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 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 #define EXPORT_SYMTAB
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #include <linux/module.h>
  28 #include <linux/slab.h>
  29 #include <linux/lustre_dlm.h>
  30 #include <linux/init.h>
  31 #include <linux/obd_class.h>
  32
  33 extern kmem_cache_t *ldlm_resource_slab;
  34 extern kmem_cache_t *ldlm_lock_slab;
  35 extern struct list_head ldlm_namespace_list;
  36 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  37 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  38
  39 inline unsigned long round_timeout(unsigned long timeout)
  40 {
  41         return ((timeout / HZ) + 1) * HZ;
  42 }
  43
  44 static struct list_head waiting_locks_list;
  45 static spinlock_t waiting_locks_spinlock;
  46 static struct timer_list waiting_locks_timer;
  47 static int ldlm_already_setup = 0;
  48
  49 static void waiting_locks_callback(unsigned long unused)
  50 {
  51         struct list_head *liter, *n;
  52
  53         spin_lock_bh(&waiting_locks_spinlock);
  54         list_for_each_safe(liter, n, &waiting_locks_list) {
  55                 struct ldlm_lock *l = list_entry(liter, struct ldlm_lock,
  56                                                  l_pending_chain);
  57                 if (l->l_callback_timeout > jiffies)
  58                         break;
  59                 LDLM_DEBUG(l, "timer expired, recovering exp %p on conn %p",
  60                            l->l_export, l->l_export->exp_connection);
  61                 recovd_conn_fail(l->l_export->exp_connection);
  62         }
  63         spin_unlock_bh(&waiting_locks_spinlock);
  64 }
  65
  66 /*
  67  * Indicate that we're waiting for a client to call us back cancelling a given
  68  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
  69  * timer to fire appropriately.  (We round up to the next second, to avoid
  70  * floods of timer firings during periods of high lock contention and traffic).
  71  */
  72 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
  73 {
  74         unsigned long timeout_rounded;
  75         ENTRY;
  76
  77         LASSERT(list_empty(&lock->l_pending_chain));
  78
  79         spin_lock_bh(&waiting_locks_spinlock);
  80         lock->l_callback_timeout = jiffies + (obd_timeout * HZ);
  81
  82         timeout_rounded = round_timeout(lock->l_callback_timeout);
  83
  84         if (timeout_rounded < waiting_locks_timer.expires ||
  85             !timer_pending(&waiting_locks_timer)) {
  86                 mod_timer(&waiting_locks_timer, timeout_rounded);
  87         }
  88         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
  89         spin_unlock_bh(&waiting_locks_spinlock);
  90         RETURN(1);
  91 }
  92
  93 /*
  94  * Remove a lock from the pending list, likely because it had its cancellation
  95  * callback arrive without incident.  This adjusts the lock-timeout timer if
  96  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
  97  */
  98 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
  99 {
 100         struct list_head *list_next;
 101
 102         ENTRY;
 103
 104         spin_lock_bh(&waiting_locks_spinlock);
 105
 106         if (list_empty(&lock->l_pending_chain)) {
 107                 spin_unlock_bh(&waiting_locks_spinlock);
 108                 RETURN(0);
 109         }
 110
 111         list_next = lock->l_pending_chain.next;
 112         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 113                 /* Removing the head of the list, adjust timer. */
 114                 if (list_next == &waiting_locks_list) {
 115                         /* No more, just cancel. */
 116                         del_timer(&waiting_locks_timer);
 117                 } else {
 118                         struct ldlm_lock *next;
 119                         next = list_entry(list_next, struct ldlm_lock,
 120                                           l_pending_chain);
 121                         mod_timer(&waiting_locks_timer,
 122                                   round_timeout(next->l_callback_timeout));
 123                 }
 124         }
 125         list_del_init(&lock->l_pending_chain);
 126         spin_unlock_bh(&waiting_locks_spinlock);
 127         RETURN(1);
 128 }
 129
 130 static int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 131                                     struct ldlm_lock_desc *desc,
 132                                     void *data, __u32 data_len, int flag)
 133 {
 134         struct ldlm_request *body;
 135         struct ptlrpc_request *req;
 136         int rc = 0, size = sizeof(*body);
 137         ENTRY;
 138
 139         if (flag == LDLM_CB_CANCELING) {
 140                 /* Don't need to do anything here. */
 141                 RETURN(0);
 142         }
 143
 144         req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
