)
1400                 RETURN(-EALREADY);
1401
1402         OBD_ALLOC(ldlm_state, sizeof(*ldlm_state));
1403         if (ldlm_state == NULL)
1404                 RETURN(-ENOMEM);
1405
1406 #ifdef __KERNEL__
1407         rc = ldlm_proc_setup();
1408         if (rc != 0)
1409                 GOTO(out_free, rc);
1410 #endif
1411
1412         ldlm_state->ldlm_cb_service =
1413                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1414                                 LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1415                                 1500, ldlm_callback_handler, "ldlm_cbd",
1416                                 ldlm_svc_proc_dir);
1417
1418         if (!ldlm_state->ldlm_cb_service) {
1419                 CERROR("failed to start service\n");
1420                 GOTO(out_proc, rc = -ENOMEM);
1421         }
1422
1423         ldlm_state->ldlm_cancel_service =
1424                 ptlrpc_init_svc(LDLM_NBUFS, LDLM_BUFSIZE, LDLM_MAXREQSIZE,
1425                                 LDLM_CANCEL_REQUEST_PORTAL,
1426                                 LDLM_CANCEL_REPLY_PORTAL, 30000,
1427                                 ldlm_cancel_handler, "ldlm_canceld",
1428                                 ldlm_svc_proc_dir);
1429
1430         if (!ldlm_state->ldlm_cancel_service) {
1431                 CERROR("failed to start service\n");
1432                 GOTO(out_proc, rc = -ENOMEM);
1433         }
1434
1435         OBD_ALLOC(blp, sizeof(*blp));
1436         if (blp == NULL)
1437                 GOTO(out_proc, rc = -ENOMEM);
1438         ldlm_state->ldlm_bl_pool = blp;
1439
1440         atomic_set(&blp->blp_num_threads, 0);
1441         init_waitqueue_head(&blp->blp_waitq);
1442         spin_lock_init(&blp->blp_lock);
1443
1444         INIT_LIST_HEAD(&blp->blp_list);
1445
1446 #ifdef __KERNEL__
1447         for (i = 0; i < LDLM_NUM_THREADS; i++) {
1448                 struct ldlm_bl_thread_data bltd = {
1449                         .bltd_num = i,
1450                         .bltd_blp = blp,
1451                 };
1452                 init_completion(&blp->blp_comp);
1453                 rc = kernel_thread(ldlm_bl_thread_main, &bltd, 0);
1454                 if (rc < 0) {
1455                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
1456                         GOTO(out_thread, rc);
1457                 }
1458                 wait_for_completion(&blp->blp_comp);
1459         }
1460
1461         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cancel_service,
1462                                     LDLM_NUM_THREADS, "ldlm_cn");
1463         if (rc)
1464                 GOTO(out_thread, rc);
1465
1466         rc = ptlrpc_start_n_threads(NULL, ldlm_state->ldlm_cb_service,
1467                                     LDLM_NUM_THREADS, "ldlm_cb");
1468         if (rc)
1469                 GOTO(out_thread, rc);
1470
1471         INIT_LIST_HEAD(&expired_lock_thread.elt_expired_locks);
1472         spin_lock_init(&expired_lock_thread.elt_lock);
1473         expired_lock_thread.elt_state = ELT_STOPPED;
1474         init_waitqueue_head(&expired_lock_thread.elt_waitq);
1475
1476         rc = kernel_thread(expired_lock_main, NULL, CLONE_VM | CLONE_FS);
1477         if (rc < 0) {
1478                 CERROR("Cannot start ldlm expired-lock thread: %d\n", rc);
1479                 GOTO(out_thread, rc);
1480         }
1481
1482         wait_event(expired_lock_thread.elt_waitq,
1483                    expired_lock_thread.elt_state == ELT_READY);
1484
1485         INIT_LIST_HEAD(&waiting_locks_list);
