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 #define EXPORT_SYMTAB
  25 #define DEBUG_SUBSYSTEM S_LDLM
  26
  27 #ifdef __KERNEL__
  28 # include <linux/module.h>
  29 # include <linux/slab.h>
  30 # include <linux/init.h>
  31 #else
  32 # include <liblustre.h>
  33 #endif
  34
  35 #include <linux/lustre_dlm.h>
  36 #include <linux/obd_class.h>
  37
  38 extern kmem_cache_t *ldlm_resource_slab;
  39 extern kmem_cache_t *ldlm_lock_slab;
  40 extern struct lustre_lock ldlm_handle_lock;
  41 extern struct list_head ldlm_namespace_list;
  42 extern int (*mds_reint_p)(int offset, struct ptlrpc_request *req);
  43 extern int (*mds_getattr_name_p)(int offset, struct ptlrpc_request *req);
  44
  45 inline unsigned long round_timeout(unsigned long timeout)
  46 {
  47         return ((timeout / HZ) + 1) * HZ;
  48 }
  49
  50 /* XXX should this be per-ldlm? */
  51 static struct list_head waiting_locks_list;
  52 static spinlock_t waiting_locks_spinlock;
  53 static struct timer_list waiting_locks_timer;
  54 static int ldlm_already_setup = 0;
  55
  56 static void waiting_locks_callback(unsigned long unused)
  57 {
  58         struct list_head *liter, *n;
  59
  60         spin_lock_bh(&waiting_locks_spinlock);
  61         list_for_each_safe(liter, n, &waiting_locks_list) {
  62                 struct ldlm_lock *l = list_entry(liter, struct ldlm_lock,
  63                                                  l_pending_chain);
  64                 if (l->l_callback_timeout > jiffies)
  65                         break;
  66                 CERROR("lock timer expired, lock %p\n", l);
  67                 LDLM_DEBUG(l, "timer expired, recovering exp %p on conn %p",
  68                            l->l_export, l->l_export->exp_connection);
  69                 recovd_conn_fail(l->l_export->exp_connection);
  70         }
  71         spin_unlock_bh(&waiting_locks_spinlock);
  72 }
  73
  74 /*
  75  * Indicate that we're waiting for a client to call us back cancelling a given
  76  * lock.  We add it to the pending-callback chain, and schedule the lock-timeout
  77  * timer to fire appropriately.  (We round up to the next second, to avoid
  78  * floods of timer firings during periods of high lock contention and traffic).
  79  */
  80 static int ldlm_add_waiting_lock(struct ldlm_lock *lock)
  81 {
  82         unsigned long timeout_rounded;
  83         ENTRY;
  84
  85         LASSERT(list_empty(&lock->l_pending_chain));
  86
  87         spin_lock_bh(&waiting_locks_spinlock);
  88         lock->l_callback_timeout = jiffies + (obd_timeout * HZ / 2);
  89
  90         timeout_rounded = round_timeout(lock->l_callback_timeout);
  91
  92         if (timeout_rounded < waiting_locks_timer.expires ||
  93             !timer_pending(&waiting_locks_timer)) {
  94                 mod_timer(&waiting_locks_timer, timeout_rounded);
  95         }
  96         list_add_tail(&lock->l_pending_chain, &waiting_locks_list); /* FIFO */
  97         spin_unlock_bh(&waiting_locks_spinlock);
  98         RETURN(1);
  99 }
 100
 101 /*
 102  * Remove a lock from the pending list, likely because it had its cancellation
 103  * callback arrive without incident.  This adjusts the lock-timeout timer if
 104  * needed.  Returns 0 if the lock wasn't pending after all, 1 if it was.