 145                               LDLM_BL_CALLBACK, 1, &size, NULL);
 146         if (!req)
 147                 RETURN(-ENOMEM);
 148
 149         body = lustre_msg_buf(req->rq_reqmsg, 0);
 150         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 151                sizeof(body->lock_handle1));
 152         memcpy(&body->lock_desc, desc, sizeof(*desc));
 153
 154         LDLM_DEBUG(lock, "server preparing blocking AST");
 155         req->rq_replen = 0; /* no reply needed */
 156
 157         ldlm_add_waiting_lock(lock);
 158         (void)ptl_send_rpc(req);
 159
 160         /* no commit, and no waiting for reply, so 2x decref now */
 161         ptlrpc_req_finished(req);
 162         ptlrpc_req_finished(req);
 163
 164         RETURN(rc);
 165 }
 166
 167 static int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags)
 168 {
 169         struct ldlm_request *body;
 170         struct ptlrpc_request *req;
 171         int rc = 0, size = sizeof(*body);
 172         ENTRY;
 173
 174         if (lock == NULL) {
 175                 LBUG();
 176                 RETURN(-EINVAL);
 177         }
 178
 179         req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
 180                               LDLM_CP_CALLBACK, 1, &size, NULL);
 181         if (!req)
 182                 RETURN(-ENOMEM);
 183
 184         body = lustre_msg_buf(req->rq_reqmsg, 0);
 185         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 186                sizeof(body->lock_handle1));
 187         body->lock_flags = flags;
 188         ldlm_lock2desc(lock, &body->lock_desc);
 189
 190         LDLM_DEBUG(lock, "server preparing completion AST");
 191         req->rq_replen = 0; /* no reply needed */
 192
 193         (void)ptl_send_rpc(req);
 194         /* no commit, and no waiting for reply, so 2x decref now */
 195         ptlrpc_req_finished(req);
 196         ptlrpc_req_finished(req);
 197
 198         RETURN(rc);
 199 }
 200
 201 int ldlm_handle_enqueue(struct ptlrpc_request *req)
 202 {
 203         struct obd_device *obddev = req->rq_export->exp_obd;
 204         struct ldlm_reply *dlm_rep;
 205         struct ldlm_request *dlm_req;
 206         int rc, size = sizeof(*dlm_rep), cookielen = 0;
 207         __u32 flags;
 208         ldlm_error_t err;
 209         struct ldlm_lock *lock = NULL;
 210         void *cookie = NULL;
 211         ENTRY;
 212
 213         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 214
 215         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 216         flags = dlm_req->lock_flags;
 217         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN &&
 218             (flags & LDLM_FL_HAS_INTENT)) {
 219                 /* In this case, the reply buffer is allocated deep in
 220                  * local_lock_enqueue by the policy function. */
 221                 cookie = req;
 222                 cookielen = sizeof(*req);
 223         } else {
 224                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 225                                      &req->rq_repmsg);
 226                 if (rc) {
 227                         CERROR("out of memory\n");
 228                         RETURN(-ENOMEM);
 229                 }
 230                 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
 231                         cookie = &dlm_req->lock_desc.l_extent;
 232                         cookielen = sizeof(struct ldlm_extent);
 233                 }
 234         }
 235
 236         /* XXX notice that this lock has no callback data: of course the
 237            export would be exactly what we may want to use here... */
 238         lock = ldlm_lock_create(obddev->obd_namespace,
 239                                 &dlm_req->lock_handle2,
 240                                 dlm_req->lock_desc.l_resource.lr_name,
 241                                 dlm_req->lock_desc.l_resource.lr_type,
 242                                 dlm_req->lock_desc.l_req_mode, NULL, 0);
 243         if (!lock)
 244                 GOTO(out, err = -ENOMEM);
 245
 246         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 247                sizeof(lock->l_remote_handle));
 248         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 249
 250         err = ldlm_lock_enqueue(lock, cookie, cookielen, &flags,
 251                                 ldlm_server_completion_ast,
 252                                 ldlm_server_blocking_ast);
 253         if (err != ELDLM_OK)
 254                 GOTO(out, err);
 255
 256         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 257         dlm_rep->lock_flags = flags;
 258
 259         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 260         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 261                 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
 262                        sizeof(lock->l_extent));
 263         if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