1486         spin_lock_init(&waiting_locks_spinlock);
1487         waiting_locks_timer.function = waiting_locks_callback;
1488         waiting_locks_timer.data = 0;
1489         init_timer(&waiting_locks_timer);
1490 #endif
1491
1492         RETURN(0);
1493
1494 #ifdef __KERNEL__
1495  out_thread:
1496         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1497         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1498 #endif
1499
1500  out_proc:
1501 #ifdef __KERNEL__
1502         ldlm_proc_cleanup();
1503  out_free:
1504 #endif
1505         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1506         ldlm_state = NULL;
1507         return rc;
1508 }
1509
1510 static int ldlm_cleanup(int force)
1511 {
1512 #ifdef __KERNEL__
1513         struct ldlm_bl_pool *blp = ldlm_state->ldlm_bl_pool;
1514 #endif
1515         ENTRY;
1516
1517         if (!list_empty(&ldlm_namespace_list)) {
1518                 CERROR("ldlm still has namespaces; clean these up first.\n");
1519                 ldlm_dump_all_namespaces(D_DLMTRACE);
1520                 RETURN(-EBUSY);
1521         }
1522
1523 #ifdef __KERNEL__
1524         while (atomic_read(&blp->blp_num_threads) > 0) {
1525                 struct ldlm_bl_work_item blwi = { .blwi_ns = NULL };
1526
1527                 init_completion(&blp->blp_comp);
1528
1529                 spin_lock(&blp->blp_lock);
1530                 list_add_tail(&blwi.blwi_entry, &blp->blp_list);
1531                 wake_up(&blp->blp_waitq);
1532                 spin_unlock(&blp->blp_lock);
1533
1534                 wait_for_completion(&blp->blp_comp);
1535         }
1536         OBD_FREE(blp, sizeof(*blp));
1537
1538         ptlrpc_stop_all_threads(ldlm_state->ldlm_cb_service);
1539         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1540         ptlrpc_stop_all_threads(ldlm_state->ldlm_cancel_service);
1541         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1542         ldlm_proc_cleanup();
1543
1544         expired_lock_thread.elt_state = ELT_TERMINATE;
1545         wake_up(&expired_lock_thread.elt_waitq);
1546         wait_event(expired_lock_thread.elt_waitq,
1547                    expired_lock_thread.elt_state == ELT_STOPPED);
1548 #else
1549         ptlrpc_unregister_service(ldlm_state->ldlm_cb_service);
1550         ptlrpc_unregister_service(ldlm_state->ldlm_cancel_service);
1551 #endif
1552
1553         OBD_FREE(ldlm_state, sizeof(*ldlm_state));
1554         ldlm_state = NULL;
1555
1556         RETURN(0);
1557 }
1558
1559 int __init ldlm_init(void)
1560 {
1561         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
1562                                                sizeof(struct ldlm_resource), 0,
1563                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
1564         if (ldlm_resource_slab == NULL)
1565                 return -ENOMEM;
1566
1567         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
1568                                            sizeof(struct ldlm_lock), 0,
1569                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
1570         if (ldlm_lock_slab == NULL) {
1571                 kmem_cache_destroy(ldlm_resource_slab);
1572                 return -ENOMEM;
1573         }
1574
1575         l_lock_init(&ldlm_handle_lock);
1576
1577         return 0;
1578 }
1579
1580 void __exit ldlm_exit(void)
1581 {
1582         if ( ldlm_refcount )
1583                 CERROR("ldlm_refcount is %d in ldlm_exit!\n", ldlm_refcount);
1584         LASSERTF(kmem_cache_destroy(ldlm_resource_slab) == 0,