 105  */
 106 int ldlm_del_waiting_lock(struct ldlm_lock *lock)
 107 {
 108         struct list_head *list_next;
 109
 110         ENTRY;
 111
 112         spin_lock_bh(&waiting_locks_spinlock);
 113
 114         if (list_empty(&lock->l_pending_chain)) {
 115                 spin_unlock_bh(&waiting_locks_spinlock);
 116                 RETURN(0);
 117         }
 118
 119         list_next = lock->l_pending_chain.next;
 120         if (lock->l_pending_chain.prev == &waiting_locks_list) {
 121                 /* Removing the head of the list, adjust timer. */
 122                 if (list_next == &waiting_locks_list) {
 123                         /* No more, just cancel. */
 124                         del_timer(&waiting_locks_timer);
 125                 } else {
 126                         struct ldlm_lock *next;
 127                         next = list_entry(list_next, struct ldlm_lock,
 128                                           l_pending_chain);
 129                         mod_timer(&waiting_locks_timer,
 130                                   round_timeout(next->l_callback_timeout));
 131                 }
 132         }
 133         list_del_init(&lock->l_pending_chain);
 134         spin_unlock_bh(&waiting_locks_spinlock);
 135         RETURN(1);
 136 }
 137
 138 static inline void ldlm_failed_ast(struct ldlm_lock *lock)
 139 {
 140         /* XXX diagnostic */
 141         recovd_conn_fail(lock->l_export->exp_connection);
 142 }
 143
 144 int ldlm_server_blocking_ast(struct ldlm_lock *lock,
 145                              struct ldlm_lock_desc *desc,
 146                              void *data, int flag)
 147 {
 148         struct ldlm_request *body;
 149         struct ptlrpc_request *req;
 150         int rc = 0, size = sizeof(*body);
 151         ENTRY;
 152
 153         if (flag == LDLM_CB_CANCELING) {
 154                 /* Don't need to do anything here. */
 155                 RETURN(0);
 156         }
 157
 158         LASSERT(lock);
 159
 160         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 161         /* XXX This is necessary because, with the lock re-tasking, we actually
 162          * _can_ get called in here twice.  (bug 830) */
 163         if (!list_empty(&lock->l_pending_chain)) {
 164                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 165                 RETURN(0);
 166         }
 167
 168         if (lock->l_destroyed) {
 169                 /* What's the point? */
 170                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 171                 RETURN(0);
 172         }
 173
 174         req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
 175                               LDLM_BL_CALLBACK, 1, &size, NULL);
 176         if (!req)
 177                 RETURN(-ENOMEM);
 178
 179         body = lustre_msg_buf(req->rq_reqmsg, 0);
 180         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 181                sizeof(body->lock_handle1));
 182         memcpy(&body->lock_desc, desc, sizeof(*desc));
 183
 184         LDLM_DEBUG(lock, "server preparing blocking AST");
 185         req->rq_replen = lustre_msg_size(0, NULL);
 186
 187         ldlm_add_waiting_lock(lock);
 188         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 189
 190         req->rq_level = LUSTRE_CONN_RECOVD;
 191         req->rq_timeout = 2;
 192         rc = ptlrpc_queue_wait(req);
 193         if (rc == -ETIMEDOUT || rc == -EINTR) {
 194                 ldlm_del_waiting_lock(lock);
 195                 ldlm_failed_ast(lock);
 196         } else if (rc) {
 197                 CERROR("client returned %d from blocking AST for lock %p\n",
 198                        req->rq_status, lock);
 199                 LDLM_DEBUG(lock, "client returned error %d from blocking AST",
 200                            req->rq_status);
 201                 ldlm_lock_cancel(lock);
 202                 /* Server-side AST functions are called from ldlm_reprocess_all,
 203                  * which needs to be told to please restart its reprocessing. */
 204                 rc = -ERESTART;
 205         }
 206
 207         ptlrpc_req_finished(req);
 208
 209         RETURN(rc);
 210 }
 211
 212 int ldlm_server_completion_ast(struct ldlm_lock *lock, int flags, void *data)
 213 {
 214         struct ldlm_request *body;
 215         struct ptlrpc_request *req;
 216         int rc = 0, size = sizeof(*body);
 217         ENTRY;
 218
 219         if (lock == NULL) {
 220                 LBUG();
 221                 RETURN(-EINVAL);
 222         }
 223
 224         req = ptlrpc_prep_req(&lock->l_export->exp_ldlm_data.led_import,