 264                 memcpy(dlm_rep->lock_resource_name, lock->l_resource->lr_name,
 265                        sizeof(dlm_rep->lock_resource_name));
 266                 dlm_rep->lock_mode = lock->l_req_mode;
 267         }
 268
 269         lock->l_export = req->rq_export;
 270         if (lock->l_export) {
 271                 l_lock(&lock->l_resource->lr_namespace->ns_lock);
 272                 list_add(&lock->l_export_chain,
 273                          &lock->l_export->exp_ldlm_data.led_held_locks);
 274                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 275         }
 276
 277         EXIT;
 278  out:
 279         if (lock)
 280                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 281                            "(err=%d)", err);
 282         req->rq_status = err;
 283
 284         if (lock) {
 285                 if (!err)
 286                         ldlm_reprocess_all(lock->l_resource);
 287                 LDLM_LOCK_PUT(lock);
 288         }
 289         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
 290
 291         return 0;
 292 }
 293
 294 int ldlm_handle_convert(struct ptlrpc_request *req)
 295 {
 296         struct ldlm_request *dlm_req;
 297         struct ldlm_reply *dlm_rep;
 298         struct ldlm_lock *lock;
 299         int rc, size = sizeof(*dlm_rep);
 300         ENTRY;
 301
 302         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 303         if (rc) {
 304                 CERROR("out of memory\n");
 305                 RETURN(-ENOMEM);
 306         }
 307         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 308         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 309         dlm_rep->lock_flags = dlm_req->lock_flags;
 310
 311         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 312         if (!lock) {
 313                 req->rq_status = EINVAL;
 314         } else {
 315                 LDLM_DEBUG(lock, "server-side convert handler START");
 316                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 317                                   &dlm_rep->lock_flags);
 318                 if (ldlm_del_waiting_lock(lock))
 319                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 320                 req->rq_status = 0;
 321         }
 322
 323         if (lock) {
 324                 ldlm_reprocess_all(lock->l_resource);
 325                 LDLM_DEBUG(lock, "server-side convert handler END");
 326                 LDLM_LOCK_PUT(lock);
 327         } else
 328                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 329
 330         RETURN(0);
 331 }
 332
 333 int ldlm_handle_cancel(struct ptlrpc_request *req)
 334 {
 335         struct ldlm_request *dlm_req;
 336         struct ldlm_lock *lock;
 337         int rc;
 338         ENTRY;
 339
 340         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 341         if (rc) {
 342                 CERROR("out of memory\n");
 343                 RETURN(-ENOMEM);
 344         }
 345         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 346         if (!dlm_req) {
 347                 CERROR("bad request buffer for cancel\n");
 348                 RETURN(-EINVAL);
 349         }
 350
 351         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 352         if (!lock) {
 353                 CERROR("bad lock handle\n");
 354                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock (lock "
 355                                   "%p)", (void *)(unsigned long)
 356                                   dlm_req->lock_handle1.addr);
 357                 req->rq_status = ESTALE;
 358         } else {
 359                 LDLM_DEBUG(lock, "server-side cancel handler START");
 360                 ldlm_lock_cancel(lock);
 361                 if (ldlm_del_waiting_lock(lock))
 362                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 363                 req->rq_status = 0;
 364         }
 365
 366         if (ptlrpc_reply(req->rq_svc, req) != 0)
 367                 LBUG();
 368
 369         if (lock) {
 370                 ldlm_reprocess_all(lock->l_resource);
 371                 LDLM_DEBUG(lock, "server-side cancel handler END");
 372                 LDLM_LOCK_PUT(lock);
 373         }
 374
 375         RETURN(0);
 376 }
 377
 378 static int ldlm_handle_bl_callback(struct ptlrpc_request *req)
 379 {
 380         struct ldlm_request *dlm_req;
 381         struct ldlm_lock *lock;
 382         int do_ast;
 383         ENTRY;
 384
 385         OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_BL_AST, 0);
 386
 387         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 388
 389         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 390         if (!lock) {
 391                 CERROR("blocking callback on lock "LPX64" - lock disappeared\n",
 392                        dlm_req->lock_handle1.addr);
 393                 RETURN(0);