1585                  "couldn't free ldlm resource slab\n");
1586         LASSERTF(kmem_cache_destroy(ldlm_lock_slab) == 0,
1587                  "couldn't free ldlm lock slab\n");
1588
1589 }
1590
1591 /* ldlm_flock.c */
1592 EXPORT_SYMBOL(ldlm_flock_completion_ast);
1593
1594 /* ldlm_extent.c */
1595 EXPORT_SYMBOL(ldlm_extent_shift_kms);
1596
1597 /* ldlm_lock.c */
1598 EXPORT_SYMBOL(ldlm_get_processing_policy);
1599 EXPORT_SYMBOL(ldlm_lock2desc);
1600 EXPORT_SYMBOL(ldlm_register_intent);
1601 EXPORT_SYMBOL(ldlm_lockname);
1602 EXPORT_SYMBOL(ldlm_typename);
1603 EXPORT_SYMBOL(ldlm_lock2handle);
1604 EXPORT_SYMBOL(__ldlm_handle2lock);
1605 EXPORT_SYMBOL(ldlm_lock_get);
1606 EXPORT_SYMBOL(ldlm_lock_put);
1607 EXPORT_SYMBOL(ldlm_lock_match);
1608 EXPORT_SYMBOL(ldlm_lock_cancel);
1609 EXPORT_SYMBOL(ldlm_lock_addref);
1610 EXPORT_SYMBOL(ldlm_lock_decref);
1611 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
1612 EXPORT_SYMBOL(ldlm_lock_change_resource);
1613 EXPORT_SYMBOL(ldlm_lock_set_data);
1614 EXPORT_SYMBOL(ldlm_it2str);
1615 EXPORT_SYMBOL(ldlm_lock_dump);
1616 EXPORT_SYMBOL(ldlm_lock_dump_handle);
1617 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
1618 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
1619 EXPORT_SYMBOL(ldlm_lock_allow_match);
1620
1621 /* ldlm_request.c */
1622 EXPORT_SYMBOL(ldlm_completion_ast);
1623 EXPORT_SYMBOL(ldlm_expired_completion_wait);
1624 EXPORT_SYMBOL(ldlm_cli_convert);
1625 EXPORT_SYMBOL(ldlm_cli_enqueue);
1626 EXPORT_SYMBOL(ldlm_cli_cancel);
1627 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
1628 EXPORT_SYMBOL(ldlm_replay_locks);
1629 EXPORT_SYMBOL(ldlm_resource_foreach);
1630 EXPORT_SYMBOL(ldlm_namespace_foreach);
1631 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
1632 EXPORT_SYMBOL(ldlm_change_cbdata);
1633
1634 /* ldlm_lockd.c */
1635 EXPORT_SYMBOL(ldlm_server_blocking_ast);
1636 EXPORT_SYMBOL(ldlm_server_completion_ast);
1637 EXPORT_SYMBOL(ldlm_server_glimpse_ast);
1638 EXPORT_SYMBOL(ldlm_handle_enqueue);
1639 EXPORT_SYMBOL(ldlm_handle_cancel);
1640 EXPORT_SYMBOL(ldlm_handle_convert);
1641 EXPORT_SYMBOL(ldlm_del_waiting_lock);
1642 EXPORT_SYMBOL(ldlm_get_ref);
1643 EXPORT_SYMBOL(ldlm_put_ref);
1644
1645 #if 0
1646 /* ldlm_test.c */
1647 EXPORT_SYMBOL(ldlm_test);
1648 EXPORT_SYMBOL(ldlm_regression_start);
1649 EXPORT_SYMBOL(ldlm_regression_stop);
1650 #endif
1651
1652 /* ldlm_resource.c */
1653 EXPORT_SYMBOL(ldlm_namespace_new);
1654 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1655 EXPORT_SYMBOL(ldlm_namespace_free);
1656 EXPORT_SYMBOL(ldlm_namespace_dump);
1657 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1658 EXPORT_SYMBOL(ldlm_resource_get);
1659 EXPORT_SYMBOL(ldlm_resource_putref);
1660
1661 /* l_lock.c */
1662 EXPORT_SYMBOL(l_lock);
1663 EXPORT_SYMBOL(l_unlock);
1664
1665 /* ldlm_lib.c */
1666 EXPORT_SYMBOL(client_import_add_conn);
1667 EXPORT_SYMBOL(client_import_del_conn);
1668 EXPORT_SYMBOL(client_obd_setup);
1669 EXPORT_SYMBOL(client_obd_cleanup);
1670 EXPORT_SYMBOL(client_connect_import);
1671 EXPORT_SYMBOL(client_disconnect_export);
1672 EXPORT_SYMBOL(target_start_recovery_thread);
1673 EXPORT_SYMBOL(target_stop_recovery_thread);
1674 EXPORT_SYMBOL(target_handle_connect);
1675 EXPORT_SYMBOL(target_cleanup_recovery);
1676 EXPORT_SYMBOL(target_destroy_export);
1677 EXPORT_SYMBOL(target_cancel_recovery_timer);
1678 EXPORT_SYMBOL(target_send_reply);
1679 EXPORT_SYMBOL(target_queue_recovery_request);
1680 EXPORT_SYMBOL(target_handle_ping);
1681 EXPORT_SYMBOL(target_handle_disconnect);