 225                               LDLM_CP_CALLBACK, 1, &size, NULL);
 226         if (!req)
 227                 RETURN(-ENOMEM);
 228
 229         body = lustre_msg_buf(req->rq_reqmsg, 0);
 230         memcpy(&body->lock_handle1, &lock->l_remote_handle,
 231                sizeof(body->lock_handle1));
 232         body->lock_flags = flags;
 233         ldlm_lock2desc(lock, &body->lock_desc);
 234
 235         LDLM_DEBUG(lock, "server preparing completion AST");
 236         req->rq_replen = lustre_msg_size(0, NULL);
 237
 238         req->rq_level = LUSTRE_CONN_RECOVD;
 239         req->rq_timeout = 2;
 240         rc = ptlrpc_queue_wait(req);
 241         if (rc == -ETIMEDOUT || rc == -EINTR) {
 242                 ldlm_del_waiting_lock(lock);
 243                 ldlm_failed_ast(lock);
 244         } else if (rc) {
 245                 CERROR("client returned %d from completion AST for lock %p\n",
 246                        req->rq_status, lock);
 247                 LDLM_DEBUG(lock, "client returned error %d from completion AST",
 248                            req->rq_status);
 249                 ldlm_lock_cancel(lock);
 250                 /* Server-side AST functions are called from ldlm_reprocess_all,
 251                  * which needs to be told to please restart its reprocessing. */
 252                 rc = -ERESTART;
 253         }
 254         ptlrpc_req_finished(req);
 255
 256         RETURN(rc);
 257 }
 258
 259 int ldlm_handle_enqueue(struct ptlrpc_request *req,
 260                         ldlm_completion_callback completion_callback,
 261                         ldlm_blocking_callback blocking_callback)
 262 {
 263         struct obd_device *obddev = req->rq_export->exp_obd;
 264         struct ldlm_reply *dlm_rep;
 265         struct ldlm_request *dlm_req;
 266         int rc, size = sizeof(*dlm_rep), cookielen = 0;
 267         __u32 flags;
 268         ldlm_error_t err;
 269         struct ldlm_lock *lock = NULL;
 270         void *cookie = NULL;
 271         ENTRY;
 272
 273         LDLM_DEBUG_NOLOCK("server-side enqueue handler START");
 274
 275         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 276         flags = dlm_req->lock_flags;
 277         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_PLAIN &&
 278             (flags & LDLM_FL_HAS_INTENT)) {
 279                 /* In this case, the reply buffer is allocated deep in
 280                  * local_lock_enqueue by the policy function. */
 281                 cookie = req;
 282                 cookielen = sizeof(*req);
 283         } else {
 284                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 285                                      &req->rq_repmsg);
 286                 if (rc) {
 287                         CERROR("out of memory\n");
 288                         RETURN(-ENOMEM);
 289                 }
 290                 if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT) {
 291                         cookie = &dlm_req->lock_desc.l_extent;
 292                         cookielen = sizeof(struct ldlm_extent);
 293                 }
 294         }
 295
 296         /* The lock's callback data might be set in the policy function */
 297         lock = ldlm_lock_create(obddev->obd_namespace,
 298                                 &dlm_req->lock_handle2,
 299                                 dlm_req->lock_desc.l_resource.lr_name,
 300                                 dlm_req->lock_desc.l_resource.lr_type,
 301                                 dlm_req->lock_desc.l_req_mode, NULL, 0);
 302         if (!lock)
 303                 GOTO(out, err = -ENOMEM);
 304
 305         memcpy(&lock->l_remote_handle, &dlm_req->lock_handle1,
 306                sizeof(lock->l_remote_handle));
 307         LDLM_DEBUG(lock, "server-side enqueue handler, new lock created");
 308
 309         LASSERT(req->rq_export);
 310         lock->l_export = req->rq_export;
 311         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 312         list_add(&lock->l_export_chain,
 313                  &lock->l_export->exp_ldlm_data.led_held_locks);
 314         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 315
 316         err = ldlm_lock_enqueue(obddev->obd_namespace, &lock, cookie, cookielen,
 317                                 &flags, completion_callback, blocking_callback);
 318         if (err)
 319                 GOTO(out, err);
 320
 321         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 322         dlm_rep->lock_flags = flags;
 323
 324         ldlm_lock2handle(lock, &dlm_rep->lock_handle);