 394         }
 395
 396         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 397
 398         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 399         lock->l_flags |= LDLM_FL_CBPENDING;
 400         do_ast = (!lock->l_readers && !lock->l_writers);
 401         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 402
 403         if (do_ast) {
 404                 LDLM_DEBUG(lock, "already unused, calling "
 405                            "callback (%p)", lock->l_blocking_ast);
 406                 if (lock->l_blocking_ast != NULL) {
 407                         lock->l_blocking_ast(lock, &dlm_req->lock_desc,
 408                                              lock->l_data, lock->l_data_len,
 409                                              LDLM_CB_BLOCKING);
 410                 }
 411         } else
 412                 LDLM_DEBUG(lock, "Lock still has references, will be"
 413                            " cancelled later");
 414
 415         LDLM_DEBUG(lock, "client blocking callback handler END");
 416         LDLM_LOCK_PUT(lock);
 417         RETURN(0);
 418 }
 419
 420 static int ldlm_handle_cp_callback(struct ptlrpc_request *req)
 421 {
 422         struct list_head ast_list = LIST_HEAD_INIT(ast_list);
 423         struct ldlm_request *dlm_req;
 424         struct ldlm_lock *lock;
 425         ENTRY;
 426
 427         OBD_FAIL_RETURN(OBD_FAIL_OSC_LOCK_CP_AST, 0);
 428
 429         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 430
 431         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 432         if (!lock) {
 433                 CERROR("completion callback on lock "LPX64" - lock "
 434                        "disappeared\n", dlm_req->lock_handle1.addr);
 435                 RETURN(0);
 436         }
 437
 438         LDLM_DEBUG(lock, "client completion callback handler START");
 439
 440         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 441
 442         /* If we receive the completion AST before the actual enqueue returned,
 443          * then we might need to switch lock modes, resources, or extents. */
 444         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 445                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 446                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 447         }
 448         if (lock->l_resource->lr_type == LDLM_EXTENT)
 449                 memcpy(&lock->l_extent, &dlm_req->lock_desc.l_extent,
 450                        sizeof(lock->l_extent));
 451         ldlm_resource_unlink_lock(lock);
 452         if (memcmp(dlm_req->lock_desc.l_resource.lr_name,
 453                    lock->l_resource->lr_name,
 454                    sizeof(__u64) * RES_NAME_SIZE) != 0) {
 455                 ldlm_lock_change_resource(lock,
 456                                          dlm_req->lock_desc.l_resource.lr_name);
 457                 LDLM_DEBUG(lock, "completion AST, new resource");
 458         }
 459         lock->l_resource->lr_tmp = &ast_list;
 460         ldlm_grant_lock(lock);
 461         lock->l_resource->lr_tmp = NULL;
 462         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 463         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 464         LDLM_LOCK_PUT(lock);
 465
 466         ldlm_run_ast_work(&ast_list);
 467
 468         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 469                           lock);
 470         RETURN(0);
 471 }
 472
 473 static int ldlm_callback_handler(struct ptlrpc_request *req)
 474 {
 475         int rc;
 476         ENTRY;
 477
 478         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 479         if (rc) {
 480                 CERROR("lustre_ldlm: Invalid request: %d\n", rc);
 481                 RETURN(rc);
 482         }
 483
 484         if (req->rq_export == NULL) {
 485                 CERROR("lustre_dlm: operation %d with missing/invalid export\n",
 486                        req->rq_reqmsg->opc);
 487                 RETURN(-ENOTCONN);
 488         }
 489
 490         switch (req->rq_reqmsg->opc) {
 491         case LDLM_BL_CALLBACK:
 492                 CDEBUG(D_INODE, "blocking ast\n");
 493                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 494                 rc = ldlm_handle_bl_callback(req);
 495                 RETURN(rc);
 496         case LDLM_CP_CALLBACK:
 497                 CDEBUG(D_INODE, "completion ast\n");
 498                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 499                 rc = ldlm_handle_cp_callback(req);
 500                 RETURN(rc);
 501
 502         default:
 503                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
 504                 RETURN(-EINVAL);
 505         }
 506
 507         RETURN(0);
 508 }
 509
 510
 511 static int ldlm_cancel_handler(struct ptlrpc_request *req)
 512 {
 513         int rc;
 514         ENTRY;
 515