 325         if (dlm_req->lock_desc.l_resource.lr_type == LDLM_EXTENT)
 326                 memcpy(&dlm_rep->lock_extent, &lock->l_extent,
 327                        sizeof(lock->l_extent));
 328         if (dlm_rep->lock_flags & LDLM_FL_LOCK_CHANGED) {
 329                 memcpy(&dlm_rep->lock_resource_name, &lock->l_resource->lr_name,
 330                        sizeof(dlm_rep->lock_resource_name));
 331                 dlm_rep->lock_mode = lock->l_req_mode;
 332         }
 333
 334         EXIT;
 335  out:
 336         if (lock)
 337                 LDLM_DEBUG(lock, "server-side enqueue handler, sending reply"
 338                            "(err=%d)", err);
 339         req->rq_status = err;
 340
 341         /* The LOCK_CHANGED code in ldlm_lock_enqueue depends on this
 342          * ldlm_reprocess_all.  If this moves, revisit that code. -phil */
 343         if (lock) {
 344                 if (!err)
 345                         ldlm_reprocess_all(lock->l_resource);
 346                 LDLM_LOCK_PUT(lock);
 347         }
 348         LDLM_DEBUG_NOLOCK("server-side enqueue handler END (lock %p)", lock);
 349
 350         return 0;
 351 }
 352
 353 int ldlm_handle_convert(struct ptlrpc_request *req)
 354 {
 355         struct ldlm_request *dlm_req;
 356         struct ldlm_reply *dlm_rep;
 357         struct ldlm_lock *lock;
 358         int rc, size = sizeof(*dlm_rep);
 359         ENTRY;
 360
 361         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 362         if (rc) {
 363                 CERROR("out of memory\n");
 364                 RETURN(-ENOMEM);
 365         }
 366         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 367         dlm_rep = lustre_msg_buf(req->rq_repmsg, 0);
 368         dlm_rep->lock_flags = dlm_req->lock_flags;
 369
 370         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 371         if (!lock) {
 372                 req->rq_status = EINVAL;
 373         } else {
 374                 LDLM_DEBUG(lock, "server-side convert handler START");
 375                 ldlm_lock_convert(lock, dlm_req->lock_desc.l_req_mode,
 376                                   &dlm_rep->lock_flags);
 377                 if (ldlm_del_waiting_lock(lock))
 378                         CDEBUG(D_DLMTRACE, "converted waiting lock %p\n", lock);
 379                 req->rq_status = 0;
 380         }
 381
 382         if (lock) {
 383                 ldlm_reprocess_all(lock->l_resource);
 384                 LDLM_DEBUG(lock, "server-side convert handler END");
 385                 LDLM_LOCK_PUT(lock);
 386         } else
 387                 LDLM_DEBUG_NOLOCK("server-side convert handler END");
 388
 389         RETURN(0);
 390 }
 391
 392 int ldlm_handle_cancel(struct ptlrpc_request *req)
 393 {
 394         struct ldlm_request *dlm_req;
 395         struct ldlm_lock *lock;
 396         int rc;
 397         ENTRY;
 398
 399         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 400         if (rc) {
 401                 CERROR("out of memory\n");
 402                 RETURN(-ENOMEM);
 403         }
 404         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 405         if (!dlm_req) {
 406                 CERROR("bad request buffer for cancel\n");
 407                 RETURN(-EINVAL);
 408         }
 409
 410         lock = ldlm_handle2lock(&dlm_req->lock_handle1);
 411         if (!lock) {
 412                 CERROR("received cancel for unknown lock cookie "LPX64"\n",
 413                        dlm_req->lock_handle1.cookie);
 414                 LDLM_DEBUG_NOLOCK("server-side cancel handler stale lock "
 415                                   "(cookie "LPU64")",
 416                                   dlm_req->lock_handle1.cookie);
 417                 req->rq_status = ESTALE;
 418         } else {
 419                 LDLM_DEBUG(lock, "server-side cancel handler START");
 420                 ldlm_lock_cancel(lock);
 421                 if (ldlm_del_waiting_lock(lock))
 422                         CDEBUG(D_DLMTRACE, "cancelled waiting lock %p\n", lock);
 423                 req->rq_status = 0;
 424         }
 425
 426         if (ptlrpc_reply(req->rq_svc, req) != 0)
 427                 LBUG();
 428
 429         if (lock) {
 430                 ldlm_reprocess_all(lock->l_resource);
 431                 LDLM_DEBUG(lock, "server-side cancel handler END");
 432                 LDLM_LOCK_PUT(lock);
 433         }
 434
 435         RETURN(0);
 436 }
 437
 438 static void ldlm_handle_bl_callback(struct ptlrpc_request *req,
 439                                     struct ldlm_namespace *ns,