 516         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 517         if (rc) {
 518                 CERROR("lustre_ldlm: Invalid request: %d\n", rc);
 519                 RETURN(rc);
 520         }
 521
 522         if (req->rq_export == NULL) {
 523                 CERROR("lustre_dlm: operation %d with missing/invalid export\n",
 524                        req->rq_reqmsg->opc);
 525                 RETURN(-ENOTCONN);
 526         }
 527
 528         switch (req->rq_reqmsg->opc) {
 529
 530         /* XXX FIXME move this back to mds/handler.c, bug 625069 */
 531         case LDLM_CANCEL:
 532                 CDEBUG(D_INODE, "cancel\n");
 533                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 534                 rc = ldlm_handle_cancel(req);
 535                 if (rc)
 536                         break;
 537                 RETURN(0);
 538
 539         default:
 540                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
 541                 RETURN(-EINVAL);
 542         }
 543
 544         RETURN(0);
 545 }
 546
 547
 548 static int ldlm_iocontrol(long cmd, struct lustre_handle *conn, int len,
 549                           void *karg, void *uarg)
 550 {
 551         struct obd_device *obddev = class_conn2obd(conn);
 552         struct ptlrpc_connection *connection;
 553         int err = 0;
 554         ENTRY;
 555
 556         if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
 557             _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
 558                 CDEBUG(D_IOCTL, "invalid ioctl (type %ld, nr %ld, size %ld)\n",
 559                        _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
 560                 RETURN(-EINVAL);
 561         }
 562
 563         OBD_ALLOC(obddev->u.ldlm.ldlm_client,
 564                   sizeof(*obddev->u.ldlm.ldlm_client));
 565         connection = ptlrpc_uuid_to_connection("ldlm");
 566         if (!connection)
 567                 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
 568
 569         switch (cmd) {
 570         case IOC_LDLM_TEST:
 571                 err = ldlm_test(obddev, conn);
 572                 CERROR("-- done err %d\n", err);
 573                 GOTO(out, err);
 574         case IOC_LDLM_DUMP:
 575                 ldlm_dump_all_namespaces();
 576                 GOTO(out, err);
 577         default:
 578                 GOTO(out, err = -EINVAL);
 579         }
 580
 581  out:
 582         if (connection)
 583                 ptlrpc_put_connection(connection);
 584         OBD_FREE(obddev->u.ldlm.ldlm_client,
 585                  sizeof(*obddev->u.ldlm.ldlm_client));
 586         return err;
 587 }
 588
 589 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
 590 {
 591         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 592         int rc, i;
 593         ENTRY;
 594
 595         if (ldlm_already_setup)
 596                 RETURN(-EALREADY);
 597
 598         MOD_INC_USE_COUNT;
 599
 600         rc = ldlm_proc_setup(obddev);
 601         if (rc != 0)
 602                 GOTO(out_dec, rc);
 603
 604         ldlm->ldlm_cb_service =
 605                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
 606                                 LDLM_MAXREQSIZE, LDLM_CB_REQUEST_PORTAL,
 607                                 LDLM_CB_REPLY_PORTAL, "self",
 608                                 ldlm_callback_handler, "ldlm_cbd");
 609
 610         if (!ldlm->ldlm_cb_service)
 611                 GOTO(out_proc, rc = -ENOMEM);
 612
 613         ldlm->ldlm_cancel_service =
 614                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
 615                                 LDLM_MAXREQSIZE, LDLM_CANCEL_REQUEST_PORTAL,
 616                                 LDLM_CANCEL_REPLY_PORTAL, "self",
 617                                 ldlm_cancel_handler, "ldlm_canceld");
 618
 619         if (!ldlm->ldlm_cancel_service)
 620                 GOTO(out_proc, rc = -ENOMEM);
 621
 622         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 623                 char name[32];
 624                 sprintf(name, "ldlm_%02d", i);
 625                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_cancel_service,
 626                                          name);
 627                 if (rc) {
 628                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 629                         LBUG();
 630                         GOTO(out_thread, rc);
 631                 }
 632         }
 633
 634         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 635                 char name[32];
 636                 sprintf(name, "ldlm_%02d", i);
 637                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_cb_service, name);
 638                 if (rc) {
 639                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 640                         LBUG();
 641                         GOTO(out_thread, rc);
 642                 }
 643         }
 644
 645         INIT_LIST_HEAD(&waiting_locks_list);
 646         spin_lock_init(&waiting_locks_spinlock);
 647         waiting_locks_timer.function = waiting_locks_callback;
 648         waiting_locks_timer.data = 0;
 649         init_timer(&waiting_locks_timer);
 650
 651         ldlm_already_setup = 1;
 652
 653         RETURN(0);
 654
 655  out_thread:
 656         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
 657         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
 658         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
 659         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
 660
 661  out_proc:
 662         ldlm_proc_cleanup(obddev);
 663
 664  out_dec:
 665         MOD_DEC_USE_COUNT;
 666         return rc;
 667 }
 668
 669 static int ldlm_cleanup(struct obd_device *obddev)
 670 {
 671         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 672         ENTRY;
 673
 674         if (!list_empty(&ldlm_namespace_list)) {
 675                 CERROR("ldlm still has namespaces; clean these up first.\n");
 676                 RETURN(-EBUSY);
 677         }
 678
 679         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
 680         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
 681         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
 682         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
 683         ldlm_proc_cleanup(obddev);
 684
 685         ldlm_already_setup = 0;
 686         MOD_DEC_USE_COUNT;
 687         RETURN(0);
 688 }
 689
 690 static int ldlm_connect(struct lustre_handle *conn, struct obd_device *src,
 691                         obd_uuid_t cluuid, struct recovd_obd *recovd,
 692                         ptlrpc_recovery_cb_t recover)
 693 {
 694         return class_connect(conn, src, cluuid);
 695 }
 696
 697 struct obd_ops ldlm_obd_ops = {
 698         o_iocontrol:   ldlm_iocontrol,
 699         o_setup:       ldlm_setup,
 700         o_cleanup:     ldlm_cleanup,
 701         o_connect:     ldlm_connect,
 702         o_disconnect:  class_disconnect
 703 };
 704
 705 static int __init ldlm_init(void)
 706 {
 707         int rc = class_register_type(&ldlm_obd_ops, 0, OBD_LDLM_DEVICENAME);
 708         if (rc != 0)
 709                 return rc;
 710
 711         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
 712                                                sizeof(struct ldlm_resource), 0,
 713                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 714         if (ldlm_resource_slab == NULL)
 715                 return -ENOMEM;
 716
 717         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
 718                                            sizeof(struct ldlm_lock), 0,
 719                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
 720         if (ldlm_lock_slab == NULL) {
 721                 kmem_cache_destroy(ldlm_resource_slab);
 722                 return -ENOMEM;
 723         }
 724
 725         return 0;
 726 }
 727
 728 static void __exit ldlm_exit(void)
 729 {
 730         class_unregister_type(OBD_LDLM_DEVICENAME);
 731         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
 732                 CERROR("couldn't free ldlm resource slab\n");
 733         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
 734                 CERROR("couldn't free ldlm lock slab\n");
 735 }
 736
 737 EXPORT_SYMBOL(ldlm_completion_ast);
 738 EXPORT_SYMBOL(ldlm_handle_enqueue);
 739 EXPORT_SYMBOL(ldlm_handle_cancel);
 740 EXPORT_SYMBOL(ldlm_handle_convert);
 741 EXPORT_SYMBOL(ldlm_register_intent);
 742 EXPORT_SYMBOL(ldlm_unregister_intent);
 743 EXPORT_SYMBOL(ldlm_lockname);
 744 EXPORT_SYMBOL(ldlm_typename);
 745 EXPORT_SYMBOL(__ldlm_handle2lock);
 746 EXPORT_SYMBOL(ldlm_lock2handle);
 747 EXPORT_SYMBOL(ldlm_lock_put);
 748 EXPORT_SYMBOL(ldlm_lock_match);
 749 EXPORT_SYMBOL(ldlm_lock_addref);
 750 EXPORT_SYMBOL(ldlm_lock_decref);
 751 EXPORT_SYMBOL(ldlm_lock_change_resource);
 752 EXPORT_SYMBOL(ldlm_lock_set_data);
 753 EXPORT_SYMBOL(ldlm_cli_convert);
 754 EXPORT_SYMBOL(ldlm_cli_enqueue);
 755 EXPORT_SYMBOL(ldlm_cli_cancel);
 756 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
 757 EXPORT_SYMBOL(ldlm_match_or_enqueue);
 758 EXPORT_SYMBOL(ldlm_it2str);
 759 EXPORT_SYMBOL(ldlm_test);
 760 EXPORT_SYMBOL(ldlm_regression_start);
 761 EXPORT_SYMBOL(ldlm_regression_stop);
 762 EXPORT_SYMBOL(ldlm_lock_dump);
 763 EXPORT_SYMBOL(ldlm_namespace_new);
 764 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 765 EXPORT_SYMBOL(ldlm_namespace_free);
 766 EXPORT_SYMBOL(ldlm_namespace_dump);
 767 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
 768 EXPORT_SYMBOL(l_lock);
 769 EXPORT_SYMBOL(l_unlock);
 770
 771 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 772 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
 773 MODULE_LICENSE("GPL");
 774
 775 module_init(ldlm_init);
 776 module_exit(ldlm_exit);