 440                                     struct ldlm_request *dlm_req,
 441                                     struct ldlm_lock *lock)
 442 {
 443         int do_ast;
 444         ENTRY;
 445
 446         /* Try to narrow down this damn iozone bug */
 447         if (lock->l_resource == NULL)
 448                 CERROR("lock %p resource NULL\n", lock);
 449         if (lock->l_resource->lr_type != LDLM_EXTENT)
 450                 if (lock->l_resource->lr_namespace != ns)
 451                         CERROR("lock %p namespace %p != passed ns %p\n", lock,
 452                                lock->l_resource->lr_namespace, ns);
 453         LDLM_DEBUG(lock, "client blocking AST callback handler START");
 454
 455         l_lock(&ns->ns_lock);
 456         lock->l_flags |= LDLM_FL_CBPENDING;
 457         do_ast = (!lock->l_readers && !lock->l_writers);
 458         l_unlock(&ns->ns_lock);
 459
 460         if (do_ast) {
 461                 LDLM_DEBUG(lock, "already unused, calling "
 462                            "callback (%p)", lock->l_blocking_ast);
 463                 if (lock->l_blocking_ast != NULL)
 464                         lock->l_blocking_ast(lock, &dlm_req->lock_desc,
 465                                              lock->l_data, LDLM_CB_BLOCKING);
 466         } else {
 467                 LDLM_DEBUG(lock, "Lock still has references, will be"
 468                            " cancelled later");
 469         }
 470
 471         LDLM_DEBUG(lock, "client blocking callback handler END");
 472         LDLM_LOCK_PUT(lock);
 473         EXIT;
 474 }
 475
 476 static void ldlm_handle_cp_callback(struct ptlrpc_request *req,
 477                                     struct ldlm_namespace *ns,
 478                                     struct ldlm_request *dlm_req,
 479                                     struct ldlm_lock *lock)
 480 {
 481         LIST_HEAD(ast_list);
 482         ENTRY;
 483
 484         LDLM_DEBUG(lock, "client completion callback handler START");
 485
 486         l_lock(&ns->ns_lock);
 487
 488         /* If we receive the completion AST before the actual enqueue returned,
 489          * then we might need to switch lock modes, resources, or extents. */
 490         if (dlm_req->lock_desc.l_granted_mode != lock->l_req_mode) {
 491                 lock->l_req_mode = dlm_req->lock_desc.l_granted_mode;
 492                 LDLM_DEBUG(lock, "completion AST, new lock mode");
 493         }
 494         if (lock->l_resource->lr_type == LDLM_EXTENT)
 495                 memcpy(&lock->l_extent, &dlm_req->lock_desc.l_extent,
 496                        sizeof(lock->l_extent));
 497         ldlm_resource_unlink_lock(lock);
 498         if (memcmp(&dlm_req->lock_desc.l_resource.lr_name,
 499                    &lock->l_resource->lr_name,
 500                    sizeof(lock->l_resource->lr_name)) != 0) {
 501                 ldlm_lock_change_resource(ns, lock,
 502                                          dlm_req->lock_desc.l_resource.lr_name);
 503                 LDLM_DEBUG(lock, "completion AST, new resource");
 504         }
 505         lock->l_resource->lr_tmp = &ast_list;
 506         ldlm_grant_lock(lock, req, sizeof(*req));
 507         lock->l_resource->lr_tmp = NULL;
 508         l_unlock(&ns->ns_lock);
 509         LDLM_DEBUG(lock, "callback handler finished, about to run_ast_work");
 510         LDLM_LOCK_PUT(lock);
 511
 512         ldlm_run_ast_work(&ast_list);
 513
 514         LDLM_DEBUG_NOLOCK("client completion callback handler END (lock %p)",
 515                           lock);
 516         EXIT;
 517 }
 518
 519 static int ldlm_callback_reply(struct ptlrpc_request *req, int rc)
 520 {
 521         req->rq_status = rc;
 522         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
 523                              &req->rq_repmsg);
 524         if (rc)
 525                 return rc;
 526         return ptlrpc_reply(req->rq_svc, req);
 527 }
 528
 529 static int ldlm_callback_handler(struct ptlrpc_request *req)
 530 {
 531         struct ldlm_namespace *ns;
 532         struct ldlm_request *dlm_req;
 533         struct ldlm_lock *lock;
 534         int rc;
 535         ENTRY;
 536
 537         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 538         if (rc) {
 539                 CERROR("Invalid request: %d\n", rc);
 540                 RETURN(rc);
 541         }
 542
 543         if (req->rq_export == NULL) {
 544                 struct ldlm_request *dlm_req;
 545
 546                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
 547                        req->rq_reqmsg->opc, req->rq_request_portal,
 548                        req->rq_reply_portal);
 549                 CERROR("--> export addr: "LPX64", cookie: "LPX64"\n",
 550                        req->rq_reqmsg->addr, req->rq_reqmsg->cookie);
 551                 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 552                 CERROR("--> lock addr: "LPX64", cookie: "LPX64"\n",
 553                        dlm_req->lock_handle1.addr,dlm_req->lock_handle1.cookie);
 554                 ldlm_callback_reply(req, -ENOTCONN);
 555                 RETURN(0);
 556         }
 557
 558         if (req->rq_reqmsg->opc == LDLM_BL_CALLBACK) {
 559                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 560         } else if (req->rq_reqmsg->opc == LDLM_CP_CALLBACK) {
 561                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CP_CALLBACK, 0);
 562         } else {
 563                 ldlm_callback_reply(req, -EIO);
 564                 RETURN(0);
 565         }
 566
 567         LASSERT(req->rq_export != NULL);
 568         LASSERT(req->rq_export->exp_obd != NULL);
 569         ns = req->rq_export->exp_obd->obd_namespace;
 570         LASSERT(ns != NULL);
 571
 572         dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 573         lock = ldlm_handle2lock_ns(ns, &dlm_req->lock_handle1);
 574         if (!lock) {
 575                 CDEBUG(D_INODE, "callback on lock "LPX64" - lock disappeared\n",
 576                        dlm_req->lock_handle1.cookie);
 577                 ldlm_callback_reply(req, -EINVAL);
 578                 RETURN(0);
 579         }
 580
 581         /* we want the ost thread to get this reply so that it can respond
 582          * to ost requests (write cache writeback) that might be triggered
 583          * in the callback */
 584         ldlm_callback_reply(req, 0);
 585
 586         switch (req->rq_reqmsg->opc) {
 587         case LDLM_BL_CALLBACK:
 588                 CDEBUG(D_INODE, "blocking ast\n");
 589                 ldlm_handle_bl_callback(req, ns, dlm_req, lock);
 590                 break;
 591         case LDLM_CP_CALLBACK:
 592                 CDEBUG(D_INODE, "completion ast\n");
 593                 ldlm_handle_cp_callback(req, ns, dlm_req, lock);
 594                 break;
 595         }
 596
 597         RETURN(0);
 598 }
 599
 600 static int ldlm_cancel_handler(struct ptlrpc_request *req)
 601 {
 602         int rc;
 603         ENTRY;
 604
 605         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 606         if (rc) {
 607                 CERROR("lustre_ldlm: Invalid request: %d\n", rc);
 608                 RETURN(rc);
 609         }
 610
 611         if (req->rq_export == NULL) {
 612                 struct ldlm_request *dlm_req;
 613                 CERROR("operation %d with bad export (ptl req %d/rep %d)\n",
 614                        req->rq_reqmsg->opc, req->rq_request_portal,
 615                        req->rq_reply_portal);
 616                 CERROR("--> export addr: "LPX64", cookie: "LPX64"\n",
 617                        req->rq_reqmsg->addr, req->rq_reqmsg->cookie);
 618                 dlm_req = lustre_msg_buf(req->rq_reqmsg, 0);
 619                 ldlm_lock_dump_handle(D_ERROR, &dlm_req->lock_handle1);
 620                 RETURN(-ENOTCONN);
 621         }
 622
 623         switch (req->rq_reqmsg->opc) {
 624
 625         /* XXX FIXME move this back to mds/handler.c, bug 625069 */
 626         case LDLM_CANCEL:
 627                 CDEBUG(D_INODE, "cancel\n");
 628                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 629                 rc = ldlm_handle_cancel(req);
 630                 if (rc)
 631                         break;
 632                 RETURN(0);
 633
 634         default:
 635                 CERROR("invalid opcode %d\n", req->rq_reqmsg->opc);
 636                 RETURN(-EINVAL);
 637         }
 638
 639         RETURN(0);
 640 }
 641
 642 static int ldlm_iocontrol(unsigned int cmd, struct lustre_handle *conn, int len,
 643                           void *karg, void *uarg)
 644 {
 645         struct obd_device *obddev = class_conn2obd(conn);
 646         struct ptlrpc_connection *connection;
 647         struct obd_uuid uuid = { "ldlm" };
 648         int err = 0;
 649         ENTRY;
 650
 651         if (_IOC_TYPE(cmd) != IOC_LDLM_TYPE || _IOC_NR(cmd) < IOC_LDLM_MIN_NR ||
 652             _IOC_NR(cmd) > IOC_LDLM_MAX_NR) {
 653                 CDEBUG(D_IOCTL, "invalid ioctl (type %d, nr %d, size %d)\n",
 654                        _IOC_TYPE(cmd), _IOC_NR(cmd), _IOC_SIZE(cmd));
 655                 RETURN(-EINVAL);
 656         }
 657
 658         OBD_ALLOC(obddev->u.ldlm.ldlm_client,
 659                   sizeof(*obddev->u.ldlm.ldlm_client));
 660         connection = ptlrpc_uuid_to_connection(&uuid);
 661         if (!connection)
 662                 CERROR("No LDLM UUID found: assuming ldlm is local.\n");
 663
 664         switch (cmd) {
 665         case IOC_LDLM_TEST:
 666                 //err = ldlm_test(obddev, conn);
 667                 err = 0;
 668                 CERROR("-- NO TESTS WERE RUN done err %d\n", err);
 669                 GOTO(out, err);
 670         case IOC_LDLM_DUMP:
 671                 ldlm_dump_all_namespaces();
 672                 GOTO(out, err);
 673         default:
 674                 GOTO(out, err = -EINVAL);
 675         }
 676
 677  out:
 678         if (connection)
 679                 ptlrpc_put_connection(connection);
 680         OBD_FREE(obddev->u.ldlm.ldlm_client,
 681                  sizeof(*obddev->u.ldlm.ldlm_client));
 682         return err;
 683 }
 684
 685 static int ldlm_setup(struct obd_device *obddev, obd_count len, void *buf)
 686 {
 687         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 688         int rc, i;
 689         ENTRY;
 690
 691         if (ldlm_already_setup)
 692                 RETURN(-EALREADY);
 693
 694         rc = ldlm_proc_setup(obddev);
 695         if (rc != 0)
 696                 RETURN(rc);
 697
 698 #ifdef __KERNEL__
 699         ldlm->ldlm_cb_service =
 700                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
 701                                 LDLM_MAXREQSIZE, LDLM_CB_REQUEST_PORTAL,
 702                                 LDLM_CB_REPLY_PORTAL,
 703                                 ldlm_callback_handler, "ldlm_cbd");
 704
 705         if (!ldlm->ldlm_cb_service) {
 706                 CERROR("failed to start service\n");
 707                 GOTO(out_proc, rc = -ENOMEM);
 708         }
 709
 710         ldlm->ldlm_cancel_service =
 711                 ptlrpc_init_svc(LDLM_NEVENTS, LDLM_NBUFS, LDLM_BUFSIZE,
 712                                 LDLM_MAXREQSIZE, LDLM_CANCEL_REQUEST_PORTAL,
 713                                 LDLM_CANCEL_REPLY_PORTAL,
 714                                 ldlm_cancel_handler, "ldlm_canceld");
 715
 716         if (!ldlm->ldlm_cancel_service) {
 717                 CERROR("failed to start service\n");
 718                 GOTO(out_proc, rc = -ENOMEM);
 719         }
 720
 721         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 722                 char name[32];
 723                 sprintf(name, "ldlm_cn_%02d", i);
 724                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_cancel_service,
 725                                          name);
 726                 if (rc) {
 727                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 728                         LBUG();
 729                         GOTO(out_thread, rc);
 730                 }
 731         }
 732
 733         for (i = 0; i < LDLM_NUM_THREADS; i++) {
 734                 char name[32];
 735                 sprintf(name, "ldlm_cb_%02d", i);
 736                 rc = ptlrpc_start_thread(obddev, ldlm->ldlm_cb_service, name);
 737                 if (rc) {
 738                         CERROR("cannot start LDLM thread #%d: rc %d\n", i, rc);
 739                         LBUG();
 740                         GOTO(out_thread, rc);
 741                 }
 742         }
 743
 744 #endif
 745         INIT_LIST_HEAD(&waiting_locks_list);
 746         spin_lock_init(&waiting_locks_spinlock);
 747         waiting_locks_timer.function = waiting_locks_callback;
 748         waiting_locks_timer.data = 0;
 749         init_timer(&waiting_locks_timer);
 750
 751         ldlm_already_setup = 1;
 752
 753         RETURN(0);
 754
 755  out_thread:
 756 #ifdef __KERNEL__
 757         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
 758         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
 759         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
 760         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
 761 #endif
 762  out_proc:
 763         ldlm_proc_cleanup(obddev);
 764
 765         return rc;
 766 }
 767
 768 static int ldlm_cleanup(struct obd_device *obddev)
 769 {
 770         struct ldlm_obd *ldlm = &obddev->u.ldlm;
 771         ENTRY;
 772
 773         if (!list_empty(&ldlm_namespace_list)) {
 774                 CERROR("ldlm still has namespaces; clean these up first.\n");
 775                 RETURN(-EBUSY);
 776         }
 777
 778 #ifdef __KERNEL__
 779         ptlrpc_stop_all_threads(ldlm->ldlm_cb_service);
 780         ptlrpc_unregister_service(ldlm->ldlm_cb_service);
 781         ptlrpc_stop_all_threads(ldlm->ldlm_cancel_service);
 782         ptlrpc_unregister_service(ldlm->ldlm_cancel_service);
 783         ldlm_proc_cleanup(obddev);
 784 #endif
 785         ldlm_already_setup = 0;
 786         RETURN(0);
 787 }
 788
 789 static int ldlm_connect(struct lustre_handle *conn, struct obd_device *src,
 790                         struct obd_uuid *cluuid, struct recovd_obd *recovd,
 791                         ptlrpc_recovery_cb_t recover)
 792 {
 793         return class_connect(conn, src, cluuid);
 794 }
 795
 796 struct obd_ops ldlm_obd_ops = {
 797         o_owner:       THIS_MODULE,
 798         o_iocontrol:   ldlm_iocontrol,
 799         o_setup:       ldlm_setup,
 800         o_cleanup:     ldlm_cleanup,
 801         o_connect:     ldlm_connect,
 802         o_disconnect:  class_disconnect
 803 };
 804
 805 int __init ldlm_init(void)
 806 {
 807         int rc = class_register_type(&ldlm_obd_ops, 0, OBD_LDLM_DEVICENAME);
 808         if (rc != 0)
 809                 return rc;
 810
 811         ldlm_resource_slab = kmem_cache_create("ldlm_resources",
 812                                                sizeof(struct ldlm_resource), 0,
 813                                                SLAB_HWCACHE_ALIGN, NULL, NULL);
 814         if (ldlm_resource_slab == NULL)
 815                 return -ENOMEM;
 816
 817         ldlm_lock_slab = kmem_cache_create("ldlm_locks",
 818                                            sizeof(struct ldlm_lock), 0,
 819                                            SLAB_HWCACHE_ALIGN, NULL, NULL);
 820         if (ldlm_lock_slab == NULL) {
 821                 kmem_cache_destroy(ldlm_resource_slab);
 822                 return -ENOMEM;
 823         }
 824
 825         l_lock_init(&ldlm_handle_lock);
 826
 827         return 0;
 828 }
 829
 830 static void __exit ldlm_exit(void)
 831 {
 832         class_unregister_type(OBD_LDLM_DEVICENAME);
 833         if (kmem_cache_destroy(ldlm_resource_slab) != 0)
 834                 CERROR("couldn't free ldlm resource slab\n");
 835         if (kmem_cache_destroy(ldlm_lock_slab) != 0)
 836                 CERROR("couldn't free ldlm lock slab\n");
 837 }
 838
 839 /* ldlm_lock.c */
 840 EXPORT_SYMBOL(ldlm_lock2desc);
 841 EXPORT_SYMBOL(ldlm_register_intent);
 842 EXPORT_SYMBOL(ldlm_unregister_intent);
 843 EXPORT_SYMBOL(ldlm_lockname);
 844 EXPORT_SYMBOL(ldlm_typename);
 845 EXPORT_SYMBOL(ldlm_lock2handle);
 846 EXPORT_SYMBOL(__ldlm_handle2lock);
 847 EXPORT_SYMBOL(ldlm_lock_put);
 848 EXPORT_SYMBOL(ldlm_lock_match);
 849 EXPORT_SYMBOL(ldlm_lock_cancel);
 850 EXPORT_SYMBOL(ldlm_lock_addref);
 851 EXPORT_SYMBOL(ldlm_lock_decref);
 852 EXPORT_SYMBOL(ldlm_lock_decref_and_cancel);
 853 EXPORT_SYMBOL(ldlm_lock_change_resource);
 854 EXPORT_SYMBOL(ldlm_lock_set_data);
 855 EXPORT_SYMBOL(ldlm_it2str);
 856 EXPORT_SYMBOL(ldlm_lock_dump);
 857 EXPORT_SYMBOL(ldlm_lock_dump_handle);
 858 EXPORT_SYMBOL(ldlm_cancel_locks_for_export);
 859 EXPORT_SYMBOL(ldlm_reprocess_all_ns);
 860
 861 /* ldlm_request.c */
 862 EXPORT_SYMBOL(ldlm_completion_ast);
 863 EXPORT_SYMBOL(ldlm_expired_completion_wait);
 864 EXPORT_SYMBOL(ldlm_cli_convert);
 865 EXPORT_SYMBOL(ldlm_cli_enqueue);
 866 EXPORT_SYMBOL(ldlm_cli_cancel);
 867 EXPORT_SYMBOL(ldlm_cli_cancel_unused);
 868 EXPORT_SYMBOL(ldlm_match_or_enqueue);
 869 EXPORT_SYMBOL(ldlm_replay_locks);
 870 EXPORT_SYMBOL(ldlm_resource_foreach);
 871 EXPORT_SYMBOL(ldlm_namespace_foreach);
 872 EXPORT_SYMBOL(ldlm_namespace_foreach_res);
 873
 874 /* ldlm_lockd.c */
 875 EXPORT_SYMBOL(ldlm_server_blocking_ast);
 876 EXPORT_SYMBOL(ldlm_server_completion_ast);
 877 EXPORT_SYMBOL(ldlm_handle_enqueue);
 878 EXPORT_SYMBOL(ldlm_handle_cancel);
 879 EXPORT_SYMBOL(ldlm_handle_convert);
 880 EXPORT_SYMBOL(ldlm_del_waiting_lock);
 881
 882 #if 0
 883 /* ldlm_test.c */
 884 EXPORT_SYMBOL(ldlm_test);
 885 EXPORT_SYMBOL(ldlm_regression_start);
 886 EXPORT_SYMBOL(ldlm_regression_stop);
 887 #endif
 888
 889 /* ldlm_resource.c */
 890 EXPORT_SYMBOL(ldlm_namespace_new);
 891 EXPORT_SYMBOL(ldlm_namespace_cleanup);
 892 EXPORT_SYMBOL(ldlm_namespace_free);
 893 EXPORT_SYMBOL(ldlm_namespace_dump);
 894
 895 /* l_lock.c */
 896 EXPORT_SYMBOL(l_lock);
 897 EXPORT_SYMBOL(l_unlock);
 898
 899 #ifdef __KERNEL__
 900 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
 901 MODULE_DESCRIPTION("Lustre Lock Management Module v0.1");
 902 MODULE_LICENSE("GPL");
 903
 904 module_init(ldlm_init);
 905 module_exit(ldlm_exit);
 906 #endif